Search is not available for this dataset
text
stringlengths 1
134k
| id
stringlengths 11
16
| article_id
stringlengths 8
10
| section_title
stringlengths 1
1.35k
| educational_score
float64 0.52
5.16
| domain
stringclasses 3
values | document_type
stringclasses 4
values | domain_scores
listlengths 3
3
| document_type_scores
listlengths 4
4
| language
stringclasses 38
values | language_score
float64 0
1
|
|---|---|---|---|---|---|---|---|---|---|---|
In this study, we employ various classical machine learning classifiers and a convolutional neural network (CNN) to distinguish ion channels from non-ion channels using representations generated by ProtBERT, ProtBERT-BFD, and MembraneBERT. We optimize each classifier’s performance through grid search over a set of hyperparameters on the training set (see Table 3 ).
|
39572876_p22
|
39572876
|
Machine learning classifiers
| 4.02661 |
biomedical
|
Study
|
[
0.9994040727615356,
0.00015119834279175848,
0.00044474456808529794
] |
[
0.999442994594574,
0.00024634465808048844,
0.00026932614855468273,
0.000041256556869484484
] |
en
| 0.999996 |
In our study, we employ a diverse set of classical machine learning classifiers from the scikit-learn library , each chosen for its proven effectiveness in various bioinformatics tasks and ability to capture different aspects of the data. Logistic Regression (LR) is utilized for its capacity to estimate class probabilities based on linear combinations of input features. We also implement Support Vector Machine (SVM) , which excels in finding an optimal hyperplane to separate classes in a high-dimensional space. The Random Forest (RF) classifier is included as an ensemble method, combining predictions from multiple decision trees through majority voting. For instance-based learning, we employ the k-Nearest Neighbor (kNN) algorithm , which classifies based on the majority class of k nearest neighbors in the feature space. Lastly, we incorporate a Feed-Forward Neural Network (FFNN) , a multi-layer perceptron capable of learning complex non-linear relationships in the data. This comprehensive selection of classifiers allows us to thoroughly evaluate and compare different machine learning approaches in the context of ion channel prediction.
|
39572876_p23
|
39572876
|
Classical machine learning classifiers
| 4.150632 |
biomedical
|
Study
|
[
0.9994370341300964,
0.00029542355332523584,
0.0002674746501725167
] |
[
0.9986361861228943,
0.00019425404025241733,
0.001105276052840054,
0.00006429663335438818
] |
en
| 0.999997 |
We implement a CNN using PyTorch to classify ion channels and non-ion channels, as well as to fine-tune ProtBERT, ProtBERT-BFD, and MembraneBERT concurrently during training. The CNN architecture consists of multiple convolutional layers, a dropout layer for regularization , and three fully connected layers. This architecture is well-suited for processing sequential data like protein sequences .
|
39572876_p24
|
39572876
|
Convolutional Neural Network (CNN)
| 4.037421 |
biomedical
|
Study
|
[
0.9987859129905701,
0.00016295503883156925,
0.001051146537065506
] |
[
0.9305176138877869,
0.06772119551897049,
0.001441774656996131,
0.0003193487646058202
] |
en
| 0.999995 |
Our proposed method for protein classification is illustrated in Figure 3 . It involves using both traditional machine learning algorithms and a deep learning-based approach to classify ion channel proteins from non-ion channel membrane proteins.
|
39572876_p25
|
39572876
|
Proposed method
| 3.448509 |
biomedical
|
Study
|
[
0.9985759258270264,
0.0002077387471217662,
0.0012163043720647693
] |
[
0.8612698316574097,
0.13469575345516205,
0.003455651458352804,
0.0005788269336335361
] |
en
| 0.999995 |
For classical machine learning classifiers, the BERT-based models are first fine-tuned using a one-layer feed-forward classifier. The learned representations are then extracted and used as input for the classifiers.
|
39572876_p26
|
39572876
|
Proposed method
| 2.162048 |
other
|
Other
|
[
0.17976586520671844,
0.0010594150517135859,
0.8191747069358826
] |
[
0.017097104340791702,
0.9810596108436584,
0.0015567558584734797,
0.000286547263385728
] |
en
| 0.999998 |
For the CNN-based approach, the BERT models are fine-tuned concurrently with the CNN during training, allowing the models to learn task-specific representations. These representations are then used as input for the CNN to classify the proteins.
|
39572876_p27
|
39572876
|
Proposed method
| 3.217797 |
biomedical
|
Other
|
[
0.983994722366333,
0.0007393315900117159,
0.015265979804098606
] |
[
0.19447124004364014,
0.8018097281455994,
0.0032184869050979614,
0.0005005263956263661
] |
en
| 0.999999 |
During training, the classifiers are trained on the training and validation sets (10 % of the training data), which are used to fine-tune the BERT-based models. The models are then evaluated on the test set to assess their generalization ability on unseen data.
|
39572876_p28
|
39572876
|
Proposed method
| 2.154922 |
biomedical
|
Other
|
[
0.6084740161895752,
0.0011892064940184355,
0.3903367519378662
] |
[
0.4084699749946594,
0.587509274482727,
0.0031753559596836567,
0.0008453668560832739
] |
en
| 0.999996 |
In this work, we used the grid search method to find the optimal hyperparameter values for each classifier (see Table 3 ). The grid search method involves training and evaluating a model on a grid of hyperparameter values, using cross-validation (CV) to estimate the model’s performance, in order to find the combination of hyperparameter values that results in the best performance on the data. By applying the grid search method, we were able to identify the classifier and hyperparameter combination that achieved the best performance on the data.
|
39572876_p29
|
39572876
|
Grid search
| 3.55521 |
biomedical
|
Study
|
[
0.9472385048866272,
0.00038999150274321437,
0.052371442317962646
] |
[
0.9976758360862732,
0.0019885175861418247,
0.00027661118656396866,
0.00005897642768104561
] |
en
| 0.999998 |
In our study, we used 5-fold CV to evaluate the performance of all of the classifiers. CV is a technique that involves dividing the dataset into a number of folds, training the model on some of the folds, and evaluating the model on the remaining folds. This process is repeated a number of times, with different combinations of training and evaluation folds, in order to obtain a more robust estimate of the model’s performance. The optimal hyperparameters for each classifier were determined based on the results of the CV, and are shown in Table 3 .
|
39572876_p30
|
39572876
|
Evaluation methods
| 3.716445 |
biomedical
|
Study
|
[
0.9755339622497559,
0.0003845498140435666,
0.024081457406282425
] |
[
0.9991268515586853,
0.000619163503870368,
0.0002138757990906015,
0.000040134822484105825
] |
en
| 0.999996 |
We utilize McNemar’s test to evaluate the differences in performance between our proposed method, TooT-BERT-CNN-C, and the state-of-the-art model TooT-BERT-C. McNemar’s test is a statistical test used to compare the performance of two different classifiers on a binary classification task. It is typically used when the same set of samples has been classified by both classifiers, and the goal is to determine whether the performance of one classifier is significantly better than the other.
|
39572876_p31
|
39572876
|
McNemar’s test
| 3.000735 |
other
|
Study
|
[
0.43674615025520325,
0.0006434483220800757,
0.562610387802124
] |
[
0.8148481845855713,
0.18230019509792328,
0.0023792798165231943,
0.0004723367455881089
] |
en
| 0.999996 |
The test is based on the contingency table, which is a 2 × 2 table that compares the outcomes of two classification methods. The formula for McNemar’s test is: (3) χ 2 = ( b − c ) 2 b + c where b represents the number of cases where the first model made an incorrect prediction, while the second model made a correct prediction. c represents the number of cases where the first model made a correct prediction, while the second model made an incorrect prediction.
|
39572876_p32
|
39572876
|
McNemar’s test
| 3.758444 |
biomedical
|
Study
|
[
0.9894994497299194,
0.00037992605939507484,
0.010120593942701817
] |
[
0.7170404195785522,
0.28077489137649536,
0.0018841163255274296,
0.00030053930822759867
] |
en
| 0.999998 |
To interpret the results of McNemar’s test, a p -value is calculated based on the chi-square statistic. If the p -value is less than a predetermined threshold (usually 0.05), then it is considered statistically significant and the null hypothesis (that there is no difference in performance between the two classifiers) is rejected. This indicates that there is a statistically significant difference in the performance of the two classifiers.
|
39572876_p33
|
39572876
|
McNemar’s test
| 3.777471 |
biomedical
|
Study
|
[
0.9645153284072876,
0.0006926776841282845,
0.0347919762134552
] |
[
0.5260199308395386,
0.4691706895828247,
0.00449000671505928,
0.00031932906131260097
] |
en
| 0.999997 |
For this paper, we used a variety of performance metrics to evaluate the effectiveness of our approach for predicting ion channels. These metrics included accuracy (Acc), sensitivity (Sen), specificity (Spc), and the Matthew’s correlation coefficient (MCC).
|
39572876_p34
|
39572876
|
Evaluation metrics
| 3.760332 |
biomedical
|
Study
|
[
0.9993079900741577,
0.0001604872668394819,
0.0005315249436534941
] |
[
0.9985252022743225,
0.0007163038826547563,
0.0006875927210785449,
0.00007095487671904266
] |
en
| 0.999999 |
The dataset used in this work, consists of 301 ion channels and 4,263 non ion-channels, which is an imbalanced dataset. This means that the number of samples in each class is not equal, and this can affect the performance of machine learning algorithms. In imbalanced datasets, accuracy can be misleading as it does not take into account the relative frequencies of the different classes. This can lead to the model achieving high accuracy by simply predicting the majority class all the time, even if it has poor performance on the minority class. Therefore, it is often recommended to use metrics that consider all classes, such as the MCC, which takes into account true and false positives and negatives .
|
39572876_p35
|
39572876
|
Evaluation metrics
| 3.832653 |
biomedical
|
Study
|
[
0.9951536655426025,
0.0001471438881708309,
0.0046992129646241665
] |
[
0.9669392704963684,
0.03160950914025307,
0.0013078551273792982,
0.00014342876966111362
] |
en
| 0.999998 |
Accuracy is a measure of the overall correct classification rate and is calculated as the number of correct predictions divided by the total number of predictions. It is expressed as a percentage and can be calculated using the following formula: (4) A c c u r a c y = T P + T N T P + T N + F P + F N
|
39572876_p36
|
39572876
|
Evaluation metrics
| 2.990875 |
biomedical
|
Other
|
[
0.8909138441085815,
0.0011581456055864692,
0.10792798548936844
] |
[
0.09171456098556519,
0.9056202173233032,
0.002371749607846141,
0.00029348389944061637
] |
en
| 0.999997 |
Sensitivity, also known as the true positive rate, is a measure of the proportion of actual positive cases that are correctly identified as such. It is calculated using the following formula: (5) S e n s i t i v i t y = T P T P + F N
|
39572876_p37
|
39572876
|
Evaluation metrics
| 3.507413 |
biomedical
|
Other
|
[
0.9966637492179871,
0.0006091875839047134,
0.0027271367143839598
] |
[
0.2625437080860138,
0.733730673789978,
0.0031304466538131237,
0.0005951399798505008
] |
en
| 0.999997 |
Specificity, also known as the true negative rate, is a measure of the proportion of actual negative cases that are correctly identified as such. It is calculated using the following formula: (6) S p e c i fi c i t y = T N T N + F P
|
39572876_p38
|
39572876
|
Evaluation metrics
| 3.504926 |
biomedical
|
Other
|
[
0.9958337545394897,
0.0006152933929115534,
0.00355097115971148
] |
[
0.21432071924209595,
0.7824127078056335,
0.002739351475611329,
0.0005273039569146931
] |
en
| 0.999997 |
MCC is a measure of the overall accuracy of a binary classifier, taking into account both the true and false positive and negative rates. It can range from −1 (perfectly incorrect) to 1 (perfectly correct) and is calculated using the following formula: (7) M C C = T P ⋅ T N − F P ⋅ F N ( T P + F P ) ( T P + F N ) ( T N + F P ) ( T N + F N ) where TP (True Positive) is a case where the classifier correctly predicts the positive class, TN (True Negative) is a case where the classifier correctly predicts the negative class, FP (False Positive) is a case where the classifier incorrectly predicts the positive class, and FN (False Negative) is a case where the classifier incorrectly predicts the negative class.
|
39572876_p39
|
39572876
|
Evaluation metrics
| 3.953038 |
biomedical
|
Other
|
[
0.9604173898696899,
0.0007857690216042101,
0.03879689425230026
] |
[
0.41174307465553284,
0.5788717269897461,
0.009025371633470058,
0.0003597730246838182
] |
en
| 0.999997 |
A primary challenge in employing protein language models is the constraint of fixed sequence lengths, which may lead to the omission of crucial structural or functional details. In our methodology, we limited the protein sequences to a length of 1,024 due to computational constraints while fine-tuning ProtBERT, ProtBERT-BFD, and MembraneBERT. For this, we utilized the tokenizer from the Transformers Python library, setting a maximum length parameter of 1,024 and enabling automatic truncation. This approach ensures that any sequence exceeding this limit is truncated, while retaining the first 1,024 amino acids for analysis.
|
39572876_p40
|
39572876
|
Protein sequence evaluation
| 4.120821 |
biomedical
|
Study
|
[
0.9995724558830261,
0.00015054522373247892,
0.00027700437931343913
] |
[
0.9983008503913879,
0.0012394781224429607,
0.000395153125282377,
0.00006459519499912858
] |
en
| 0.999999 |
We conducted an analysis, depicted in Figure 5 , to evaluate the implications of this truncation on our dataset. The histogram demonstrates the range of protein sequence lengths in the dataset. Interestingly, most of the sequences fall below the truncation limit of 1,024 amino acids, suggesting that the truncation during fine-tuning of the BERT models is unlikely to result in substantial information loss.
|
39572876_p41
|
39572876
|
Protein sequence evaluation
| 3.794294 |
biomedical
|
Study
|
[
0.9991047978401184,
0.0002351567818550393,
0.0006600093329325318
] |
[
0.9993005990982056,
0.00047394377179443836,
0.00016846183279994875,
0.000057037446822505444
] |
en
| 0.999997 |
To visualize the feature representations from ProtBERT, ProtBERT-BFD, and MembraneBERT for ion channels and non-ion channels, we employ t-SNE, or t-Distributed Stochastic Neighbor Embedding, as outlined by van der Maaten et al. (). This technique serves as a powerful tool for reducing dimensionality while maintaining the relationships among high-dimensional data points. This approach is particularly useful for capturing intricate, non-linear relationships, making it widely used in areas like machine learning and data visualization.
|
39572876_p42
|
39572876
|
Protein sequence evaluation
| 3.943612 |
biomedical
|
Study
|
[
0.9966965913772583,
0.0001563824334880337,
0.0031470651738345623
] |
[
0.9621323943138123,
0.0326758548617363,
0.005038391798734665,
0.0001532964815851301
] |
en
| 0.999997 |
The t-SNE plot, as shown in figure Figure 6 , highlights the proficiency of ProtBERT, ProtBERT-BFD, and MembraneBERT in differentiating important features of ion channels from those of non-ion channels. In the plot, ion channels are marked in blue and non-ion channels in orange. The separation between the two categories in this two-dimensional representation suggests that the models effectively capture essential distinctions between the two groups. These distinctions may encompass variations in sequence composition or structural features that serve as hallmarks for ion channel proteins.
|
39572876_p43
|
39572876
|
Protein sequence evaluation
| 4.081589 |
biomedical
|
Study
|
[
0.9996428489685059,
0.00015637274191249162,
0.0002007649454753846
] |
[
0.9983392953872681,
0.0007430395344272256,
0.0008500576368533075,
0.00006759199459338561
] |
en
| 0.999998 |
In this section, we assess the computational overhead associated with fine-tuning ProtBERT-BFD, contrasting setups with and without the inclusion of CNN layers. Although incorporating CNN layers typically escalates both computational demands and execution time, our empirical data suggest that the improvements in model performance justify this additional investment.
|
39572876_p44
|
39572876
|
Execution time analysis
| 1.984242 |
other
|
Study
|
[
0.28598102927207947,
0.0013314089737832546,
0.712687611579895
] |
[
0.9250149130821228,
0.07272379845380783,
0.0015701117226853967,
0.000691135530360043
] |
en
| 0.999997 |
Specifically, fine-tuning ProtBERT-BFD without CNN integration took approximately 2 h when executed on a Tesla V100 GPU equipped with 32 GB RAM. The incorporation of CNN layers extended this time frame to roughly 2 h and 30 min. This increment in training duration can be directly attributed to the computational complexity introduced by the added CNN layers.
|
39572876_p45
|
39572876
|
Execution time analysis
| 1.895144 |
other
|
Other
|
[
0.1838488131761551,
0.001061560120433569,
0.8150897026062012
] |
[
0.4488455355167389,
0.5485752820968628,
0.0014259639428928494,
0.0011532598873600364
] |
en
| 0.999998 |
This section provides a comprehensive analysis of the performance metrics for various Protein Language Models (PLMs) and classifiers. The evaluation metrics include Accuracy, Sensitivity, Specificity, and the Matthews Correlation Coefficient (MCC). Results from both cross-validation (CV) and independent tests are presented in Table 4 and visualized in Figure 7 .
|
39572876_p46
|
39572876
|
Comparative analysis of classifier and PLM performances
| 3.922492 |
biomedical
|
Study
|
[
0.998415470123291,
0.00018499050929676741,
0.001399557339027524
] |
[
0.996751070022583,
0.000559959327802062,
0.0026305443607270718,
0.000058400815760251135
] |
en
| 0.999997 |
ProtBERT-BFD emerges as the strongest performer, particularly in Sensitivity, with an average of 75.84 % on independent test sets. This demonstrates its superior ability to identify true positive cases, crucial in bioinformatics applications. In contrast, MembraneBERT shows high Sensitivity in CV tests but lower performance on independent test sets, suggesting potential overfitting issues.
|
39572876_p47
|
39572876
|
Overall performance
| 3.775094 |
biomedical
|
Study
|
[
0.9975183010101318,
0.00024176306033041328,
0.002239956520497799
] |
[
0.9953954815864563,
0.0036086863838136196,
0.000888268172275275,
0.00010752027446869761
] |
en
| 0.999998 |
The kNN algorithm performs exceptionally well when paired with ProtBERT and MembraneBERT, especially in terms of Sensitivity. This suggests that kNN’s instance-based learning approach synergizes well with these PLM’s feature representations. The CNN, when used with ProtBERT-BFD, achieves the highest MCC on independent test sets, indicating a well-balanced performance across classes.
|
39572876_p48
|
39572876
|
Classifier performance
| 3.034539 |
biomedical
|
Study
|
[
0.5393747091293335,
0.0008032746845856309,
0.45982205867767334
] |
[
0.9260681867599487,
0.07207667082548141,
0.001483039348386228,
0.0003721072571352124
] |
en
| 0.999997 |
The kNN classifier demonstrates the most stable performance across different PLMs, exhibiting the least variance in CV metrics. This stability makes kNN a robust choice for protein classification tasks. Conversely, MembraneBERT shows significant inconsistencies between CV and independent test set performances, particularly in Sensitivity, warranting further investigation into its generalization capabilities.
|
39572876_p49
|
39572876
|
Stability and consistency
| 3.701668 |
biomedical
|
Study
|
[
0.9979816675186157,
0.00020783065701834857,
0.001810496556572616
] |
[
0.9901397824287415,
0.008325747214257717,
0.0013777486747130752,
0.000156777270603925
] |
en
| 0.999997 |
The combination of ProtBERT with SVM achieves the highest accuracy (98.24 %) and MCC among independent tests, suggesting this pairing as a promising configuration for further exploration in protein classification tasks.
|
39572876_p50
|
39572876
|
Best individual performance
| 3.469708 |
biomedical
|
Study
|
[
0.9974623918533325,
0.000261755776591599,
0.0022758205886930227
] |
[
0.989933431148529,
0.008745463564991951,
0.0011359708150848746,
0.00018508944776840508
] |
en
| 0.999997 |
To further validate our approach and assess its generalization capabilities, we evaluated our models on the newly curated dataset, DS-Cv2. Table 5 presents the performance metrics for various classifiers, including CNN, and protein language models on this updated dataset.
|
39572876_p51
|
39572876
|
Performance analysis on updated dataset (DS-Cv2)
| 2.461089 |
biomedical
|
Study
|
[
0.9614959359169006,
0.0008152949158102274,
0.037688784301280975
] |
[
0.9923297762870789,
0.0068996320478618145,
0.0005643280455842614,
0.00020622051670216024
] |
en
| 0.999997 |
The results on DS-Cv2 demonstrate the robustness and generalization capabilities of our approach across different datasets. CNN-based models generally outperform classical machine learning classifiers, particularly when combined with ProtBERT and ProtBERT-BFD. Notably, the ProtBERT-BFD + CNN combination achieves the highest performance on the independent test set, with an MCC of 0.9492 and ROC AUC of 0.9968, underscoring the effectiveness of combining pre-trained language models with deep learning architectures for ion channel classification.
|
39572876_p52
|
39572876
|
Performance analysis on updated dataset (DS-Cv2)
| 4.109998 |
biomedical
|
Study
|
[
0.9990907907485962,
0.00025979371275752783,
0.0006494217086583376
] |
[
0.9993010759353638,
0.0002576217520982027,
0.0003935332642868161,
0.00004764255209011026
] |
en
| 0.999996 |
MembraneBERT shows strong performance with classical classifiers, often outperforming ProtBERT and ProtBERT-BFD. However, its performance with CNN is unexpectedly lower, particularly in terms of sensitivity on the independent test set, suggesting that MembraneBERT’s embeddings may have characteristics better suited to classical machine learning approaches.
|
39572876_p53
|
39572876
|
Performance analysis on updated dataset (DS-Cv2)
| 2.00211 |
other
|
Study
|
[
0.3627794682979584,
0.0009996090084314346,
0.6362208724021912
] |
[
0.5007438659667969,
0.49333691596984863,
0.004778591450303793,
0.0011406547855585814
] |
en
| 0.999996 |
Among classical classifiers, Random Forest (RF) and Logistic Regression (LR) show the most consistent performance across different PLMs, indicating their robustness to variations in input representations. SVM, on the other hand, exhibits high variability in performance, particularly with ProtBERT-BFD and MembraneBERT.
|
39572876_p54
|
39572876
|
Performance analysis on updated dataset (DS-Cv2)
| 2.827787 |
biomedical
|
Study
|
[
0.5897376537322998,
0.0006003935704939067,
0.4096619486808777
] |
[
0.5452210307121277,
0.4359435737133026,
0.018267136067152023,
0.0005682673654519022
] |
en
| 0.999996 |
We observed some discrepancies between cross-validation and independent test performance, particularly for sensitivity. This suggests that while the models generally generalize well, there might be some overfitting or dataset-specific characteristics influencing the results. Notably, all models achieve very high specificity ( > 99 %) on the independent test set, indicating excellent performance in correctly identifying non-ion channel proteins.
|
39572876_p55
|
39572876
|
Performance analysis on updated dataset (DS-Cv2)
| 3.886142 |
biomedical
|
Study
|
[
0.9990054965019226,
0.00021112911053933203,
0.0007833891431801021
] |
[
0.9989970326423645,
0.0005044671124778688,
0.0004419023171067238,
0.00005658221198245883
] |
en
| 0.999996 |
The overall improved performance on DS-Cv2 compared to the original dataset (DS-C) can be attributed to several factors. First, DS-Cv2 contains significantly more samples, providing more diverse and representative training data. Second, the updated dataset incorporates the most recent protein annotations, potentially offering more accurate and refined information for classification. Lastly, the new dataset may provide a more balanced representation of different types of ion channels and membrane proteins, leading to improved generalization.
|
39572876_p56
|
39572876
|
Performance analysis on updated dataset (DS-Cv2)
| 3.904871 |
biomedical
|
Study
|
[
0.9990391731262207,
0.00014143093721941113,
0.0008193990797735751
] |
[
0.9965642094612122,
0.0024503369349986315,
0.0009096658322960138,
0.00007586486026411876
] |
en
| 0.999997 |
The ROC curves illustrated in Figure 8 provide a visual comparison of the performance of our models in distinguishing ion channels from non-ion channels. All three models – TooT-BERT-CNN-C, ProtBERT, and MembraneBERT – demonstrate exceptional performance, with curves that hug the top-left corner of the plot, indicating high true positive rates even at low false positive rates. TooT-BERT-CNN-C, with an AUC of 0.9968, shows a slight but noticeable superiority over the other models, particularly in the lower false positive rate range. This suggests that TooT-BERT-CNN-C maintains high sensitivity without sacrificing specificity, a crucial characteristic for reliable ion channel classification. ProtBERT and MembraneBERT also show strong performance with AUCs of 0.9903 and 0.9936 respectively, but their curves fall slightly below that of TooT-BERT-CNN-C, especially in the early stages of the curve. All models significantly outperform the random classifier baseline (represented by the diagonal line), underscoring the effectiveness of our approach in leveraging protein language models for this classification task. The minimal differences between these high-performing models highlight the robustness of our methodology across different protein language model architectures, while also demonstrating the incremental improvement achieved by TooT-BERT-CNN-C.
|
39572876_p57
|
39572876
|
Performance analysis on updated dataset (DS-Cv2)
| 4.181207 |
biomedical
|
Study
|
[
0.9993700385093689,
0.0003278286021668464,
0.0003021830343641341
] |
[
0.9982993006706238,
0.0002347750123590231,
0.0013884931104257703,
0.00007753052341286093
] |
en
| 0.999997 |
Figure 9 presents confusion matrices for our binary classification models: TooT-BERT-C and TooT-BERT-CNN-C. TooT-BERT-C achieves 46 True Positives (TP), 848 True Negatives (TN), 2 False Positives (FP), and 14 False Negatives (FN). TooT-BERT-CNN-C slightly outperforms this with 45 TP, 850 TN, 0 FP, and 15 FN, primarily due to its increased TN and reduced FP counts.
|
39572876_p58
|
39572876
|
Comparison to state-of-the-art
| 2.7678 |
biomedical
|
Study
|
[
0.6201242208480835,
0.0007702382281422615,
0.3791055381298065
] |
[
0.9450390338897705,
0.053290270268917084,
0.0013414840213954449,
0.0003291191824246198
] |
en
| 0.999996 |
Table 6 compares TooT-BERT-CNN-C with three established methodologies: DeepIon , MFPS_CNN , and TooT-BERT-C . TooT-BERT-CNN-C consistently outperforms these approaches across most metrics. On the independent test set, it achieves the highest scores in accuracy (98.35 %), specificity (100 %), and MCC (0.86). In cross-validation, it leads in accuracy (99.39 %) and MCC (0.95), while matching the highest specificity (99.82 %).
|
39572876_p59
|
39572876
|
Comparison to state-of-the-art
| 2.679451 |
other
|
Study
|
[
0.3025384247303009,
0.0007000654586590827,
0.6967615485191345
] |
[
0.9530059695243835,
0.04483356699347496,
0.0017785070231184363,
0.00038200151175260544
] |
en
| 0.999996 |
The superior performance of TooT-BERT-CNN-C can be attributed to its hybrid architecture, which combines BERT-based embeddings with a CNN classifier. This allows it to capture both context-sensitive sequence nuances and hierarchical data representations effectively.
|
39572876_p60
|
39572876
|
Comparison to state-of-the-art
| 1.701625 |
other
|
Other
|
[
0.13870792090892792,
0.0010198340751230717,
0.860272228717804
] |
[
0.05262775719165802,
0.9449066519737244,
0.0017115016235038638,
0.0007541858358308673
] |
en
| 0.999996 |
A McNemar’s test comparing TooT-BERT-CNN-C and TooT-BERT-C yields a p -value of 0.0625, suggesting a statistically significant improvement in performance.
|
39572876_p61
|
39572876
|
Comparison to state-of-the-art
| 3.061962 |
biomedical
|
Study
|
[
0.9437452554702759,
0.0009370105108246207,
0.055317748337984085
] |
[
0.9826346039772034,
0.016389155760407448,
0.0008162524900399148,
0.00016000693722162396
] |
en
| 0.999997 |
In this research, we have made significant strides in advancing the accurate classification of ion channels and non-ion channels, a challenge with profound implications for both biology and medicine. Building on our prior work with TooT-BERT-C, we extended our investigation to encompass a broader array of classical classifiers and introduced a Convolutional Neural Network for comparative assessment.
|
39572876_p62
|
39572876
|
Conclusions
| 3.513431 |
biomedical
|
Study
|
[
0.9987528324127197,
0.0001381069450872019,
0.0011091041378676891
] |
[
0.9943121671676636,
0.004511526320129633,
0.0010631321929395199,
0.00011322915088385344
] |
en
| 0.999998 |
Our empirical findings substantiate that the newly proposed method, TooT-BERT-CNN-C, exceeds both the state-of-the-art and our preceding approach in performance metrics. On our original dataset, we observed a boost in the Matthews Correlation Coefficient from 0.8486 to 0.8584, along with an uptick in accuracy rates from 98.24 % to 98.35 %. Even more impressively, on our newly curated dataset DS-Cv2, the ProtBERT-BFD + CNN combination achieved an MCC of 0.9492 and ROC AUC of 0.9968 on the independent test set. These results not only affirm the efficacy of our methodology in discerning ion channels from non-ion channels but also underscore the promise of fusing pre-trained language models with deep learning architectures for this application. The superior performance on the larger, more recent DS-Cv2 dataset further validates the robustness and generalizability of our approach.
|
39572876_p63
|
39572876
|
Conclusions
| 4.109606 |
biomedical
|
Study
|
[
0.9992971420288086,
0.00026508362498134375,
0.00043783578439615667
] |
[
0.9990487694740295,
0.00020999003027100116,
0.000688306288793683,
0.000052938932640245184
] |
en
| 0.999996 |
Looking ahead, several promising avenues for future work emerge from our findings. We envision exploring alternative pre-trained language models and deep learning architectures to potentially further improve performance. Additionally, investigating feature engineering techniques could provide deeper insights into the molecular characteristics that distinguish ion channels from other membrane proteins. The development of a web server stands out as a crucial next step, as it would make our tool more accessible to the broader scientific community. This would involve creating a user-friendly interface, implementing backend infrastructure, integrating database functionality, ensuring scalability, and implementing robust security measures. Furthermore, extending our approach to more fine-grained classification tasks, such as distinguishing between different types of ion channels, could enhance the utility of our tool. Integrating structural information, where available, also holds potential for enhancing prediction accuracy.
|
39572876_p64
|
39572876
|
Conclusions
| 4.016285 |
biomedical
|
Study
|
[
0.9992533326148987,
0.0001643374125706032,
0.0005823167157359421
] |
[
0.9258120656013489,
0.03851044923067093,
0.03532720357179642,
0.0003503253392409533
] |
en
| 0.999996 |
Based on our findings, we recommend that future research in this field focus on further exploring the synergy between protein language models and deep learning architectures, as this combination has shown exceptional promise in our study. Investigating the specific features learned by our models, particularly the CNN, could yield valuable insights into the molecular signatures of ion channels, potentially informing our understanding of their structure and function beyond mere classification.
|
39572876_p65
|
39572876
|
Conclusions
| 4.035682 |
biomedical
|
Study
|
[
0.9996100068092346,
0.00011179522698512301,
0.00027823614072985947
] |
[
0.994172990322113,
0.0025858741719275713,
0.0031369258649647236,
0.00010417451267130673
] |
en
| 0.999997 |
In conclusion, our work represents a significant advancement in computational methods for ion channel classification. By leveraging the power of protein language models and deep learning, we have developed a highly accurate tool that could accelerate research in ion channel biology and potentially aid in drug discovery efforts targeting these crucial membrane proteins. The implementation of these recommendations could further enhance the utility and impact of this approach in the field of bioinformatics and molecular biology.
|
39572876_p66
|
39572876
|
Conclusions
| 3.985893 |
biomedical
|
Study
|
[
0.9996904134750366,
0.0001584007841302082,
0.0001512612943770364
] |
[
0.98046875,
0.004926870111376047,
0.014364403672516346,
0.0002399603254161775
] |
en
| 0.999998 |
The use of multiple-choice questions (MCQs) to assess higher order learning is standard practice for assessing clinical knowledge in the health professions . However, the creation of such questions is not straightforward, often requiring ‘Item Writing Workshops’ to train staff in the creation of clinical scenarios for such questions . There is a need to expand the use of higher order MCQs, beyond the evaluation of clinical scenarios and into medical science education more generally, where assessments of ‘higher order learning’ may be currently given in the form of written coursework such as essays and dissertations. These assessments tend to show limited coverage of the curriculum and are vulnerable to a number of forms of misconduct, such as plagiarism , outsourcing to commercial writers , and the use of Chatbot AI tools such as ChatGPT .
|
PMC11698704_p0
|
PMC11698704
|
Introduction
| 3.470967 |
biomedical
|
Other
|
[
0.9336475729942322,
0.017949262633919716,
0.04840317368507385
] |
[
0.008609985001385212,
0.9760575890541077,
0.014742548577487469,
0.0005898000672459602
] |
en
| 0.999998 |
Many authors have written guidance for the use of MCQs to assess higher order learning, and these were recently reviewed into a set of guidance principles for educators wishing to write their own questions . A key feature of MCQs that assess higher-order learning is the use of problem-type scenarios, which the learner then solves, rather than the recall of a standalone fact . The scenario should contain a lot of information, as is the case in real-world problems, and the test-taker has to critically appraise the information to identify key relevant features . Another feature of MCQs which assess higher order learning is the use of ‘assumed knowledge’; i.e., the student is required to have specific subject knowledge or skills which are missing from the problem scenario and thus can act as a cognitive bridge between the problem/question and the answer options . Students who do not have this knowledge will not be able to answer the question.
|
PMC11698704_p1
|
PMC11698704
|
Introduction
| 1.966552 |
other
|
Other
|
[
0.023813553154468536,
0.0012697974452748895,
0.9749166369438171
] |
[
0.006937488913536072,
0.9798021912574768,
0.012681746855378151,
0.0005785672110505402
] |
en
| 0.999997 |
Higher order learning is usually defined with reference to Bloom’s taxonomy or some other hierarchy (e.g. [ 1 , 2 , 9 , 11 – 14 ]). Bloom’s taxonomy is usually presented as a hierarchy of verbs for the creation of learning outcomes, with the principle that outcomes created using verbs from the base of the hierarchy (e.g. list, recall) are ‘lower order’ learning, whereas verbs at the higher end of the taxonomy (e.g. ‘evaluate’, ‘justify’) represent ‘higher order learning’ . However, Bloom’s taxonomy has faced many decades of criticism from multiple angles, including a concern that it cannot meaningfully identify ‘higher order learning’ [ 18 – 21 ], and recent studies in the UK and the USA have shown that the presentation of Bloom’s taxonomy is remarkably inconsistent between universities and other sources. That is, one presentation of Bloom’s may place a verb at the very bottom of the taxonomy, whereas another will place that same verb at the very top .
|
PMC11698704_p2
|
PMC11698704
|
Introduction
| 1.072823 |
other
|
Other
|
[
0.004714627750217915,
0.0004026966053061187,
0.9948827028274536
] |
[
0.09143451601266861,
0.8639000654220581,
0.04209611937403679,
0.002569297095760703
] |
en
| 0.999998 |
Despite these criticisms, it is still common for Bloom’s taxonomy to be used as the reference point when defining higher order assessment items. One approach is to ask participants (e.g. academics or students) to assign assessment items to the relevant level of Bloom’s taxonomy, or to identify relevant action verbs when writing the items . These sorts of validations were undertaken in the studies which were reviewed to develop the higher-order item writing guidance being tested here . However, this sort of validation normally requires collapsing the six tiers of the taxonomy into just two or three, and then asking subject experts to make subjective ratings about the level of learning which is being assessed by the question. Even then it has been repeatedly shown that academics find it difficult to do this reliably (e.g. the ratings of an assessment item as ‘higher’ or ‘lower order’ will vary between academics) [ 24 – 27 ]. Other taxonomies of learning have also been used to classify learning by cognitive level; for example, the ‘Structure of Observed Learning Outcomes’ (SOLO) Taxonomy, popular in higher education in the UK, has been used to design MCQs which aim at assessing ‘higher order’ learning , while in North America, clinical question banks may be classified into 1st, 2nd, and 3rd orders, with 1st-order questions assessing factual recall, while 3rd-order questions require critical thinking and problem-solving . Again though, while there are guidelines for the creation of questions which map to these levels, there is a paucity of objective evidence showing that the assigning of questions to these levels is reliable.
|
PMC11698704_p3
|
PMC11698704
|
Introduction
| 3.630408 |
other
|
Study
|
[
0.31438255310058594,
0.001341387745924294,
0.6842761039733887
] |
[
0.7603545188903809,
0.06384653598070145,
0.17506709694862366,
0.0007318853167816997
] |
en
| 0.999998 |
An alternate approach is to generate objective data about the ability of students to correctly answer the questions, dependent on their level of expertise, on the basis that students with existing lower order knowledge will find it easier to answer the higher-order questions than subject-novices who do not have that knowledge. These hypotheses are supported by some prior research (e.g. ), but still the majority of research in this area appears to rely on subjective ratings of question difficulty. Here then, we tested guidance designed to help educators write higher-order MCQs , by generating objective data on the ability of subject novices vs experts to answer questions which had been written using the guidance.
|
PMC11698704_p4
|
PMC11698704
|
Introduction
| 2.477122 |
other
|
Study
|
[
0.21372109651565552,
0.0007266393513418734,
0.7855523824691772
] |
[
0.9767720103263855,
0.02239721454679966,
0.0005957597168162465,
0.00023509903985541314
] |
en
| 0.999999 |
This guidance has been published previously, in a detailed form , based on a number of papers describing the use of MCQs to assess higher order learning [ 10 , 12 , 14 , 31 – 46 ]. A summary of the guidelines is shown in Table 1 . Examples of higher order questions written using the guidance, and their lower order equivalents, are given in ESM Appendix 1 . Table 1 Summary of principles for the creation of higher-order bridge MCQs. Derived from • Start with a lower order MCQ that assesses factual knowledge. Identify an existing question, or write a simple MCQ that tests factual knowledge • Problem scenario . (Re)write the question stem into a problem that needs to be solved rather than a question to be answered. One simple way to do this is to write a scenario that describes the correct answer, but in non-technical terms • Identify assumed knowledge (bridge). The student should have some prior knowledge in order to answer the question, without which the question cannot be answered. This prior knowledge must be relevant to interpreting both the question and the answer options. One way to identify assumed knowledge is to identify a sequence of steps/knowledge needed to solve a problem, and take out those in the middle • Use common language . Using everyday language, as a more authentic problem presentation. Avoid technical terms, as these can signpost correct answers. Where technical terms are unavoidable, try to put them in only one of either the scenario, or the answer options • Use active answers . Having the answer options as actions, or describing actions, makes them less likely to be a list of facts • Use annotated images. (optional) • Number of answers . Increase these to ~ 8, to allow for more granularity in the answer options
|
PMC11698704_p5
|
PMC11698704
|
Guidance for the Creation of Higher Order MCQs
| 3.052684 |
other
|
Other
|
[
0.12482959032058716,
0.0028497802559286356,
0.8723205924034119
] |
[
0.03609367832541466,
0.9496428966522217,
0.013631153851747513,
0.0006323098787106574
] |
en
| 0.999997 |
We conducted two different sets of experiments to compare questions written using the guidance with their lower order counterparts. A summary of the key differences between these experiments is given in Table 2 . Table 2 Summary of the two sets of experiments conducted here Feature Experiment 1 Experiment 2 Aim Pilot proof of principle Full testing of guidance Participants Novices only Novices vs experts Lower order questions Author generated Existing questions from textbooks Payment Flat fee Flat fee plus bonus per question Answer options Different between higher order Consistent between higher order and lower order Subject Neuroscience Genetics
|
PMC11698704_p6
|
PMC11698704
|
Guidance for the Creation of Higher Order MCQs
| 1.977988 |
other
|
Study
|
[
0.484959214925766,
0.0011095955269411206,
0.5139312148094177
] |
[
0.9122563600540161,
0.08565524965524673,
0.0015029394999146461,
0.0005853998009115458
] |
en
| 0.999998 |
All studies were carried out using Qualtrics surveys with student participants recruited via the online labour market Prolific ( www.Prolific.com ). The fee was at an estimated hourly rate of £8, advised by Prolific to be a ‘good’ rate at the time of the study. Participants were screened into expert or novice groups on the basis of the subjects being studied at university. The full list of subjects which were used to screen participants for each study is shown in ESM Appendix 1 . Before beginning the study, all participants were given information about their data protection rights, their right to withdraw at any time, and a contact email for any questions. We were not aware of any similar studies in the literature which might give us a reasonable expectation of the possible sizes of any differences between the experimental groups, and so we were unable to undertake a meaningful power analysis and sample size calculation. Thus, the sample sizes are based on the experience of the authors.
|
PMC11698704_p7
|
PMC11698704
|
Guidance for the Creation of Higher Order MCQs
| 2.053921 |
biomedical
|
Study
|
[
0.7537810206413269,
0.0016372912796214223,
0.2445816695690155
] |
[
0.9852805733680725,
0.014035380445420742,
0.000437576585682109,
0.0002465300785843283
] |
en
| 0.999997 |
Ethical Approval.
|
PMC11698704_p8
|
PMC11698704
|
Guidance for the Creation of Higher Order MCQs
| 1.151377 |
biomedical
|
Other
|
[
0.7707632780075073,
0.006955417804419994,
0.22228138148784637
] |
[
0.024311712011694908,
0.971585750579834,
0.0023844668176025152,
0.0017181608127430081
] |
en
| 0.999996 |
Both experiments were approved by the ethics committee of the Swansea University Medical School ref SUMS RESC 2022-0042A.
|
PMC11698704_p9
|
PMC11698704
|
Guidance for the Creation of Higher Order MCQs
| 0.765302 |
biomedical
|
Other
|
[
0.7846062183380127,
0.0068980189971625805,
0.20849579572677612
] |
[
0.013814649544656277,
0.9832941293716431,
0.0013829084346070886,
0.0015083085745573044
] |
en
| 0.999996 |
An initial study was conducted to determine whether subject matter novices could meaningfully answer lower order questions, under time-limited conditions. In this same experiment, we then evaluated whether this ability would be reduced when the questions were rewritten using the guidance from Table 1 . Questions were tested on novice students, meaning they were studying subjects unrelated to the subject matter of the questions (neuroscience). Each question was asked in both higher and lower-order formats, and under both closed book and open book conditions.
|
PMC11698704_p10
|
PMC11698704
|
Experiment 1. Novices Only, Proof of Principle
| 2.098161 |
other
|
Study
|
[
0.27434322237968445,
0.0011596905533224344,
0.7244970798492432
] |
[
0.9824714064598083,
0.016641484573483467,
0.0006149365217424929,
0.00027221933123655617
] |
en
| 0.999996 |
We evaluated the performance of the questions under two conditions: ‘closed book’ wherein participants were asked to try and answer the questions without referral to any other sources, and then an ‘open book’ condition, wherein students were encouraged to ‘cheat’ by any means, for example by using Google or Alexa. Participants were asked to ‘cheat’ rather than to use ‘an open book condition’ on the basis that this would be more familiar terminology and so easier for the participants to correctly follow the instructions. To minimise confusion in the instructions for participants, each participant was asked to answer in only one condition (closed or open book), but was asked questions from both formats (lower order or higher order), although the format difference was not explained. Thus, a participant only saw one version of an individual question. The same questions were asked, in the same order, in each group, but in different formats (higher order or lower order). Therefore, the experimental groups were as follows: Higher-lower (closed book) Lower-higher (closed book) Higher-lower (open book) Lower-higher (open book)
|
PMC11698704_p11
|
PMC11698704
|
Procedure
| 2.178875 |
other
|
Study
|
[
0.18059960007667542,
0.001152348006144166,
0.8182479739189148
] |
[
0.9882705211639404,
0.011063582263886929,
0.00045401969691738486,
0.00021186275989748538
] |
en
| 0.999995 |
Thus, groups 1 and 3 would start with Q1 in a higher-order format, followed by Q2 in a lower-order format, then Q3 in a higher order format, and so on. Groups 2 and 4 would start with Q1 in a lower-order format, Q2 in a higher order format, and so on. Dropout and motivation were a concern, since it seemed likely that participants would be unable to answer the questions, and so the study was conducted in two parts: an initial proof-of-principle pilot evaluated the performance of 12 questions, and then the second part evaluated the performance of 5 further questions. Twelve participants were recruited to each group, except for group 3 which had 13 participants in the initial pilot, due to a technical error. The tasks on Prolific were set so that any one participant could only be recruited to one study.
|
PMC11698704_p12
|
PMC11698704
|
Procedure
| 2.074012 |
biomedical
|
Study
|
[
0.7856848835945129,
0.0025334295351058245,
0.21178168058395386
] |
[
0.9803186058998108,
0.01917167939245701,
0.00026102690026164055,
0.00024865090381354094
] |
en
| 0.999995 |
The task was advertised on Prolific as ‘A Study about University Assessment Methods’. Participants were given the instructions shown in Fig. 1 . Fig. 1 Instructions given to participants in A the closed book condition and B the open book condition experiment of Experiment 1
|
PMC11698704_p13
|
PMC11698704
|
Procedure
| 1.495531 |
other
|
Study
|
[
0.11284276098012924,
0.0010240132687613368,
0.8861331939697266
] |
[
0.764918863773346,
0.23207484185695648,
0.001895326655358076,
0.0011109784245491028
] |
en
| 0.999999 |
The participants were then given a maximum of 90 s to answer each question, after which the instrument automatically forwarded on to the next question. At the end of the survey, the purpose of the study was explained in more detail and those who had participated in the open-book group were asked what methods they had used.
|
PMC11698704_p14
|
PMC11698704
|
Procedure
| 1.91347 |
biomedical
|
Study
|
[
0.791384756565094,
0.0035332846455276012,
0.20508193969726562
] |
[
0.9484978318214417,
0.05016665533185005,
0.0007889086264185607,
0.000546592113096267
] |
en
| 0.999998 |
The first experiment used both lower and higher order questions written by the authors, which potentially leads to a conflict when we also designed the research. Experiment 2 was designed to test the guidance in a more general context that reflects the practical application of the guidance, by starting with lower-order questions that had not been written by the authors, but which subject novices could obtain the correct answer through simple Googling. These questions were identified as described below, and then rewritten by the authors following the guidance in Table 1 . We hypothesised that, once rewritten subject experts would still be able to answer the questions, but novices would not, even in the open book condition. In Experiment 1, participants were also not provided with any specific motivation to find the correct answer in the open book condition. We did not want participants to simply give up, especially since this seemed more likely for novices and so could artificially bias the data. This was addressed in Experiment 2 by giving participants a small financial incentive for answering correctly under the open book condition. A final difference between the two studies was that, in Experiment 1, the answer options were different between some of the lower and higher order formats of the same question, in accordance with the guidance (e.g. to make the answer options active). Here we retained the same answer options with the lower and higher order versions, including keeping the same correct answer. This was to ensure that the revised questions were assessing the same learning outcome as the original lower-order question, although it did present a challenge with utilising some aspects of the guidance (e.g., to make the answer options active, and to present them in plain language).
|
PMC11698704_p15
|
PMC11698704
|
Experiment 2. Novice vs Expert, Testing of Full Guidance
| 3.229518 |
other
|
Study
|
[
0.3929925560951233,
0.0012162372004240751,
0.6057912111282349
] |
[
0.9950976967811584,
0.004339596256613731,
0.0004318761348258704,
0.00013076070172246546
] |
en
| 0.999998 |
A pilot study was undertaken to identify ‘lower order’ questions which novices could only answer under open-book conditions. Twenty questions were selected from the introductory chapter of a genetics textbook and from past-papers of the United Kingdom Biology A-level exam (these are the exams taken by students aged ~ 18, in part as the basis for entry to higher education). They were selected by the authors on the basis that they should be questions whose answers would be reasonably familiar to current undergraduates studying the life sciences.
|
PMC11698704_p16
|
PMC11698704
|
Experiment 2. Novice vs Expert, Testing of Full Guidance
| 2.349405 |
biomedical
|
Study
|
[
0.9767041206359863,
0.0007158538792282343,
0.022580064833164215
] |
[
0.9740144610404968,
0.02500285767018795,
0.0006683385581709445,
0.0003143545472994447
] |
en
| 0.999996 |
Participants were then given the instructions shown in Fig. 2 . Fig. 2 Instructions given to all participants at A the beginning (closed book) and B halfway (the open book condition) of Experiment 2. For the second part of Experiment 2, only 10 questions were used, and so the bonus payments and instructions identified in section B were modified accordingly
|
PMC11698704_p17
|
PMC11698704
|
Experiment 2. Novice vs Expert, Testing of Full Guidance
| 1.641079 |
other
|
Study
|
[
0.22393235564231873,
0.0013671470806002617,
0.774700403213501
] |
[
0.8523551225662231,
0.14518190920352936,
0.0016186351422220469,
0.0008443325641565025
] |
en
| 0.999996 |
Twenty participants were recruited to each of four groups. This experiment also included two attention-check questions, one in each condition. These were questions which appeared to be regular multiple-choice questions formatted in the same way as others in the study, but where the question stem included an instruction to select a specific answer. In keeping with the Prolific.co guidance on attention checks, participants who failed both attention checks by selecting the incorrect answer were not paid and their data were not included. Additional participants were then added as required. Novices, closed book Novices, open book Experts, closed book Experts, open book
|
PMC11698704_p18
|
PMC11698704
|
Experiment 2. Novice vs Expert, Testing of Full Guidance
| 1.815146 |
other
|
Study
|
[
0.2930397689342499,
0.0016378520522266626,
0.7053224444389343
] |
[
0.9736087322235107,
0.025463268160820007,
0.000551986217033118,
0.000375960924429819
] |
en
| 0.999998 |
Using these pilot data, we identified questions for further development, based upon the following criteria: Large difference in performance on open book vs closed book for novices (i.e. questions in which the open book condition allowed the participants to score highly) Experts scored highly under both conditions Experts scored higher than non-subject specialists under the closed book condition, to ensure the expertise tested was not common knowledge.
|
PMC11698704_p19
|
PMC11698704
|
Experiment 2. Novice vs Expert, Testing of Full Guidance
| 2.024747 |
biomedical
|
Study
|
[
0.5636958479881287,
0.0018277909839525819,
0.43447640538215637
] |
[
0.8773193955421448,
0.12124419957399368,
0.0008853752515278757,
0.000550937547814101
] |
en
| 0.999998 |
The precise metrics for identifying questions were not formalised—each question was considered individually by both authors, and selected questions were then drafted into the higher order format using the guidance in Table 1 . Both authors drafted higher-order questions and then a final version of each question was agreed through discussion. The experiment was then repeated but with ten higher-order questions.
|
PMC11698704_p20
|
PMC11698704
|
Experiment 2. Novice vs Expert, Testing of Full Guidance
| 1.911914 |
biomedical
|
Study
|
[
0.5150108933448792,
0.0020196966361254454,
0.4829694330692291
] |
[
0.9109333753585815,
0.08658196032047272,
0.001706180744804442,
0.0007784454501233995
] |
en
| 0.999996 |
For both experiments, analysis was undertaken by question, with the dependent variable being the percentage of participants who answered a question correctly. Specific statistical tests are identified in the relevant section of the results.
|
PMC11698704_p21
|
PMC11698704
|
Analysis
| 2.003677 |
biomedical
|
Study
|
[
0.9066656231880188,
0.0014611309161409736,
0.09187326580286026
] |
[
0.9711088538169861,
0.026429856196045876,
0.0020383712835609913,
0.00042297542677260935
] |
en
| 0.999999 |
For the methods used by students in the open book condition, a simple quantitative content analysis was performed on the very brief free-text comments left by participants.
|
PMC11698704_p22
|
PMC11698704
|
Analysis
| 1.808187 |
other
|
Study
|
[
0.3365265429019928,
0.0014834852190688252,
0.6619900465011597
] |
[
0.7946338653564453,
0.20276957750320435,
0.0017424577381461859,
0.0008541884017176926
] |
en
| 0.999997 |
Example questions from Experiment 1 and Experiment 2 are shown in ESM Appendix 1 (the full set of questions is available upon reasonable request, from the corresponding author). Statistical analysis and figure creation were undertaken using GraphPad Prism V10 (San Diego, CA).
|
PMC11698704_p23
|
PMC11698704
|
Analysis
| 1.625175 |
biomedical
|
Study
|
[
0.9740927219390869,
0.001251153415068984,
0.024656057357788086
] |
[
0.6690866351127625,
0.32520368695259094,
0.004114421084523201,
0.0015952893299981952
] |
en
| 0.999995 |
There was a clear effect of condition, where an average of 53.4% of participants were able to answer the questions in the open-book condition when the questions were written in the lower order format, compared to 18.6% when the questions where in the higher order format. In the closed book condition, an average of 24% of participants answered correctly in the lower order format with 13.2% answering correctly in the higher format. The percentage of participants able to answer each question in the lower order, open-book format was significantly higher than for every other format/condition when analysed using a two-way repeated measures ANOVA, with percentage of participants answering correctly as the dependent variable, and test format (closed book vs open book) and question format (higher order vs lower order) as the conditions. There was a significant effect of question format ( F (1, 16) = 25.03, P = 0.0002), and test format ( F (1, 16) = 14.59, P < 0.0001) and a significant interaction between the two ( F (1, 16) = 6.95, P = 0.0018). Post hoc Bonferroni tests revealed a significant difference between the lower-order, open-book condition, and all other conditions. No other significant differences were observed. The results are shown in Fig. 3 . Fig. 3 Rewriting lower order questions into higher order makes it harder for subject matter novices to answer. Different groups of participants were given the same question in two different formats, lower order and higher order, and under two different conditions, open book and closed book. Participants were able to answer the lower order questions in the open-book condition but were not able to answer the questions in the closed-book condition or when they were rewritten into the higher order format, even under the open-book condition. * P , 0.05 when compared to all other conditions by post hoc Bonferroni tests following two-way repeated measures ANOVA (see text for details)
|
PMC11698704_p24
|
PMC11698704
|
Experiment 1
| 4.005941 |
biomedical
|
Study
|
[
0.830172061920166,
0.0010575223714113235,
0.1687704622745514
] |
[
0.9987665414810181,
0.0008424673578701913,
0.0003373248619027436,
0.000053663414291804656
] |
en
| 0.999998 |
In 25 participants, 22 left comments, most very brief (the total corpus was 542 words). All 22 identified ‘searching the internet’ as their strategy, with 20/22 naming Google directly, and one asking ‘Siri’. In 22 participants, 6 identified the time limit as a factor which made it difficult to answer the questions. In 22 participants, 4 identified that there were questions that were more ‘difficult’ (the participants were not told that the questions were in two different formats).
|
PMC11698704_p25
|
PMC11698704
|
Methods Used in the Open-Book Condition
| 1.274176 |
other
|
Other
|
[
0.04435053840279579,
0.0010544745018705726,
0.9545950293540955
] |
[
0.38767075538635254,
0.6086754202842712,
0.002028241055086255,
0.001625568838790059
] |
en
| 0.999996 |
The results from Experiment 1 were clear: novices used Google to successfully answer questions written in a lower-order format, but this approach was not successful when the questions were in a higher-order format. However, the design of the study contained some potential confounds; (1) the participants had no obvious motivation to try and answer correctly. This was potentially more significant in the higher order condition due to the extra work required to successfully answer the question, and the increased length and complexity of the questions. (2) Both lower and higher-order questions were written by the authors. (3) There is no within-study positive control since Experiment 1 only used subject matter novices, and thus it is not clear that subject experts could still answer the higher-order questions. This is important to demonstrate that the rewritten questions remain a valid form of assessment for the subject matter content. Finally, the answer options were often completely different between the lower order and higher order question formats.
|
PMC11698704_p26
|
PMC11698704
|
Experiment 2
| 2.08983 |
other
|
Study
|
[
0.10820566117763519,
0.000832404475659132,
0.8909619450569153
] |
[
0.9880331754684448,
0.01096601877361536,
0.0007363377371802926,
0.0002644169144332409
] |
en
| 0.999997 |
Thus, in Experiment 2, we included a within-experiment positive control. We also started with existing lower order questions, in the public domain, that had not been written by either of the authors, using the original answer options and same correct answer, but with additional answer options added.
|
PMC11698704_p27
|
PMC11698704
|
Experiment 2
| 1.621156 |
other
|
Study
|
[
0.31389039754867554,
0.0016412334516644478,
0.6844683885574341
] |
[
0.9277065992355347,
0.07028059661388397,
0.0012921489542350173,
0.0007205934962257743
] |
en
| 0.999997 |
The results are shown in Fig. 4 . Novices were again able to successfully answer lower order questions by Googling the answers, an average of 73.3% answering correctly but this dropped to 23% when the questions were rewritten into a higher order format. A two-way repeated measures ANOVA was conducted, with percentage of participants answering correctly as the dependent variable, and test format (closed book vs open book) and expertise (novice vs expert) as the conditions. For lower order questions , there was a significant effect of question format ( F (1, 9) = 163.0, P < 0.0001) and expertise ( F (1, 9) = 17.09, P = 0.0025), with a significant interaction between these two conditions ( F (1, 9) = 61.91, P < 0.0001). A post hoc Bonferroni multiple comparisons test showed a significant effect of expertise in the closed-book condition but not the open-book condition . For the higher order questions , a two-way repeated measures ANOVA again showed a significant effect of question format ( F (1, 9) = 6.553, P = 0.0307) and expertise ( F (1, 9) = 18.40, P = 0.0020), but no interaction between the two conditions ( F (1, 9) = 0.2687, P = 0.6167). Fig. 4 Rewriting publicly available lower order questions into higher order questions makes it harder for novices to answer them, but experts can still answer under closed-book conditions. Different groups of participants were given the same question in two different formats, lower order and higher order, and under two different conditions, open book and closed book. Novices were able to answer the lower order questions in the open-book condition but not when the questions when rewritten into a higher order format using the guidance tested here. Experts showed good ability to answer questions under all conditions. See text for statistical analysis
|
PMC11698704_p28
|
PMC11698704
|
Experiment 2
| 4.032603 |
biomedical
|
Study
|
[
0.8333771824836731,
0.0011432432802394032,
0.16547955572605133
] |
[
0.9986107349395752,
0.0008834991604089737,
0.0004465272941160947,
0.00005919987233937718
] |
en
| 0.999996 |
If the higher order questions are truly assessing higher order learning, then they should be harder, and so the percentage of experts answering them would be lower. An average of 53.3% of experts were able to answer questions in the lower order format under closed book conditions; this dropped to 40% when rewritten into the higher order format. This same comparison was 67.5% and 51.5% in the open book format. When analysed using a paired t- test, the difference between lower and higher order was significant for both the closed book and the open book condition .
|
PMC11698704_p29
|
PMC11698704
|
Experiment 2
| 1.795548 |
other
|
Study
|
[
0.07969173043966293,
0.0007555472548119724,
0.919552743434906
] |
[
0.8495455980300903,
0.14782194793224335,
0.0017475569620728493,
0.0008848222205415368
] |
en
| 0.999997 |
Here we conducted a validation test on guidance which is designed to help educators write higher order MCQs . When questions were written using a traditional lower order format, subject-matter novices were able to Google their way to the correct answer, to the same extent as experts, even though the novices were studying subjects unrelated to the topic. When those same questions were rewritten into a higher-order format using the guidance, subject-novices found it significantly harder to answer using Google, while experts were still able to answer the questions. These findings suggest that questions written using the summary guidance in Table 1 do indeed assess higher-order learning. The summary guidance in Table 1 is largely complementary to, and intended to be used alongside, a large body of existing literature on what makes an effective multiple-choice question, regardless of whether they assess lower or higher-order learning .
|
PMC11698704_p30
|
PMC11698704
|
Discussion
| 1.932655 |
other
|
Study
|
[
0.06237373873591423,
0.00048595829866826534,
0.9371402859687805
] |
[
0.9086558222770691,
0.08907034993171692,
0.001623339019715786,
0.0006505520432256162
] |
en
| 0.999999 |
There are a number of factors and potential limitations that need to be considered when interpreting our findings.
|
PMC11698704_p31
|
PMC11698704
|
Discussion
| 1.735123 |
biomedical
|
Study
|
[
0.9663891792297363,
0.0016745172906666994,
0.03193630650639534
] |
[
0.7673903107643127,
0.21720664203166962,
0.014060904271900654,
0.0013421286130324006
] |
en
| 0.999997 |
The guidance shown in Table 1 contains a number of different elements which combine to make the question an active, problem-solving exercise. On the basis of this study, it is not currently possible to determine which, if any, of the individual elements is the most important for ensuring that an MCQ assesses ‘higher order learning’. This is the subject of ongoing work where each element is subject to a systematic experimental appraisal. These analyses may result in a revised and condensed set of guidelines which could prioritise individual elements for the creations of higher-order MCQs.
|
PMC11698704_p32
|
PMC11698704
|
Discussion
| 1.785976 |
other
|
Other
|
[
0.16529597342014313,
0.0013812233228236437,
0.8333227634429932
] |
[
0.26808497309684753,
0.7268404364585876,
0.0038105605635792017,
0.0012639894848689437
] |
en
| 0.999996 |
When using online labour markets for opinion surveys, there is potentially an issue participants could simply give random, or minimal, answers , or that the participants may in fact be ‘bots’ . Here we had objective outcome measures with right and wrong answers. Although there was no incentive to answer correctly, we did see a clear and expected difference in those experimental conditions which would indicate that participants are real, valid, and following instructions (for example between subject experts and novices).
|
PMC11698704_p33
|
PMC11698704
|
Discussion
| 1.181498 |
other
|
Study
|
[
0.032064542174339294,
0.0006308884476311505,
0.9673046469688416
] |
[
0.5433326363563538,
0.4536673426628113,
0.0015156021108850837,
0.0014844454126432538
] |
en
| 0.999997 |
Participants in the open-book conditions were instructed to ‘cheat’ by using whatever sources necessary. Even so, subject novices struggled to answer the higher order questions. However, it is important to be clear that this is not specifically a study of ‘cheating’ and that these higher-order questions are not ‘cheat-proof’, especially in the new era of ChatGPT, which can answer very complex problem-solving MCQs , and when cheating in online open-book exams is already high . Indeed, the ability of ChatGPT to answer questions written using these guidelines has been recently tested and ChatGPT answers almost all the questions correctly, apart from where novel, labelled images are included . However, these higher order questions should be more resilient to cheating in invigilated examinations, and offer a way to assess higher order learning in such exams. In this way, a supervised exam based on these questions would likely be more resistant to misconduct than other assessment formats which are currently used to assess higher order learning, such as essays and other asynchronous written coursework; these are open to multiple sources of misconduct such as plagiarism contract cheating and can be completed to a high standard using tools such as ChatGPT .
|
PMC11698704_p34
|
PMC11698704
|
Discussion
| 1.875579 |
other
|
Study
|
[
0.05525219067931175,
0.000598361948505044,
0.9441494941711426
] |
[
0.6750589609146118,
0.3200257122516632,
0.0037886120844632387,
0.0011266948422417045
] |
en
| 0.999995 |
Guidelines developed here have been tested in only two subject areas, both of which are from human physiology and medical science. It seems reasonable to propose the guidelines could be used to test higher order learning in other subjects, including those outside of medical science education. This will require careful validation and is also the subject of future work.
|
PMC11698704_p35
|
PMC11698704
|
Discussion
| 1.918521 |
biomedical
|
Other
|
[
0.9396257996559143,
0.00235421909019351,
0.05801996961236
] |
[
0.07841134071350098,
0.9124524593353271,
0.008064137771725655,
0.001072046346962452
] |
en
| 0.999997 |
In summary, we have validated guidance for item-writing, or rewriting, based on the literature for the use of MCQs to assess higher order learning. In an experimental situation, questions that were rewritten using this guidance were much more challenging to answer by simple Googling but were still answerable by students who were studying relevant subjects, although those experts found the questions harder. These findings suggest that the guidance could be used by educators and institutions to develop MCQ-based exams to assess higher order learning, to partially replace asynchronous written coursework such as essays.
|
PMC11698704_p36
|
PMC11698704
|
Discussion
| 2.060385 |
other
|
Other
|
[
0.058060288429260254,
0.0004823571362067014,
0.9414573311805725
] |
[
0.4501047432422638,
0.5428159832954407,
0.005939132999628782,
0.0011400720104575157
] |
en
| 0.999996 |
Below is the link to the electronic supplementary material. Supplementary file1 (DOCX 18 KB)
|
PMC11698704_p37
|
PMC11698704
|
Supplementary Information
| 1.036319 |
other
|
Other
|
[
0.1836632639169693,
0.0026532600168138742,
0.8136835098266602
] |
[
0.009041821584105492,
0.9892666339874268,
0.0011328585678711534,
0.0005587530322372913
] |
en
| 0.999997 |
Among cancers, colorectal cancer (CRC) ranks 3rd in males and 2nd d in females, making it a serious issue in world health. Males are more likely to develop and die from colorectal cancer than females, and the disease makes up around 10% of all cancer cases globally, according to the GLOBOCAN database . In Egypt, CRC accounts for around 3.9% of all cancer cases, ranking seventh in the area .
|
39751895_p0
|
39751895
|
Introduction
| 2.284023 |
biomedical
|
Other
|
[
0.9933033585548401,
0.001714065088890493,
0.0049825445748865604
] |
[
0.0128479627892375,
0.9701444506645203,
0.01575835607945919,
0.00124927272554487
] |
en
| 0.999996 |
For patients with locally advanced rectal cancer (LARC), surgical intervention stands as the cornerstone of treatment. However, rectal cancer is more prone to local recurrence than colon cancer, highlighting the importance of neoadjuvant chemoradiotherapy in curative treatment plans . The conventional treatment for LARC (T3-4N-ve or T1-4N + ve) involves neoadjuvant CRT and total mesorectal excision (TME) and, in many cases, adjuvant chemotherapy with fluoropyrimidine with or without oxaliplatin. Neoadjuvant treatment has been shown to improve sphincter preservation and promote tumor downstaging . Furthermore, ypT0N0, or a pathologic complete response (pCR), is attained by about 15% of patients and is linked to excellent long-term survival results . Twenty percent of patients also get tumor downstaging to ypT1/T2N0 after neoadjuvant treatment .
|
39751895_p1
|
39751895
|
Introduction
| 4.090234 |
biomedical
|
Review
|
[
0.997343122959137,
0.0018934250110760331,
0.0007634408539161086
] |
[
0.19317518174648285,
0.004451637621968985,
0.801617443561554,
0.0007557232747785747
] |
en
| 0.999995 |
Despite these advances, about 1/3 of rectal cancer patients eventually develop distant metastasis after surgery . One possible solution to this problem is adjuvant chemotherapy, which can eradicate micrometastases, decrease recurrence, and improve survival [ 11 – 13 ]. The benefit of adjuvant chemotherapy in ypT0-2N0 disease following neoadjuvant CRT and surgery is still debatable, as it does not always achieve positive results [ 14 – 17 ]. The National Comprehensive Cancer Network (NCCN) states that adjuvant chemotherapy decisions should depend on pre-treatment staging rather than post-surgical pathology. This uncertainty, combined with low compliance with the NCCN guidelines, reflects the complexity of decision-making in this setting .
|
39751895_p2
|
39751895
|
Introduction
| 4.023155 |
biomedical
|
Review
|
[
0.9934106469154358,
0.004816871602088213,
0.0017724352655932307
] |
[
0.02353423461318016,
0.007471929304301739,
0.9684364199638367,
0.0005574380047619343
] |
en
| 0.999996 |
This research aims to evaluate the effect of adjuvant chemotherapy on the survival of rectal cancer cases who were downstaged to ypT0-2N0 after neoadjuvant CRT and surgical intervention.We believe our study provides significant value for several reasons. First, while the practice of total neoadjuvant therapy (TNT) is gaining traction globally, its implementation remains highly variable across institutions, as evidenced by recent surveys from Germany and China showing diverse treatment protocols and practices . This variability underscores the importance of continuing research to refine therapeutic strategies for rectal cancer. Importantly, growing evidence suggests that some rectal cancer patients achieving ypT0-2N0 after neoadjuvant chemoradiation may fare well without additional chemotherapy, potentially avoiding overtreatment and the associated treatment-related toxicities . This is particularly relevant for elderly and frail patients with limited tolerance for intensive treatment regimens. Rather than adopting a "one-size-fits-all" approach with TNT, further research is crucial to identify patient subgroups who could benefit from de-escalated treatment strategies. Moreover, our study is the first to address this topic in the Egyptian, Arab, or African context, where healthcare infrastructure and treatment paradigms differ significantly from those in high-income countries. As no regional data currently exists, our findings can serve as a benchmark for future research and guide clinical practice in similar settings. We hope this perspective highlights the relevance of our study in enriching the global discourse on rectal cancer management and personalized treatment approaches.
|
39751895_p3
|
39751895
|
Introduction
| 4.136679 |
biomedical
|
Study
|
[
0.9983912110328674,
0.0013741357252001762,
0.0002346081892028451
] |
[
0.9971901774406433,
0.0007037415634840727,
0.001922582508996129,
0.00018357580120209605
] |
en
| 0.999998 |
This retrospective analysis was conducted at Shefaa El Orman Cancer Hospital and concentrated on patients with rectal adenocarcinoma from January 2016 to December 2020. The Institutional Review Board (IRB) at Shefaa El Orman Cancer Hospital approved the study, which was conducted according to the Declaration of Helsinki. Because of the retrospective nature of the study, informed consent was waived. However, patient confidentiality was protected by anonymizing all data that were securely stored and accessed only by authorized personnel. All analyses were performed on aggregated, anonymized data to maintain confidentiality.This study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for reporting observational studies. The completed STROBE checklist is included as supplementary material.
|
39751895_p4
|
39751895
|
Study design
| 3.859402 |
biomedical
|
Study
|
[
0.9985125660896301,
0.0011141011491417885,
0.0003732366021722555
] |
[
0.999042809009552,
0.0006173938745632768,
0.00020811018475797027,
0.00013169681187719107
] |
en
| 0.999996 |
Patients had to be at least 18 years old and diagnosed with rectal adenocarcinoma confirmed histologically to be eligible to participate. They had to have completed neoadjuvant concurrent chemoradiotherapy (CRT) and radical surgical resection, and they had to have an ECOG performance level of 0–2. Inclusion in the study was restricted to patients who had a pathological staging of ypT0-2N0 based on postoperative pathological examination of the surgical specimen.
|
39751895_p5
|
39751895
|
Criteria for participant selection
| 3.858604 |
biomedical
|
Study
|
[
0.8201266527175903,
0.1773192137479782,
0.002554137259721756
] |
[
0.8113654851913452,
0.18048956990242004,
0.0015639171469956636,
0.006581056397408247
] |
en
| 0.999997 |
The sample size was determined by the number of eligible patients available during the study period, ensuring adequate power for detecting clinically significant differences.
|
39751895_p6
|
39751895
|
Criteria for participant selection
| 2.427261 |
biomedical
|
Study
|
[
0.9899411797523499,
0.006896020378917456,
0.0031628140714019537
] |
[
0.9706546068191528,
0.028249578550457954,
0.0005067610763944685,
0.0005889240419492126
] |
en
| 0.999996 |
Data were retrieved retrospectively from hospital medical records, including both demographic and clinical characteristics. To minimize bias, consistent data collection procedures were followed, and only validated clinical records were used. Patient anonymity was maintained throughout the study by removing all identifiable information. Selection bias was minimized by including all eligible patients meeting the inclusion criteria during the study period. Demographic data included age, sex, smoking status, and family history of cancer. Clinical data collected comprised ECOG performance status, comorbidities (such as hypertension and diabetes), tumor-related details (e.g., tumor grade, distance from the anal verge), initial carcinoembryonic antigen (CEA) levels, and pretreatment clinical staging (both T and N stages).
|
39751895_p7
|
39751895
|
Data retrieval process
| 4.006942 |
biomedical
|
Study
|
[
0.9935179948806763,
0.006042999215424061,
0.00043909158557653427
] |
[
0.9986494183540344,
0.0006024883477948606,
0.0004576185019686818,
0.0002905144647229463
] |
en
| 0.999998 |
Treatment-related data were also collected, including details on neoadjuvant CRT, the type of surgery performed (abdominoperineal resection [APR] or low anterior resection [LAR]), and the surgical technique used (laparoscopic or open surgery). The administration of adjuvant chemotherapy was also recorded, including the type of regimen used (capecitabine or CAPOX/FOLFOX). Follow-up data were obtained on local recurrence, distant metastasis, and survival.
|
39751895_p8
|
39751895
|
Data retrieval process
| 3.804179 |
biomedical
|
Study
|
[
0.8569306135177612,
0.1419333964586258,
0.0011359693016856909
] |
[
0.9654029011726379,
0.02832525037229061,
0.002214462962001562,
0.004057329148054123
] |
en
| 0.999998 |
Data on certain risk factors, such as extramural venous invasion, tumor deposits, and circumferential margin status, were not consistently available due to limitations in preoperative imaging and postoperative histopathology reporting practices at the time of this study. Missing data for key variables were assessed. If data were missing, they were handled using listwise deletion to ensure accuracy in analysis.
|
39751895_p9
|
39751895
|
Data retrieval process
| 2.622297 |
biomedical
|
Study
|
[
0.9965207576751709,
0.0022391597740352154,
0.0012400582199916244
] |
[
0.9966570138931274,
0.0028356679249554873,
0.00022765975154470652,
0.00027963213506154716
] |
en
| 0.999996 |
Patients received neoadjuvant concurrent chemoradiotherapy (CRT) with either three-dimensional conformal (3DCRT) or intensity-modulated (IMRT) radiotherapy in combination with oral capecitabine. The prescribed radiation dose was 45–50.4 Gy in 25–28 daily fractions over five weeks. Patients concurrently received capecitabine 825 mg/m 2 twice daily during radiotherapy days. Surgical resection was performed 4 to 12 weeks after the end of CRT. The type of surgery—either low anterior resection (LAR) or abdominoperineal resection (APR)—was chosen depending on tumour location and patient factors. All patients had total mesorectal excision (TME) as part of the surgical procedure. Based on clinical judgment, adjuvant chemotherapy was recommended in selected patients following surgery. Adjuvant chemotherapy, usually given for four months after surgery, protocols included either capecitabine or CAPOX/FOLFOX. The CAPOX protocol included intravenous oxaliplatin (130 mg/m 2 on day 1) and oral capecitabine . The FOLFOX regimen consisted of intravenous oxaliplatin (85 mg/m 2 ), leucovorin (400 mg/m 2 ), and 5-fluorouracil (5-FU), followed by a continuous 46-h infusion of 2400 mg/m 2 5-FU, repeated every two weeks.
|
39751895_p10
|
39751895
|
Treatment protocols
| 4.074841 |
biomedical
|
Study
|
[
0.8023845553398132,
0.19577372074127197,
0.001841707737185061
] |
[
0.715155839920044,
0.25640663504600525,
0.010186150670051575,
0.018251409754157066
] |
en
| 0.999998 |
In line with hospital policy, every treatment decision was thoroughly reviewed and discussed during a multidisciplinary team (MDT) meeting. These meetings involved input from various medical professionals, including oncologists, surgeons, radiologists, pathologists, and other relevant specialists, ensuring a comprehensive and collaborative approach to patient care. The decision to administer adjuvant chemotherapy and the choice of regimen (capecitabine vs. CAPOX/FOLFOX) were determined during multidisciplinary team (MDT) meetings. These decisions were based on patient-related factors such as age, comorbidities, and performance status, as well as tumor characteristics and clinical response to neoadjuvant therapy.
|
39751895_p11
|
39751895
|
Treatment protocols
| 3.652593 |
clinical
|
Other
|
[
0.4578602612018585,
0.5383095145225525,
0.0038302442990243435
] |
[
0.19990527629852295,
0.7723535895347595,
0.006448020227253437,
0.021293150261044502
] |
en
| 0.999995 |
After completing therapy, patients had to attend regular follow-up appointments. These visits were planned every 3 months for the first 2 years, every 6 months up to 5 years, and then annually. Physical exams, carcinoembryonic antigen (CEA) tests, and imaging procedures like chest, liver, and pelvis CT scans were all part of each patient's routine assessments. A colonoscopy was performed every 3–5 years after one year, provided no abnormalities were detected.
|
39751895_p12
|
39751895
|
Monitoring and evaluation of patient outcomes
| 2.795447 |
clinical
|
Other
|
[
0.28177887201309204,
0.7151744961738586,
0.0030465854797512293
] |
[
0.33681201934814453,
0.5615302920341492,
0.006217597983777523,
0.09544004499912262
] |
en
| 0.999997 |
All statistical analyses were performed using IBM SPSS Statistics, version 26. Continuous variables were summarized as means and standard deviations (SD) for normally distributed data, or as medians and interquartile ranges (IQR) for non-normally distributed data. Categorical variables were presented as frequencies and percentages. Comparisons between groups were made using chi-square or Fisher’s exact tests for categorical variables, and continuous variables were analyzed using Mann–Whitney U tests or independent t-tests as appropriate.
|
39751895_p13
|
39751895
|
Statistical analysis
| 3.736476 |
biomedical
|
Study
|
[
0.9993982315063477,
0.00024292497255373746,
0.00035888771526515484
] |
[
0.9946948885917664,
0.004398694261908531,
0.0007949628052301705,
0.00011145430471515283
] |
en
| 0.999998 |
Survival analyses for DFS and OS were conducted using Kaplan–Meier methods, with differences between groups compared using the log-rank test. DFS was defined as the time from surgery to either local or distant recurrence, while OS was defined as the time from diagnosis to death from any cause or last follow-up. Multivariate analysis was performed using the Cox proportional hazards regression model to identify independent prognostic factors associated with DFS and OS. Hazard ratios (HR) and 95% confidence intervals (CI) were reported. A p-value of ≤ 0.05 was considered statistically significant for all analyses.
|
39751895_p14
|
39751895
|
Statistical analysis
| 4.103982 |
biomedical
|
Study
|
[
0.9991377592086792,
0.0006875252583995461,
0.00017470031161792576
] |
[
0.9986035227775574,
0.0006020373548381031,
0.0006787618040107191,
0.00011571458890102804
] |
en
| 0.999998 |
The findings were compared with prior studies to assess consistency and enhance validity. Differences were critically analyzed in the discussion section.
|
39751895_p15
|
39751895
|
Statistical analysis
| 2.010227 |
biomedical
|
Study
|
[
0.947413444519043,
0.0018222543876618147,
0.050764307379722595
] |
[
0.9358117580413818,
0.05135294049978256,
0.011908011510968208,
0.0009273569448851049
] |
en
| 0.999998 |
The study reviewed the clinicopathological features of 85 rectal cancer patients treated with neoadjuvant chemoradiotherapy (CRT) and surgery, as presented in Table 1 , 55 of whom received adjuvant chemotherapy, while 30 did not. Key findings revealed that age was a significant factor in the decision to receive adjuvant chemotherapy, with the median age of patients in the no-adjuvant group being significantly higher (60 years, IQR 48–64.75) compared to those who received adjuvant therapy (47 years, IQR 40–57.5; P = 0.006). Gender distribution, tumor location, grade, initial CEA levels, clinical T and N stages, and type of surgery were balanced between the groups, suggesting no systematic differences in these characteristics. Tumors were predominantly located in the middle rectum (54.1%), and most were grade II (90.6%). Clinical staging showed that 65.88% were T3, and 54.12% were N1. Surgical approaches included low anterior resection (LAR) in 60% of cases and abdominoperineal resection (APR) in 40%, with no significant difference between groups. Open surgery was more common (74.12%) than laparoscopic surgery (25.88%), and this distribution was consistent across the groups. Among the 85 patients included in the study, 33 (37.5%) achieved ypT0N0, 4 (4.5%) ypT1N0, and 48 (58%) ypT2N0. These proportions highlight the significant representation of ypT0N0 cases in our cohort. Table 1 Patient profile and tumor characteristics Characteristic Total ( n = 85) Adjuvant Chemotherapy ( n = 55) No Adjuvant Chemotherapy ( n = 30) P -value Age in years (median, IQR) 52 (42–61) 47 (40–57.5) 60 (48–64.75) 0.006* Male (%) 39 (45.9%) 26 (47.3%) 13 (43.3%) 0.728 Female 46 (54.12%) 29 (52.73%) 17 (56.67%) 0.728 Lower rectum (0–5 cm) 34 (40%) 22 (40%) 12 (40%) 0.753 Middle rectum (6–10 cm) 46 (54.1%) 29 (52.7%) 17 (56.7%) 0.753 Upper rectum (11–15 cm) 5 (5.9%) 4 (7.3%) 1 (3.3%) 0.753 Grade I 1 (1.2%) 0 (0%) 1 (3.3%) 0.352 Grade II 77 (90.6%) 51 (92.7%) 26 (86.7%) 0.352 Grade III 7 (8.2%) 4 (7.3%) 3 (10.0%) 0.352 Initial CEA < 5 ng/mL 22 (25.9%) 13 (23.6%) 9 (30%) 0.796 Initial CEA ≥ 5 ng/mL 13 (15.3%) 9 (16.4%) 4 (13.3%) 0.796 Unknown CEA 50 (58.8%) 33 (60%) 17 (56.7%) 0.796 Clinical T2 11 (12.94%) 5 (9.09%) 6 (20%) 0.462 Clinical T3 56 (65.88%) 38 (69.09%) 18 (60%) 0.462 Clinical T4 17 (20%) 11 (20%) 6 (20%) 0.462 Unknown Clinical T status 1 (1.18%) 1 (1.82%) 0 (0%) 0.462 Clinical N0 9 (10.59%) 6 (10.91%) 3 (10%) 0.759 Clinical N1 46 (54.12%) 31 (56.36%) 15 (50%) 0.759 Clinical N2 29 (34.12%) 17 (30.91%) 12 (40%) 0.759 Unknown Clinical T status 1 (1.18%) 1 (1.82%) 0 (0%) 0.759 LAR 51 (60%) 30 (54.55%) 21 (70%) 0.165 APR 34 (40%) 25 (45.45%) 9 (30%) 0.165 Open Surgery 63 (74.12%) 13 (23.64%) 9 (30%) 0.522 Laparoscopic 22 (25.88%) 42 (76.36%) 21 (70%) 0.522 Median Nodal Harvest (median, IQR) 13 (10–16) 14 (11–17) 12 (9–15) 0.118 CRM Involvement (%) 8% (7 patients) 7.3% (4 patients) 10% (3 patients) 0.482 Median Follow-Up (months) 46 months (IQR: 36–58) 45 months (IQR: 36–55) 44 months (IQR: 32–50) 0.645
|
39751895_p16
|
39751895
|
Clinical and tumor characteristics
| 4.182958 |
biomedical
|
Study
|
[
0.9984858632087708,
0.0012562077026814222,
0.0002579693973530084
] |
[
0.9970526695251465,
0.00031779654091224074,
0.002441599266603589,
0.000187906131031923
] |
en
| 0.999996 |
Subsets and Splits
SQL Console for rntc/test-pp-aa
The query retrieves a sample of documents that are clinical cases with an educational score above 3, providing limited analytical value.
Clinical Cases Sample
Returns a sample of 100 clinical case documents, providing a basic overview of the document type's content.