2.4. Model Training and Testing

We used 12 CiPA training drugs to train the model (quinidine, sotalol, dofetilide, bepridil, cisapride, terfenadine, chlorpromazine, ondansetron, verapamil, ranolazine, diltiazem, and mexiletine). As the number of AP shapes for each drug was 2000, the total number of AP shapes for training was 24,000. The model training worked out 100 epochs through 10-fold cross-validation to determine the optimal model with hyperparameters for assessing proarrhythmic risks of drugs. Then, we determined the hyperparameters of the final model through comparison of the classification performances from the validation sets to the training sets, which were randomly distinguished from 12 training drugs. The final model was validated using 16 CiPA test drugs: disopyramide, ibutilide, vandetanib, azimilide, clarithromycin, clozapine, domperidone, droperidol, pimozide, risperidone, asemizole, metoprolol, nifedipine, nitrendipine, tamoxifen, and loratadine. The number of AP shapes for the test was 32,000.

The proposed model was validated through a 10,000-times-repeated testing method in which we randomly extracted samples from each drug set, generating 10,000 test sets [8]. We then plotted the receiver operating curves for 10,000 test sets and evaluated model performance via calculating the area under the curve (AUC), sensitivity, specificity, and likelihood (LR) values. Each value was calculated for individual TdP-risk categories based on the 10,000 AUCs;

where TP and TN are true positives and true negatives, respectively, indicating that the model correctly answers actual positive and negative problems. An FP is a “false positive”, indicating that the model mispredicts an actual negative problem as positive. An FN is a “false negative”, representing the mispredicted case of an actual positive problem as negative. In the calculation of LR+, we set a small number, close to zero, in the denominator to prevent the result from becoming infinite.