Figure 1.
Drug differential analysis. Phase 1 (data processing): The raw dataset was preprocessed; split into a test, train, and validation set; and then fed through a feature selection procedure to produce a separate final dataset. This final dataset was our train and test set combined and only included our final features. Phase 2 (neural network development): We configured our neural network and used our final dataset to perform k-fold validation to produce metrics. During training, we iteratively optimized our neural network tuning parameters. Phase 3 (model training and testing): We then used our model configuration from Phase 2 and our train and test set from Phase 1 with our top features also from Phase 1 to train and validate a model. Phase 4 (predicted improvement analysis): We iterated through each subject of our differential analysis set. For each patient, we used our neural network with each possible drug to find the probability of remission with that drug. Once the patient had a probability of remission for each drug, the drug with the highest probability was retained. Once all 200 patients had a probability of remission, we took an average of those probabilities. This process was run five times, and the average was computed. Phase 5 (actual improvement analysis): We used k-fold validation with our entire dataset. This entire dataset is a combination of the training set and both holdout sets. The patients within the test set for each fold were used to perform the differential analysis. This analysis was conservative, as we only retained subjects if the drug for which our neural network produced the highest probability of remission was actually the drug they received in the study. We then took the average of all patients kept after all folds from the k-fold validation process.
Figure 2.
The 17 most important features in the differential treatment prediction model.