AI-assisted prediction of differential response to antidepressant classes using electronic health records

Abstract

Antidepressant selection is largely a trial-and-error process. We used electronic health record (EHR) data and artificial intelligence (AI) to predict response to four antidepressants classes (SSRI, SNRI, bupropion, and mirtazapine) 4 to 12 weeks after antidepressant initiation. The final data set comprised 17,556 patients. Predictors were derived from both structured and unstructured EHR data and models accounted for features predictive of treatment selection to minimize confounding by indication. Outcome labels were derived through expert chart review and AI-automated imputation. Regularized generalized linear model (GLM), random forest, gradient boosting machine (GBM), and deep neural network (DNN) models were trained and their performance compared. Predictor importance scores were derived using SHapley Additive exPlanations (SHAP). All models demonstrated similarly good prediction performance (AUROCs ≥ 0.70, AUPRCs ≥ 0.68). The models can estimate differential treatment response probabilities both between patients and between antidepressant classes for the same patient. In addition, patient-specific factors driving response probabilities for each antidepressant class can be generated. We show that antidepressant response can be accurately predicted from real-world EHR data with AI modeling, and our approach could inform further development of clinical decision support systems for more effective treatment selection.

Fig. 1. Flowchart for sample selection after sequential application of exclusion criteria.

Text boxes denote each step for sampleselection. The numbers next to the arrows represent the number of patients remaining after applying the selection step in the preceding text box.

Fig. 2. Top 15 features by global SHAP score.

Global SHAP scores were calculated by averaging the absolute values of local (i.e., individual) level SHAP scores, and represent overall contribution of the predictors to the model.

Fig. 3. Illustration of differential treatment response predictions and local predictor importance for different treatment scenarios.

a Predicted response for actual and alternative treatments for three patients randomly drawn from strata that were modeled to be moderately likely (Patient 1), highly likely (Patient 2), and less likely (Patient 3) to respond, respectively, to the actual antidepressant class prescribed. Results are shown for the representative model (feed-forward DNN). b “Force plots” illustrating local predictor importance (by SHAP score) for each of the treatment scenarios for Patient 1 from a. The patient was a 49-year-old male with co-morbid alcohol used disorder, depressed mood/anhedonia, anxiety symptoms, fatigue/loss of energy, and 8 co-occurring medications prescribed in the past 3 months, who was started on an SSRI. The directions and strengths of each predictor are shown in directed bars in either blue or red. Only predictors with strengths greater than a threshold level are captioned. The bars denote predictors that decrease (blue) or increase (red) the likelihood of response from base value (i.e., sample mean response probability). Longer bars indicate stronger contributions to predicted response.

Fig. 4. Diagrams for prediction modeling designs.

a Schematic diagram for modeling antidepressant treatment response (improved vs no evidence for improvement). Model inputs comprised structured and unstructured EHR data regarding demographics and clinical history, choice of antidepressant class, and response outcome labels. Labels were based on chart review by a psychiatrist, plus in some cases by deep learning imputed labels as described in the text. Prediction model outputs modeled probabilities of treatment response, which can be further binarized to a modeled improvement (yes/no) label. Hexagonal boxes indicate data components that were experimentally evaluated for their effect on prediction performance. Yellow boxes indicate data that are used as inputs for every model. Light green and cyan boxes are inputs used only with prediction models shown in matching colors. b Schematic diagram for the Transformer + feed-forward DNN model. The model first takes in the vectorized clinical notes through the Transformer, transforms them into a fixed-sized vector, which is concatenated with the other features and then passed through additional feed-forward layers.