Prediction of short-term antidepressant response using probabilistic graphical models with replication across multiple drugs and treatment settings

Abstract

Heterogeneity in the clinical presentation of major depressive disorder and response to antidepressants limits clinicians' ability to accurately predict a specific patient's eventual response to therapy. Validated depressive symptom profiles may be an important tool for identifying poor outcomes early in the course of treatment. To derive these symptom profiles, we first examined data from 947 depressed subjects treated with selective serotonin reuptake inhibitors (SSRIs) to delineate the heterogeneity of antidepressant response using probabilistic graphical models (PGMs). We then used unsupervised machine learning to identify specific depressive symptoms and thresholds of improvement that were predictive of antidepressant response by 4 weeks for a patient to achieve remission, response, or nonresponse by 8 weeks. Four depressive symptoms (depressed mood, guilt feelings and delusion, work and activities and psychic anxiety) and specific thresholds of change in each at 4 weeks predicted eventual outcome at 8 weeks to SSRI therapy with an average accuracy of 77% (p = 5.5E-08). The same four symptoms and prognostic thresholds derived from patients treated with SSRIs correctly predicted outcomes in 72% (p = 1.25E-05) of 1996 patients treated with other antidepressants in both inpatient and outpatient settings in independent publicly-available datasets. These predictive accuracies were higher than the accuracy of 53% for predicting SSRI response achieved using approaches that (i) incorporated only baseline clinical and sociodemographic factors, or (ii) used 4-week nonresponse status to predict likely outcomes at 8 weeks. The present findings suggest that PGMs providing interpretable predictions have the potential to enhance clinical treatment of depression and reduce the time burden associated with trials of ineffective antidepressants. Prospective trials examining this approach are forthcoming.

Fig. 1. Study overview.

A Measurement-based psychiatry using validated rating scales such as the 17-item Hamilton Depression Rating Scale (HDRS) to measure severity of depression symptoms. HDRS total score is sum of severity of individual HDRS item (depressive symptom). B Heterogeneity of symptom severity in the training datasets (Mayo Clinic PGRN-AMPS and ISPC subjects) with HDRS total score of 25 at baseline. C Heterogeneity in longitudinal variations of HDRS total score in Mayo Clinic PGRN-AMPS and ISPC subjects treated with citalopram/escitalopram. D Proposed analyses workflow to build probabilistic graphical model (PGM) and derive individualized prognoses of treatment outcomes at 8 weeks using changes in severity of focused set of depressive symptoms between baseline and after 4 weeks of antidepressant treatment.

Fig. 2. Schematic of symptom dynamic paths and prognostic depressive symptoms.

A Symptom dynamic paths in patients in the training datasets (Mayo PGRN-AMPS and ISPC subjects). B Longitudinal variation in severity score of depressed mood (HDRS item) in patients starting in the A3 stratum at baseline. C Symptom clusters within patient strata (e.g., A1 at baseline), illustrating the grouping of prognostic symptoms. Fig. B, D, E, and F depict variations in depressed mood (prognostic symptom) and suicide ideation (nonprognostic symptom) in patients with antidepressants or placebo on symptom dynamic paths A3 → B3 → C3 (nonresponders at 8 weeks), A3 → B2 → C2 (responders without remission at 8 weeks), and A3 → B1 → C1 (remission at 8 weeks). In Fig. B, D, E and F, the solid blue lines in each figure represent the variations (mean changes) in prognostic symptom scores, and shaded regions around the mean illustrate their 95% confidence intervals (CIs). The boxplots and error bars represent the overall variability in prognostic symptom severity scores at each timepoint. Fig. B and D: Comprise all patients originating in stratum A3 in training and placebo datasets. The variations in prognostic and nonprognostic symptoms in testing data cohorts are visualized in Fig. E and F, respectively.

Fig. 3. Prognosis rules and their predictive accuracies.

A Demonstration of the operationalization of prognoses rules to predict 8-week treatment outcome. B For each of the baseline and 4-week strata, we illustrate the accuracy of the prognoses in comparison with the average prediction accuracy (53% in dashed red line) that is achieved when using only baseline clinical and sociodemographic factors as predictors.