Early prediction of SSRI treatment response in adolescents with depression using multimodal machine learning
Presenter Status
Staff
Abstract Type
Clinical Research
Primary Mentor or Principal Investigator
Catherine Koertje, MD
Presentation Type
Poster
Start Date
19-5-2026 11:00 AM
End Date
19-5-2026 12:00 PM
Abstract Text
Background:
Major Depressive Disorder (MDD) in adolescents is commonly treated with Selective Serotonin Reuptake Inhibitors (SSRIs). Treatment response is typically assessed every 4-6 weeks after initiating pharmacotherapy, and medication adjustments may be made if the initial regimen is ineffective. This standard-of-care delay can prolong or exacerbate symptom burden and functional impairment. Early prediction of treatment response could support more timely and personalized medication management in adolescents with MDD.
Objectives/Goal:
To develop and evaluate multimodal machine-learning models for early prediction of SSRI treatment response in adolescents with MDD, assessing the predictive value of clinical, wearable, and proteomic features both individually and in integrated models.
Methods/Design:
Adolescents aged 12-18 initiating pharmacotherapy with fluoxetine or escitalopram for MDD were enrolled in an open-label naturalistic study. Clinical data were collected at baseline and two weeks after treatment initiation. Baseline proteomic data was generated using the Olink® Explore HT platform and normalized using Olink NPX procedures. Daily summaries of biometric data from Garmin vívosmart® 4 devices were captured over the first two weeks of treatment. All analyses and model development were conducted in R. Treatment response was evaluated at 4-6 weeks and dichotomized into Responder and Non-Responder categories (a threshold of ≥ 50% reduction in CDRS-R raw score from baseline). Feature engineering included calculating averages and change-from-baseline values for clinical and wearable features. Feature selection was performed using univariate tests. Models were developed using XGBoost with 10-fold cross validation and ROSE upsampling to address class imbalance (17 Non-Responders/11 Responders). Model performance was evaluated using AUC, sensitivity, and specificity estimates and 95% confidence intervals, with classification thresholds optimized using Youden’s J statistic and visualized with ROC curves. Feature importance was derived from XGBoost gain scores to identify the most influential features from each modality contributing to treatment prediction. Participants who reached ≥ 70% medication adherence in the first 14 days of treatment were included in analysis.
Results:
The final dataset consisted of N=28 participants (median age=16.0, IQR: 2.65, and 64% Female). Across single modality models, proteomic features provide the strongest predictive performance (AUC=0.94[0.85-1], sensitivity=0.91[0.59-1], specificity=0.88[0.64-0.99]), followed by the moderate yet unstable performance from the clinical only model (0.87[0.74-1]), and the wearable-only model (AUC= 0.74[0.55-0.93]). Integration of wearable and clinical features resulted in high but unstable metrics (AUC= 0.9[0.77-1]). Integrating all three modalities into a full model generated the best performance, (AUC=0.99 [0.98-1], sensitivity=1[0.72-1], specificity=0.94[0.7-1]). Overall, the full and proteomics-only models achieved the best performance metrics, with proteomic features being most influential.
Conclusions
These findings highlight a potential use-case for baseline proteomic profiling and early activity-based biometrics to identify likely Responders and Non-Responders as early as two weeks into an SSRI treatment course for adolescent MDD. Small sample sizes often result in machine learning models to overfit with unstable model performance estimates, therefore larger prospective studies are needed to refine and validate single and multimodal models prior to clinical application. Future work will incorporate an additional metabolomic modality into these analyses.
Early prediction of SSRI treatment response in adolescents with depression using multimodal machine learning
Background:
Major Depressive Disorder (MDD) in adolescents is commonly treated with Selective Serotonin Reuptake Inhibitors (SSRIs). Treatment response is typically assessed every 4-6 weeks after initiating pharmacotherapy, and medication adjustments may be made if the initial regimen is ineffective. This standard-of-care delay can prolong or exacerbate symptom burden and functional impairment. Early prediction of treatment response could support more timely and personalized medication management in adolescents with MDD.
Objectives/Goal:
To develop and evaluate multimodal machine-learning models for early prediction of SSRI treatment response in adolescents with MDD, assessing the predictive value of clinical, wearable, and proteomic features both individually and in integrated models.
Methods/Design:
Adolescents aged 12-18 initiating pharmacotherapy with fluoxetine or escitalopram for MDD were enrolled in an open-label naturalistic study. Clinical data were collected at baseline and two weeks after treatment initiation. Baseline proteomic data was generated using the Olink® Explore HT platform and normalized using Olink NPX procedures. Daily summaries of biometric data from Garmin vívosmart® 4 devices were captured over the first two weeks of treatment. All analyses and model development were conducted in R. Treatment response was evaluated at 4-6 weeks and dichotomized into Responder and Non-Responder categories (a threshold of ≥ 50% reduction in CDRS-R raw score from baseline). Feature engineering included calculating averages and change-from-baseline values for clinical and wearable features. Feature selection was performed using univariate tests. Models were developed using XGBoost with 10-fold cross validation and ROSE upsampling to address class imbalance (17 Non-Responders/11 Responders). Model performance was evaluated using AUC, sensitivity, and specificity estimates and 95% confidence intervals, with classification thresholds optimized using Youden’s J statistic and visualized with ROC curves. Feature importance was derived from XGBoost gain scores to identify the most influential features from each modality contributing to treatment prediction. Participants who reached ≥ 70% medication adherence in the first 14 days of treatment were included in analysis.
Results:
The final dataset consisted of N=28 participants (median age=16.0, IQR: 2.65, and 64% Female). Across single modality models, proteomic features provide the strongest predictive performance (AUC=0.94[0.85-1], sensitivity=0.91[0.59-1], specificity=0.88[0.64-0.99]), followed by the moderate yet unstable performance from the clinical only model (0.87[0.74-1]), and the wearable-only model (AUC= 0.74[0.55-0.93]). Integration of wearable and clinical features resulted in high but unstable metrics (AUC= 0.9[0.77-1]). Integrating all three modalities into a full model generated the best performance, (AUC=0.99 [0.98-1], sensitivity=1[0.72-1], specificity=0.94[0.7-1]). Overall, the full and proteomics-only models achieved the best performance metrics, with proteomic features being most influential.
Conclusions
These findings highlight a potential use-case for baseline proteomic profiling and early activity-based biometrics to identify likely Responders and Non-Responders as early as two weeks into an SSRI treatment course for adolescent MDD. Small sample sizes often result in machine learning models to overfit with unstable model performance estimates, therefore larger prospective studies are needed to refine and validate single and multimodal models prior to clinical application. Future work will incorporate an additional metabolomic modality into these analyses.
Comments
Full text not provided by primary author
Poster Board Number: 9