Predicting Depression and Anxiety Progression in Multiple Sclerosis from Longitudinal Clinical Data Using Machine Learning

Technical Summary: Predicting Depression and Anxiety Progression in Multiple Sclerosis from Longitudinal Clinical Data Using Machine Learning

Problem Statement
Depression and anxiety are highly prevalent in patients with Multiple Sclerosis (MS), affecting approximately 30.5% and 22% of the population respectively, yet tools for predicting mental health trajectories from clinical data remain limited. While machine learning has shown mixed success in cross-sectional screening (AUCs of 0.67–0.73), predicting future symptom severity or treatment response is more challenging. A consistent finding in existing literature is that baseline symptom severity dominates predictions, often reflecting statistical artifacts like regression to the mean rather than actionable clinical drivers. This study investigates what structured electronic health record (EHR) data can predict regarding the progression of depression (PHQ-9) and anxiety (GAD-7) in MS, and identifies the limits of such predictions.

Methodology
The authors conducted a retrospective cohort study using data from 18,252 MS patients at an academic medical center.

• Cohort Selection: The study utilized 2,163 patients with 7,327 valid PHQ-9 assessment pairs and 1,465 patients with 3,319 valid GAD-7 pairs. Inclusion required a confirmed MS diagnosis and at least one baseline-follow-up pair with a 0.5–1.5 year interval.
• Temporal Design: Baseline ($t_0$) features were extracted from a $\pm30$ day window, while interval features were collected between $t_0$ and the follow-up ($t_1$). The outcome was defined as the score change ($\Delta = \text{score}_{t1} - \text{score}_{t0}$), where positive values indicate worsening and negative values indicate improvement.
• Feature Engineering: From thousands of structured EHR candidates (demographics, labs, medications, utilization, vitals), features were filtered for missingness (>50% removed) and correlation (Pearson $r > 0.9$). The final models utilized approximately 17–19 clinical features plus the baseline score.
• Model Architecture: Gradient boosting models (CatBoost) were employed to predict score change. The data was split 80/20 at the patient level to prevent leakage. Hyperparameters were tuned via Optuna to minimize Root Mean Squared Error (RMSE).
• Interpretability: SHAP (SHapley Additive exPlanations) values were used to quantify feature importance and interpret model behavior. To protect privacy, SHAP values were aggregated into bins of 10 patients for visualization.

Key Contributions

1. Model Development: Development and validation of gradient boosting models specifically for predicting depression and anxiety score change (trajectory) rather than static future scores in a large MS cohort.
2. Quantification of Regression to the Mean: Characterization of baseline severity as the dominant predictor, demonstrating that while it drives prediction, a significant portion of this effect reflects statistical regression to the mean (high scores tend to improve; low scores tend to worsen).
3. Distinct Predictive Profiles: Identification of divergent feature importance between depression and anxiety models, revealing that depression trajectories rely on symptom subscales, while anxiety trajectories incorporate physical disease factors.

Results

• Predictive Performance: The models achieved modest performance, explaining 22% of the variance in PHQ-9 score change ($R^2=0.22$) and 28% in GAD-7 score change ($R^2=0.28$). The remaining variance is attributed to factors not captured in structured EHRs (e.g., life events, social support).
• Baseline Score Dominance: The baseline score was the single most important feature (mean $|SHAP|$ of 0.97 for PHQ-9 and 2.16 for GAD-7). However, SHAP decomposition revealed that baseline score accounted for only 31% of total predictive attribution in the PHQ-9 model and 59% in the GAD-7 model.
• Age as a Secondary Predictor: Age emerged as a consistent secondary predictor across both models. Younger patients showed positive SHAP values, indicating smaller improvements or greater worsening independent of baseline severity.
• Feature Specificity:
  • PHQ-9: Relied heavily on symptom subscales beyond the total score, specifically self-worth, energy, motor symptoms, and sleep.
  • GAD-7: Incorporated clinical factors such as pain scores, disease duration, heart rate, and PHQ mood items, suggesting a link between physical burden and anxiety trajectories.

Significance and Claims
The paper claims that structured clinical data alone captures only a fraction of the drivers behind mental health trajectories in MS. While baseline severity is the strongest statistical predictor, it largely reflects regression to the mean rather than a modifiable clinical mechanism. The study concludes that meaningful individual-level prediction requires richer data sources beyond structured EHRs, specifically:

• Clinical notes (via NLP) to capture patient concerns and social circumstances.
• Patient-reported outcomes collected outside clinical encounters (e.g., digital phenotyping, ecological momentary assessments).
• Integration of passive sensing data (smartphones, fitness trackers) to track mood and activity between visits.

The authors emphasize that while the models identify risk factors (such as younger age), the dominant predictors are not modifiable. Therefore, the clinical value lies in identifying patients who need intervention rather than discovering new intervention targets. The study serves as a benchmark for the current limits of EHR-based mental health prediction in MS.