Predicting Long-Term Depression Progression in Parkinson's Disease: A Machine-Learning Survival Analysis and Risk Score.
The authors developed an explainable machine-learning survival model and integer risk score using baseline clinical, autonomic, cognitive and mood measures in de novo PD patients to predict long-term progression of depression (test C-index 0.744) and stratify patients into low/moderate/high risk.
What the AI sees
The authors developed an explainable machine-learning survival model and integer risk score using baseline clinical, autonomic, cognitive and mood measures in de novo PD patients to predict long-term progression of depression (test C-index 0.744) and stratify patients into low/moderate/high risk.
Research significance
Although not mechanistic, the tool enables early risk stratification and trial enrichment and highlights autonomic, sleep, cognitive and gut-related features that can guide personalized monitoring and targeted nonpharmacologic or pharmacologic interventions to prevent or mitigate depression in…
Source abstract
BACKGROUND: Depression in Parkinson's disease (dPD) is common and heterogeneous, impairs quality of life, and may accelerate disease progression. Tools that predict long-term dPD progression are lacking. METHODS: We retrospectively analyzed de novo, drug-naïve Parkinson's disease (PD) participants in the Parkinson's Progression Markers Initiative (PPMI; 2011-2024). The primary outcome was depressive progression, defined as a sustained worsening in Geriatric Depression Scale-15 (GDS-15) category over 12 months. Candidate predictors included demographic, motor, and non-motor variables at both total and sub-item levels. Four survival machine learning models, Random Survival Forests (RSF), Extreme Gradient Boosting, Support Vector Survival Machines, and Gradient Boosting Survival Analysis, were evaluated using concordance index (C-index). Shapley Additive exPlanations were applied to identify key predictors and construct an integer-based risk score. RESULTS: Of 1819 eligible participants, 496 met inclusion criteria (median age 62 years [IQR: 55-69]; 61.3% male); 94 (19.0%) progressed over a median 6 year follow-up. RSF achieved the best discrimination (test-set C-index 0.744). Key predictors included age, baseline GDS-15; SCOPA-AUT subscores (thermoregulatory, gastrointestinal, cardiovascular); cognition (BJLOT, SDMT); impulse control disorder (QUIP-CS score), and MDS-UPDRS I (sleep problems night, pain and other sensations). The SHAP-derived score stratified patients into low (progression 7.3%), moderate (14.7%), and high-risk (36.5%) groups with clear Kaplan-Meier separation (log-rank p < 0.001). Time-dependent AUCs were 0.721, 0.770, 0.794, 0.792, and 0.812 at 2, 4, 6, 8, and 10 years. CONCLUSIONS: An explainable survival model and integer-based risk score using routinely collected measures predicted long-term dPD progression and enabled pragmatic risk stratification to support early, personalized management.