A Panel Framework for Large-Scale Multimodal Digital Phenotyping of Depression in a Real-World Cohort

Description

This record contains the code, derived analysis datasets, and result files used for a study on PHQ-9 binary depression prediction using digital phenotyping data from the COMPASS project. The analysis used 14-day participant-level panel datasets integrating baseline characteristics, ecological momentary assessment (EMA) variables, lifelog features, and wearable sensor-derived measures. The primary final model dataset included 1,307 participants, each observed over a fixed 14-day window, yielding 18,298 participant-day rows. Among these participants, 396 were classified as PHQ-9 positive, corresponding to a prevalence of 30.30%. The uploaded materials include Python scripts for data preprocessing, feature selection, model comparison, final model training, and sensitivity/ablation analyses. The repository also includes derived analysis datasets, selected feature lists, model performance metrics, SHAP-based feature importance outputs, temporal occlusion results, and figure source files. Folder structure: - code/: Python analysis scripts. - data/final_analysis/: Derived datasets used for final feature-selection and model analyses. - data/intermediate_panels/: Intermediate 14-day panel datasets used in earlier analyses. - results/: Generated model metrics, feature-selection outputs, SHAP results, and figure files. These files are provided to support transparency and reproducibility of the analyses reported in the associated manuscript.

Institutions

• Dankook University Hospital

  Chungcheongnam-do, Cheonan

Categories

Licence