DSC 291 Final Project: Enhance Mental Issue Predictions With HRV Wearable Data

We use the dataset from Baigutanova, A. et al. Research Question: Does continuous daytime sensor data improve the prediction of depression (PHQ9) and anxiety (GAD7) scores compared to a baseline model using only sleep diaries and participant surveys?

---

Dataset Overview

The dataset combines multimodal data from 49 participants over four weeks, located in the data folder.

• Surveys (survey.csv): Participant demographics, lifestyle habits, and biweekly clinical scores (PHQ9, GAD7, ISI).
• Sleep Diaries (sleep_diary.csv): Daily self-reported sleep metrics like sleep_duration, sleep_efficiency, and sleep_latency.
• Processed Sensor Data (sensor_hrv_filtered.csv): 5-minute daytime aggregates of physiological (HR, HRV), motion (steps), and environmental (light) data.
• Raw Sensor Data (raw_data/): High-frequency (10 Hz) time-series signals from the smartwatch sensors.

---

High-Level Approach

Our analysis is a retrospective comparison of three machine learning models to isolate the predictive value of passive sensor data.

Core Modules

1. Data Ingestion:

   • Merge survey, sleep, and processed sensor data tables.
   • Aggregate daily and 5-minute data to align with biweekly PHQ9/GAD7 scores.

2. Feature Engineering:

   • Extract static features (demographics, lifestyle).
   • Engineer aggregate sleep features from diaries (e.g., mean_sleep_duration).
   • Engineer aggregate sensor features from daytime data (e.g., mean_daytime_rmssd).

3. Comparative Modeling:

   • Train and evaluate three models using cross-validation:
     1. Baseline Model: Static + Sleep Features
     2. Sensor-Only Model: Static + Sensor Features
     3. Combined Model: Static + Sleep + Sensor Features

---

Running the Final Model Notebook

The final analysis is implemented as an interactive marimo notebook located at src/notebooks/final_model_notebook.py.

Prerequisites

• Python >= 3.10
• Install dependencies using uv:

uv sync

Running the Notebook

To launch the interactive notebook:

uv run marimo run src/notebooks/final_model_notebook.py

To edit the notebook in the marimo editor:

uv run marimo edit src/notebooks/final_model_notebook.py

The notebook provides interactive controls for:

• Exploring the dataset (surveys, sleep diaries, sensor data)
• Configuring UMAP dimensionality reduction parameters
• Adjusting regularization strength for Ridge/Lasso models
• Visualizing predicted vs actual scores
• Viewing patient-level time series trends