HiddenVis

HiddenVis visualizes hidden states inferred by deep learning models from accelerometer time series to support interpretation of temporal profiles associated with health outcomes.

Key Features:

• Visualization of Hidden States: Generates visual representations of latent states produced by deep learning models from time series accelerometer inputs.
• Comparative Cohort Analysis: Compares outcome-associated temporal profiles across cohorts to identify differences in activity patterns.
• Pre-trained Deep Learning Model Integration: Integrates outputs from pre-trained deep learning models to infer hidden states from input time series.
• Accelerometer Time Series Ingestion: Processes continuous accelerometer time series data as input for state inference and downstream analysis.
• Temporal Profile Identification: Identifies critical time periods within daily activity profiles that correlate with health outcomes.

Scientific Applications:

• Health Outcome Prediction: Applied to NHANES accelerometer data to identify differences in 5-year mortality associated with daily activity patterns.
• Temporal Profile Analysis: Detects time-of-day periods within activity data that correlate with health outcomes for clinical and public health research.
• Human Movement Monitoring: Enables analysis of continuous accelerometer measurements to monitor and interpret movement-related health signals.

Methodology:

Implemented using JavaScript, Python, and the ECharts framework; explicitly described steps include importing accelerometer time series data, utilizing pre-trained deep learning models to infer hidden states, and generating visualizations for interpretation of those states in relation to health outcomes.