Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AttentiveGRUAE: An Attention-Based GRU Autoencoder for Temporal Clustering and Behavioral Characterization of Depression from Wearable Data

Nidhi Soley, Vishal M Patel, Casey O Taylor

TL;DR

AttentiveGRUAE addresses the problem of finding interpretable, reproducible behavioral subtypes from low-frequency wearable sleep data to predict depression risk. It combines an attention-enhanced GRU encoder–decoder with a binary outcome head and soft clustering of embeddings via a Gaussian Mixture Model, trained end-to-end with a joint objective $ abla\mathcal{L}=\nabla\mathcal{L}_{AE}+\nabla\mathcal{L}_{BCE}$ and gradient-surgery to mitigate conflict. The approach yields superior clustering and prediction performance on the GLOBEM discovery data and demonstrates robust cross-cohort reproducibility with high ARI stability, while enabling temporally interpretable insights through attention-weighted trajectory analysis. These findings support clinically relevant risk stratification from passively collected sleep patterns and highlight the value of outcome-guided, attention-based temporal modeling for real-world wearable data.

Abstract

In this study, we present AttentiveGRUAE, a novel attention-based gated recurrent unit (GRU) autoencoder designed for temporal clustering and prediction of outcome from longitudinal wearable data. Our model jointly optimizes three objectives: (1) learning a compact latent representation of daily behavioral features via sequence reconstruction, (2) predicting end-of-period depression rate through a binary classification head, and (3) identifying behavioral subtypes through Gaussian Mixture Model (GMM) based soft clustering of learned embeddings. We evaluate AttentiveGRUAE on longitudinal sleep data from 372 participants (GLOBEM 2018-2019), and it demonstrates superior performance over baseline clustering, domain-aligned self-supervised, and ablated models in both clustering quality (silhouette score = 0.70 vs 0.32-0.70) and depression classification (AUC = 0.74 vs 0.50-0.67). Additionally, external validation on cross-year cohorts from 332 participants (GLOBEM 2020-2021) confirms cluster reproducibility (silhouette score = 0.63, AUC = 0.61) and stability. We further perform subtype analysis and visualize temporal attention, which highlights sleep-related differences between clusters and identifies salient time windows that align with changes in sleep regularity, yielding clinically interpretable explanations of risk.

AttentiveGRUAE: An Attention-Based GRU Autoencoder for Temporal Clustering and Behavioral Characterization of Depression from Wearable Data

TL;DR

AttentiveGRUAE addresses the problem of finding interpretable, reproducible behavioral subtypes from low-frequency wearable sleep data to predict depression risk. It combines an attention-enhanced GRU encoder–decoder with a binary outcome head and soft clustering of embeddings via a Gaussian Mixture Model, trained end-to-end with a joint objective and gradient-surgery to mitigate conflict. The approach yields superior clustering and prediction performance on the GLOBEM discovery data and demonstrates robust cross-cohort reproducibility with high ARI stability, while enabling temporally interpretable insights through attention-weighted trajectory analysis. These findings support clinically relevant risk stratification from passively collected sleep patterns and highlight the value of outcome-guided, attention-based temporal modeling for real-world wearable data.

Abstract

In this study, we present AttentiveGRUAE, a novel attention-based gated recurrent unit (GRU) autoencoder designed for temporal clustering and prediction of outcome from longitudinal wearable data. Our model jointly optimizes three objectives: (1) learning a compact latent representation of daily behavioral features via sequence reconstruction, (2) predicting end-of-period depression rate through a binary classification head, and (3) identifying behavioral subtypes through Gaussian Mixture Model (GMM) based soft clustering of learned embeddings. We evaluate AttentiveGRUAE on longitudinal sleep data from 372 participants (GLOBEM 2018-2019), and it demonstrates superior performance over baseline clustering, domain-aligned self-supervised, and ablated models in both clustering quality (silhouette score = 0.70 vs 0.32-0.70) and depression classification (AUC = 0.74 vs 0.50-0.67). Additionally, external validation on cross-year cohorts from 332 participants (GLOBEM 2020-2021) confirms cluster reproducibility (silhouette score = 0.63, AUC = 0.61) and stability. We further perform subtype analysis and visualize temporal attention, which highlights sleep-related differences between clusters and identifies salient time windows that align with changes in sleep regularity, yielding clinically interpretable explanations of risk.

Paper Structure

This paper contains 13 sections, 7 equations, 8 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Results
  5. Performance on discovery data and ablation
  6. Reproducibility Across Cohorts and Cluster Stability
  7. Cluster subtype analysis and temporal interpretation
  8. Discussion
  9. Supplementary Material
  10. Notation.
  11. Optimization.
  12. External validation.
  13. Selection for the number of clusters K.

Figures (8)

  • Figure 1: Outcome Guided-GRU encoder–decoder with attention; embeddings clustered by GMM (BIC for $K$). We report AUC/MSE, Silhouette/DBI, and assess cluster reproducibility via ARI on external validation data.
  • Figure 2: t-SNE of all the model variants and baselines (Cluster 1 = yellow, Cluster 0 = purple).
  • Figure 3: Violin plots for key differentiators.
  • Figure 4: BIC across $K$. The elbow/minimum occurs at $K{=}2$. Models with $K{>}2$ did not improve AUC/Silhouette and produced small, unstable clusters.
  • Figure 5: ARI over 200 resamples on DS3+DS4 (mean $\approx 0.89$, dashed line).
  • ...and 3 more figures