Analyzing Patient Daily Movement Behavior Dynamics Using Two-Stage Encoding Model

TL;DR

This paper tackles the challenge of extracting meaningful patterns from high-frequency, unlabeled in-home movement data of people with dementia. It introduces a two-stage, self-supervised framework that first encodes time-series activity as text and then learns latent states via dimensionality reduction, clustering, and PageRank-based transition analysis to reveal activity biases. The authors apply this to a dementia cohort, identifying five latent states and demonstrating how state transitions relate to clinical measures, with implications for personalized care. The work lays groundwork for interpretable, data-efficient representations of patient behavior that could support clinical decision-making and data augmentation for sensitive medical datasets.

Abstract

In the analysis of remote healthcare monitoring data, time series representation learning offers substantial value in uncovering deeper patterns of patient behavior, especially given the fine temporal granularity of the data. In this study, we focus on a dataset of home activity records from people living with Dementia. We propose a two-stage self-supervised learning approach. The first stage involves converting time-series activities into text strings, which are then encoded by a fine-tuned language model. In the second stage, these time-series vectors are bi-dimensionalized for applying PageRank method, to analyze latent state transitions to quantitatively assess participants behavioral patterns and identify activity biases. These insights, combined with diagnostic data, aim to support personalized care interventions.

Figures (11)

• Figure 1: Flowchart of the framework.
• Figure 2: Flowchart of data preprocessing. The figure illustrates the monitoring data for a single participant over the course of one day. The left graph displays the raw, unprocessed measurements. In the middle graph, the data is rectified into 20-minute intervals, where periods of inactivity are labeled as "nowhere." Within each window, the most frequent location, excluding "nowhere," is identified and recorded. The right graph presents the corresponding text strings, which are formatted for interpretation by the language model.
• Figure 3: Daily histogram.
• Figure 4: Location Histogram.
• Figure 5: Three month distribution.