A Foundation Model for Patient Behavior Monitoring and Suicide Detection

Rodrigo Oliver, Josué Pérez Sabater, Diego Herrero-Quevedo, Antonio Artés, Alejandro Lancho, Pablo M. Olmos

Keywords: Foundation models for patient monitoring, Suicide detection via deep learning

Abstract:

Foundation models have achieved remarkable success across various domains, yet their adoption in healthcare remains limited, particularly in areas requiring the analysis of smaller and more complex datasets. While foundation models have made significant advances in medical imaging, genetic biomarkers, and time series from electronic health records, the potential for patient behavior monitoring through wearable devices remains underexplored. Wearable device datasets are inherently heterogeneous and multisource and often exhibit high rates of missing data, presenting unique challenges. Notably, missing patterns in these datasets are frequently not-at-random, and when adequately modeled, these patterns can reveal crucial insights into patient behavior.
This paper introduces a novel foundation model based on a modified vector quantized variational autoencoder (VQ-VAE), specifically designed to process real-world data from wearable devices. Our model excels at reconstructing heterogeneous multisource time-series data and effectively models missing data patterns. We demonstrate that our pretrained model, trained on a broad cohort of psychiatric patients with diverse mental health issues, can perform downstream tasks without fine-tuning on a held-out cohort of suicidal patients. This is illustrated through the use of a change-point detection algorithm that identifies suicide attempts with high accuracy, matching or surpassing patient-specific methods, thereby highlighting the potential of VQ-VAE as a versatile tool for behavioral analysis in healthcare.

Primary Area: applications to neuroscience & cognitive science

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6904