Unsupervised Machine Learning for Visual Frailty Prediction in Mice

Dustin M. Miao

Document Type

Article

Publication Date

2025

Keywords

JMG

Abstract

Frailty quantifies biological aging and predicts adverse health outcomes, but manual frailty index (FI) assessments are labor-intensive and variable across scorers. We previously developed the visual frailty index (vFI), which uses machine vision to estimate FI from video-based behavioral features. Here, we developed the Unsupervised Visual Frailty Index (uvFI), which uses unsupervised behavioral segmentation of open-field videos from 851 mice (540 C57BL/6J, 311 Diversity Outbred; 1126 trials). We used an unsupervised behavior segmentation model to extracted data-driven features from behavioral syllables, transitions, and poses to predict frailty (uvFI), age (uvFRIGHT), and proportion of life lived (uvPLL). We achieved a mean absolute error (MAE) of 1.37 ± 0.122 for frailty and 14.9 ± 1.18 weeks for age, with accuracy further improved by combining supervised and unsupervised features. These results show that unsupervised behavioral features capture aging signatures, enabling scalable and reproducible frailty assessment.

Please contact the Joan Staats Library for information regarding this document.

COinS