A linear regression-based machine learning pipeline for the discovery of clinically relevant correlates of gait speed reserve from multiple physiological systems

Abstract

Frailty in older adults is characterized by reduced physiological reserve. Gait speed reserve (GSR: maximum minus usual gait speed) could help identify frailty and act as a proxy for physiological reserve. Utilizing data from 2397 participants aged 50+ from wave 3 of The Irish Longitudinal Study on Ageing, we developed a stepwise linear regression-based machine learning pipeline to select the most important GSR predictors from 34 manually selected features across multiple domains. Variables were selected one at a time such that they maximized the mean adjusted r-squared score from a 5-fold cross-validation. A peak score of (0.16 +/- 0.03) was achieved with 14 variables (giving adjusted-r-squared of 0.18 and 0.16 on 80% training and 20% test data, respectively). The first 7 variables explained 95% of the peak score: grip strength, MOCA cognitive score, third level education, chair stands time, sex, age, and body mass index (BMI). Of the 14 selected features, 11 had statistically significant (p<0.05) effects in the model: sex, MOCA, third level education, chair stands time, age, BMI, grip strength, cardiac output, number of medications, fear of falling, and mean choice reaction time. Associations between GSR and multi-domain features suggest that a network physiology approach is necessary for assessing physiological reserve.