Assessing pre-frailty in the elderly using explainable AI

Abstract

Frailty characterized by a rapid decline in physical and cognitive functions is highly associated with an increased prevalence of chronic diseases and higher risks of disability and mortality. Particularly, a pre-frailty stage marks the initial phase of transitioning from a healthy state to a frail state. Timely and appropriate therapeutic interventions during this period can provide a critical opportunity to delay or prevent the progression of frailty. This study aims to develop a model for assessing the pre-frailty stage using explainable AI (XAI) and to explore key indicators that significantly contribute to the detection of the pre-frailty stage. The GSTRIDE public dataset was used as the input for this study. Data of elderly individuals aged between 70 and 98 years were included, among which 62 individuals were healthy and 59 individuals were in the pre-frailty stage. Input data included demographic information and measures from five physical and cognitive function assessment tools (4-meter gait speed test, Time Up and Go (TUG) test, Short Physical Performance Battery (SPPB) test, Short Falls Efficacy Scale International (FES-I), and Global Deterioration Scale (GDS)), and gait analysis indicators collected with inertial measurement unit (IMU) sensors. XGBoost was utilized as the XAI for classifying pre-frailty stages. Shapley values were calculated to evaluate indicators significantly influencing classification of the pre-frailty stage. The pre-frailty assessment model demonstrated an area under the ROC curve (AUC) of 0.81 (0.73–0.88) when using the minimal feature set data. Short FES-I was identified as the most influential predictor of pre-frailty by SHAP analysis, followed by the time duration of the 4-meter gait speed test, clearance, time duration of the TUG test, strides, and time. This study reveals that traditional frailty assessment tools remain effective in evaluating the pre-frailty stage. Additionally, gait analysis indicators, which are non-consciously measured and derived from IMU sensors that enable daily walking assessments, prove to be valuable for pre-frailty evaluations. This highlights the possibility of employing these measurements collected during routine activities to effectively evaluate the onset of pre-frailty.