A Clinically Validated Multi-Model Fusion Framework Integrating Machine Learning and LSTM Networks for Real-Time Geriatric Frailty Assessment

Detecting frailty early and objectively is crucial for preventing falls and maintaining independence in older adults, yet traditional assessments remain subjective and resource intensive. We developed a portable, real time frailty assessment system using a consumer grade depth camera (RealSense D435) with Nuitrack for skeletal tracking. From 3D gait data, temporal and kinematic features were extracted and evaluated using five classical machine learning models and an Long Short Term Memory (LSTM) network. To address limitations of prior work such as reliance on single model classifiers, wearable sensor constraints, and depth camera studies limited to Parkinson’s or fall-risk tasks we introduce a hierarchical multi-model fusion framework that integrates LSTM derived temporal representations with lower limb skeletal features. This architecture enhances robustness to noise, improves generalization on small clinical datasets, and enables real time non-wearable frailty assessment using only a monocular depth sensor. In a clinical study involving 120 gait samples across three sites, CatBoost and ensemble models achieved 97% accuracy. With multi-model fusion, performance improved to an F1-score of 0.98 and reached 100% accuracy across four frailty levels. This clinically validated framework provides a fast, objective, and scalable solution for geriatric frailty assessment in hospitals, homes, and tele-health settings.