PFIMPA: A Multimodal Approach to Predict Physical Function Impairment in Older Adults Using Physical Activity From Wrist-Worn Accelerometer

Diagnosing physical function impairment (PFI) in older adults is challenging due to delayed diagnosis and cumbersome clinical assessments. We propose PFIMPA, a novel artificial intelligence method that enables seamless remote PFI assessment using physical activity (PA) data provided by wrist-worn accelerometers for the first time. To capture the complex cumulative and dynamic characteristics of PA, we designed a comprehensive feature representation method that integrates demographic information, clinical PA features, time-frequency domain features, and PA sequence conversion image features (PASCIs) based on Gramian angular field (GAF), Markov transfer field, and recursive plot transformations. We developed a lightweight convolutional neural network (CNN) enhanced by convolutional block attention module (CBAM) to extract spatio-temporal features from PASCIs, further encapsulating the correlation, accumulation, and transition properties of PA. PFI detection is achieved through a three-stage feature-level fusion method combined with machine learning (ML) and hyperparameter optimization. The PASCIs module achieves an AUC of 83.72% in PFI prediction, while the multimodal approach combining clinical, time-frequency, and PASCIs features improves the AUC to 87.16%, surpassing state-of-the-art (SOTA) time series classification models. PFIMPA represents a significant advancement in PFI assessment for older adults, enabling seamless remote monitoring and timely intervention.