Implementing an AI That Automatically Extracts Standardized Functional Patterns From Wearable Sensor Data

Standardized functional tests (SFTs) are frequently used to quantify a person’s ability to perform activities of daily living (ADLs) against expected values. SFTs provide valuable information to aid a clinician’s understanding of a patient’s status and direction of improvement. Such information may support clinicians when assessing a wide range of conditions such as musculoskeletal or neurological diseases or age-related frailty. If these SFTs could be performed reliably by patients in their own home without supervision, they could track and help to guide rehabilitation. Wearable devices have the potential to collect reliable data from the wearer when performing SFTs. However, the data generated from sensors for long-term recordings of movement can become very large, making it difficult to manually extract movement information with any level of accuracy. Hence, it is important to evaluate whether it is possible to implement an automated system capable of extracting SFT patterns from long-term sensor data without manual data processing. This article describes an artificial neural network (ANN) system that was trained to extract specific SFTs such as the 30-s chair stand test (CST) (30s-CST) and the 40-m fast-paced walk test (40m-FPWT). The resultant model obtained 99.7% accuracy in 30s-CST pattern recognition, 99.3% accuracy in 40m-FPWT pattern recognition, and 97.3% accuracy in detecting false patterns. The system provided an overall accuracy of 98.76% with supervised clinical trial data. The ambulatory data testing phase obtained an overall accuracy of 90.18%. Specifically, it achieved 92.73% accuracy in detecting 30s-CST patterns and 86.67% accuracy in recognizing 40m-FPWT patterns.