Predicting prosthesis use and mobility needs in lower limb amputees: a machine learning approach using clinical and actigraphy data

Background: Despite advances in prosthetic technology, many individuals with lower limb amputation (LLA) fall short of achieving their full functional potential. The Medicare Functional Classification Level (K-level) system remains a standard for guiding prosthetic prescriptions; however, it has historically reflected anticipated mobility potential rather than actual performance in daily life. Similarly, standardized clinical assessments provide only episodic snapshots and are limited in their ability to determine whether individuals will meet their real-world mobility goals. Consequently, there is a pressing need for objective, scalable methods to capture prosthesis use and activity in the community. Objective: This study investigated the potential of machine learning (ML) and passively collected community activity data to predict two key outcomes in prosthetic users: (1) reassessed K-level based on real-world mobility data and expert panel review, and (2) attainment of self-defined mobility goals. Methods: Fifty-three adults with unilateral or bilateral LLA participated in three months of passive community monitoring using ActiGraph accelerometers. Participants also completed standardized performance-based, patient-reported outcome measures, and a semi-structured mobility goal interview. A multidisciplinary expert panel reviewed all data to reassess the K-level and review whether participants met their personal mobility goals. These served as ground truth for model evaluation. Gradient-boosted classifiers were trained using combinations of demographic, performance-based, patient-reported, and sensor-derived (ActiGraph) features across various community monitoring durations (3, 7, 30, and ≥ 90 days). Results: Models predicting reassessed K-levels demonstrated modest recall (range: 30.6–60.4%), with the best performance (60.4%) achieved using demographic, performance-based, patient-reported and 30 days of community ActiGraph data. Longer monitoring durations (≥ 90 days) reduced performance. In contrast, models predicting personal goal attainment achieved consistently high recall (Range: 77.0–89.9%), with the best-performing model using demographic, performance-based, patient-reported, and ActiGraph data. Notably, removing the performance-based outcomes had minimal impact (recall ≥ 82.1% with just 3 days of monitoring). Conclusions: ML models predicting personal goal attainment outperformed those targeting K-level classification, particularly when using passively collected community activity data. These findings support the value of real-world passive sensing as scalable, low-burden strategy to remotely identify individuals at risk of suboptimal outcomes and enable timely, goal-oriented rehabilitation. Trial registration This work includes data from a registered clinical trial (NCT03930199).