Abstract
Machine learning models informed by patient-generated wearable data can be used to predict patient-reported and performance-based outcome measures. This approach offers a promising alternative to traditional outcome monitoring, which is commonly limited by recall bias, discrete sampling, and healthcare resource constraints. The aims of this systematic review were to identify wearable-derived features strongly associated with performance-based and patient-reported outcome measures, to compare the predictive performance across machine learning approaches, and to consolidate methodological limitations and provide suggestions for future work. Following a systematic search of four databases (PubMed, Scopus, Embase, and IEEE Xplore), 18 eligible studies were identified, published between 2017 and 2024, spanning patients across eight disease categories. Most studies used wrist-worn devices measuring accelerometry, sometimes combined with heart rate, respiratory, or sleep metrics. Random forest and support vector machine models were the most common, while hidden Markov temporal models showed improved performance with access to longitudinal data. Predictive performance ranged from poor to excellent (AUC 0.56-0.92), and non-linear models generally outperformed linear models. Despite promising early results, most studies report similar limitations of small sample sizes, limited external validation, and difficulty achieving acceptable accuracy beyond binary predictions. Overall, these studies highlight the potential of wearable-informed machine learning for continuous and objective outcome assessment, but the consensus calls for further work to apply larger, more diverse longitudinal datasets and interpretable temporal modelling approaches to bridge the gap between the current proof-of-concept state and clinical translation.