The aging population faces a rising prevalence of neurological conditions, leading to significant disability. While rehabilitation interventions are beneficial, patient responses vary significantly, highlighting the need for precision rehabilitation. Traditional clinical assessments are time-consuming and infrequent, hindering real-time intervention adjustments. This study explores the use of wearable sensors and machine learning to continuously monitor motor recovery, allowing for personalized rehabilitation strategies. Wearable sensors offer the advantage of collecting data in real-world settings, providing longitudinal data on upper-limb motor function with minimal burden on patients and clinicians. Previous research by the authors demonstrated the feasibility of estimating clinical scores related to movement quality using this technology; this study aims to extend this to assessing impairment severity.

Existing literature shows the benefits of rehabilitation interventions across neurological conditions. However, the variability in patient responses to these interventions emphasizes the need for precision rehabilitation approaches tailored to individual characteristics. Researchers are exploring the use of patient genotypes and motor phenotypes to guide personalized interventions. The challenge lies in efficiently monitoring motor recovery to ensure intervention effectiveness and make necessary adjustments. The limitations of traditional clinical assessments, which are often infrequent and time-consuming, have spurred interest in wearable sensor technology for continuous monitoring and data collection in real-world settings. Prior studies by the authors have demonstrated the potential of wearable sensor data to accurately estimate clinical scores related to movement quality, and this study builds on that foundation.

This prospective longitudinal study recruited 37 participants (16 stroke survivors and 21 TBI survivors) with residual upper-limb impairments. Participants underwent two visits, baseline and 3 months later. During each visit, they completed the Fugl-Meyer Assessment (FMA) and Functional Ability Scale (FAS), clinical tests evaluating motor impairment severity and movement quality, respectively. Wearable sensors (Shimmer2) were placed on their upper limbs during the performance of eight selected functional tasks from the Wolf Motor Function Test (WMFT). Accelerometer data were processed to extract various features, and four machine learning algorithms (linear regression, random forest, balanced random forest, and a proposed technique incorporating FAS estimates) were used to estimate FMA and FAS scores from the sensor data. The accuracy of the algorithms was assessed using RMSE and R². Statistical analysis included Chi-square and t-tests to compare clinical characteristics between stroke and TBI groups, and paired t-tests to assess the significance of algorithm improvements.

The study demonstrated that accurate estimations of both FMA and FAS scores could be derived from wearable sensor data. The proposed machine learning algorithm, which incorporated both sensor data and FAS estimates, achieved a high coefficient of determination (R² = 0.86) and low RMSE (3.99 points) for FMA score estimation. For FAS score estimation, a previously developed algorithm showed an R² of 0.79 and a RMSE of 0.38 points. The results were robust despite the relatively small sample size and non-uniform distribution of clinical scores. The study showed that the estimation error could be reduced with repeated measures, leading to more reliable estimations of clinical scores.

The findings address the need for continuous monitoring of motor recovery in rehabilitation by providing a method for accurate and efficient estimation of clinically relevant scores using wearable sensors and machine learning. The high accuracy of the proposed technique, particularly for FMA score estimation, significantly advances the use of wearable technology in precision rehabilitation. The ability to monitor patients' responses continuously allows for timely adjustments to interventions, potentially maximizing motor gains and improving outcomes. This method addresses the limitations of traditional clinical assessments and opens possibilities for personalized rehabilitation strategies delivered in various settings.

This study successfully demonstrated a novel method for accurately estimating key clinical scores using wearable sensor data and machine learning algorithms. The high accuracy of the proposed technique holds significant implications for precision rehabilitation, enabling continuous monitoring of patient progress and personalized intervention adjustments. Future research could focus on refining the algorithms, expanding the dataset, and investigating the clinical efficacy of this approach to optimize rehabilitation outcomes. Further investigation into automated data analysis procedures could improve the reliability and efficiency of clinical score estimation.

The study's relatively small sample size and non-uniform distribution of clinical scores might limit the generalizability of the findings. The study focused on specific functional tasks, and future research should examine the generalizability to a broader range of activities. The study utilized a specific type of wearable sensor; further research is needed to determine whether the findings generalize to other sensor technologies.