Enhancing ALS Progression Tracking with Semi-Supervised ALSFRS-R Scores Estimated from Ambient Home Health Monitoring

Noah Marchal^1,2*William E Janes³Mihail Popescu^1,2Xing Song^1,2

• ¹University of Missouri Institute for Data Science and Informatics, Columbia, United States
• ²University of Missouri Department of Biomedical Informatics, Biostatistics, and Medical Epidemiology, Columbia, United States
• ³University of Missouri Department of Occupational Therapy, Columbia, United States

Clinical monitoring of functional decline in ALS relies on periodic assessments that may miss critical changes between visits when timely interventions are most beneficial. To address this gap, semi-supervised regression models with pseudo-labeling were developed to estimate rates of decline in a three-patient ALS case series targeting ALSFRS-R trajectories with continuous in-home sensor data. We compared three model paradigms (individual batch learning and cohort-level batch versus incremental fine-tuned transfer learning) across linear slope, cubic polynomial, and ensembled self-attention pseudo-label interpolations. Results showed cohort-level homogeneity across functional domains, with transfer learning reducing prediction error for ALSFRS-R subscales in 28 of 34 contrasts (mean RMSE=0.20(0.14–0.25)). For composite ALSFRS-R scores, individual batch learning was optimal in 2 of 3 participants (mean RMSE=3.15(2.24–4.05)). Self-attention interpolation best captured nonlinear progression, providing the lowest subscale-level error (mean RMSE=0.19(0.15–0.23)), outperforming linear and cubic interpolations in 21 of 34 contrasts. Conversely, linear interpolation produced more accurate composite predictions (mean RMSE=3.13(2.30––3.95)). Distinct homogeneity-heterogeneity profiles were identified across domains, with respiratory and speech functions showing patient-specific progression patterns that improved with personalized incremental fine-tuning, while swallowing and dressing functions followed cohort-level trends suited for batch transfer modeling. These findings indicate that dynamically matching learning and pseudo-labeling techniques to functional domain-specific homogeneity-heterogeneity profiles enhances predictive accuracy in ALS progression tracking. As an exploratory pilot, these results reflect case-level observations rather than population-wide effects. Integrating adaptive model selection into sensor platforms may enable timely interventions as a method for scalable deployment in future multi-center studies.