TLDR: A new study explores using in-home sensors and semi-supervised machine learning to continuously track ALS progression, estimating daily ALSFRS-R scores. The research found that transfer learning, especially with incremental fine-tuning, improves prediction accuracy for individual functional subscales, while self-attention interpolation is best for subscale predictions and linear interpolation for the composite score. The findings suggest that tailoring learning methods to specific functional domains can enhance predictive accuracy, paving the way for more timely and personalized interventions in ALS care.
Amyotrophic Lateral Sclerosis (ALS) is a challenging neurodegenerative disease that progressively weakens muscles and impacts physical function. Traditionally, tracking the decline in ALS relies on periodic clinical assessments using tools like the Revised Amyotrophic Lateral Sclerosis Functional Rating Scale (ALSFRS-R). While valuable, these infrequent visits can miss crucial changes in a patient’s condition that occur between appointments, potentially delaying beneficial interventions.
A recent study, titled “Enhancing ALS Progression Tracking with Semi-Supervised ALSFRS-R Scores Estimated from Ambient Home Health Monitoring,” explores a novel approach to address this gap. Researchers Noah Marchal, William E. Janes, Mihail Popescu, and Xing Song investigated how continuous, unobtrusive monitoring using in-home sensors, combined with advanced machine learning, could provide more frequent and accurate insights into ALS progression. You can read the full paper here.
The ALSFRS-R scale is a self-reported measure that assesses a patient’s ability to perform various functional tasks across different domains, including speech, swallowing, handwriting, walking, and breathing. Scores range from 0 (dependence) to 4 (no difficulty), and a composite score summarizes overall function. However, its subjective nature and the infrequency of its collection in clinical settings limit its ability to capture dynamic changes in real-time.
This research leveraged ambient home health monitoring systems, which include bed mattress sensors to track respiration, pulse, and sleep restlessness, as well as privacy-preserving thermal depth sensors and passive infrared (PIR) motion sensors to detect activity and movement patterns. These sensors collect continuous data without requiring patients to wear devices or actively participate, making the monitoring seamless and non-intrusive.
To bridge the gap between infrequent clinical ALSFRS-R scores and continuous sensor data, the researchers developed semi-supervised regression models. A key technique used was “pseudo-labeling,” where daily ALSFRS-R scores were estimated for periods between actual clinic visits using different interpolation methods: linear, cubic polynomial, and a more advanced self-attention interpolation. This allowed the sensor data to be aligned with estimated functional scores on a daily basis.
The study compared three different machine learning approaches: individual batch learning (models trained on each patient’s data separately), and two types of transfer learning (models initialized on a group of patients and then fine-tuned for individual patients, either in a batch or incrementally). The goal was to see which method best predicted ALSFRS-R trajectories from the in-home sensor data.
The findings revealed several important insights. For predicting individual ALSFRS-R subscales (like speech or handwriting), transfer learning, particularly with incremental fine-tuning, generally improved prediction accuracy and correlation. This suggests that while ALS progression has shared patterns across patients, adapting models to individual patient data is crucial for capturing unique decline trajectories. Interestingly, the choice of pseudo-labeling technique also mattered: self-attention interpolation performed best for subscale-level predictions, capturing complex, non-linear patterns. However, for the overall composite ALSFRS-R score, simpler linear interpolation proved more stable and accurate, aligning with how clinicians often track overall progression.
The research also highlighted that different functional domains exhibit varying degrees of patient-specific patterns versus cohort-level homogeneity. For example, respiratory and speech functions showed more patient-specific patterns, benefiting from personalized model adaptation. In contrast, functions like swallowing and dressing tended to follow more consistent cohort-level trajectories. This suggests that a tailored approach, matching the learning method and pseudo-labeling technique to the specific functional domain, can enhance predictive accuracy.
Also Read:
- Supporting Independent Living: A New Framework for Monitoring Daily Activities of Older Adults
- Advanced Fall Detection: Combining Wearables, AI, and Robots for Senior Care
While the study involved a small cohort of three patients, its findings are promising. They suggest that integrating adaptive model selection within home sensor monitoring platforms could enable more timely interventions and facilitate scalable deployment in larger, multi-center studies in the future. This approach could transform how ALS progression is tracked, moving towards more personalized and proactive care strategies.
