Objective: To develop a predictive model of amyotrophic lateral sclerosis (ALS) progression using self-administered digital assessments.

Background: Remote fit-for-purpose sensor-based assessments stand to enhance the frequency and precision of patient monitoring during clinical trials. We evaluated the potential of these assessments to predict progression of key clinical features in the ALS functional rating scale (ALSFRS).

Method: ALS patients (n=19) remotely completed weekly cognitive, speech, and mobility assessments, and validated digital ALSFRS scales, on a smartphone and smartwatch during one-year observational study. Continuous raw data collected from assessments included touchscreen, voice, and triaxial accelerometer and gyroscope recordings. Time- and frequency-dependent signal processing extracted digital features from raw data streams. Features selected for further analysis demonstrated significant: (1) test-retest reliability per Spearman correlations; and (2) association with ALSFRS scores and progression per linear mixed effects models. Machine learning methods were used to construct models of ALSFRS scores. Hold-n-out subject-wise cross-validation procedures sorted participants into test (15% of sample) and training sets. Forward stepwise feature selection was performed on training data to reduce multi-collinearity among features. Lasso regression models were constructed from training data and model-based predictions were generated for test data. Spearman correlations evaluated model predictions of ALSFRS scores. Progression profiles of observed and predicted ALSFRS scores were evaluated to assess potential for progression monitoring.

Results: ALSFRS scores showing significant progression were speech, handwriting, walking, self-care, and total score. Of the 3,654 features engineered, 346 showed both significant test-retest reliability and association with ALSFRS scores, which were distributed across gait (n=148), balance (n=75), reading (n=46), articulation (n=28), phonation (n=33), and finger tapping (n=15) assessments. Models significantly predicted ALSFRS total (r=0.592, p<0.00001) and sum of scores (r=0.649, p<0.00001). Comparison of predicted and observed ALSFRS scores over time revealed similar progression profiles.

Conclusion: Remotely monitored sensor-based assessments produced digital endpoints that predicted ALS severity and progression. Further work is necessary to evaluate model validity and therapeutic sensitivity.

« Back to 2023 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/predicting-als-progression-using-remote-sensor-based-assessments/