Dynamic modelling of the ALSFRS-R: leveraging population-based scores using neural networks

Abstract

Background Amyotrophic lateral sclerosis (ALS) is a rapidly progressive neurodegenerative disorder with highly heterogeneous trajectories. The Revised ALS Functional Rating Scale (ALSFRS-R) is challenging to model due to irregularly spaced data and patient-level variability. Here we sought to develop and validate a short-horizon prediction tool leveraging a fully connected neural network (FCNN) to forecast individual ALSFRS-R trajectories, providing a natural history benchmark for trials and clinical practice. Methods We retrospectively analysed 29,992 ALSFRS-R measurements from 5319 people living with ALS (plwALS) in the population-based PRECISION-ALS dataset. plwALS were randomised (80:20) into a training and test cohort using group-based splitting. A three-layer FCNN was built in TensorFlow to predict a third ALSFRS-R score given two historical scores and their respective time intervals. Performance was evaluated on the PRECISION-ALS test set and externally on the PROACT database. Linear extrapolation served as a baseline comparator. Findings On the PRECISION-ALS test set, the FCNN achieved a mean absolute error (MAE) of 0.0552 (95% CI 0.0547-0.0576) on a normalised 0-1 scale, corresponding to 2.65 (2.63, 2.76) points on the 48-point ALSFRS-R. This remained consistent across all post-diagnostic periods. The model generalised well to the PROACT dataset, with an improved MAE of 0.0485 (95% CI 0.0481, 0.0489). Linear extrapolation performed significantly worse across all metrics. Error remained consistent across all clinical groups investigated, such as sex, genotype, site of onset, age at diagnosis, age at onset and diagnostic delay. Interpretation A short-horizon FCNN can provide clinically interpretable, individualised ALSFRS-R forecasts from sparse, irregularly spaced data. By supporting rapid identification of those who step outside of the model, this approach holds promise for optimising patient counselling, clinical trial monitoring, and early intervention strategies. This approach allows us to better utilise our growing bank of ALS patient data to support decision making. Funding R McFarlane is supported by a grant from Target ALS, Precision ALS is funded by Taighde & Eacute;ireann (Research Ireland, formerly Science Foundation Ireland). Copyright (c) 2025 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).