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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Amyotrophic lateral sclerosis (ALS) presents heterogeneous patient trajectories, complicating clinical management and research.
• The ALS Functional Rating Scale-Revised (ALSFRS-R) is difficult to model due to irregular data spacing and individual variability.
• Developing accurate prediction tools is crucial for benchmarking clinical trials and informing patient care.

Purpose of the Study:

• To develop and validate a short-horizon prediction tool using a fully connected neural network (FCNN).
• To forecast individual ALSFRS-R trajectories for people living with ALS (plwALS).
• To establish a natural history benchmark for ALS clinical trials and practice.

Main Methods:

• Retrospective analysis of 29,992 ALSFRS-R measurements from 5,319 plwALS in the PRECISION-ALS dataset.
• Development of a three-layer FCNN in TensorFlow to predict future ALSFRS-R scores from historical data.
• Validation on the PRECISION-ALS test set and external validation using the PROACT database, with linear extrapolation as a baseline.

Main Results:

• The FCNN achieved a mean absolute error (MAE) of 0.0552 on the PRECISION-ALS test set (2.65 points on the 48-point ALSFRS-R scale).
• The model demonstrated good generalizability with an improved MAE of 0.0485 on the PROACT dataset.
• Performance was consistent across various clinical subgroups, and the FCNN outperformed linear extrapolation.

Conclusions:

• A short-horizon FCNN effectively provides interpretable, individualized ALSFRS-R forecasts from sparse, irregular data.
• This approach can help identify patients deviating from predicted trajectories, optimizing counselling and monitoring.
• The FCNN model enhances the utilization of ALS patient data for improved decision-making in clinical settings and research.