A Temporal Machine Learning-Based Time-to-Event Model for Predicting ALS Progression and Healthcare Utilization

Amyotrophic lateral sclerosis (ALS) is a highly heterogeneous disease, making it difficult for clinicians to predict when patients will reach specific milestones, such as the need for a wheelchair or other assistive devices. This study aimed to create a scalable, interpretable, and dynamic predictive framework that uses longitudinal patient data to forecast functional decline and healthcare utilization, providing a tool for personalized clinical decision support.

The authors constructed a harmonized dataset from the ALS Natural History Consortium, integrating diagnosis records, demographic information, and longitudinal ALS Functional Rating Scale-Revised (ALSFRS-R) assessments. They used correlation-based clustering to identify five coherent functional domains (bulbar, fine motor, gross motor, walking, and respiratory).

To model disease progression, the team employed two primary methods:

1. Generalized Additive Mixed Models (GAMMs): Used to characterize nonlinear, domain-specific functional decline while accounting for repeated measurements.
2. Attention-based Stage Transition Predictor (ASTP): A temporal machine learning model that processes sequences of clinical visits to predict the time-to-next-transition for each functional domain.

Finally, they applied a Cox proportional hazards model to quantify the association between specific functional impairments and the time to wheelchair use, and deployed the resulting framework as an interactive, web-based clinical decision-support tool.

The study found that lower limb function—specifically walking and stair-climbing ability—is the most significant predictor of earlier wheelchair access in ALS patients. The developed ASTP model successfully captures stage-dependent disease progression, allowing for the generation of individualized survival curves that can be updated dynamically as new clinical data becomes available. This approach provides a more nuanced, patient-specific view of disease trajectory compared to static prognostic models.