Deep Sequence Modelling of Non-Stationary Disease Progression in Amyotrophic Lateral Sclerosis from Simulated Longitudinal Data

Amyotrophic Lateral Sclerosis (ALS) is a progressive neurodegenerative disease that affects motor neurons in the brain and spinal cord, leading to the gradual loss of voluntary muscle control. Most patients experience rapid functional decline and have a life expectancy of 3 to 5 years from the first symptom’s onset. Despite extensive research, there is still no known single cause or curative treatment for ALS. In this context, improving patients’ quality of life relies heavily on timely and personalized symptom management. A key step toward this goal is the development of accurate prognostic models capable of predicting individual disease trajectories, enabling clinicians to make informed decisions about interventions and care planning. Modelling ALS progression presents substantial challenges due to limited real-world longitudinal data, varying visit frequencies, and the inherently non-stationary nature of its evolution as patients transition across disease stages. This thesis addresses these challenges through a hybrid approach that combines probabilistic modelling and deep sequence models. First, we propose a generative modelling approach to simulate realistic patient trajectories using a Dynamic Bayesian Network (DBN) trained on the PRO-ACT dataset, which is the largest publicly available ALS clinical trial dataset. The DBN captures temporal dependencies and disease phase transitions, enabling the creation of a large, synthetic cohort with diverse disease trajectories and progression patterns across different clinical phases. On top of this simulated data, we train and evaluate deep sequence models, including Recurrent Neural Networks such as LSTM and GRU networks, as well as Transformer models with multi-head attention, to predict future values of the ALS Functional Rating Scale subscores. Special attention is given to the impact of non-stationarity on model performance, as disease dynamics shift across phases. We compare the predictive capabilities of these models and their ability to generalize across heterogeneous patient trajectories. The proposed framework offers a scalable and interpretable approach for studying complex, evolving diseases and contributes a reproducible pipeline for synthetic data generation, sequence modelling, and progression prediction.