Sparse Multivariate Functional Cox Landmarking Model with Applications to ALS Progression

Evaluating disease progression over time is crucial for decision-making by patients, clinicians, and health-care providers. Dynamic prediction models update survival curves as new disease progression data become available, often by summarizing the data into scalar values. To reduce data compression, we propose a sparse multivariate functional Cox landmarking model. At each landmark time, our approach first applies sparse multivariate functional principal component analysis (FPCA) to estimate individual disease progression trajectories and then uses a linear functional Cox model to dynamically predict individual survival curves. Simulation studies demonstrate the model’s ability to recover functional coefficients in different scenarios and its superior predictive performance. Our model is applied to Amyotrophic lateral sclerosis (ALS) National History Consortium (NHC) data to dynamically predict risk incorporating the full trajectory of disease progression. External validation using data from Emory University’s ALS Clinic further confirms its predictive advantage over non-functional counterparts.