Deep learning for the joint analysis of item-level longitudinal and survival data

Patients with neurodegenerative diseases are often assessed using rating scales containing a number of items, where each item is scored based on an ordinal scale of 0 to $K$ (0 indicating normal function and $K$ indicating severe impairment). The total score, calculated as the total sum of the items, is commonly used for subsequent analysis due to its simplicity. However, the total score is treated as a continuous value and does not respect the ordinal nature of the item-level data. In addition, the total score may lead to information loss as neurodegenerative diseases are multi-faceted, and using a single numeric value may not effectively represent the disease progression. In this article, we propose a convolutional neural network (CNN) designed to take longitudinal ordinal items as input and predict patients' future survival trajectories. We demonstrate that using the item-level data improves the predictive performance in comparison to traditional joint models using the total score. These advantages are shown through both a simulation study and real data application to a Parkinson's disease study.