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

Area of Science:

• Computational neuroscience
• Medical informatics
• Machine learning in healthcare

Background:

• Neurodegenerative disease assessment commonly uses ordinal rating scales.
• Total scores derived from these scales are often treated as continuous, losing ordinal information and potentially oversimplifying disease complexity.
• Existing methods may not fully capture the multi-faceted nature of disease progression.

Purpose of the Study:

• To develop a novel deep learning model for predicting future patient survival trajectories.
• To leverage longitudinal, item-level ordinal data for improved predictive accuracy.
• To compare the performance of the proposed model against traditional methods using total scores.

Main Methods:

• A convolutional neural network (CNN) architecture was designed to process longitudinal ordinal item data.
• The CNN model was trained to predict future survival trajectories.
• A simulation study and a real-world Parkinson's disease dataset were used for validation.

Main Results:

• The proposed CNN model demonstrated superior predictive performance compared to traditional joint models.
• Utilizing item-level ordinal data significantly improved prediction accuracy over using total scores.
• The model effectively captured disease progression nuances in a Parkinson's disease cohort.

Conclusions:

• Convolutional neural networks can effectively model longitudinal ordinal data for survival prediction in neurodegenerative diseases.
• Item-level data analysis offers advantages over total score aggregation for capturing disease complexity and improving prognostic accuracy.
• This approach holds promise for more precise patient management and clinical trial design in neurodegenerative disease research.