Application of Neural-Controlled Differential Equations in Disease Progression Modeling Using Irregularly Sampled Multimodal Data

This paper presents a novel application of neural Controlled Differential Equations (Neural-CDEs) in modeling disease progression using multimodal data. Two deep learning approaches - Long short-term memory (LSTMs) and Neural-CDE models - were evaluated using structured and multimodal (combination of structured and image) time series data. The dataset used in the current study was from the Open-Source Imaging Consortium (OSIC). Using only the structured data, the test RMSE of the Neural-CDE was lower (0.2693) than the LSTM (0.3745), reflecting the better performance of the Neural-CDE approach. Furthermore, modeling the data by the addition of a CT scan to the structured data (multimodal), resulted in an improvement in Neural-CDE performance, with a lower test RMSE (0.2352). Compared to the literature reports, Neural-CDEs using multimodal data performed well.