Multimodal Recommender System in the Prediction of Disease Comorbidity

This paper presents two novel applications of deep learning algorithm-based recommender systems, Neural Collaborative Filtering (NCF) and Deep Hybrid Filtering (DHF), for disease diagnosis. Two datasets, first dataset with all diseases and the second dataset with 50 most commonly occurring diseases, were derived from the MIMIC database. The testing, validation and training accuracy of the model with reduced dataset (50 ICD code) was lower (~ 80%) than the model trained on all ICD-9 codes (~ 90%). The model using all ICD codes performed better (80%), also in terms of hit ratio@10, compared to the model with 50 ICD-9 codes (35%). Reasons for a superior performance with dataset using all ICD can be mainly attributed to the higher volume of data and the powerful nature of deep learning algorithms. Compared to literature reports, the novel approach of using deep recommender systems performed well. Results from the deep hybrid filtering model show better in training accuracy (93.75%) compared to NCF model (90.82%), indicating that the addition of text data from clinical notes provided improved performance in predicting comorbidity. The deep learning-based recommender systems have shown promise in accurately predicting subject disease co-occurrence.