Multimodal Recommender Systems in the Prediction of Disease Comorbidity

원문 초록

While deep-learning based recommender systems utilizing collaborative filtering have been commonly used for recommendation in other domains, their application in the medical domain have been limited. In addition to modeling user-item interactions, we show that deep-learning based recommender systems can be used to model subject-disease code interactions. Two novel applications of deep learning-based recommender systems using Neural Collaborative Filtering (NCF) and Deep Hybrid Filtering (DHF) were utilized for disease diagnosis based on known past patient comorbidities. Two datasets, one incorporating all subject-disease code pairs present in the MIMIC-III database, and the other incorporating the top 50 most commonly occurring diseases, were used for prediction. Accuracy and Hit Ratio@10 were utilized as metrics to estimate model performance. The performance of the NCF model making use of the reduced "top 50" ICD-9 code dataset was found to be lower (accuracy of ~80% and hit ratio@10 of 35%) as compared to the performance of the NCF model trained on all ICD-9 codes (accuracy of ~90% and hit ratio@10 of ~80%). Reasons for the superior performance of the sparser dataset with all ICD codes can be mainly attributed to the higher volume of data and the robustness of deep-learning based recommender systems with modeling sparse data. Additionally, results from the DHF models reflect better performance than the NCF models, with a better accuracy of 94.4% and hit ratio@10 of 85.36%, reflecting the importance of the incorporation of clinical note information. Additionally, compared to literature reports utilizing primarily natural language processing-based predictions for the task of ICD-9 code co-occurrence, the novel deep learning-based recommender systems approach performed better. Overall, the deep learning-based recommender systems have shown promise in predicting disease comorbidity.