A cross-domain recommender system through information transfer for medical diagnosis

Abstract

The electronic diagnostic records of patients, primarily collected by hospitals, comprise valuable data for the development of recommender systems to support physicians in predicting the risks associated with various diseases. For some diseases, the diagnostic record data are not sufficient to train a prediction model to generate recommendations; this is referred to as the data sparsity problem. Cross-domain recommender systems offer a solution to this problem by transferring knowledge from a similar domain (source domain) with sufficient data for modeling to facilitate prediction in the current domain (target domain). However, building a cross-domain recommender system for medical diagnosis presents two challenges: (1) uncertain representations, such as the symptoms characterized by interval numbers, are often used in medical records, and (2) given two different diseases, the feature spaces of the two diagnostic domains are often disparate because the diseases are only likely to share a few symptoms. This study addresses these challenges by proposing a cross-domain recommender system, named information transfer for medical diagnosis (ITMD), to provide physicians with personalized recommendations for disease risks. In ITMD, a novel dissimilarity measurement was performed for diagnosis, represented as interval numbers. The space alignment technique eliminated the feature space divergence caused by different symptoms between two diseases, and the development of collective matrix factorization enabled knowledge transfer between the source and target domains. Experiments and a case study using real-world data demonstrated that ITMD outperforms four baselines and improves the accuracy of recommendations for disease risks in patients to support physicians in determining a final medical diagnosis.