Background
Applying machine learning (ML) methods on electronic health records (EHRs) can accurately predict the occurrence of a variety of diseases or complications related to medications, which can contribute to improve healthcare quality. However, EHRs contain multiple modalities of data from heterogeneous sources, and these different information domains need to be fused properly. The deep neural network (NN) approach, which offers the ability to learn simultaneously classification and feature representation, can be used to reach the goal.

Methods
In this study, we utilize a large hospital-based EHR database to conduct a cohort for predicting 1-year gastrointestinal (GI) bleeding hospitalizations in patients using anticoagulants or antiplatelet drugs. A total of 815,499 records (16,757 unique patients) were used to construct architectures that jointly information from three different HER modalities (disease, medication, and biomarker domains). We compared the performance of 4 deep multimodal fusion models and 4 other ML approaches.

Results
NNs can result in high prediction performances than random forest (RF), gradient boosting decision tree (GBDT), and logistic regression (LR) approaches. We demonstrate that deep multimodal NNs with early fusion are capable of accurately predicting GI bleeding events (area under the receiver operator curve [AUROC] 0.876) and are significant better than the HAS-BLED score (AUROC 0.668).

Conclusions
In this work, we showed that algorithms based on deep NNs with early fusion can achieve high AUROC for prediction of 1-year GI bleeding hospitalizations and outperform the commonly used HAS-BLED score. Further research is necessary to determine whether the use of these algorithms could lead to an improved healthcare quality.