M2-DIA: Enhancing Diagnostic Capabilities in Imbalanced Disease Data using Multimodal Diagnostic Ensemble Framework

Abstract

The ADS (Automatic Diagnosis System) has garnered increasing attention for its potential to aid clinicians and enhance healthcare accessibility for patients. These systems are designed to identify a wide range of diseases, including those that may be overlooked by human doctors. However, the diagnostic data used in ADS suffer from disease imbalances due to statistical and medical rare diseases, limiting the scope of diagnosable conditions and finally reducing the accuracy of disease identification. Previous research attempted to address this imbalance using knowledge injection networks during the training phase. However, this approach relies on a training process makes it challenging to diagnose diseases absent from the training data, even with the use of knowledge data. To address this problem, we introduce Multimodal Diagnostic Ensemble Framework called M2-DIA which employs multiple representations—Text, Statistics, and One-hot vectors—derived from a patient’s medical condition. This multi-faceted approach improves its diagnostic capabilities, especially for diseases that are not present in training data. We evaluated our framework on the public dataset with multiple test datasets to verify diagnostic ability from various angles and showed our approach surpasses existing works