Transformer-based Clinical Decision Support Systems using Structured and Unstructured EHR Data

This study presents a different framework of a
Transformer-Based Clinical Decision Support System (CDSS)
that uses structured (e.g., vitals, lab results) and unstructured (e.g., clinical notes) Electronic Health Record (EHR) data to allow for clinically meaningful, accurate, and interpretable clinical predictions. The model consists of a novel dual-stream model that contains a Transformer encoder for the numerical data and a fine-tuned ClinicalBERT module for the textual data that are fused together utilizing a cross-attention layer that captures contextual dependencies of the combined data types. The proposed model was evaluated on the publicly-released multi-modal the EHR dataset - MIMIC-III, and the proposed CDSS obtained an AUROC of 93.2%, accuracy of 91.0%, and F1-score of 90.4%, outperforming all of the current state-of-the-art established baselines such as RETAIN (AUROC: 87.6%), BEHRT (AUROC: 89.8%), and
ClinicalBERT+MLP (AUROC: 90.1%). Moreover, we created
interpretability methods using attention weights and SHAP
values, which allows for clinicians to understand what clinically led to the predictions. Our research shows that when EHR data is processed with transformers as a multi-modal fusion, significant predictive and clinical usability improvements can be obtained in clinical decision support systems and can help to support widespread adoption in the healthcare landscape.