Interpretable Disease Prediction from Clinical Text by Leveraging Pattern Disentanglement

For artificial intelligence (AI) systems to be adopted in high stake human-oriented applications, they must be able to make complex decisions in an understandable and interpretable manner. While AI systems today have grown leaps and bounds in predictive power using larger datasets with more complex architectures, existing models remain ineffective at generating interpretable insights in the clinical setting. In this paper, we address the challenge of discovering interpretable insights from the clinical text for disease prediction. For this purpose, we apply the clinical notes from the electronic health records (EHR) available in the Medical Information Mart of Intensive Care III (MIMIC-III) dataset, which are labeled with the international classification of diseases (ICD9) codes. Our proposed algorithm combines interpretable text-based features with a novel pattern discovery and disentanglement algorithm. Specifically, our approach encompasses the following: (1) uncovering strong association patterns between clinical notes and diseases, (2) surpassing baseline clustering algorithms in effectively distinguishing between disease clusters, and (3) demonstrating comparable performance to baseline supervised methods in predicting diseases. Our results validate the model's capability to strike a balance between interpretability and outcome prediction accuracy. By unveiling insightful patterns between clinical notes and diseases, our approach upholds a reasonable level of diagnostic accuracy. Clinical relevance-This paper proposes a novel all-in-one clinical Natural Language Processing (NLP) knowledge base, which can be applied in healthcare systems to discover interpretable insights from the clinical text for predicting medical conditions.