Explainable zero-shot learning and transfer learning for real time Indian healthcare

Clinical note research is globally recognized, but work on real-time data, particularly from India, is still lagging. This study initiated by training models on medical information mart for intensive care (MIMIC) clinical notes, focusing on conditions like chronic kidney disease (CKD), myocardial infarction (MI), and asthma using the structured medical domain bidirectional encoder representations from transformers (SMDBERT) model. Subsequently, these models were applied to an Indian dataset obtained from two hospitals. The key difference between publicly available datasets and real-time data lies in the prevalence of certain diseases. For example, in a real-time setting, tuberculosis may exist, but the MIMIC dataset lacks corresponding clinical notes. Thus, an innovative approach was developed by combining a fine-tuned SMDBERT model with a customized zero-shot learning method to effectively analyze tuberculosis-related clinical notes. Another research gap is the lack of explainability because deep learning (DL) models are inherently black-box. To further strengthen the reliability of the models, local interpretable model-agnostic explanations (LIME) and shapley additive explanations (SHAP) explanations were projected along with narrative explanations which generated explanations in a natural language format. Thus, the research provides a significant contribution with ensemble technique of zero-shot learning and SMDBERT model with an accuracy of 0.92 as against the specialized models like scientific BERT (SCIBERT), biomedical BERT (BIOBERT) and clinical BioBERT.