PREDICTIVE MODELLING OF DIAGNOSTIC ERRORS USING EHR

Diagnostic errors such as missed, delayed, or incorrect diagnoses remain a major concern in healthcare systems worldwide. Despite the widespread adoption of Electronic Health Records (EHR), clinicians often face challenges in interpreting large volumes of patient data effectively, which can lead to diagnostic mismatches. This research proposes a predictive modelling framework for identifying diagnostic errors using EHR data through an Explainable Artificial Intelligence (XAI) based Clinical Decision Support System (CDSS). The system integrates structured data such as laboratory values and vital signs, unstructured clinical notes, and temporal patient records to detect potential diagnostic discordance. Structured patient information is processed using the XGBoost algorithm to identify key diagnostic indicators, while temporal trends in patient vitals are captured using a Long Short-Term Memory (LSTM) neural network with an attention mechanism. Additionally, natural language processing techniques such as TFIDF and BioBERT are used to analyze clinical notes and extract meaningful patterns from textual medical data. The outputs from these models are combined using an ensemble learning approach to produce a Diagnostic Discordance Score that highlights potential mismatches between clinical evidence and physician diagnoses. Explainability is ensured through SHAP feature importance and attention visualization, allowing clinicians to understand model reasoning. A full-stack webbased implementation using FastAPI, React, and MySQL provides role-based access for doctors, reviewers, administrators, and researchers. The system also includes a human-in-the-loop validation mechanism where experts review predictions to improve model performance over time. The proposed framework enhances diagnostic safety, improves clinical decision-making, and demonstrates how explainable AI can support healthcare professionals in reducing diagnostic errors and improving patient outcomes.