Healing Through Language: Transformer-Based Gen AI Pipeline with T5 for Clinical Knowledge Extraction from EHRs

Abstract

In today's scenario, a disease comes with a lot of burden both financially and mentally on the individual starting from doctors checkup, medicines, lab tests. Family doctors are very helpful at such times. The proposed paper introduces SymptoScan, an AI-powered medical assistant chatbot that leverages the strengths of transformer-based language models specifically FLAN-T5 and Qwen2.5-1.5B-Instruct-to assist users in understanding medical conditions, symptoms, treatments, and drug side effects. SymptoScan features a dual-mode interface: (1) A Symptom Scanner for structured condition prediction and drug recommendation based on user-selected body parts and symptoms, and (2) A Medical Q&A module for handling free-form natural language queries related to health concerns. The system integrates a hybrid architecture that combines model-generated reasoning with a structured medical dataset to enhance the accuracy and relevance of predictions. FLAN-T5 is primarily used for interpreting and answering natural language questions, while Qwen2.5-1.5B-Instruct is employed for interactive dialogue and condition/drug explanation. The backend uses pandas for data processing and Gradio to create a user-friendly web interface with dynamic symptom selection. By matching user input with dataset entries and augmenting it with model-generated insights, SymptoScan achieves robust performance, with preliminary evaluation indicating a condition prediction accuracy of approximately 85 %. The system also provides detailed side effect information and dynamically adapts its options based on user selections. SymptoScan represents a scalable and accessible tool in AI-driven healthcare support, merging large language model capabilities with domain-specific data to facilitate informed preliminary diagnosis and health awareness, while underscoring the importance of professional medical consultation.

Ähnliche Arbeiten