Interoperable web platform based on large language models for medical data analysis

Abstract

This Thesis presents an interoperable web platform for managing medical data, prioritizing security and integration of information from multiple sources using FHIR (Fast Healthcare Interoperability Resources). The objective is to optimize medical record analysis with artificial intelligence (AI) and machine learning, offering automatic alerts and preventive recommendations while complying with the General Data Protection Law (LGPD). The platform facilitates efficient sharing of data among hospitals, clinics, remote devices, and healthcare systems, improving diagnostic and treatment accuracy. The methodology involved creating a secure, LGPD-compliant web platform, the integration of data through FHIR to ensure interoperability. AI algorithms analyze medical data, generate alerts, and provide personalized recommendations. Performance was assessed in controlled and stress tests, focusing on scalability and security. Results highlighted promising performance of the Retrieve Augmentation Generation (RAG) technique with BAAI/bge-small-en embedding models. Metrics such as BertF1, BertP, and BertR ranged from 0.389 to 0.538, averaging 0.43, indicating moderate consistency. The average Bleu score was 0.442, reflecting diverse response quality, while Rouge metrics averaged 0.326, indicating lower precision. Performance with Chest X-rays and MedQA datasets showed better results with Chest X-rays, achieving higher scores but higher perplexity (3.635e4), indicating challenges in generating clinical text. MedQA showed greater response diversity (0.807) but lower precision. In qualitative analysis, Chest X-rays demonstrated higher semantic similarity (mean 0.767) compared to MedQA (mean 0.754). During load testing, the platform remained stable as user numbers increased, but response times grew under stress, suggesting bottlenecks in high-demand scenarios. In conclusion, the platform is a promising tool for integrating medical data and supporting clinical decisions. The FHIR standard ensured interoperability, while AI effectively analyzed records and issued alerts. Adjustments are needed in response times under heavy loads and improvements in infrastructure and mobile experience to encourage greater patient adoption.