From Lab Results to Clinical Reasoning: Evaluating and Advancing AI-Driven Diagnosis, Causal Understanding and Patient Interpretation in Healthcare

Speaker:

Balu Bhasuran is a Research Faculty member at the School of Information, Florida State University, working in the eHealth Lab. His research focuses on the application of artificial intelligence, natural language processing, and large-scale clinical data to improve healthcare outcomes. He has contributed to multiple interdisciplinary studies involving electronic health records, clinical decision support systems, and large language models in medicine. Prior to this, he has been associated with leading research environments including the Bakar Computational Health Sciences Institute.

Abstract:

Electronic health records (EHRs) contain rich longitudinal clinical data but are difficult to mine due to class imbalance, unstructured text, abbreviations, and selection biases. We developed machine-learning frameworks that integrate structured EHR data, clinical notes, and biomedical knowledge graphs to support clinical reasoning and early diagnosis of rare diseases.

In a multi-center study (UCSF/UCLA), we trained separate referral and diagnosis models for acute hepatic porphyria (AHP)—a rare, treatable condition with ~15-year diagnostic delays. The best models achieved F-scores of 86–92% and could have identified 71% of cases earlier, saving an average of 1.2 years per patient.

We further applied automated machine learning and natural language processing to abstract the Ulcerative Colitis Mayo Endoscopic Subscore from free-text colonoscopy reports, reaching 97% accuracy with strong generalizability across health systems.

These results demonstrate how AI can overcome EHR challenges to accelerate rare-disease diagnosis and automate clinical phenotyping.