Next-Gen Data Analytics and Intelligent Automation
A Probabilistic Graph-Theoretic Framework for Reducing Medical Hallucinations in Large Language Models
Authors: Samvar Harshil Shah
Publishing Date: 24-12-2025
ISBN: 978-81-975670-6-3
Abstract
Large Language Models (LLMs) show great potential for transforming medical knowledge discovery and clinical decision making. However, their broader adoption is limited by a persistent issue: hallucinations or statements that sound confident but are factually incorrect. Current methods for detecting hallucinations in LLMs rely on deterministic knowledge graphs. That’s a problem — LLM outputs are probablistic and answered my maximum likelihood, not certainity. Thus these system can’t handle ambiguity or partial truth well. They end up failing in subtle but important ways. This project proposes a probabilistic graph-theoretic framework instead. We build a fuzzy similarity graph over medical concepts using semantic embeddings. Edge weights are derived from sigmoid-scaled cosine similarity and interpreted as conditional probabilities. This turns the graph into a Bayesian network that captures uncertainty and dependencies. When an LLM generates a medical claim, we extract it as a subject–predicate–object triple. Then we search the graph for supporting or contradictory evidence. We chain probabilities along paths and compute a final confidence score. Low scores mean a higher risk of hallucinations. We also add a symbolic overlap check using Jaccard similarity to catch paraphrased errors. The end goal: give every LLM-generated medical statement a hallucination risk score for which we have a working proof of concept.
Keywords
Large Language Models (LLMs), medical knowledge discovery, hallucinations, semantic embeddings, sigmoid-scaled cosine, Jaccard similarity.
Cite as
Samvar Harshil Shah, "A Probabilistic Graph-Theoretic Framework for Reducing Medical Hallucinations in Large Language Models", In: Kusum Kumari Bharti and Noor Firdoos Jahan (eds), Next-Gen Data Analytics and Intelligent Automation, SCRS, India, 2025, pp. 23-37. https://doi.org/10.56155/978-81-975670-6-3-3