Agentic and Explainable Hybrid Random Forest and LLM Framework for Detecting Healthcare Insurance Fraud

Abstract

One of the biggest challenges in healthcare is healthcare insurance fraud which puts a serious financial burden on healthcare systems worldwide by draining vast sums annually through duplicate billing, upcoding, systematic provider misconduct, and unnecessary procedures. Automated detection systems have become a necessity but, the two leading variations each have a fundamental weakness: traditional machine learning models can detect anomalies in statistical data but fail to explain them, while large language models (LLMs) are good at reasoning in natural language but struggle to stay rooted to hard numerical data when applied to structured fraud classification tasks. To address both the limitations, this paper proposes a hybrid Agentic AI framework, a ReAct-based LLM agent actively reasons over claim evidence along with direct application of domain specific fraud detection rules, with Random Forest (RF) integration acting as the classification agent through an adaptive trust calibration system. The ReAct agent is unlike passive prediction models or post-hoc explanation systems as it dynamically cycles through 4 stages; Thought, Action, Observation, and Final Answer stages which coordinates statistical anomalies with clinical domain knowledge at every step. The framework is evaluated on a dataset of 36,000 synthetic healthcare claims, it achieves an accuracy of 83% and an F1-score of 0.80, which outperforms a zero-shot LLM baseline (Gemini Flash lite) by 34 percentage points in accuracy as well as recall. To demonstrate the practical value of agentic reasoning in fraud auditing, qualitative analysis is used to confirm that the agent is able to distinguish between genuinely suspicious claims and clinically justified ones.