This invention discloses a method, medium, and device for identifying DRG over-compensation based on
invoice reverse engineering and causal imprinting. The method includes: acquiring detailed medical billing data from the medical records to be audited; extracting features of treatment items and mapping them to a
medical knowledge graph to obtain a set of feature atoms; inputting the feature atom set into a large
language model, outputting real
disease phenotypes as pseudo-labels, and performing semantic comparison with the reported primary diagnosis to identify over-compensation of
the primary diagnosis; classifying the sensitivity of other reported diagnoses by
score, screening high-value complication entities, and verifying their
resource consumption matching degree through a
clinical pathway constraint
library to identify fictitious complications; extracting the main billing items and their low-value billing imprint rule
library, and identifying falsely inflated primary intervention items through counterfactual causal comparison; finally, calculating a comprehensive anomaly index based on deviation, matching degree, and
divergence, and outputting an audit report. This invention automatically identifies DRG over-compensation anomalies from multiple dimensions, improving the accuracy and comprehensiveness of audits.