AGGREGATION OF PREDICTIONS FROM LARGE LANGUAGE MODELS VIA GROUP DECISION METHODS FOR AUTOMATIC CODING OF CLINICAL DIAGNOSES

Large language models (LLMs) have been used for the automatic coding of clinical diagnoses, translating textual descriptions into codes from the International Classification of Diseases (ICD-10). However, different LLM architectures and configurations often produce divergent predictions for the same case, generating uncertainty regarding the most appropriate code. This work proposes an approach based on group decision-making methods to consolidate predictions generated by multiple LLMs. Each model is treated as a decision-maker that provides a ranking of the most likely codes, and aggregation techniques such as the Borda count and variants of Condorcet voting are applied to identify the codes with the highest consensus. The approach is evaluated using real hospital discharge summaries from a maternity and neonatal health care facility. The results indicate that aggregating predictions significantly improves coding accuracy, reduces inter-model variability, and outperforms individual LLMs. The findings highlight the potential of group decision-making methods as robust metamodels for complex classification tasks, fostering integration between artificial intelligence and decision support in healthcare systems.