Personalized rTMS treatment recommendation with retrieval-augmented LLM reasoning.

Abstract

Repetitive transcranial magnetic stimulation (rTMS) is an effective, non-invasive neuromodulation therapy for major depressive disorder (MDD) and treatment-resistant depression (TRD). However, current clinical practice typically relies on standardized protocols that may not adequately account for individual patient variability. To address this gap, we propose a novel, interpretable framework for personalized rTMS treatment recommendations that combines a pretrained sentence embedding model with large language model (LLM)-based reasoning in a retrieval-augmented generation (RAG) setting. Specifically, our approach leverages a pretrained sentence embedding model to encode structured patient profiles into a dense semantic representation, enabling the retrieval of clinically similar cases. These retrieved examples serve as few-shot prompts for in-context learning (ICL), enabling the LLM to reason over these examples and synthesize customized rTMS treatment parameters. Unlike previous approaches that focus solely on individual aspects of personalization, our framework integrates all key parameters (frequency, intensity, and stimulation mode) into a comprehensive recommendation. We systematically evaluated various sentence embedding models and LLMs. Among them, Bge-large-en-v1.5 for few-shot retrieval and GPT-4o-mini for reasoning achieved the highest rTMS protocol matching accuracy of 78.18% using 15 few-shot examples. Our approach is fine-tuning-free, interpretable, and adaptable to real-world, resource-poor clinical settings, providing a promising step forward in data-driven personalized neurostimulation therapy.