JMIR Publications

ABSTRACT

Background:

Traditional pulmonary nodule follow-up management relies heavily on manual workflows, leading to inefficiencies and inconsistencies that hinder early lung cancer detection.

Objective:

This study aimed to develop and validate Nodule-AFM, an intelligent, automated system designed to enhance both the efficiency and personalization of pulmonary nodule follow-up management.

Methods:

We developed Nodule-AFM, an AI-driven active follow-up management system based on a multi-agent collaborative framework comprising an Information Agent (IA) for patient data retrieval and synthesis, an Active Dialogue Agent (ADA) for adaptive goal-oriented conversations using Chain-of-Thought Self-Consistency (CoT-SC), and a Summary and Validation Agent (SVA) for quality control and structured reporting. The system was developed using 2,460 follow-up recordings for knowledge construction and inquiry categorization. Prospective evaluation utilized 283 real-world follow-up dialogues, with 85 recordings allocated to validation for prompt engineering and 198 recordings to independent testing. Four experienced physicians independently assessed the 198 test cases.

Results:

Nodule-AFM achieved a 94% conversation completion rate, 84% accuracy in dialogue information extraction, and 79% accuracy in summarizing patient intentions. The system maintained high output quality, with LLM-based evaluation scores exceeding 4.9 across multiple safety and relevance dimensions. Furthermore, it demonstrated generalizability across different LLM backbones and outperformed stand-alone Chain-of-Thought or Self-Consistency methods.

Conclusions:

The findings indicate that a collaborative multi-agent framework can support scalable, clinically reliable, and patient-specific follow-up communication. Nodule-AFM represents a novel LLM-driven approach to pulmonary nodule follow-up management that improves efficiency, accuracy, and patient engagement while reducing clinician workload.