Objective  To develop an AI agent dedicated to assisting the diagnosis and treatment of pneumoconiosis, improve clinical communication outcomes and treatment adherence, and promote precision prevention and control of pneumoconiosis.

Methods A cross-sectional study was conducted by recruiting 388 patients with pneumoconiosis in Guangdong Province. Behavioral characteristics and difficulties in this population were analyzed and incorporated into workflow scripts. Unstructured interviews were conducted with 4 patients, 2 respiratory physicians, 1 volunteer from a public welfare organization, and 1 occupational disease litigation attorney. Based on the questionnaire and interview results, an AI agent was developed on the Coze platform, and its effectiveness and advantages were evaluated through specialist physician assessment and comparison with the use of general-purpose large models.

Results A total of 357 patients were ultimately included, of whom 344 were male (96.4%). Questionnaire results showed that, regarding disease cognition, patients’ treatment cognition was positively influenced only by educational level (B = 0.110, β = 0.163, P ＜ 0.01). Symptom cognition was influenced not only positively by educational level (B = 0.082, β = 0.112, P ＜ 0.05) but also negatively by hope level (B = −0.480, β = −0.190, P ＜ 0.001). The total cognition score was positively influenced by educational level (B = 0.192, β = 0.115, P ＜ 0.05) and negatively influenced by hope level (B = −0.754, β = −0.131, P ＜ 0.05). Regarding self-management capacity, female patients had higher self-management capacity than male patients (B = 6.875, β = 0.156, P ＜ 0.01). Educational level, disease stage, degree of media exposure, and hope level all showed positive effects (B = 0.519, β = 0.110, P ＜ 0.05; B = 1.134, β = 0.125, P ＜ 0.05; B = 0.782, β = 0.323, P ＜ 0.001; B = 2.090, β = 0.128, P ＜ 0.05). Unstructured interview results showed that many patients had passive information behaviors, lacked health education support from primary-level organizations, and had a structural imbalance in disease cognition; interviewees without pneumoconiosis generally considered that patients had certain problems in obtaining disease information and in physician-patient communication. The AI agent “E Xiaozhu”, developed to address these issues, includes the following core functions: knowledge-graph-based intelligent Q&A, standardized guidance for symptom self-assessment, generation of personalized health management plans, and pre-visit communication assistance. Two specialist physicians reported that “E Xiaozhu” has adequate reliability and scientific rigor. A comparison of human-AI interaction in response to specific diagnostic and therapeutic questions between the AI agent “E Xiaozhu” and general-purpose large models such as DeepSeek and Xinghuo showed that the AI agent can provide diagnostic and therapeutic information more directly from a specialist physician perspective, and offers emotional support, thereby facilitating patients’ active disease management. Therefore, the AI agent “E Xiaozhu” can optimize the diagnostic and therapeutic communication model by improving treatment efficiency and information completeness, considering patient needs, and analyzing complex lesions.

Conclusion The AI agent developed based on the cognitive and behavioral characteristics of patients with pneumoconiosis can assist in providing more comprehensive diagnosis and life advice, and to improve treatment and communication outcomes.