JMIR Publications

ABSTRACT

Background:

For medical diagnosis, clinicians typically begin with a patient's chief complaint, followed by questions about symptoms and medical history, physical examinations, and requests for necessary auxiliary examinations, in order to gather comprehensive medical information. This complex medical investigation process has yet to be modelled by existing artificial intelligence (AI) methodologies. The variety of questions that can be asked during the medical investigation process, especially when it incorporates physical examinations and auxiliary examinations, poses a significant challenge. Additionally, patients vary in their understanding of their own health conditions and their ability to articulate it as chief complaints. Therefore, devising a personalized and efficient inquiry strategy to gather complete medical information from various patients is highly challenging.

Objective:

The aim of this study was to develop an AI-driven medical inquiry assistant for clinical diagnosis, which provided inquiry recommendations by simulating clinicians' medical investigating logic via reinforcement learning. A retrospective analysis was then performed to evaluate the AI model.

Methods:

We compiled de-identified outpatient electronic health records from 76 hospitals in Shenzhen, China, spanning the period from July to November 2021. These records consist of both unstructured textual information and structured laboratory test results. We first performed feature extraction and standardization using natural language processing techniques, and then employed a reinforcement learning actor-critic framework to explore the intrinsic logic of physicians' inquiries. To align the inquiry process with actual clinical practice, we segmented the inquiry into four stages: inquiring about symptoms and medical history, conducting physical examinations, requesting auxiliary examinations, and terminating the inquiry with diagnosis.

Results:

This study focused on two distinct inquiry-and-diagnosis tasks in the emergency and pediatrics departments. The emergency departments provided 339,020 records, and the pediatrics department provided 561,659 records. When conducting its own inquiries in both scenarios, the AI model demonstrated high diagnostic performance with 0.955 (95% CI 0.953-0.956) and 0.943 (95% CI 0.941-0.944) area under the receiver operating characteristic curve (AUROC), respectively. When the AI model was used in simulated collaboration with physicians, it notably reduced the number of physicians inquiries to 46% (6.037/13.26, 95% CI 6.009-6.064) and 43% (6.245/14.364, 95% CI 6.225-6.269), respectively, while achieving 0.972 (95% CI 0.970-0.973) and 0.968 (95% CI 0.967-0.969) AUROC in both scenarios. Furthermore, we analyzed the inquiry sequences of the AI model and conducted external validation to demonstrate its clear and reasonable inquiry logic.

Conclusions:

The experimental results in emergency, pediatrics, and doctor-patient interaction settings demonstrated high diagnostic performance of our AI model in both standalone and AI-clinician collaborative inquiry scenarios. Its investigation process was found to be consistent with the clinicians' medical investigation logic. These findings highlight its promising potential in assisting the decision-making processes of healthcare professionals, ultimately advancing patient care.