Development and Application of a Machine Learning-Based Screening Model for Non-Alcoholic Fatty Liver Disease and Economic Benefit Evaluation: A Real-World Study

BACKGROUND:In recent years, machine learning (ML) predictive models for nonalcoholic fatty liver disease (NAFLD) have been extensively studied. However, their real-world application and evaluation of cost-effectiveness remain relatively limited.OBJECTIVE:This study aimed to develop and validate a ML-based diagnostic system for NAFLD, evaluating its accuracy and cost-effectiveness through real-world data validation, with the goal of improving NAFLD screening in primary healthcare settings.METHODS:We recruited participants in two batches between July 2024 and October 2024 from those undergoing medical examinations at Suzhou Municipal Hospital to perform transient elastography (TE) testing. Data from the first batch were used to develop a NAFLD screening model based on the Random Forest (RF) algorithm, with a focus on distinguishing between no, mild, and moderate-to-severe NAFLD. The parameters are specially optimized for screening set. The optimized RF model was then integrated into a web-based platform named AI Plus and applied to a second cohort to assess its real-world diagnostic accuracy. To evaluate the cost-effectiveness of our AI-based diagnostic approach, we compared its economic benefits in a suburban hospital with those of the tertiary hospital clinic during the same period between October 2024, and November 2024.RESULTS: A total of 369 participants were enrolled across four communities affiliated with Suzhou Municipal Hospital, including 101 with mild NAFLD and 110 with moderate-to-severe NAFLD. Body-related indicators such as BMI, WC, and WHR increased consistently with the severity of steatosis. Serum levels of FPG, ALT, AST, hs-CRP, and GGT were significantly higher in patients with moderate-to-severe NAFLD compared to those with mild or no NAFLD (P < 0.05). The 295 cases collected in the first batch were used for model development and software creation, and the developed model was then evaluated in a prospective cohort of 74 individuals in the second batch, achieving AUROC values of 0.82, 0.59, and 0.84 for predicting no, mild, and moderate-to-severe NAFLD, respectively. In a suburban hospital, we issued 106 AI-predicted moderate-to-severe NAFLD reports based on their free residents’ health check-up, and 27 participants completed follow-up visits and TE testing within five weeks. A total of 20 patients were diagnosed with moderate-to-severe NAFLD, with an average cost of 365.85 CYN per confirmed case. During the same period, 23 initial visits were made to the NAFLD specialty clinic at Suzhou Municipal Hospital, resulting in 15 diagnoses of moderate-to-severe NAFLD, with an average cost of 693.20 CNY per case. With a slightly higher level of detection rate (20/27, 74.1% vs 15/23, 65.2%), our ML-based approach reduced consumption by 47.2% (365.85 CNY vs 693.20 CNY) per person, meaning about 40–50 USD.CONCLUSION:AI-based diagnostic models demonstrate significant advantages in enhancing screening efficiency and reducing costs. We hope that our software will be widely adopted in China’s primary healthcare system, especially in medically underserved areas.