Artificial intelligence-driven precision treatment of reproductive medicine-related diseases: the optimal protocol choice for IVF-ET.

INTRODUCTION: Optimal ovarian stimulation (OS) selection is critical for IVF success, but expert-based decisions often lack consistency in outcomes, cost-efficiency, and personalization, highlighting the need for more individualized and data-driven approaches. OBJECTIVES: This study propose an artificial intelligence (AI) system that analyzes extensive IVF-ET cycles to uncover OS-pregnancy outcome relationships, enabling personalized treatment recommendations while improving success rates and minimizing unnecessary costs. METHODS: This study analyzed anonymized data from 17,791 patients undergoing OS and IVF/ICSI at Tongji Hospital between May 2015 and May 2019. An adaptive AI model was developed to predict key indicators-including progesterone (P), number of oocytes retrieved (NOR), estradiol (E2), and endometrial thickness (EMT) on the hCG day-by integrating personal characteristics, ovarian reserve, and etiological factors. This model facilitated personalized OS selection, pregnancy outcome grading, and the development of an AI-driven clinical decision support system (CDSS). RESULTS: The key indicators-progesterone (P), number of oocytes retrieved (NOR), estradiol (E2), and endometrial thickness (EMT) on the hCG day-were used to establish a pregnancy grading system. Pregnancy rates are stratified as follows: Level IV (Total Score 15-16), 0.55; Level III (Total Score 13-14), 0.44; Level II (Total Score 11-12), 0.24; and Level I (Total Score 4-10), 0.07. After OS optimization, 1,355 patients who were initially at level I were elevated to a better level. Of the 2,341 patients initially in level II, 2,290 improved, and of the 3,839 initially in level III, 1,448 improved. Patients elevated to level IV accounted for 80 percent of all cases. The CDSS prioritized a GnRH antagonist regimen for 54.64 % of patients, resulting in per-patient time savings of 15.39-33.48 days and cost reductions of ¥989-¥2,623 compared to non-optimal to antagonist. Scaled to China's > 1 million ART cycles annually, this corresponds to projected direct savings of approximately ¥0.54-1.43 billion per year. In the new evaluation datasets (n = 4,251), implementation of CDSS recommendations increased the clinical pregnancy rate from 0.452 to 0.512 (p < 0.001) and reduced mean per-cycle cost from ¥7,385 to ¥7,242 (p = 0.018), demonstarting cost-effectiveness dominance with ICER saving of ¥2,383 per additional clinical pregnancy. CONCLUSION: This AI-assisted CDSS streamlines clinicians' decision-making by enabling efficient and accurate initial judgments on OS, standardizing and personalizing recommendations, and optimizing OS for effectiveness and cost-efficiency.