P-742 Using large-scale data to create predictive diagnostic algorithms for 18 reproductive conditions
Study question Can an online health assessment tool, supported by machine learning, accurately screen for 18 reproductive health conditions with high diagnostic confidence? Summary answer Our tool, combining machine learning algorithms and tailored blood tests, achieves 98-99% confidence in diagnosing 18 reproductive pathologies, ensuring clinician-comparable accuracy. What is known already Accurate diagnosis of reproductive conditions like polycystic ovary syndrome (PCOS) and endometriosis is crucial for improving women’s health outcomes. Traditional diagnostic approaches often involve lengthy, invasive, and costly procedures, leading to delayed or missed diagnoses. Machine learning (ML) models have been increasingly applied in healthcare for predictive diagnostics, enhancing accuracy and efficiency. However, few large-scale datasets exist for reproductive health, and most AI-driven diagnostic tools lack continuous learning capabilities. Our study builds on the largest dataset in this field, integrating ML and natural language processing (NLP) models to provide a high-confidence, automated diagnostic platform for reproductive health screening. Study design, size, duration This study utilises the largest dataset of its kind with data from 552,801 women’s health assessments and blood test results from over 48,291 individuals. Our online health assessment (OHA) tool employs 25 fine-tuned XGBoost classification models and NLP techniques to diagnose 18 reproductive conditions. The machine learning models were developed and validated using retrospective data, with ongoing updates to enhance accuracy. The study is designed to continuously improve performance as new data is incorporated. Participants/materials, setting, methods The OHA is built using international diagnostic criteria and guidelines for each condition and gathers comprehensive information on menstrual cycle, medical history, demographics, symptoms, medications and lifestyle habits, integrated with biochemical analysis of serum for up to ten hormones collected via capillary samples. NLP and large language models (LLMs) expedite triage by extracting key clinical details. Diagnostic performance is benchmarked against clinician assessments to ensure reliability and accuracy. Main results and the role of chance Our ML-based diagnostic tool screens for 18 reproductive pathologies with 98-99% confidence in precision and recall. By integrating questionnaire-based assessments with tailored blood tests, it achieves specific, clinician-comparable diagnoses across a broad spectrum of conditions, including PCOS, endometriosis, and hypothyroidism. The system’s probabilistic approach allows for nuanced, condition-specific insights, reducing diagnostic uncertainty. The NLP models facilitate efficient triage by processing and categorising patient-reported symptoms, while LLMs extract essential clinical details, enhancing care pathways. Our continuously learning algorithms improve diagnostic accuracy over time as more data becomes available. This dynamic adaptation ensures the tool remains relevant across diverse populations and evolving clinical guidelines. The model’s performance was validated against clinician-confirmed cases, demonstrating equivalent diagnostic reliability. Variability was minimised through extensive dataset diversity, with rigorous validation ensuring robust predictions. The combination of ML classification and NLP-driven data extraction optimises both accuracy and usability, offering a scalable solution for reproductive health screening. Limitations, reasons for caution While the tool shows high diagnostic accuracy, it relies on self-reported symptoms and blood test availability, which may introduce bias. Continuous model updates mitigate potential inaccuracies, but external clinical validation is necessary to ensure broader applicability. Further studies are needed to assess real-world clinical impact across diverse populations. Wider implications of the findings This AI-driven diagnostic tool enhances accessibility and accuracy in reproductive healthcare, reducing delays in diagnosis and treatment. Its scalable, continuously learning framework supports early intervention, potentially improving health outcomes globally. Future integration into clinical workflows could optimise personalised care and expand AI applications in reproductive medicine. Trial registration number No
Read more