Application of Artificial Intelligence and Machine Learning Was Not Able to Reliably Predict Poor Outcomes in People With Hemophilia.

Background Artificial intelligence (AI) and machine learning (ML) are currently used in the clinical field to improve the outcome predictions on disease diagnosis and prognosis. However, to date, few AI/ML applications have been reported in rare diseases, such as hemophilia. In this study, taking advantage of the ATHNdataset, an extensive repository of hemostasis and thrombosis data, we aimed to demonstrate the application of AI/ML approaches to build predictive models to identify persons with hemophilia (PwH) who are at risk of poor outcome and to inform providers in clinical decision-making towards helping patients prevent long-term complications. Materials and methods This project was carried out in two steps. First, the data were mined from ATHN 7, a subset study of the ATHNdataset, to determine markers that defined "poor outcome." Second, we applied multiple AI/ML approaches on the ATHNdataset to validate our findings and to develop predictive models to identify PwH at risk of poor outcomes. The classical regression-based predictive model was used as a reference to evaluate the performance of various AI/ML models. Results Our models included features similarly distributed to response variables of interest, resulting in a limited ability to distinguish poor outcomes. Low recall (<53%) resulted in no single model reliably predicting poor outcomes out of all actual positive cases. Our results suggest that, to build a more useful AI/ML model, we may need a larger dataset size along with additional features. Furthermore, our results showed that most of the AI/ML models outperformed the classical logistic regression model in both model accuracy and precision. Conclusions Our AI and ML model showed limited ability to predict poor outcomes in people with hemophilia.