Disease Diagnosis and Severity Prediction: Artificial Intelligence Applications in Hemophilia A

OBJECTIVES: Despite recent advances, the hemophilia A diagnosis accuracy in claims/clinical settings remains unknown. Artificial intelligence (AI) is an emerging reality that has the potential to bring a paradigm shift in hemophilia A. Machine learning (ML) has encouraging results in hemophilia diagnosis, severity prediction, user-centered app, gene therapy, myocardial infarction risk estimation, identification of factor V and CRISPR/Cas9 nuclease off-treatment target. The objective of this review was to explore the AI use in hemophilia A diagnosis and severity prediction.

METHODS: PubMed, EMBASE and secondary searches were conducted on the AI in hemophilia A in clinical trials/real-world settings, from inception till June 2023. Outcomes included diagnosis and severity prediction aiding clinical decision-making using AI.

RESULTS: Of the 108 hits, ten studies utilized AI for diagnostic decision-making (n=5), severity prediction (n=4), and both (n=1) in hemophilia A. Different AI models included ML (Case-Based Reasoning, Expert system, Haemaxpert, Hema-class) and Deep learning (DL) (Graph-based neural network). Validation was described by four studies. Reference standards included biological thrombin generation assay and genetic testing. ML model accuracy for severity prediction ranged between 62%-73.86%; for diagnosis ranged between 80-95.57% with 65% positive predictive value. At a probability threshold of 0.6, 94.4% sensitivity, and 90.1% specificity were seen. A cascade of ML models accurately diagnosed hemophilia (99.18%), its type (98.1%), and severity (96.23%). DL model accuracy for severity prediction was 69%.

CONCLUSIONS: Though AI models for hemophilia A can deal with highly abstract data features and different data types assisting diagnosis and severity prediction, they lack comprehensive quality datasets and face operational, ethical, interpretability, clinically irrelevant performance metrics, and methodological research concerns. To endorse best practices for AI in hemophilia A, it is crucial to develop critical safeguards, transparent policies, and robust data infrastructure. Further research on adapting AI models for implementation in clinical practice is warranted.