Skip to main content
ISTH 2023 Main banner
Back

    Oral Communication

    Hemophilia and Rare Bleeding Disorders

    Hemophilia Clinical - Inhibitors

    OC 50.5 - Prediction of inhibitor development in previously untreated children with hemophilia A using a machine-learning network and information theory quantifiers: Results from the HEMFIL Cohort Study

    Tuesday, June 27, 2023
    11:45 – 12:00 ET
    Room: 516
    Session Available As: Live-Stream and On-Demand
    Background: Predicting inhibitor development in patients with haemophilia A (HA) is not yet possible.

    Aims: To build a prediction model for inhibitor development by using a network of clinical variables and biological biomarkers.

    Methods: Previously untreated children (PUC) with severe/moderately-severe HA (FVIII < 2%) from a prospective cohort (HEMFIL) study were followed until 75 exposure days (ED) or inhibitor development. Clinical data and biological samples were collected before first factor (F) VIII infusion (T0). A predictive model was used to compare the network and potential global topological differences of the patients’ profiles at T0 and their association with inhibitor development. The method transforms related variables into structural patterns via complex networks by maximizing the Network Node Dispersion (Figure 1). Variables are listed in Figure 2A. FVIII genotype was classified as high-risk (inversion, nonsense, frameshift mutations, large deletions and insertions) or low-risk mutations (the remaining).

    Results: We included 95 PUC, all treated with recombinant FVIII who have completed the follow-up [median age at inclusion: 10.0 months (IQR,6.0-15.0)]. Inhibitor was detected in 31(33%) patients of whom 22(71%) were high-titer (>5UB). The algorithm identified a topological association between the PUC by projecting information of the most relevant variables (Figure 2A) and its behavior (Figure 2B). By using cross-validation (80/20), there was an accuracy of 74.2% for PUC who developed inhibitors and 78.1% for the ones who did not. Interestingly, after restricting the analysis to groups with high-risk mutations, the accuracy increased to 82.1% to identify PUC who developed inhibitor and 87.8% to the ones who did not.

    Conclusion(s): Our machine-learning algorithm had a high accuracy (average 76,8%) to predict inhibitor development, increasing to 85.5% (average) when the analysis was restricted to high-risk mutation. We are currently working on a prediction score for inhibitor development based on this analysis, which shall be completed to be presented at the congress.

    Learning Objectives:

    • At the end of this session, participants will be able to discuss the patterns of clinical variables and biological biomarkers in patients with haemophilia.
    • At the end of this session, participants will be able to discuss inhibitor development.
    • At the end of this session, participants will be able to discuss the use of machine-learning methods in haemophilia patients.