OC 24.4 - Evaluation of an explainable, decision tree-based AI model for the optimization of thrombophilia diagnosis and thrombosis risk stratification in the outpatient setting

Background: Thrombophilia diagnosis and management is often convoluted as it is affected by many factors. This provides the opportunity for optimization using AI.

Aims: To evaluate an explainable, tree-based AI model for thrombophilia diagnosis and thrombosis risk stratification.

Methods: Clinical and laboratory data were retrospectively assessed from 250 patients who underwent thrombophilia evaluation at our ambulatory coagulation clinic from 11/2019 to 2/2023. XGBoost, a type of supervised learning algorithm, was used to perform a randomized search over a predefined set of tree parameters to find a well-performing configuration. The dataset contained 12 clinical and 26 laboratory parameters. The target variable included thrombophilia probability score (TPS), calculated based on clinical data, and thrombophilia risk factors (TRF), which assign a risk of venous thromboembolism recurrence to each thrombophilia diagnosis based on the current literature (Table). To ensure a representative distribution across the evaluation process, the dataset was divided into three subsets: 200 training, 25 validation, and 25 testing.

Results: The algorithm revealed decisive factors for thrombophilia detection and suggested additional and finer thresholds (e.g. age at first thrombosis was used 20 times, suggesting a mean+/-SD of 47.25+/-10.63 years) for an objective/standardized diagnostic procedure. Of the 19 most relevant features, a larger proportion showed high correlation to TPS than TRF (Figure). A misclassification of +/-1 point was assumed to be no error. We found a sensitivity of 100% and specificity of 100% for TPS, and a sensitivity of 75% and specificity of 98% for TRF for an error tolerance of 1 off.

Conclusion(s): The higher correlation of features to TPS can be explained by the established poor correlation of thrombophilia test results to thrombosis risk. Our results highlight the potential for optimization/standardization of thrombophilia diagnosis and thrombosis risk stratification using AI. This study is currently ongoing with continued patient enrollment, and we expect improved algorithm performance accordingly.