Forecasting Hematological Activity in Antiphospholipid Syndrome Using Predictive Models

Introduction:

Hematological activity in antiphospholipid syndrome (APS) involves thrombocytopenia and autoimmune hemolytic anemia (AIHA). Only thrombocytopenia was recently included in the new ACR/EULAR 2023 classification criteria with platelet counts > 20 x 10⁹/L which may not adequately identify some patients with this presentation. In a previous study, differences were found regarding titers and positivity of antibodies between hematologic and thrombotic phenotype. To our knowledge, currently there are no models that help predict hematological activity in APS.

Aim:

To train various machine learning models to predict the development of hematological activity in APS defined as moderate to severe thrombocytopenia and/or AIHA.

Materials and methods:

We perform a retrospective study that consecutively included patients from 1990 to 2023 with primary APS according to the Sapporo criteria and who had hematological activity. Variables such as age, sex, blood cell counts at diagnosis (hemoglobin, leukocytes, neutrophils, lymphocytes, platelets), mean platelet volume (MPV), positivity for lupus anticoagulant, and titers (IgG and IgM) for anti-cardiolipin (aCL), anti-β2-glycoprotein-I (anti-β2GPI), and antiphosphatidyl serine/prothrombin antibodies were evaluated.

Numerical variables were standardized and normalized with imputation performed using the K-nearest neighbors' method; the optimal k value was 4 using F1 scores.

We selected six machine learning models for prediction: Logistic Regression, Support Vector Machine, Random Forest Classifier, AdaBoost Classifier, Gradient Boosting Machine (GBM), and Extreme Gradient Boosting Machine (XGBoost). The dataset was split into training and testing sets with an 80-20 split to evaluate the models' performance on unseen data. To ensure the robustness and reliability of the models, cross-validation with 5 folds was performed.

The models were evaluated using several performance metrics: accuracy, precision, recall, and F1 score to assess the predictive capabilities. Additionally, the Receiver Operating Characteristic Area Under the Curve (ROC-AUC) was calculated to evaluate the models' ability to distinguish between classes. SHAP (Shapley Additive exPlanations) values were used to interpret feature importance for the models, and they were compared to patients with APS without hematological activity.

Results:

A total of 97 patients with primary APS were included and 66 (68%) had hematological activity. From the group with hematological activity, the mean age was 32.9 + 10.8 years; 48 (72.2%) were women and had a median follow-up of 131 (12-360) months. Regarding patients with hematological activity, 62 (93%) had thrombocytopenia, 15 (22%) was moderate with a median platelet count (MPC) of 64 x 10⁹/L (61-85), and 33 (50%) had severe with a MPC of 11 x 10⁹/L (1-45). Eleven (16%) patients had < 20 x 10⁹/L platelets. A total of 13 (19.6%) had AIHA, and 9 (13.6%) patients had both thrombocytopenia and AIHA.

Of the six models evaluated, GBM and XGBoost demonstrated the best performance in predicting hematologic activity. The GBM model achieved the highest accuracy at 75.37%, with a precision of 77.61%, recall of 76.42%, F1 score of 77.84%, and ROC-AUC score of 0.75. The most significant features using SHAP values were higher levels of IgM anti-β2GPI, MPV, and basal platelet counts at diagnosis, with mean absolute SHAP values of 1.264, 0.932, and 0.619, respectively.

The XGBoost model had an accuracy of 74.32%, precision of 75.80%, recall of 74.32%, F1 score of 73.80% and ROC-AUC of 0.75. This model highlighted higher IgM anti-β2GPI and basal platelet counts at diagnosis and lower IgM aCL as key features, with mean absolute SHAP values of 1.627, 0.710, and 0.687, respectively.

Conclusion:

These findings show the robustness and reliability of the GBM and XGBoost models in predicting hematologic activity and its development in APS with specific laboratory and clinical characteristics identified as crucial predictors. Predictive models can help create risk calculators as a next step. These results could help with monitoring and clinical management of APS patients.

No relevant conflicts of interest to declare.