Machine Learning of Laboratory Data in Predicting 30-Day Mortality for Adult Hemophagocytic Lymphohistiocytosis

Background: Hemophagocytic Lymphohistiocytosis (HLH) carries a high mortality rate. Current existing risk-evaluation methodologies fall short and improved predictive methods are needed. This study aimed to forecast 30-day mortality in adult HLH patients using 11 distinct machine learning (ML) algorithms.

Methods: A retrospective analysis on 431 adult HLH patients from January 2015 to September 2021 was conducted. Feature selection was executed using the least absolute shrinkage and selection operator. We employed 11 ML algorithms to create prediction models. The area under the curve (AUC), sensitivity, specificity, positive predictive value, negative predictive value, F1 score, calibration curve and decision curve analysis were used to evaluate these models. We assessed feature importance using the SHapley Additive exPlanation (SHAP) approach.

Results: Seven independent predictors emerged as the most valuable features. An AUC between 0.65 and 1.00 was noted among the eleven ML algorithms. The gradient boosting decision tree (GBDT) algorithms demonstrated the most optimal performance (1.00 in the training cohort and 0.80 in the validation cohort). By employing the SHAP method, we identified the variables that contributed to the model and their correlation with 30-day mortality. The AUC of the GBDT algorithms was the highest when using the top 4 (ferritin, UREA, age and thrombin time (TT)) features, reaching 0.99 in the training cohort and 0.83 in the validation cohort. Additionally, we developed a web-based calculator to estimate the risk of 30-day mortality.

Conclusions: With GBDT algorithms applied to laboratory data, accurate prediction of 30-day mortality is achievable. Integrating these algorithms into clinical practice could potentially improve 30-day outcomes.