Diagnosing Hemophagocytic Lymphohistiocytosis with Machine Learning: A Proof of Concept

Abstract

Hemophagocytic lymphohistiocytosis is a hyperinflammatory syndrome characterized by uncontrolled activation of immune cells and mediators. Two diagnostic tools are widely used in clinical practice: the HLH-2004 criteria and the Hscore. Despite their good diagnostic performance, these scores were constructed after a selection of variables based on expert consensus. We propose here a machine learning approach to build a classification model for HLH in a cohort of patients selected by glycosylated ferritin dosage in our tertiary center in Lyon, France. On a dataset of 207 adult patients with 26 variables, our model showed good overall diagnostic performances with a sensitivity of 71.4% and high specificity, and positive and negative predictive values which were 100%, 100%, and 96.9%, respectively. Although generalization is difficult on a selected population, this is the first study to date to provide a machine-learning model for HLH detection. Further studies will be required to improve the machine learning model performances with a large number of HLH cases and with appropriate controls.

Keywords: HLH-2004; Hscore; hemophagocytic lymphohistiocytosis; inflammation; machine learning.

Figure 1

Receiver operating characteristic curves for the five different algorithms used for HLH detection. (A) AutoML ensemble model, (B) ridge classifier, (C) CatBoost classifier, (D) K nearest neighbors (KNN) classifier, and (E) stacking classifier with XGBoost as a meta-estimator. Red dots are used to mark the true positive rate and the false positive rate at the threshold used for predictions. Abbreviations: AUC: area under curve; HLH: hemophagocytic lymphohistiocytosis; and ROC: receiver operating characteristic.

Figure 2

Line plot representing the main metrics values of each model used in the study. Abbreviations: KNN: K nearest neighbors; ML: machine learning; PPV: positive predictive value; and NPV: negative predictive value.

Figure 3

Beeswarm plot of Shap values for all patients and for each feature. The right side of the plot favors the output “HLH” and the left side favors the output “not HLH”. The gradient color corresponds to the feature value (dark pink if high or present for categorical features and blue if low or absent for categorical features). Abbreviations: BMI: body mass index; CRP: C reactive protein; PR: prothrombin rate; SGOT/AST: serum glutamic-oxaloacetic transaminase; and SGPT/ALT: serum glutamic-pyruvate transaminase.