Predicting Extravascular Hemolysis in Paroxysmal Nocturnal Hemoglobinuria

Introduction Paroxysmal Nocturnal Hemoglobinuria (PNH) is a life-threatening blood disorder characterized by the destruction of red blood cells due to complement system dysregulation. The advent of C5 complement inhibitors has markedly improved the outlook for PNH patients by mitigating intravascular hemolytic crises and thrombotic events. However, many patients undergoing C5-inhibitor treatment experience C3-mediated extravascular hemolysis (EVH), which can lead to transfusion dependence, lower quality of life, and poorer health outcomes. The development of proximal complement inhibitors has heightened the need to identify patients at high risk for EVH resulting from C5 inhibitor therapy. Predictive models that identify these high-risk PNH patients could enable personalized, physician-supervised treatment selection, improving clinical outcomes and reducing healthcare costs. We describe a machine learning model, trained on demographic, laboratory, and Next-Generation Sequencing (NGS) data, designed to predict the risk of EVH in PNH patients. Methods We analyzed the medical records of 172 PNH patients treated between 2000 and 2022 at the University of Texas Southwestern and Cleveland Clinic Foundation. The dataset included clinical, laboratory, and NGS sequencing information, with specific clinical variables such as the type and duration of complement inhibitor(s) used, PNH clone size and distribution, EVH occurrence, antecedent aplastic anemia, and laboratory markers of hemolysis. Laboratory markers included lactate dehydrogenase (LDH), hemoglobin, absolute reticulocyte count and percentage, total bilirubin, d-dimer, AST, ALT, direct antiglobulin test (DAT), WBC, MCV, and platelet count. Missing data were imputed using the K-Nearest Neighbors algorithm (K=10). Feature selection through a Random Forest algorithm identified 23 significant clinical markers. A 5-layer multilayer perceptron classification algorithm trained using Leave-One-Out Cross Validation (LOOCV), achieved 87% sensitivity, 75% specificity, and an Area Under the Curve (AUC) of 0.80. Results Out of the 172 patients, 104 started on C5 complement inhibitors, while one began on a C3 inhibitor; eventually, nine were on a C3 inhibitor. About 26% of patients (n=27) were tested for EVH based on persistent anemia, with 15 testing positive for C3 complement activation at DAT evaluation. Of these, 71% were treated with eculizumab, and 28% with ravulizumab. Significant differences in clinical markers, such as Type II RBC levels (predictive score of 0.72), hemoglobin levels (predictive score of 0.42), LDH levels (predictive score of 0.50), and reticulocyte counts (predictive score of 0.46), were observed between EVH-positive and EVH-negative patients. Discussion Our findings are promising, demonstrating a machine learning model capable of predicting EVH with high accuracy. Given the recent approval of C3 and factor B inhibitors and ongoing development of proximal complement inhibitors, predicting EVH in PNH patients on C5 inhibitor therapy is increasingly relevant. Our model represents a significant advancement in identifying the risk of EVH at the time of diagnosis using accessible clinical variables. This predictive capability could enhance treatment monitoring, personalize treatment strategies based on patient risk profiles, and reduce healthcare costs. Future retrospective cohort studies to validate our model on patient data from other institutions would be valuable.