Predicting 30-day mortality in hemophagocytic lymphohistiocytosis: clinical features, biochemical parameters, and machine learning insights

This study aims to evaluate the clinical characteristics and biochemical parameters of hemophagocytic lymphohistiocytosis (HLH) patients to predict 30-day mortality. Parameters analyzed include lymphocyte count (L), platelet count (PLT), total protein (TP), albumin (ALB), blood urea nitrogen (BUN), and activated partial thromboplastin time (APTT). Machine learning (ML) approaches, including LASSO, random forest (RF), and support vector machine (SVM), were employed alongside meta-analysis and sensitivity analysis to validate the prognostic potential of these indicators. A retrospective analysis of 151 HLH patients was conducted to identify key predictive variables. Receiver operating characteristic (ROC) analysis, Kaplan-Meier (K-M) survival curves, and Cox regression analysis were used to evaluate the predictive capabilities of these parameters. ML algorithms determined optimal cut-off values to classify patients into high-risk and low-risk groups. A survival nomogram and risk scoring system were developed to provide individualized prognostic assessments. Meta-analysis aggregated data from existing literature to further validate differences in PLT, ALB, and APTT between deceased and surviving patients. Older age, low L, low PLT, low ALB, elevated BUN, and prolonged APTT were strongly associated with higher 30-day mortality risk in HLH patients. Six key indicators—TP, L, APTT, BUN, ALB, and PLT—were identified as critical predictors. ROC and K-M survival analyses highlighted the significance of these parameters. The survival nomogram and risk scoring system demonstrated high accuracy in predicting individualized mortality risk. Meta-analysis confirmed significant differences in PLT, ALB, and APTT between deceased and surviving patients, reinforcing the clinical value of these indicators. This study underscores the prognostic importance of specific clinical and biochemical parameters in predicting 30-day mortality in HLH patients. By integrating ML methodologies, a survival nomogram and risk scoring system were developed, offering valuable tools for early diagnosis, prognosis assessment, and personalized treatment planning in clinical practice.