The HLH-Risk-Calculator is a machine learning-based tool to predict course & mortality of secondary hemophagocytic lymphohistiocytosis

Abstract

Purpose

Secondary hemophagocytic lymphohistiocytosis (sHLH) is a life-threatening hyperinflammatory condition. While few diagnostic scores are established, none exist to predict both clinical course and time-point specific outcome of sHLH patients so far. We present a machine learning (ML)-based tool to predict Initial Disease Severity (IDS; defined as admission to intensive care units (ICU) OR death < 90 days without ICU admission) and mortality across different time points in sHLH patients.

Methods

167 adult sHLH patients from six study centers across three European countries were included retrospectively. Clinical and demographic features, course, survival, and laboratory data were assessed. Random forest models were trained with two sets of eight clinical and laboratory features: one to predict IDS, and five to predict mortality at distinct time points (30, 60, 90, 180 or 365 days). After calibration, the models were tested against hold-out test sets containing n = 32 (IDS) or n = 43 (mortality) sHLH patients.

Results

Overall, the models demonstrated strong discriminatory ability, overall performance, and accurate prediction of risk. Serum levels of the soluble interleukin-2 receptor (sIL-2R) and albumin (for IDS) or sIL-2R and platelet counts (for mortality prediction) showed the strongest contributions to the models’ predictions.

Conclusion

The HLH-Risk-Calculator is an exploratory tool predicting the clinical course of sHLH. External validation is critical to assess its validity, applicability, and robustness for real-world use. To this end, the calculator is available at www.hlh-risk-calculator.com for research use only, and is currently not intended for clinical decision-making.

Visual abstract

Supplementary Information

The online version contains supplementary material available at 10.1007/s00134-026-08515-1.

Take-Home Message

The HLH-Risk-Calculator predicts initial disease severity as well as time-point specific mortality in adult patients with secondary hemophagocytic lymphohistiocytosis. The strongest predictors for adverse outcome were low platelet counts and high levels of the soluble interleukin-2 receptor. The HLH-Risk-Calculator may facilitate rapid identification of high-risk patients and predict long-term outcome. Additional external validation is critical before clinical implementation.

Introduction

Hemophagocytic lymphohistiocytosis (HLH) is a rare life-threatening syndrome characterized by excessive hyperinflammation. Commonly, two forms are distinguished: a primary form (pHLH; genetic), which mainly affects pediatric patients, and a secondary form (sHLH; acquired), which is predominantly found in adults [1].

Although incompletely understood, the pathophysiology of sHLH is considered multifactorial. Various triggers have been described, including infectious diseases (e.g., Epstein-Barr virus (EBV) infection), malignancies, autoimmune disorders, immunotherapies, and others. Genetic predispositions impairing the down-regulation of immune responses, background inflammatory disorders, underlying immunosuppression, and infectious agents are considered to synergistically drive hyperinflammation beyond critical thresholds, potentially resulting in multi-organ failure, hypercoagulation, and death [1, 2].

Clinically, sHLH is difficult to distinguish from sepsis [3]. To guide diagnostic considerations, the HLH-2004 criteria were established [4]. Although initially designed for pediatric pHLH patients, a revised version is frequently applied to adults with suspected sHLH. Criteria include fever, splenomegaly, cytopenia of ≥ 2 lineages, hypertriglyceridemia and/or hypofibrinogenemia, hemophagocytosis, hyperferritinemia, and elevated serum levels of soluble interleukin-2-receptor (sIL-2R) [5]. Other established diagnostic scores include the HScore and the optimized HLH inflammatory index (OHI) [6, 7].

Mortality rates for sHLH patients range between 26 and 75% [8–12], underscoring the importance of swift diagnosis and treatment initiation. Few studies have investigated prognostic markers. Among others, associations with adverse outcomes have been described for male sex, prolonged Partial Prothrombin Time (PTT), low platelet counts, elevated levels of lactate dehydrogenase (LDH), C-reactive protein (CRP), sIL-2R, ferritin, urea, aspartate aminotransferase (ASAT), and malignant disease [13–17]. Few scoring systems stratifying patients into three to four risk groups regarding long-term (2- or 5-year survival) [13, 17] or short-term outcome (30 days) have been presented [18]. At the time of writing, clinical practice still lacks established tools for early recognition of sHLH patients at high risk, granular estimates of prognosis over time, and reliable means to assess disease severity including the need for intensive care management. Such tools would allow better guidance of clinical decisions, timely therapeutic interventions, and prognosis-based stratification of patients for randomized treatment trials.

To this end, we developed the HLH-Risk-Calculator, a machine learning (ML)-based approach composed of six Random Forest models (RFMs). RFMs aggregate predictions from multiple decision trees. In contrast to traditional prognostic models, RFMs can capture complex nonlinear relationships between variables, which is particularly valuable in sHLH given its heterogeneous physiology. Further, their robustness to multicollinearity and missing data is advantageous in a condition that is frequently diagnosed late [15], and values may thus derive from different time points or be missing entirely. The ensemble-learning structure also helps reduce overfitting and typically improves predictive accuracy relative to conventional regression-based scores. Lastly, the ability to perform feature-importance analyses enhances interpretability by providing information on each variable’s individual contribution to the model’s predictions.

Using laboratory and demographic features, the HLH-Risk-calculator predicts Initial Disease Severity (IDS; defined as admission to intensive care units (ICU) OR death < 90 days if not admitted to ICU), as well as time-point specific mortality (30, 60, 90, 180, or 365 days after treatment initiation).

Patients and methods

Patient inclusion & follow-up

167 patients ≥ 18 years of age were included into this study retrospectively if a diagnosis of sHLH was confirmed (fulfilling ≥ 5 HLH-2004 criteria and/or Hscore ≥ 170). Patients derived from six study centers across three European countries (Germany, Switzerland, Austria). 102 patients were included in up to two of three previous publications [14, 15, 19]. 65 patients had not been assessed previously.

The follow-up period was measured from the initiation of effective sHLH treatment, including high-dose corticosteroids administered for alternative diagnostic considerations.

Laboratory parameters

All laboratory parameters derive from the time point of treatment initiation (accepted tolerance: ± 7 days). If repeated measures were performed within this time frame, values with the most significant pathological deviation were retained.

Statistical analyses & model development

Software

Statistical analyses and model training were conducted with Python 3.12.7. Libraries are specified in Supplementary Materials.

Survival analyses

Kaplan–Meier survival curves were generated for all patients based on the most probable HLH trigger. Group comparisons were conducted using the log-rank test. Hazard Ratios (HR) were calculated by means of the univariable Cox proportional hazards model.

Comparison of laboratory parameters

The distribution of laboratory values between groups was compared with a two-sided nonparametric Mann–Whitney U test. Markers with ≤ 30 valid observations per group were excluded.

False discovery rates

P-values were adjusted for multiple testing using the Benjamini–Hochberg method to control the false discovery rate (FDR) if applicable. Adjusted p-values < 0.05 were considered statistically significant.

Imputation

Missing values in the training dataset were imputed using an iterative multivariate approach (IterativeImputer from scikit-learn). 46 variables (presented in supplementary Table 5) were used as auxiliary variables in the imputation, excluding outcome variables (ICU admission, mortality). The study site was not considered due to the small number of patients included at some centers. A RFM was used as the estimator (100 trees, fixed random state, 10 iterations, no posterior sampling). Binary variables were re-binarized following imputation applying a 0.5 threshold. Training and test data sets were defined before imputation.

Random forest model training

First, features were standardized using means and standard deviations from the training set (StandardScaler by scikit-learn). For validation of the final models, the same transformation was applied to the test data. Model training was conducted by means of the RandomForestClassifier (scikit-learn) with pre-specified hyperparameters (300 trees; maximum depth 5; minimum node-split size 10; minimum leaf size 5). Patients with survival data censored at < 365 days were excluded for training and validation of the RFM predicting mortality at 365 days.

Further details on patients, methods, and definitions are provided in the Supplemental Data.

Results

Patient and disease characteristics

167 sHLH patients were included in the study. The median follow-up period with regard to surviving patients was 532.5 [150.0–1328.0] days. Most patients were male (n = 104; 62.3%) with a median age of 53.0 [38.5–68.0] years. 86 (51.1%) patients suffered from an immunosuppressive condition or were on immunosuppressive treatment. 88 (52.7%) patients had a malignant comorbidity with T-cell (n = 26; 15.6%) and B-cell (n = 25; 15.0%) neoplasias being the most frequent ones. The most likely underlying sHLH triggers were of infectious (46.0%), malignant (33.7%), autoimmune (10.2%), or idiopathic (6.4%) nature, followed by vaccines or immunotherapies (3.7%) (supplementary Fig. 1). 86 (52.4%) patients were admitted to ICU at a mean of 5.0 days after treatment initiation (SD 11.1; max. 60 days). Further details are depicted in supplementary Tables 1–3.

Survival

The 1-year overall survival (1y-OS) across all groups was 49.2%. Outcome depended on the underlying HLH trigger (Fig. 1). Patients with malignant triggers (1y-OS = 30.5%) showed significantly worse survival than those with infectious causes (1y-OS = 52.5%; hazard ratio (HR): 1.8, p = 0.014), followed by idiopathic cases (1y-OS = 58.4%; HR: 2.1, p = 0.153). Cases triggered by autoimmune disease (1y-OS = 88.5%) took most favorable courses (HR: 5.6, p < 0.001). The proportional hazards assumption was met in all Cox models (supplementary Table 4). Further comparisons are shown in Fig. 1.

Fig. 1.

Survival of sHLH patients depending on the respective disease trigger. Kaplan-Meier survival curves are plotted based on the most probable underlying sHLH cause. For clarity, survival data were censored after a period of five years. The log-rank test was used for curve comparison. p-values < 0.05 were considered statistically significant. Due to the low case number and short follow-up period, no comparisons were drawn for patients with immunotherapy/vaccines as most likely triggers

Predictors of course and outcome

To identify predictors of severe disease courses, we compared laboratory parameters of patients who were admitted to ICU with patients who were not (Fig. 2A). We found that higher International Normalized Ratios (INR), PTTs, serum levels of bilirubin, ASAT, creatinine, and alkaline phosphatase (ALP), as well as lower levels of serum albumin were significantly associated with ICU admission.

Fig. 2.

A + B Comparison of laboratory values based on clinical course and outcome. Patients were grouped depending on necessity for ICU admission (A) or clinical outcome (B), and their laboratory results are displayed as combined violin and box plots. Each panel depicts individual value distributions via the violin plot. Box width reflects interquartile ranges (IQR), and bold horizontal lines indicate group medians. Whiskers mark data points within 1.5 times the IQR from the lower and upper quartiles (Q1 or Q3), respectively, representing data spread while excluding outliers. Reference ranges are shaded in blue. Comparisons between groups were conducted using Mann–Whitney U tests with FDR-correction for multiple testing. FDR-adjusted p-values are indicated in the respective panels. Only statistically significant comparisons (FDR-adjusted p < 0.05) for laboratory markers with an availability of n ≥ 30 for each group are shown

We next compared laboratory markers of patients who survived (at the end of the individual follow-up time) and those who deceased (Fig. 2B). Lower platelet counts as well as higher PTTs, INRs, and serum levels of bilirubin or sIL-2R were found among patients with adverse outcomes.

To explore correlations between an extended set of variables (n = 46, supplementary Table 5) and ICU admission or outcome, respectively, we conducted Spearman rank correlation analyses. We found no further variables to show statistically significant correlations with ICU admission (Fig. 3A) in addition to those displayed in Fig. 2A. Regarding outcome, we identified 12 variables with significant prognostic meaning (Fig. 3B).

Fig. 3.

A + B. Variables with prognostic value for disease severity or outcome. 46 variables were tested for association with ICU admission (A) or adverse outcome (B) using Spearman rank correlation. The Spearman correlation coefficient (r) is plotted on the x-axis with blue bars reflecting negative and red bars positive correlations. FDR-adjusted p-values are indicated by asterisks (* for p < 0.05; ** for p < 0.01; *** for p < 0.001). Only statistically significant comparisons are shown

Data preparation

45 variables (supplementary Table 5; excluding hemodialysis) were considered for the final models. Patients with missing information on ≥ 20% of the variables, outcome, or ICU admission were excluded, leaving a cohort of n = 152 patients for further data processing.

Imputation

Missing values were imputed as described in the Methods section. Holdout test sets (for IDS: n = 32, 21.1%; for mortality: n = 43, 28.3%) were generated prior to imputation, containing only patients with complete data on all features relevant to the final models. The number of patients with complete information on those features were n = 64 for IDS and n = 62 for mortality.

To assess a possible bias of imputation on the feature-outcome association in the training data, we compared Spearman’s rank correlations between each variable and outcome (or ICU admission, respectively) before and after imputation (supplementary Fig. 2). The maximum change in correlation (Δr) did not exceed 0.1 for any variable.

Definition of outcome variables

The first RFM was trained to predict high Initial Disease Severity (IDS) based on the need for ICU treatment. To minimize potential bias from patients with severe disease courses who were not admitted to the ICU due to treatment limitations (e.g., unfavorable oncological prognosis, frailty, patient preference), we defined the binary outcome variable as follows: admission to ICU (n = 83) OR death within < 90 days if not admitted to ICU (n = 15). This time frame was chosen due to the finding that most deaths and all ICU admissions occurred within this period, and we hence consider it the most dynamic phase of the condition.

The second endpoint was defined as death vs survival at distinct time points (30, 60, 90, 180 and 365 days after treatment initiation).

Identification of predictive features

Using all 45 variables from the imputed training data set, we trained one RFM to predict IDS and five separate RFMs to predict time-point specific mortality. In exploratory 5-fold cross-validation (CV) approaches, we then compared various reduced feature sets (comprised of 4–12 of the most influential predictors for either outcome), and chose the sets with the highest mean area under the curve (AUC) across all folds (and time points, respectively) for training of the final models (Table 1). Further details on feature selection are provided in Supplementary Methods.

Table 1.

Selected features for final model training

Shared features
Platelet count
sIL-2R
Bilirubin
Albumin

Individual features (IDS)	Individual features (mortality)
Creatinine	Age
CRP	PTT
ASAT	INR
Ferritin	Fibrinogen

Final feature sets derive from the explorative 5-fold CV approach. For prediction of either outcome (IDS or time-point specific mortality), a total of eight features was selected, with four features being shared

Model development and evaluation—IDS

We trained a RF classifier to predict the probability of the composite binary target variable IDS using the feature set defined in Table 1. The model was trained on the imputed training dataset. Calibration and threshold tuning were conducted as specified in Supplementary Methods.

Performance of the calibrated model was tested against the held-out test dataset (n = 32). The receiver operating characteristic (ROC) curve (Fig. 4A) demonstrated strong overall discrimination with an AUC of 0.845. The model achieved an accuracy of 81.3%, sensitivity of 91.3%, specificity of 55.6%, with a positive predictive value (PPV) of 84.0%, and negative predictive value (NPV) of 71.4%. The Brier score was 0.159, and the estimated expected calibration error (ECE) over five equal-width probability bins was 0.163 (supplementary Table 7). The calibration plot (supplementary Fig. 3A) showed close agreement between predicted and observed risks without systematic bias. Feature-importance analysis indicated that albumin and sIL-2R were the strongest contributors to the model’s predictions (Fig. 4B).

Fig. 4.

A–E Discriminatory ability, performance metrics, and feature importance of the calibrated models. A ROC curve depicting the IDS model’s performance in predicting the binary outcome in the test dataset. True positive versus false positive rates are plotted at different classification thresholds. The area under the curve (AUC) reflects the overall ability of the model to discriminate between outcomes. B Feature-importance scores quantifying each feature’s average contribution to the IDS model’s decision-making process are ranked in descending order. Higher values correspond to greater impact on the model’s outcome. C ROC curves depicting the calibrated mortality models’ performances on the test dataset in predicting the binary outcome at each time point. D Performance metrics of the calibrated mortality models’ ability to predict the binary outcome (survival vs death) in the test dataset at each time point. E Feature-importance scores of the mortality models were aggregated across all time points, and the mean score for each feature was calculated. Time-point specific feature importance scores are provided in supplementary Fig. 4. A, C, D Detailed metrics alongside 95% confidence intervals are provided in supplementary tables 7+8.

Model development and evaluation—time-point specific mortality

For each time point, we trained a RF classifier to predict the binary outcome (survival vs death) using the feature set specified in Table 1. The models were trained on the imputed training dataset. Calibration and threshold tuning were performed as described in Supplementary Methods.

Model performance was tested against the held-out test cohort (n = 43). The ROC-curves demonstrated strong discriminatory ability at all time points (Fig. 4C). Time-point specific performance metrics are displayed graphically in Fig. 4D and numerically in supplementary Table 8. The mean performance metrics across all time points were as follows: accuracy = 79.3%, PPV = 72.7%, NPV = 85.0%, sensitivity = 82.8%, specificity = 78.2%, AUC = 0.866, Brier score = 0.145, ECE = 0.068. The mean feature-importance scores across all time points are depicted in Fig. 4E. Calibration curves indicated accurate prediction of risk, as observed and expected values aligned closely without systematic over- or underestimation (supplementary Fig. 3B).

Sensitivity analyses

Sensitivity analyses excluding sIL-2R as a predictor, predicting ICU admission as the only endpoint, or using non-imputed data for training are provided in supplementary tables 7 and 9 and supplementary Figs. 5–9.

Discussion

sHLH is a severe condition, which is reflected by the high fraction of patients admitted to ICU (52.4%) and dismal 1y-OS (49.2%). Identifying patients at risk of high IDS early on is of substantial clinical value. While exploratory and not validated for clinical use, the HLH-Risk-Calculator achieved high sensitivity (91.3%) in detecting such patients. At the same time, a PPV of 84.0% and preserved NPV (71.4%) demonstrated balanced performance. The composite endpoint (IDS) was superior to the prediction of ICU admission alone, which is likely attributable to bias introduced by patients who were not admitted to ICU due to treatment limitations. It is important to acknowledge that ICU admission policies vary across institutions, which may introduce site-specific bias. Further, the composite endpoint combines two clinically related, but non-equivalent outcomes, which may complicate interpretation compared with a single, clearly defined endpoint. Accordingly, IDS should be interpreted as a marker of an adverse clinical course rather than as a direct ICU triage tool, and model performance for the composite endpoint should not be extrapolated to the individual endpoint components.

Treatment recommendations for sHLH include pharmacologic regimens that often carry a considerable risk of severe adverse events [1, 20]. Mortality estimates may eventually allow for therapeutic decisions to be weighed against a patient’s individual prognosis, helping to avoid both overly superficial and inappropriately toxic interventions. Often, prognostic considerations are based on the underlying disease (e.g., malignancy), which frequently remains unknown for a long period. The ability to estimate prognosis based on objectively measurable baseline parameters at an early stage may be of substantial clinical value. Stratification of sHLH patients based on prognosis may also represent a critical factor in the design of future prospective randomized treatment trials. Across all predefined time points, the HLH-Risk-Calculator provided reliable predictions with a particularly high negative predictive value (up to 95.7%). Further, it achieved a high sensitivity across all timepoints, thereby also allowing recognition of high-risk patients. Nevertheless, external validation of the models’ performances is critical to assess their real-world applicability and robustness before they can be considered for clinical decision-making.

A major strength of the presented RFMs is the use of feature sets restricted to highly objective variables (i.e., laboratory values, age) as opposed to variables potentially susceptible to bias—e.g., the presumed underlying HLH trigger. The final models may therefore be applied across all sHLH subsets without requiring operator-dependent interpretation of clinical constellations.

Limitations of this study include the small patient number at some study centers, and the asymmetrical distribution and high missingness of sIL-2R values across sites, the latter requiring imputation in a considerable fraction of patients. While our approach closely preserved the feature-outcome association, imputation holds the potential to introduce bias and limit the transferability and robustness of predictive models. Broader availability of sIL-2R would likely further enhance performance of our models, particularly since sIL-2R showed the strongest correlation with survival of all variables assessed. This limitation reflects clinical reality: sIL-2R measurements are not yet broadly established [21] and frequently reported only in a binary fashion, restricting their use in more granular modeling approaches. The implementation of sIL-2R into our models adds further value to our approach, as it has not been considered for any of the predictive scores presented to date [13, 16–18]. Sensitivity analyses demonstrated that despite its limitations, sIL-2R enhances the overall accuracy of our predictions. We hope that our findings will help accelerate the broader implementation of sIL-2R testing in clinical practice, enabling future studies to determine whether models incorporating sIL-2R remain robust, well calibrated, and transportable. In particular, we advocate for high-resolution measurement instead of binary cutoff reporting.

The comparably low PPV (60.0%) at 30 days for mortality prediction may be considered as another limitation. Of note, with a sensitivity of 92.3%, identification of high-risk patients at 30 days is very reliable. In our cohort, most deaths occurred within the first 90 days, which likely explains why patients wrongly identified as high risk at 30 days still carried a substantial risk shortly thereafter. Distinguishing those patients based solely on baseline markers is inherently challenging, and may account for the weaker PPV at this early time point. It may further imply that the underlying disease (i.e., malignancy) significantly drives prognosis. Nevertheless, the model predicting mortality at 30 days achieved a particularly high NPV (95.7%), providing important reassurance in identifying patients at low risk, which is of great value regarding the overall high mortality within the cohort. We therefore retained the 30-day prediction in the calculator, as it adds value despite its lower PPV. To account for this limitation, the entire time horizon of 365 days as well as the probability of high IDS calculated by the complementary model should be considered and evaluated in context.

Due to the orphan disease status [22, 23], clinical data on sHLH remain scarce. Our cohort size was sufficient to train and calibrate models that showed strong performance in the test dataset. Moreover, the multicentric and international nature of our study strengthen the models’ validity. Nevertheless, the limited patient number restricts the capacity of the RFMs to fully leverage their ability to capture complex interactions. Despite careful model design, RFMs still carry a risk of overfitting when trained on small training datasets, and their inability to extrapolate beyond the range of training data may impair predictive performance in extreme clinical presentations. Lastly, the high missingness of sIL-2R values and the necessity to impute missing values can restrict model robustness. External validation in independent cohorts is essential to assess the models’ real-world applicability and eventually support clinical implementation. Until then, the HLH-Risk-Calculator must be viewed as an exploratory tool.

To facilitate external validation and foster collaboration, a web-based application of the HLH-Risk-Calculator will be made publicly available (www.hlh-risk-calculator.com). The authors emphasize that this application is not designed to guide clinical decisions and is destined for research use only.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

MR, JF, MT, TW, LH, MM, CS, MS, PB, KEN, KSG and FB gathered the data. MR conducted statistical analyses. MR, HCS and KS analyzed the data. MR performed model training, calibration and evaluation. HCS advised on model development. MR designed the web application. MR and SJE co-wrote the manuscript. HJS, SSS, HSK, FH, CMW, MH, SJE, MS, AP, DW, MvB and KS reviewed the data and gave conceptual advice. All authors reviewed and commented on previous versions of the manuscript. The final manuscript was approved by all authors.

Funding

Open Access funding enabled and organized by Projekt DEAL. This project did not receive any funding

Data availability

All results are included in the article and supplement. Code and model artifacts are archived at Zenodo (version-specific DOI: 10.5281/zenodo.20810066). Subsequent software updates are tracked through the Zenodo concept DOI: (10.5281/zenodo.17795536) and maintained in the GitHub repository (https://github.com/micruzicka/HLH-Risk-Calculator). Pseudonymized patient data are available from the corresponding author (michael.ruzicka@med.uni-muenchen.de) on reasonable request and subject to ethics and data protection approval.

Declarations

Conflicts of interest

The authors have no competing financial or non-financial interests to declare that are relevant in the context of this article.

Ethical approval

This study and the participation of all study centers involved was approved by the institutional ethics committees of the medical faculty of the LMU Munich (project-numbers: 24-0440, 24-0441, 20-0932) and Medical University of Innsbruck (1324/2021), and performed in accordance with the 1964 Helsinki Declaration and its later amendments.