Development and Validation of A Lasso-Logistic Regression Model for Predicting Disseminated Intravascular Coagulation in Pediatric Hemophagocytic Lymphohistiocytosis

Abstract

Background:

Pediatric hemophagocytic lymphohistiocytosis (HLH) patients who develop disseminated intravascular coagulation (DIC) experience rapid disease progression and substantially elevated mortality. Currently, no validated early identification strategies exist for this life-threatening complication. We aimed to develop a prediction model for early DIC detection in pediatric HLH patients.

Methods:

This retrospective cohort study included patients from the Hunan Children’s Hospital HLH database (January 2018–August 2024). Data imbalance was addressed through combined oversampling and undersampling techniques, including Synthetic Minority Oversampling Technique. The cohort was divided into training and validation sets (7:3 ratio). A Least Absolute Shrinkage and Selection Operator (LASSO)-logistic regression model was developed and validated internally, with external validation using prospectively collected cases (January–August 2025). Model performance was evaluated using receiver operating characteristic curves, calibration plots, and decision curve analysis.

Results:

Of 265 included patients, 217 cases were analyzed after the Synthetic Minority Oversampling Technique application. DIC incidence was 42.1% (64/152) and 44.6% (29/65) in training and validation cohorts, respectively. LASSO regression (lambda.1se = 0.08) identified six potential predictors: C-reactive protein, globulin, cholesterol, high-density lipoprotein cholesterol, prothrombin time, and interferon-γ. Multivariable logistic regression confirmed three independent predictors: C-reactive protein prothrombin time, and interferon-γ (all P < 0.05). The model demonstrated robust discriminative performance with an area under the receiver operating characteristic curve (AUROC) of 0.865 (95% confidence interval [CI] 0.809–0.922) in the training cohort and a bootstrap-corrected C-index of 0.857 (95% CI 0.828–0.886). Internal validation yielded an AUROC of 0.797 (95% CI 0.686–0.908), whereas external validation achieved an AUROC of 0.950. Decision curve analysis showed positive net benefit across threshold probabilities of 0% - 99% in training and 0% - 88% in validation sets.

Conclusions:

The LASSO-logistic prediction model, incorporating three readily available biomarkers, demonstrated promising discriminative ability for predicting DIC risk in pediatric HLH. This tool may facilitate early risk stratification and timely therapeutic interventions to improve clinical outcomes.

INTRODUCTION

Hemophagocytic lymphohistiocytosis (HLH) represents a life-threatening hyperinflammatory syndrome characterized by dysregulated immune activation and excessive cytokine production (1). This disorder carries substantial mortality, particularly in pediatric populations (2). Among the various complications of HLH, coagulation dysfunction emerges as a critical determinant of patient outcomes, with disseminated intravascular coagulation (DIC) representing one of the most severe and fatal manifestations, significantly impacting patient survival (3).

DIC is an acquired syndrome characterized by systemic activation of coagulation, resulting in microvascular thrombosis, consumption of coagulation factors and platelets, and paradoxical bleeding (4,5). When DIC develops in HLH patients, the synergistic interaction between inflammatory and coagulation cascades complicates clinical management and worsens prognosis (6). Recent studies indicate DIC occurs in 54.8% of critically ill pediatric HLH patients (7), with associated mortality rates approaching 77% in those with concurrent organ dysfunction (8).

The pathophysiology of DIC in HLH involves multiple interconnected mechanisms, including cytokine storm-induced endothelial damage, platelet dysfunction, tissue factor expression, and hepatic dysfunction with consequent decreased synthesis of coagulation factors (9–11). This complex pathogenesis, combined with overlapping clinical features between HLH and DIC, poses significant diagnostic challenges. Traditional DIC scoring systems, including the International Society on Thrombosis and Haemostasis (ISTH) criteria, demonstrate limited applicability in HLH patients due to confounding effects of the underlying disease on coagulation parameters (12). Current approaches rely primarily on conventional laboratory markers, including activated partial thromboplastin time (APTT), platelet count (PLT), fibrinogen (FIB) levels, and serum ferritin (13–18). While these indicators provide valuable information, they lack the sensitivity and specificity required for early DIC detection in the context of HLH. Novel biomarkers, including thrombomodulin, thrombin–antithrombin complex, and plasmin–antiplasmin complex, show promise for improving diagnostic accuracy (19,20); however, these markers remain unavailable in many clinical settings.

Identifying pediatric HLH patients at high risk for developing DIC before clinical manifestation remains challenging. Machine learning approaches, particularly Least Absolute Shrinkage and Selection Operator (LASSO) regression combined with logistic regression (21,22), have proven effective in handling high-dimensional medical data while maintaining model interpretability—a crucial feature for clinical implementation. To date, no predictive model specifically designed for early DIC identification in pediatric HLH has been validated.

Therefore, this study aimed to develop and validate a LASSO-logistic regression model for predicting DIC occurrence in pediatric HLH patients using readily available clinical and laboratory parameters. By leveraging data from a large cohort of patients with HLH treated at Hunan Children’s Hospital between 2018 and 2024, we sought to create a practical, accurate, and clinically applicable tool that could enhance early DIC detection and ultimately improve patient outcomes in this vulnerable population.

METHODS

Study design and population

This single-center retrospective cohort study analyzed the clinical data of pediatric patients with hemophagocytic lymphohistiocytosis admitted to Hunan Children’s Hospital between January 2018 and August 2024.

The inclusion criteria were as follows: (1) age ≤18 years; (2) confirmed HLH diagnosis according to HLH-2004 criteria (23); (3) absence of DIC at admission; and (4) complete laboratory data within 24 hours of admission. The exclusion criteria were as follows: (1) >20% missing key clinical data; (2) DIC diagnosed at admission; and (3) death within 24 hours of admission precluding DIC assessment.

Class imbalance between DIC and non-DIC groups was addressed using the Synthetic Minority Oversampling Technique (24) combined with undersampling. Based on the reported DIC incidence in HLH patients (30% - 50%) and an expected area under the curve of 0.80, with events per variable ≥10 (25), we calculated a minimum sample size of 150. Accounting for 20% potential data loss, we planned to include at least 180 patients.

The DIC during hospitalization was the primary outcome. Among the outcomes, “discharge without cure” was defined as patients leaving the hospital against medical advice or being transferred to other facilities before completing treatment, without achieving clinical or laboratory improvement; adverse outcomes were defined as death or discharge without cure; whereas “improvement” was defined as survival to hospital discharge with resolution of DIC and/or HLH symptoms, including normalization of coagulation parameters, inflammatory markers, and achievement of clinical stability.

The study protocol was approved by the Medical Ethics Committee of Hunan Children’s Hospital (approval number: HCHLL-2023-97) and complied with the Declaration of Helsinki. Given the retrospective nature of this study, the requirement for informed consent was waived.

Diagnostic criteria

HLH diagnosis required fulfillment of at least five of the eight HLH-2004 criteria (23): (1) fever ≥38.5°C; (2) splenomegaly; (3) cytopenias affecting ≥2 lineages (hemoglobin <90 g/L, platelets <100 × 10⁹/L, neutrophils <1.0 × 10⁹/L); (4) hypertriglyceridemia (≥3.0 mmol/L) and/or hypofibrinogenemia (≤1.5 g/L); (5) ferritin ≥500 μg/L; (6) soluble CD25 ≥2,400 U/mL; (7) decreased or absent natural killer cell activity; and (8) hemophagocytosis in bone marrow, spleen, or lymph nodes.

DIC diagnosis was based on the Chinese DIC Scoring System (CDSS) (26), validated for Chinese pediatric populations and used as the institutional standard throughout our 2018–2025 study period.

The CDSS evaluates three components: (1) underlying disease (0–2 points); (2) clinical manifestations including bleeding tendencies, microcirculatory disturbances, or organ dysfunction (0–3 points); and (3) laboratory parameters: PLT, D-dimer, PT/APTT prolongation, and FIB (0–8 points). For HLH patients, a CDSS score of ≥7 points was conservatively applied for DIC diagnosis when considering nonmalignant hematological diseases. For malignant hematological diseases, the first component (clinical manifestations) was excluded from scoring, and a CDSS score of ≥6 points was used for diagnosis. In this study, we selected CDSS over ISTH criteria (12) given its disease-specific thresholds for HLH-associated coagulopathy, which potentially offer better clinical relevance for our cohort. The main differences between ISTH and CDSS criteria are presented in Supplementary Table 1, http://links.lww.com/SHK/C724.

Data collection and variables

All laboratory data included in the analysis were obtained from the first 24 hours following hospital admission. Data were extracted from electronic medical records based on institutional admission protocols. A subset of patients had expanded laboratory assessments (cytokine panels, lymphocyte subsets) through participation in our ongoing HLH research cohort (Protocol KS2020-34), which prospectively collects research specimens at admission (day 1), day 4, and day 8 under informed consent. For this study, only admission (day 1) values were included in the analysis to ensure all predictors represent baseline risk assessment. Thirty-five candidate predictors were collected spanning hematological, coagulation, biochemical, inflammatory, immunological, and cytokine parameters: PLT, APTT, prothrombin time (PT), FIB, thrombin time, international normalized ratio, antithrombin III, fibrin degradation products, D-dimer, mean corpuscular volume, C-reactive protein (CRP), procalcitonin, serum electrolytes (potassium, sodium, phosphorus), liver function tests, renal function markers, lipid profile, serum ferritin, immunoglobulin A (IgA), natural killer cell percentage (CD3⁻CD16⁺CD56⁺), interleukins (IL-2, IL-4, IL-6, IL-10), interferon-γ (IFN-γ), and Epstein–Barr virus DNA load.

Following the Prediction model Risk Of Bias ASsessment Tool guidelines (25), continuous variables were retained in their original form to preserve potential nonlinear relationships with outcomes.

Statistical analysis

All analyses were performed using R software (version 4.3.3, R Foundation for Statistical Computing, Vienna, Austria), an open-source statistical computing environment freely available at https://www.r-project.org/. Distribution characteristics were assessed using Shapiro–Wilk tests and visual inspection of histograms. For univariate analysis in the training set, continuous variables with normal distribution were expressed as mean ± standard deviation and compared using an independent samples t test, while non-normally distributed variables were presented as median [Q1, Q3] and compared using the Wilcoxon rank sum test. Categorical variables were expressed as frequencies (percentages) and compared using the χ² test, with Fisher exact test applied when the sample size was small (<40) or expected frequencies were insufficient (R package CBCgrps version 2.8.2).

Missing data were handled using multiple imputation with the multivariate imputation by chained equations algorithm (five imputations) (27). Outliers were identified using boxplot methods with 1.5 × interquartile range (IQR) criterion and winsorized at the 1st and 99th percentiles.

The cohort was randomly divided 7:3 into training and internal validation sets using stratified sampling to maintain consistent DIC incidence.

To address potential class imbalance between DIC and non-DIC groups, we applied Synthetic Minority Over-sampling Technique (24), a data-level approach that generates synthetic samples for the minority class by interpolating between existing minority class instances and their k-nearest neighbors (k = 5), combined with random undersampling of the majority class to achieve 1:1 ratio. This technique mitigates overfitting risks associated with simple oversampling while preserving important characteristics of the minority class.

For variable selection, we employed LASSO regression (22), a penalized regression method that performs automatic variable selection by shrinking regression coefficients, with some coefficients reduced to exactly zero. LASSO applies an L1 penalty (λ) to the sum of absolute coefficient values, effectively performing feature selection while preventing overfitting in high-dimensional data. The optimal λ was determined using 10-fold cross-validation with the one-standard-error rule, which selects the most parsimonious model within one standard error of the minimum cross-validation error.

Variables with P < 0.05 in univariate analysis underwent variable selection using LASSO regression with 10-fold cross-validation to determine the optimal regularization parameter λ using the one-standard-error rule (R package glmnet version 4.1-8).

The LASSO-selected variables were incorporated into multivariable logistic regression (R package stats version 4.3.3) to develop the prediction model, and a corresponding nomogram was developed (R package rms version 6.8-1).

For comparison, an ISTH variable model was developed by incorporating the four diagnostic variables from the ISTH DIC scoring system into logistic regression.

Model performance was evaluated through (1) discrimination via receiver operating characteristic (ROC) analysis with 95% confidence intervals (CIs) (R package pROC version 1.18.5); (2) calibration using the Hosmer–Lemeshow test, calibration plots with Bootstrap method (n = 1,000), and Brier scores (R package rms version 6.8-1); and (3) clinical utility via decision curve analysis (R package rmda version 1.6).

Internal validation employed 1,000 bootstrap resamples to obtain bias-corrected performance estimates. Temporal external validation utilized prospectively collected cases between January and August 2025.

Subgroup analyses were performed to evaluate model performance across different HLH etiologies. The concordance between ISTH and CDSS criteria was evaluated. Model development and performance assessment in the primary analysis were based on the CDSS diagnostic criteria for DIC. The final model was also evaluated using the ISTH diagnostic criteria. Statistical significance was set at P < 0.05.

RESULTS

Study cohort characteristics and DIC incidence

Between January 2018 and August 2024, 318 consecutive pediatric patients fulfilling HLH-2004 diagnostic criteria underwent evaluation at Hunan Children’s Hospital. The patient selection process and study workflow are illustrated in Figure 1. In the original cohort, 230 patients (72.3%) required intensive care unit admission, 127 (39.9%) developed coagulation dysfunction, and 33 (10.4%) progressed to DIC. Clinical outcomes differed between groups: among 33 DIC patients, 18 (54.5%) died, 6 (18.2%) were discharged without cure, and 9 (27.3%) showed improvement, yielding an overall adverse outcome rate of 72.7%. In contrast, among 285 patients without DIC, 10 (3.5%) died, 33 (11.6%) were discharged without cure, and 242 (84.9%) improved, with an adverse outcome rate of 15.1%.

Fig. 1.

Study flow chart.

Following application of the exclusion criteria (53 patients excluded: 28 for >20% missing data, 19 for admission DIC, and 6 for death within 24 hours), 265 patients constituted the baseline cohort. Implementation of combined oversampling and undersampling techniques to mitigate class imbalance resulted in 217 patients for the analytical cohort. Stratified random sampling allocated patients to training (n = 152, 70%) and internal validation (n = 65, 30%) cohorts, maintaining proportional DIC representation. The external validation cohort included 24 prospectively enrolled patients between January and August 2025.

The cumulative incidences of DIC were 42.1% (64/152) and 44.6% (29/65) in the training and internal validation cohorts, respectively. Comparative demographic analysis revealed no statistically significant disparities in age distribution (non-DIC vs. DIC: median [IQR]: 3.56 [1.49–6.78] vs. 2.17 [1.43–4.07] years, P = 0.057) or sex composition (male predominance: 52.27% vs. 56.25%, P = 0.748) between the groups (Table 1). However, clinical outcomes diverged markedly, with case fatality rates of 60.9% (39/64) in DIC patients versus 6.8% (6/88) in non-DIC patients (relative risk = 8.95, 95% CI 4.01–19.97, P < 0.001). Primary HLH had a higher prevalence in the DIC cohort (20.31% vs. 9.09%, P = 0.048).

Table 1.

Baseline characteristics of HLH patients with and without DIC in the training cohort

Characteristics	Total (n = 152)	Non-DIC group (n = 88)	DIC group (n = 64)	P
Age, years	2.66 (1.46, 5.96)	3.56 (1.49, 6.78)	2.17 (1.43, 4.07)	0.057
Sex, n (%)				0.748
Male	82 (53.95%)	46 (52.27%)	36 (56.25%)
Female	70 (46.05%)	42 (47.73%)	28 (43.75%)
Prognosis, n (%)				<0.001
Improvement	74 (48.70%)	67 (76.10%)	7 (10.90%)
Noncure	33 (21.70%)	15 (17.00%)	18 (28.10%)
Death	45 (29.60%)	6 (6.80%)	39 (60.90%)
Primary etiology classification, n (%)				0.023
Primary HLH	21 (13.82%)	8 (9.09%)	13 (20.31%)
Infection-associated HLH	87 (57.24%)	55 (62.50%)	32 (50.00%)
Malignancy-associated HLH	37 (24.34%)	18 (20.45%)	19 (29.69%)
MAS-HLH	5 (3.29%)	5 (5.68%)	0 (0.00%)
Idiopathic HLH	2 (1.32%)	2 (2.27%)	0 (0.00%)
PLT, ×109/L	76.00 (34.00, 148.00)	86.00 (54.00, 170.00)	57.00 (30.00, 112.00)	0.004
APTT, seconds	48.74 (42.50, 58.88)	47.10 (41.08, 53.90)	49.46 (44.38, 61.57)	0.117
PT, seconds	14.60 (13.40, 17.10)	14.15 (13.18, 15.65)	15.39 (14.34, 19.18)	<0.001
FIB, mg/dL	209.00 (133.00, 264.00)	214.00 (137.00, 299.00)	202.00 (128.00, 239.00)	0.058
TT, seconds	20.10 (18.20, 24.73)	19.90 (18.18, 24.20)	20.51 (18.50, 25.42)	0.235
INR	1.16 (1.03, 1.42)	1.10 (1.00, 1.26)	1.23 (1.12, 1.63)	<0.001
AT3, %	92.00 ± 25.00	99.00 ± 25.00	82.00 ± 23.00	<0.001
FDP, mg/L	16.25 (7.64, 38.76)	11.77 (6.68, 27.41)	18.64 (12.73, 51.47)	0.003
D-Dimer, mg/L	1.99 (1.29, 6.23)	1.78 (0.99, 3.04)	4.26 (1.64, 9.40)	<0.001

Statistical analysis: χ² test was used for categorical variables; Fisher exact test was applied when expected cell count was <5.

APTT, activated partial thromboplastin time; AT3, antithrombin III; CRP, C-reactive protein; DIC, disseminated intravascular coagulation; FDP, fibrin degradation products; FIB, fibrinogen; HLH, hemophagocytic lymphohistiocytosis; INR, international normalized ratio; PLT, platelet count; PT, prothrombin time; TT, thrombin time.

Biomarker profiling and feature selection

Comprehensive laboratory assessments within 24 hours of admission revealed distinctive biomarker signatures that distinguished DIC evolution. Table 2 presents key laboratory parameters, including LASSO-selected predictors and essential HLH markers, with additional parameters detailed in Supplementary Table 3, http://links.lww.com/SHK/C725. The DIC cohort exhibited significantly elevated acute-phase reactants (CRP: median [IQR] in DIC vs. non-DIC: 45.73 [15.89–71.77] vs. 9.18 [3.10–27.61] mg/L, P < 0.001), prolonged clotting times (PT: 15.39 [14.34–19.18] vs. 14.15 [13.18–15.65] s, P < 0.001; APTT: 49.46 [44.38–61.57] vs. 47.10 [41.08–53.90] s, P < 0.001), and profoundly elevated pro-inflammatory cytokines (IFN-γ: 1,373.15 [160.08–2,500.00] vs. 36.62 [6.26–173.62] pg/mL, P < 0.001; IL-6: 67.19 [27.96–208.63] vs. 23.83 [10.58–38.80] pg/mL, P < 0.001).

Table 2.

Key laboratory parameters in HLH patients stratified by DIC status in the training cohort

Parameters*	Total (n = 152)	Non-DIC group (n = 88)	DIC group (n = 64)	P
CRP, mg/L	16.89 (5.45, 55.34)	9.18 (3.10, 27.61)	45.73 (15.89, 71.77)	<0.001
GLO, g/L	23.05 (19.88, 28.30)	26.00 (21.10, 32.17)	22.00 (19.02, 24.07)	<0.001
CHOL, mmol/L	2.54 (1.97, 3.15)	2.92 (2.24, 3.88)	2.21 (1.59, 2.72)	<0.001
HDL-c, mmol/L	0.41 (0.30, 0.63)	0.51 (0.36, 0.86)	0.33 (0.24, 0.41)	<0.001
SF, ng/mL	2,001.17 (896.15, 6,597.44)	1,500.00 (407.27, 5,489.35)	2,878.58 (1,494.35, 7,572.02)	0.002
NK Cells (CD3⁻CD16⁺CD56⁺), %	6.42 (3.19, 10.30)	5.13 (2.69, 8.97)	7.91 (4.45, 11.73)	0.007
IL—6.00, pg/mL	37.51 (17.97, 101.08)	23.83 (10.58, 38.80)	67.19 (27.96, 208.63)	<0.001
IL—4.00, pg/mL	3.23 (2.53, 3.81)	3.61 (2.61, 3.92)	2.85 (2.50, 3.46)	0.004
IL—2.00, pg/mL	3.87 (3.00, 4.44)	4.05 (3.77, 4.53)	3.19 (2.58, 4.12)	<0.001
IL—10.00, pg/mL	375.80 (29.84, 2120.87)	70.42 (22.09, 899.49)	1,343.28 (102.14, 2,500.00)	<0.001
IFN-γ, pg/mL	135.29 (11.10, 1,353.49)	36.62 (6.26, 173.62)	1,373.15 (160.08, 2,500.00)	<0.001
EBV-DNA†, copies/mL	180,000.00 (35,700.00, 482,000.00)	165,000.00 (10,707.50, 349,500.00)	195,000.00 (92,403.50, 666,855.10)	0.031

Continuous parameters are expressed as median (25th–75th percentile) for non-normally distributed data and as mean ± SD for normally distributed data.

Categorical parameters are presented as numbers (percentages).

^*Displayed parameters with statistically significant differences (P < 0.05) between groups.

^†EBV-DNA detection rate was 53.38%. The values shown represent viral loads among patients with detectable levels.

CHOL, cholesterol; CRP, C-reactive protein; EBV-DNA, Epstein–Barr virus DNA load; GLO, globulin; HDL-c, high-density lipoprotein cholesterol; IFN-γ, interferon-gamma; IL, interleukin; NK cells, natural killer cells; SD, standard deviation; SF, serum ferritin;.

Epstein–Barr virus DNA load was detected in 53.38% of the patients tested. Among those with detectable levels, median viral loads were 195,000 [92,403.50–666,855.10] copies/mL in DIC patients versus 165,000 [10,707.50–349,500.00] copies/mL in non-DIC patients (P = 0.031).

LASSO regularization with 10-fold cross-validation, employing the one-standard-error rule (lambda.1se = 0.08) to optimize the bias-variance trade-off (23), identified six nonzero coefficients from 35 candidate predictors: CRP, GLO, CHOL, high-density lipoprotein cholesterol, PT, and IFN-γ (Supplementary Fig. 1, http://links.lww.com/SHK/C726; Supplementary Table 4, http://links.lww.com/SHK/C727). These variables were subsequently evaluated using multivariable logistic regression. Multicollinearity diagnostics confirmed acceptable variance inflation factors (all <2.0) and condition indices (all <30), thus supporting variable independence.

Multivariable modeling and nomogram construction

Multivariable logistic regression incorporating the six LASSO-selected features revealed that three variables (CRP, PT, and IFN-γ) remained statistically significant independent risk factors for DIC development (Table 3, all P < 0.05). The final prediction model was constructed using these three variables to yield the following equation:

Table 3.

Independent predictors for DIC in pediatric HLH using multivariable LASSO-logistic regression analysis

Variable	β coefficient	OR	95% CI	SE	Wald	P
CRP	0.011	1.011	1.002–1.020	0.004	2.528	0.011
PT	0.171	1.187	1.078–1.324	0.052	3.314	0.001
IFN-γ	0.001	1.001	1.000–1.001	0.000	4.245	<0.001
Intercept	−5.827	0.003	0.000–0.016	0.910	−6.406	<0.001

CI, confidence interval; CRP, C-reactive protein; IFN-γ, interferon-gamma; OR, odds ratio; PT, prothrombin time; SE, standard error.

$$\begin{array}{cc}\mathrm{l}\mathrm{o}\mathrm{g}\mathrm{i}\mathrm{t}\left(\mathrm{P}\right)& =-5.827+0.011\times \mathrm{C}\mathrm{R}\mathrm{P}\\ & +0.171\times \mathrm{P}\mathrm{T}+0.001\times \mathrm{I}\mathrm{F}\mathrm{N}-\gamma \end{array}$$

The nomogram translation facilitated clinical implementation through visual risk stratification (Fig. 2).

Fig. 2.

Nomogram for predicting HLH-DIC based on LASSO-logistic regression. A, The nomogram consists of six horizontal axes integrating three predictors (CRP, PT, and IFN-γ) for risk assessment. The top “Points” axis (0–100) serves as the reference scale for converting predictor values to point contributions. The three predictor axes display: CRP (mg/L, 0–180), PT (seconds, 10–45), and IFN-γ (pg/mL, 0–5,500). The “Total Points” axis (0–160) represents the cumulative score from all predictors. The bottom “Risk” axis converts total points to predicted HLH-DIC probability (0.1–0.99). B, The scoring reference tables facilitate clinical application. The upper panel provides exact point assignments for CRP (0–180 mg/L, 0–63 points), PT (10–45 seconds, 0–97 points), and IFN-γ (0–5,500 pg/mL, 0–100 points). The lower panel displays the risk probability conversion table mapping total points (11–127) to risk probabilities (0.1–0.99). The blue arrow indicates the workflow: sum individual predictor points to calculate total points, then determine the corresponding HLH-DIC risk probability. CRP, C-reactive protein; DIC, disseminated intravascular coagulation; HLH, hemophagocytic lymphohistiocytosis; IL, interleukin; IFN-γ, interferon-gamma; LASSO, Least Absolute Shrinkage and Selection Operator; PT, prothrombin time.

Model performance metrics

Discrimination analysis revealed the superior performance of the LASSO-logistic regression model. In the training cohort, the area under the ROC curve (AUROC) was 0.865 (95% CI 0.809–0.922). The maximum Youden index was 0.587, corresponding to an optimal threshold of 0.269, with a sensitivity of 85.9% and a specificity of 72.7%. The bootstrap-corrected C-index was 0.857 (95% CI 0.828–0.886) based on 1,000 resamples (Table 4, Fig. 3, Supplementary Fig. 2, http://links.lww.com/SHK/C728).

Table 4.

Predictive performance of DIC screening score in pediatric HLH

Model	Set	Sen (%)	Spe (%)	FNR (%)	FPR (%)	PPV (%)	NPV (%)
Lasso-logistic	Training	85.9	72.7	14.1	27.3	69.6	87.7
ISTH variable	Training	56.2	78.4	43.8	21.6	65.5	71.1
Lasso-logistic	Validation	72.4	77.8	27.6	22.2	72.4	77.8
Lasso-logistic	External validation	100	85.0	0.0	15.0	50.0	100

The ISTH variable model was developed by incorporating the four diagnostic variables from the International Society on Thrombosis and Haemostasis (ISTH) DIC scoring system.

Sen, sensitivity; Spe, specificity; FNR, false negative rate; FPR, false positive rate; PPV, positive predictive value; NPV, negative predictive value.

Fig. 3.

ROC curve analysis comparing the LASSO-logistic regression model and the ISTH variable models. A, Training cohort: The LASSO-logistic regression model (red line, AUC = 0.865, 95% CI 0.809–0.922) outperformed the ISTH variable model (blue line, AUC = 0.696, 95% CI 0.6–0.782). B, Validation cohort: The LASSO-logistic regression model (red line, AUC = 0.797, 95% CI 0.686–0.908) and the ISTH variable model (blue line, AUC = 0.759, 95% CI 0.637–0.880) showed similar performance. Gray diagonal line represents no discrimination (AUC = 0.5). AUC, area under the curve; CI, confidence interval; ISTH, International Society on Thrombosis and Haemostasis; LASSO, Least Absolute Shrinkage and Selection Operator; ROC, receiver operating characteristic.

Internal validation demonstrated an AUROC of 0.797 (95% CI 0.686–0.908). The maximum Youden index was 0.502, corresponding to an optimal threshold of 0.468 with a sensitivity of 72.4% and specificity of 77.8%. External validation performance showed an AUROC of 0.950 (95% CI 0.89–1.00).

Calibration assessment via the Hosmer–Lemeshow statistic indicated an adequate model fit for both the training (χ² = 9.31, P = 0.32) and validation sets (P = 0.78). Bootstrap calibration (n = 1,000) showed mean absolute errors of 0.053 (training) and 0.058 (validation) with consistency indices of 0.857 and 0.823, respectively. A Brier score of 0.152 suggests an acceptable calibration accuracy. Calibration plots demonstrated a satisfactory agreement between the predicted and observed probabilities across the risk deciles (Supplementary Fig. 3, http://links.lww.com/SHK/C729).

Decision curve analysis revealed substantial clinical utility, with positive net benefit across threshold probabilities ranging from 0% to 99% in the training set and 0% to 88% in the validation set, significantly exceeding both treat-all and treat-none strategies. This broad range of clinical applicability supports the model’s potential for risk-stratified management across diverse clinical scenarios (Supplementary Fig. 4, http://links.lww.com/SHK/C730).

Stratified performance analysis

Subgroup analyses demonstrated a robust performance across HLH etiologies (Supplementary Table 5, http://links.lww.com/SHK/C731). For noninfectious HLH, the AUROC was 0.824 (95% CI 0.752–0.896), while infectious HLH showed consistent discrimination with an AUROC of 0.789 (95% CI 0.718–0.860).

Model performance across diagnostic criteria

The concordance between CDSS and ISTH diagnostic criteria was assessed in our 265-patient cohort. Among these patients, 28 were positive by both criteria, 200 were negative by both criteria, 0 were positive by CDSS but negative by ISTH, and 37 were negative by CDSS but positive by ISTH, yielding moderate agreement (Cohen’s κ = 0.53). When ISTH criteria were applied as the diagnostic standard for DIC, the LASSO-logistic regression model maintained comparable discriminative performance, with an AUC of 0.759 (95% CI 0.686–0.832). DeLong test revealed no significant difference in model performance between the two diagnostic standards (P = 0.79), suggesting robust predictive capacity regardless of the DIC diagnostic criteria employed (Supplementary Table 1, http://links.lww.com/SHK/C724).

DISCUSSION

This investigation represents the first application of LASSO-regularized logistic regression for DIC prediction in pediatric HLH, addressing a critical gap in risk stratification tools for this high-mortality complication. Through systematic feature selection from 35 candidate variables, LASSO regression identified six potential predictors, with subsequent multivariable analysis confirming three independent risk factors. The three-biomarker model achieved robust discrimination (AUROC = 0.865, bootstrap-corrected C-index = 0.857) while maintaining clinical feasibility through reliance on routinely available laboratory parameters.

The biomarkers retained through regularization reflect distinct yet interconnected pathophysiological cascades underlying HLH-associated coagulopathy. IFN-γ emerged as the most discriminative predictor, with concentrations elevated 37.5-fold in DIC patients. This pleiotropic cytokine orchestrates the hyperinflammatory response characteristic of HLH through multiple mechanisms (9,11): (1) enhanced macrophage activation via JAK-STAT signaling; (2) upregulation of tissue factor expression on monocytes and endothelium; (3) suppression of physiological anticoagulant pathways including the protein C system; and (4) promotion of neutrophil extracellular trap formation. Recognition of the central role of IFN-γ has led to the development of emapalumab, a targeted anti-IFN-γ monoclonal antibody that received FDA approval in 2018 for the treatment of refractory, relapsed, or progressive primary HLH, highlighting the clinical relevance of this pathway (28). CRP, synthesized hepatically in response to IL-6 stimulation, serves as a quantitative marker of systemic inflammation intensity (2). The observed 5-fold elevation of CRP in DIC patients corroborates previous investigations linking inflammatory burden to coagulopathy progression through multiple mechanisms, including endothelial activation, tissue factor upregulation, and natural anticoagulant consumption (6,14). PT prolongation directly quantifies consumption of vitamin K-dependent coagulation factors (II, VII, IX, X) and fibrinogen, representing the consumptive coagulopathy fundamental to DIC pathogenesis (4,5). The dose–response relationship observed (odds ratio = 1.187 per second) aligns with international consensus recognizing PT as a core DIC diagnostic parameter (12). The integration of these three biomarkers provides a comprehensive risk assessment superior to individual parameter monitoring. CRP reflects inflammatory burden, IFN-γ represents cytokine-mediated pathophysiology, and PT quantifies consumptive coagulopathy.

The practical implementation of this model in clinical workflow can be conceptualized through the nomogram-based risk stratification system (Fig. 2). Upon HLH diagnosis, clinicians obtain the three required biomarkers (CRP, PT, and IFN-γ), all of which are routinely available in tertiary pediatric centers within 24 hours of admission. Using the nomogram, each biomarker value is converted to points: for example, a patient with CRP 80 mg/L (28 points), PT 20 seconds (28 points), and IFN-γ 2,000 pg/mL (36 points) would accumulate 92 total points, corresponding to greater than 90% DIC risk. This quantitative risk score then informs a tiered monitoring and intervention strategy. Patients with low-risk scores (<34 points, <30% DIC probability) may continue with standard monitoring protocols, while intermediate-risk (34–63 points, 30% - 70% probability) and high-risk (>63 points, >70% probability) patients may warrant intensified surveillance and proactive supportive interventions, with specific protocols determined by institutional practices and clinical judgment.

Recent work by Wang et al. has explored prognostic models in adult HLH-related coagulopathy (29). Our study complements these efforts by developing a pediatric-specific tool for early DIC prediction, incorporating both inflammatory markers and coagulation parameters.

Treatment options for HLH remain limited, with the HLH-94 protocol continuing as the standard first-line approach for controlling acute inflammation (1,10,22,30). Our model does not fundamentally alter these therapeutic algorithms. Rather, its primary contributions lie in three specific domains. First, it enables risk-stratified monitoring intensity for more efficient resource allocation. Second, it facilitates refined prognostic counseling to provide families with quantitative outcome expectations. Third, it may inform the timing of supportive interventions. However, we must acknowledge that whether such risk-stratified approaches translate to improved clinical outcomes remains unproven. This is an important limitation of our study: while we demonstrate statistical discrimination, we have not established that using this model to guide clinical decisions reduces DIC-related morbidity or mortality. HLH is inherently a multisystem disorder requiring comprehensive supportive care irrespective of DIC risk, including intensive immunosuppression, management of cytopenias, infection surveillance, and organ support. The model does not suggest withholding standard HLH treatment in low-risk patients, nor does it identify specific DIC-preventive therapies to be added in high-risk patients (as such therapies remain investigational). Rather, it provides a quantitative risk assessment that may optimize resource allocation and monitoring intensity. Future prospective studies evaluating intervention protocols guided by this model are essential to determine whether early recognition genuinely improves patient outcomes or primarily serves as a prognostic tool.

This study has several methodological strengths. We adhered to transparent reporting of a multivariable prediction model for individual prognosis or diagnosis reporting guidelines (21) and employed LASSO regularization (lambda.1se = 0.08) for systematic variable selection from 35 candidates while minimizing overfitting (22). Both internal and temporal external validation confirmed model stability. Decision curve analysis demonstrated clinical utility across a broad range of threshold probabilities.

Several limitations should be acknowledged. The single-center retrospective design may limit generalizability to other populations and healthcare settings. Outcome assessment was limited to the index hospitalization without post-discharge follow-up, which may introduce some outcome misclassification for patients discharged without cure or against medical advice. The modest sample size, while adequate for model development, may constrain precision for certain subgroup analyses. The high external validation performance (AUROC = 0.950) should be interpreted cautiously, given the small validation cohort (n = 24). Additionally, we used baseline measurements rather than dynamic biomarker trajectories, and lacked genetic data to explore whether molecular HLH subtypes might require tailored prediction approaches. Most importantly, we have not yet demonstrated that implementing this model improves clinical outcomes. Prospective intervention trials are needed to determine whether risk-stratified care reduces DIC-related morbidity or mortality, and whether early risk identification translates to better outcomes or serves primarily for prognosis and resource allocation.

CONCLUSIONS

We developed and validated a LASSO-logistic prediction model incorporating three readily available biomarkers (CRP, PT, and IFN-γ) that demonstrated promising discriminative ability for identifying pediatric HLH patients at risk for DIC development. The model enables risk stratification at admission, potentially allowing clinicians to implement intensified monitoring protocols in high-risk patients. While these findings require validation in larger, multicenter cohorts and prospective implementation trials before clinical adoption, this work represents an important step toward improving early recognition of patients at risk for this life-threatening complication.

ACKNOWLEDGMENTS

The authors would like to thank Dan Yang for her assistance in data management.