Development and Internal Validation of an Inflammatory-Coagulation Biomarker–Enhanced Nomogram for 90-Day Outcome After Deep Hypertension-Related Intracerebral Hemorrhage

Abstract

Background

Early prognostic assessment after intracerebral hemorrhage (ICH) relies mainly on neurological severity and neuroimaging features. Whether routinely available inflammatory-coagulation biomarkers provide meaningful incremental information beyond these established predictors remains uncertain.

Objective

To develop and internally validate an admission-based nomogram incorporating inflammatory-coagulation biomarkers for predicting 90-day poor functional outcome in patients with deep hypertension-related ICH.

Methods

This single-center retrospective cohort study included consecutive patients with deep hypertension-related ICH admitted between January 2020 and December 2025. Poor functional outcome was defined as a 90-day modified Rankin Scale score of 4–6. Candidate predictors available within 24 hours of admission were evaluated using least absolute shrinkage and selection operator logistic regression followed by multivariable logistic regression. Model performance was assessed using discrimination, calibration, Brier score, bootstrap internal validation, reclassification metrics, and decision curve analysis.

Results

Among 462 patients, 162 (35.1%) had poor functional outcome. The final model retained age, Glasgow Coma Scale score, hematoma volume, intraventricular hemorrhage, admission blood glucose, neutrophil-to-lymphocyte ratio, and ln(D-dimer). Adding NLR and ln(D-dimer) increased the AUC from 0.765 to 0.828 (P = 0.008), representing a statistically significant but modest absolute improvement of 0.063, and reduced the Brier score from 0.178 to 0.142. The optimism-corrected AUC was 0.816. Decision curve analysis showed greater net benefit than the conventional model across threshold probabilities of 15%–65%. Sensitivity analyses yielded consistent biomarker associations and model performance.

Conclusion

Adding NLR and D-dimer to established clinical-imaging predictors provided a modest improvement in 90-day outcome prediction in deep hypertension-related ICH. These biomarkers may offer complementary prognostic information, but the clinical importance of this incremental gain remains to be established. External validation is required before broader clinical application.

Introduction

Intracerebral hemorrhage (ICH) remains a major cause of stroke-related death and disability worldwide.^1,2 Hypertension is a principal modifiable risk factor and is particularly associated with deep hemorrhages involving the basal ganglia, thalamus, brainstem, and cerebellum.³ Early prognostic assessment in this population is clinically important for risk stratification, monitoring, rehabilitation planning, and communication with patients and families.

Outcome prediction after ICH has traditionally relied on neurological severity and neuroimaging features, including level of consciousness, hematoma volume, hematoma location, and intraventricular extension. However, secondary injury also involves inflammatory and coagulation-fibrinolytic responses that may provide prognostic information beyond the mechanical burden of hemorrhage.⁴ Among routinely available blood markers, the neutrophil-to-lymphocyte ratio (NLR) and D-dimer have received particular attention. Kim et al reported associations of NLR and PLR with hematoma expansion and adverse outcomes after acute ICH,⁵ whereas Zhou et al, in a large retrospective cohort, found that admission D-dimer was independently associated with 3-month poor functional outcome and mortality and evaluated its addition to an ICH outcome score.⁶ These studies support the prognostic relevance of inflammatory and coagulation markers, but they primarily established biomarker–outcome associations or biomarker-specific predictive performance rather than directly quantifying the joint incremental value of NLR and D-dimer over a prespecified clinical-imaging model for 90-day functional outcome.

Recent prediction studies have also emphasized the importance of phenotype-specific modeling. Chen et al developed a nomogram for early acute hydrocephalus after spontaneous ICH, illustrating the value of combining routinely available clinical and imaging characteristics for a specific post-ICH complication.⁷ In parallel, work on non-vascular structural-related intraventricular hemorrhage has demonstrated substantial etiological and prognostic heterogeneity across hemorrhagic phenotypes.⁸ These findings support more focused prognostic modeling rather than treating all spontaneous hemorrhages as a single population. Nevertheless, whether routinely available inflammatory-coagulation biomarkers provide measurable incremental prognostic information beyond established admission clinical and imaging predictors specifically in deep hypertension-related ICH remains insufficiently defined. Moreover, a useful prediction model should be evaluated not only by discrimination but also by calibration, internal validation, and clinical utility.⁹

Therefore, this study aimed to develop and internally validate an admission-based prognostic model for 90-day poor functional outcome in patients with deep hypertension-related ICH. We specifically evaluated whether adding NLR, PLR, and D-dimer to a conventional clinical-imaging model improved discrimination, overall prediction error, calibration, reclassification, and clinical net benefit. Rather than evaluating these biomarkers as novel prognostic markers in isolation, the study was designed to quantify their incremental contribution beyond established predictors within a clinically and etiologically more homogeneous ICH population.

Methods

Study Design

This single-center retrospective cohort study analyzed data systematically extracted from the electronic medical record (EMR) system of the neurology, neurosurgery, and stroke center departments. We evaluated a consecutive cohort of patients diagnosed with hypertensive ICH admitted between January 2020 and December 2025. The study protocol was reviewed and approved by the Ethics Committee of Huanggang Central Hospital (approval No. HGYY-KY-2026-027). This retrospective study used pre-existing clinical records generated during routine medical care. All study-specific patient identification, eligibility screening, data extraction, database construction, and statistical analyses were initiated only after ethics approval had been obtained. The requirement for written informed consent was waived by the Ethics Committee because the study involved only de-identified retrospective data and no additional intervention or patient contact. All research procedures strictly adhered to the ethical standards set forth in the Declaration of Helsinki and were conducted in alignment with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) and the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) reporting guidelines.

Study Population

The study cohort was identified from the institutional medical record system using International Classification of Diseases, Tenth Revision (ICD-10) codes for ICH (I61.0–I61.9). Consecutive adult patients admitted to the neurology, neurosurgery, and stroke center departments between January 2020 and December 2025 were screened. Eligible records were reviewed by trained clinical researchers using clinical, laboratory, and neuroimaging data to confirm spontaneous ICH, hematoma location, and likely etiology. Initial non-contrast computed tomography was reviewed in all patients, with computed tomography angiography or venography, magnetic resonance imaging or angiography, and digital subtraction angiography additionally reviewed when available from routine clinical evaluation. Uncertain eligibility, location, or etiology was adjudicated by a senior neurologist or neurosurgeon.

Patients were eligible if they were aged ≥18 years, were admitted within 24 hours after symptom onset or last-known-well time, had spontaneous parenchymal ICH confirmed by non-contrast computed tomography, had baseline complete blood count, coagulation profile, and D-dimer measurements within 24 hours after admission, and had an available 90-day modified Rankin Scale (mRS) assessment. Deep hypertension-related ICH was operationally defined as hemorrhage centered in the basal ganglia, thalamus, brainstem, or cerebellum in a patient with documented pre-existing hypertension or pre-admission antihypertensive treatment and no identified alternative macrovascular, structural, traumatic, treatment-related, or major coagulopathic cause.^10,11 Lobar ICH was not included.

Patients were excluded for an identified secondary ICH etiology, including macrovascular or structural lesions, cerebral venous thrombosis, intracranial tumor, hemorrhagic transformation of ischemic stroke, treatment- or procedure-related hemorrhage, or trauma; primary intraventricular or other non-parenchymal intracranial hemorrhage; anticoagulation-associated ICH or severe coagulopathy requiring specific reversal therapy;¹² or major conditions likely to substantially affect inflammatory or coagulation biomarkers, including severe hepatic failure, end-stage renal disease, hematological disease, active malignancy, severe systemic infection, or major surgery, severe trauma, or immunosuppressive therapy within the preceding month. Additional exclusions were pre-stroke mRS >2, inability to confirm admission within 24 hours of onset, missing critical baseline imaging or laboratory data, and pre-existing terminal illness or palliative-care status. Pre-admission antiplatelet therapy was recorded but was not an exclusion criterion.

Data Collection and Outcome Definition

Clinical, neuroimaging, laboratory, treatment, and follow-up data were extracted from the electronic medical record, picture archiving and communication system, and laboratory information system by trained researchers using standardized electronic forms.

Baseline predictors were restricted to information available within 24 hours after admission. For laboratory biomarkers, the earliest available venous blood sample was used, with exact collection times obtained from the laboratory information system. Sampling time was compared with documented timestamps for minimally invasive hematoma aspiration, craniotomy, external ventricular drainage, and blood-product transfusion. When complete blood count and D-dimer were obtained separately, pre-intervention sampling required both specimens to precede the earliest qualifying intervention; otherwise, timing was classified as post-intervention or indeterminate.

The primary outcome was poor functional outcome at 90 days, assessed using the mRS.¹³ Assessments within 90 ± 14 days were accepted. Favorable outcome was defined as mRS 0–3 and poor outcome as mRS 4–6, with death assigned an mRS score of 6.^10,14 Outcomes were obtained from outpatient, rehabilitation, or rehospitalization records, structured telephone follow-up, or confirmed death records. Two trained assessors who were not involved in model development and were blinded to calculated inflammatory-coagulation biomarkers and predicted risks assessed the outcome, with disagreements adjudicated by a senior physician. Cases without sufficient information for reliable mRS assignment were considered to have missing outcome data.

Candidate predictors included demographic characteristics, smoking and alcohol history, relevant comorbidities and pre-admission medications; onset-to-admission time, admission temperature and blood pressure, and Glasgow Coma Scale score; and initial computed tomography findings, including hematoma location, intraventricular hemorrhage, midline shift, and hematoma volume. Hematoma volume was estimated using the ABC/2 method.¹⁵ Laboratory variables included complete blood count, admission blood glucose, serum albumin, creatinine, estimated glomerular filtration rate, prothrombin time, activated partial thromboplastin time, international normalized ratio, fibrinogen, and D-dimer. NLR and PLR were calculated as neutrophil-to-lymphocyte and platelet-to-lymphocyte ratios, respectively. D-dimer values were harmonized to mg/L fibrinogen equivalent units. Variables with potentially incomplete retrospective documentation, including hypertension duration, body mass index, smoking, and alcohol use, were assessed for data completeness before model development.

Subsequent acute interventions, including minimally invasive hematoma aspiration, craniotomy, external ventricular drainage, intensive care unit admission, and early hemostatic or blood pressure-lowering treatment, were recorded descriptively but were not included as core predictors in the admission-based model; selected interventions were evaluated in sensitivity analyses.

Sample Size Calculation

Sample size planning assumed a 31.4% incidence of 3-month poor functional outcome (mRS 4–6) based on a comparable ICH cohort.¹³ Following contemporary recommendations for prediction model development,¹⁶ and assuming a maximum of 10–12 effective predictor parameters with a pragmatic minimum of 10 outcome events per parameter, approximately 100–120 events and 319–382 evaluable patients were required. Allowing for 10%–15% missing key data or incomplete outcome ascertainment, the screening target was approximately 425–450 patients. The observed number of outcome events was subsequently checked against the planned model complexity before model development.

Statistical Analysis

All analyses were performed using R version 4.5.0 (R Foundation for Statistical Computing, Vienna, Austria), with two-sided P < 0.05 considered statistically significant. Baseline characteristics were summarized and compared between outcome groups using the independent-samples t-test or Mann–Whitney U-test for continuous variables and Pearson’s χ² or Fisher’s exact test for categorical variables, as appropriate; these comparisons were descriptive and were not used to determine predictor inclusion.

Patients with missing 90-day mRS outcomes were excluded from the primary analysis and compared with the included cohort for potential attrition bias. Predictor variables with <5% missingness were analyzed using available data, those with 5%–20% missingness were handled using multiple imputation by chained equations with 10 datasets, and variables with >20% missingness were excluded from primary model development. The imputation model included the outcome, candidate predictors, and relevant auxiliary variables. Complete-case analysis was performed as a sensitivity analysis.

Candidate predictors were prespecified according to clinical relevance, previous evidence, and availability within 24 hours after admission. LASSO logistic regression with 10-fold cross-validation and the 1-standard error rule was used for feature selection. Age, Glasgow Coma Scale score, hematoma volume, and intraventricular hemorrhage were retained as established clinical-imaging predictors for prespecified model comparison. Multicollinearity was assessed using variance inflation factors, with VIF >5 considered relevant; redundant severity measures and mathematically coupled hematological variables were not entered simultaneously.

Two logistic regression models were evaluated. The conventional model included age, Glasgow Coma Scale score, initial hematoma volume, intraventricular hemorrhage, and admission blood glucose. The extended model additionally incorporated inflammatory-coagulation biomarkers retained after feature selection. Continuous predictors were modeled without categorization, and markedly skewed D-dimer values were natural-log transformed. Regression coefficients and odds ratios with 95% confidence intervals were reported.

Discrimination was assessed using the area under the receiver operating characteristic curve (AUC), with AUCs compared using DeLong’s test. Calibration was evaluated using the calibration intercept, calibration slope, calibration plot, and Brier score, with the Hosmer–Lemeshow test reported as a supplementary measure. Internal validation used 1000 bootstrap resamples to obtain optimism-corrected AUC, calibration slope, and Brier score. Incremental performance was further examined using continuous net reclassification improvement and integrated discrimination improvement as supplementary metrics. Clinical utility was assessed using decision curve analysis.

Four sensitivity analyses were performed: redefining poor outcome as mRS 3–6; additionally adjusting for post-admission acute interventions, including minimally invasive aspiration, craniotomy, external ventricular drainage, and intensive care unit admission; repeating the primary analysis in complete cases; and restricting the final model to patients in whom both complete blood count and D-dimer sampling preceded any qualifying major acute intervention.

Results

Study Population and Baseline Characteristics

Between January 2020 and December 2025, 715 patients with ICH were screened. After excluding 253 patients due to secondary hemorrhage (n=68), pre-stroke mRS > 2 (n=45), anticoagulation-associated hemorrhage (n=52), admission >24 hours (n=41), missing baseline data (n=23), or loss to follow-up (n=24), 462 eligible patients were included in the primary analysis (Figure 1). Baseline characteristics were comparable between the 24 patients with missing 90-day outcomes and the included cohort. Hypertension duration was excluded because of substantial missingness.

Figure 1.

Flowchart of patient inclusion and exclusion.

The median interval from hospital admission to baseline biomarker sampling was 1.8 hours (IQR, 1.1–3.0 hours). Both complete blood count and D-dimer samples were obtained before any qualifying major acute intervention in 427 patients (92.4%). In 24 patients (5.2%), one or both baseline samples were collected after a qualifying intervention, while the temporal sequence could not be established reliably in 11 patients (2.4%). Overall, 162 (35.1%) patients experienced a poor functional outcome at 3 months. Baseline characteristics and acute interventions stratified by outcome are summarized in Table 1. Patients with poor functional outcome had greater baseline clinical and neurological severity, including lower GCS scores, larger hematoma volumes, and more frequent IVH and midline shift, together with higher inflammatory-coagulation biomarker levels. Acute interventions were also more frequent in this group (Table 1).

Table 1.

Baseline Clinical, Neuroimaging, Laboratory Characteristics and Subsequent Acute Interventions According to 3-Month Functional Outcome

Variable	Favorable Outcome (N=300)	Poor Outcome (N=162)	Statistic	P-value
Age (years)	61.5 ± 10.8	67.2 ± 11.4	t = −5.263	<0.001
Sex (Male)	183 (61.0%)	102 (63.0%)	χ2 = 0.174	0.676
BMI (kg/m2)	24.3 ± 3.2	24.1 ± 3.5	t = 0.620	0.536
Smoking history	120 (40.0%)	68 (42.0%)	χ2 = 0.168	0.682
Alcohol consumption	96 (32.0%)	58 (35.8%)	χ2 = 0.686	0.408
Hypertension	279 (93.0%)	154 (95.1%)	χ2 = 0.771	0.38
Diabetes mellitus	54 (18.0%)	45 (27.8%)	χ2 = 5.922	0.015
Coronary artery disease	36 (12.0%)	26 (16.0%)	χ2 = 1.455	0.228
Prior stroke	42 (14.0%)	38 (23.5%)	χ2 = 6.643	0.01
Antiplatelet agents	45 (15.0%)	31 (19.1%)	χ2 = 1.306	0.253
Antihypertensive drugs	210 (70.0%)	110 (67.9%)	χ2 = 0.215	0.643
Statins	60 (20.0%)	30 (18.5%)	χ2 = 0.145	0.703
Onset-to-admission time (h)	5.5 (3.2–9.8)	5.2 (2.8–9.5)	Z = 0.615	0.539
Admission body temperature (°C)	36.8 ± 0.4	37.1 ± 0.6	t = −6.211	<0.001
SBP (mmHg)	168.5 ± 22.4	172.3 ± 25.1	t = −1.655	0.099
DBP (mmHg)	95.4 ± 14.2	97.1 ± 15.6	t = −1.185	0.237
GCS score	13 (11–14)	9 (6–11)	Z = 11.452	<0.001
Primary hematoma location			χ2 = 5.586	0.134
− Basal ganglia	180 (60.0%)	95 (58.6%)
− Thalamus	75 (25.0%)	46 (28.4%)
− Brainstem	15 (5.0%)	14 (8.6%)
− Cerebellum	30 (10.0%)	7 (4.3%)
IVH	54 (18.0%)	78 (48.1%)	χ2 = 47.382	<0.001
Midline shift	45 (15.0%)	65 (40.1%)	χ2 = 36.564	<0.001
Initial hematoma volume (mL)	11.5 (7.2–16.8)	26.4 (16.5–40.2)	Z = −12.345	<0.001
WBC (×109/L)	8.2 ± 2.5	10.6 ± 3.4	t = −8.541	<0.001
Neutrophil count (×109/L)	5.6 (4.1–7.5)	8.4 (6.2–11.1)	Z = −8.672	<0.001
Lymphocyte count (×109/L)	1.6 (1.2–2.1)	1.2 (0.8–1.6)	Z = 7.125	<0.001
Platelet count (×109/L)	215 ± 58	205 ± 65	t = 1.688	0.092
NLR	3.5 (2.2–5.1)	7.0 (4.5–10.4)	Z = −10.551	<0.001
PLR	134.4 (95.5–179.2)	170.8 (120.5–235.6)	Z = −5.612	<0.001
Admission blood glucose (mmol/L)	6.4 (5.2–7.8)	7.8 (6.1–9.6)	Z = −5.845	<0.001
Serum albumin (g/L)	38.5 ± 4.2	36.2 ± 4.8	t = 5.291	<0.001
eGFR (mL/min/1.73m2)	88.5 ± 18.4	80.2 ± 20.6	t = 4.416	<0.001
INR	0.98 ± 0.10	1.01 ± 0.12	t = −2.839	0.005
ln(D-dimer, mg/L FEU)	−0.15 ± 0.65	0.58 ± 0.82	t = −10.435	<0.001
Subsequent Acute Interventions
Minimally invasive aspiration	30 (10.0%)	25 (15.4%)	χ2 = 2.923	0.087
Craniotomy	15 (5.0%)	35 (21.6%)	χ2 = 28.532	<0.001
External ventricular drainage	10 (3.3%)	40 (24.7%)	χ2 = 48.715	<0.001
Intensive care unit admission	90 (30.0%)	130 (80.2%)	χ2 = 106.431	<0.001

Notes: Data are expressed as mean ± standard deviation, median (interquartile range), or number (percentage). Baseline characteristics were compared using the independent samples t-test (statistic t) for normally distributed continuous variables, Mann–Whitney U-test (statistic Z) for skewed continuous variables, and Pearson’s χ² test (statistic χ²) for categorical variables. D-dimer values were harmonized to mg/L fibrinogen equivalent units (FEU) and natural-log transformed before statistical comparison to address skewed distribution.

Abbreviations: BMI, body mass index; SBP, systolic blood pressure; DBP, diastolic blood pressure; GCS, Glasgow Coma Scale; IVH, intraventricular hemorrhage; WBC, white blood cell; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; eGFR, estimated glomerular filtration rate; INR, international normalized ratio.

Predictor Selection and Multivariable Analysis

LASSO logistic regression using 10-fold cross-validation and the 1-standard error rule (λ = 0.038) retained seven predictors: age, GCS score, initial hematoma volume, IVH, admission blood glucose, NLR, and ln (D-dimer); PLR was excluded after shrinkage to zero (Figure 2).

Figure 2.

Predictor selection using least absolute shrinkage and selection operator (lasso) logistic regression. (A) LASSO coefficient profiles of the 33 baseline candidate variables. Each curve represents the trajectory of a variable coefficient as the penalty parameter λ changes. The colored curves indicate the 7 predictors retained at the final λ value, including age, GCS score, hematoma volume, IVH, admission glucose, NLR, and ln(D-dimer). (B) Ten-fold cross-validation curve for selecting the penalty parameter λ. The gray dotted vertical line indicates λmin, which yielded the minimum binomial deviance, whereas the black dashed vertical line indicates λ1se selected according to the 1-standard error rule. The final model used λ1se = 0.038 and retained 7 predictors with non-zero coefficients.

Abbreviations: GCS, Glasgow Coma Scale; IVH, intraventricular hemorrhage; NLR, neutrophil-to-lymphocyte ratio.

All seven retained predictors remained independently associated with 3-month poor functional outcome in the multivariable model, with higher GCS score associated with lower risk and the remaining predictors associated with higher risk (Table 2). No relevant multicollinearity was observed (VIF range, 1.05–1.34).

Table 2.

Multivariable Logistic Regression Analysis of Predictors Associated with 3-Month Poor Functional Outcome

Variable	β Coefficient	Standard Error	Wald Z	Odds Ratio (95% CI)	P-value	VIF
Age (per 1-year increase)	0.041	0.012	3.417	1.042 (1.018–1.066)	0.001	1.08
GCS score (per 1-point increase)	−0.294	0.052	−5.654	0.745 (0.672–0.825)	<0.001	1.34
Initial hematoma volume (per 1-mL increase)	0.05	0.009	5.556	1.051 (1.032–1.071)	<0.001	1.28
IVH (Present vs Absent)	0.896	0.291	3.079	2.450 (1.385–4.332)	0.002	1.22
Admission blood glucose (per 1-mmol/L increase)	0.144	0.049	2.939	1.155 (1.048–1.272)	0.004	1.05
NLR (per 1-unit increase)	0.168	0.044	3.818	1.183 (1.085–1.289)	<0.001	1.15
ln(D-dimer) (per 1-unit increase)	0.522	0.155	3.368	1.685 (1.242–2.285)	0.001	1.12

Notes: The multivariable logistic regression model was constructed using variables selected by the LASSO regression (λ = 0.038, 1-standard error rule). The variance inflation factor (VIF) for each variable was < 5, indicating no significant multicollinearity. The Odds Ratios (OR) represent the multiplicative change in the odds of experiencing a 3-month poor functional outcome corresponding to the specified unit change in each continuous variable or the presence of the categorical variable.

Abbreviations: CI, confidence interval; GCS, Glasgow Coma Scale; IVH, intraventricular hemorrhage; NLR, neutrophil-to-lymphocyte ratio; VIF, variance inflation factor.

Comparison of Predictive Performance Between the Conventional and Extended Models

Adding NLR and ln (D-dimer) to the conventional clinical-imaging model improved discrimination, with the AUC increasing from 0.765 (95% CI, 0.720–0.810) to 0.828 (95% CI, 0.788–0.868; DeLong P = 0.008). The extended model also showed lower overall prediction error, comparable calibration, and improved reclassification metrics (Table 3).

Table 3.

Comparison of Predictive Performance Between the Conventional Clinical-Imaging Model and the Inflammatory-Coagulation Extended Model

Performance Metric	Conventional Model	Extended Model	P-value
Discrimination
AUC (95% CI)	0.765 (0.720–0.810)	0.828 (0.788–0.868)	0.008*
Overall Accuracy & Calibration
Brier score	0.178	0.142	-
Calibration intercept	0.042	0.015	-
Calibration slope	0.945	0.982	-
Reclassification
Continuous NRI (95% CI)	Reference	0.415 (0.228–0.602)	<0.001
IDI (95% CI)	Reference	0.072 (0.035–0.109)	<0.001

Notes: The conventional model includes age, Glasgow Coma Scale score, initial hematoma volume, intraventricular hemorrhage, and admission blood glucose. The extended model includes all variables in the conventional model plus neutrophil-to-lymphocyte ratio and natural log-transformed D-dimer. *P-value for the comparison of AUCs was calculated using DeLong’s test.

Abbreviations: AUC, area under the receiver operating characteristic curve; CI, confidence interval; NRI, net reclassification improvement; IDI, integrated discrimination improvement.

Construction of the Prognostic Nomogram

To visualize individualized risk estimation based on the final inflammatory-coagulation extended model, a prognostic nomogram was constructed for predicting 3-month poor functional outcome in patients with hypertensive ICH. The nomogram incorporated the seven baseline predictors retained in the final multivariable model: age, GCS score, initial hematoma volume, IVH, admission blood glucose, NLR, and natural ln (D-dimer) (Figure 3). The PLR was not included because it was reduced to a zero coefficient during LASSO feature selection.

Figure 3.

Nomogram predicting the probability of 3-month poor functional outcome in patients with hypertensive intracerebral hemorrhage. To utilize the nomogram, locate the patient’s specific value on the respective axis for each of the seven baseline predictors: age, GCS score, initial hematoma volume, IVH (absent = 0, present = 1), admission blood glucose, NLR, and ln(D-dimer). Draw a vertical line upward to the “points” axis to determine the assigned score for each variable. Sum the scores from all variables to calculate the “total points”. Finally, locate the sum on the “total points” axis and draw a vertical line downward to the “risk of 3-month poor functional outcome” axis to ascertain the estimated individual probability of experiencing an mRS score of 4 to 6 at 90 days. This predictive tool is exclusively predicated on clinical, neuroimaging, and laboratory indicators obtained within 24 hours of hospital admission.

Abbreviations: GCS, Glasgow Coma Scale; IVH, intraventricular hemorrhage; NLR, neutrophil-to-lymphocyte ratio;mRS, modified Rankin Scale.

Model Calibration, Internal Validation, and Clinical Utility

The extended model demonstrated good calibration (Hosmer-Lemeshow χ² = 7.842, P = 0.450). Following internal validation via 1000 bootstrap resamples, the optimism-corrected AUC was 0.816 (apparent AUC = 0.828). The apparent Brier score was 0.142, with an optimism-corrected calibration slope of 0.965 and an intercept of 0.024. The calibration plot confirmed a close alignment between the bias-corrected curve and the ideal reference line (Figure 4A).

Figure 4.

Calibration curve and decision curve analysis of the prognostic models. (A) Calibration curve of the extended nomogram model evaluated via 1000 bootstrap resamples. The x-axis represents the nomogram-predicted probability of 3-month poor functional outcome, and the y-axis represents the actual observed proportion. The diagonal dotted gray line represents perfect prediction (ideal calibration). The apparent solid line represents the performance on the original dataset, and the bias-corrected dashed line represents the bootstrap-validated performance. (B) Decision curve analysis (DCA) comparing the clinical net benefit of the conventional model and the inflammatory-coagulation extended model. The x-axis indicates the threshold probability, and the y-axis indicates the net benefit. The solid red line represents the extended model, and the dashed blue line represents the conventional model. The gray line represents the strategy of classifying all patients as high risk, whereas the horizontal black line represents the strategy of classifying no patients as high risk.

DCA (Figure 4B) demonstrated that the extended model provided a positive net benefit across threshold probabilities of 10% to 75%. Within the specific threshold probability window of 15% to 65%, the extended model consistently yielded a higher net benefit compared with the conventional model.

Sensitivity Analyses

Four sensitivity analyses were performed to assess the robustness of the inflammatory-coagulation extended model (Supplementary Table 1). First, when poor functional outcome was redefined as mRS 3–6, occurring in 52.4% of patients, the associations of NLR and ln(D-dimer) remained consistent with the primary analysis, and model discrimination remained stable (AUC = 0.812; 95% CI: 0.765–0.859). Second, complete-case analysis in 418 patients yielded similar effect estimates for NLR and ln(D-dimer), with an AUC of 0.826 (95% CI: 0.785–0.867), supporting the consistency of the findings without multiple imputation.

Third, after additionally adjusting for post-admission acute interventions, including minimally invasive aspiration, craniotomy, external ventricular drainage, and intensive care unit admission, both biomarkers remained independently associated with poor outcome, although the effect estimates were modestly attenuated. The adjusted ORs were 1.145 (95% CI: 1.042–1.268; P = 0.005) for NLR and 1.520 (95% CI: 1.115–2.085; P = 0.008) for ln(D-dimer), with an AUC of 0.852 (95% CI: 0.815–0.889). Fourth, restricting the analysis to the 427 patients in whom both complete blood count and D-dimer sampling preceded any major acute intervention produced results comparable with the primary analysis. NLR remained associated with poor outcome (OR = 1.179; 95% CI: 1.075–1.293; P < 0.001), as did ln(D-dimer) (OR = 1.641; 95% CI: 1.194–2.256; P = 0.002), and model discrimination was essentially unchanged (AUC = 0.826; 95% CI: 0.784–0.868).

Discussion

In this single-center retrospective cohort of patients with deep hypertension-related ICH, 35.1% experienced poor functional outcome at 90 days. The final model incorporated age, GCS score, initial hematoma volume, IVH, admission blood glucose, NLR, and ln(D-dimer). Adding NLR and ln(D-dimer) to the conventional clinical-imaging model increased the AUC from 0.765 to 0.828, reduced the Brier score, and yielded higher net benefit across a range of clinically relevant thresholds. However, the absolute AUC gain of 0.063, although statistically significant, should be regarded as modest rather than evidence of a major improvement in clinical prediction. The incremental value of these biomarkers is likely to depend on case mix, laboratory practice, baseline risk, and the clinical consequences associated with specific risk thresholds. Accordingly, the present findings support further evaluation of routinely available inflammatory-coagulation biomarkers rather than their immediate use as stand-alone clinical decision markers.

The conventional predictors retained in the model are consistent with previous ICH prognostic studies. Age, neurological impairment, hematoma volume, and ventricular extension remain central components of models predicting functional outcome, and recent nomogram studies have similarly identified clinical severity and neuroimaging burden as dominant prognostic factors.^17,18 Admission glucose provided additional information in the present model, consistent with observational evidence linking acute hyperglycemia with poorer functional outcome after spontaneous ICH.^19,20 These findings reinforce the importance of neurological and anatomical severity as the foundation of early prognostic assessment, against which any incremental biomarker contribution should be evaluated.

NLR was independently associated with poor 90-day functional outcome after adjustment for established clinical and imaging predictors. Previous studies have linked elevated NLR and related inflammatory indices with hematoma expansion, pneumonia, disability, and mortality after ICH,^5,21 while meta-analytic evidence also supports an association between systemic inflammatory markers and adverse ICH outcomes.²² Biologically, NLR may reflect the balance between neutrophil-dominant innate inflammatory activation and relative lymphopenia after acute brain injury. Nevertheless, NLR is nonspecific and may be affected by physiological stress, infection, medication exposure, and comorbidity. Its contribution in the present study should therefore be interpreted as complementary prognostic information within a multivariable model rather than evidence of an ICH-specific biological marker.

D-dimer showed a similar incremental association. Elevated D-dimer has previously been associated with poor functional outcome and mortality after spontaneous ICH,⁶ and admission point-of-care data have also linked lower D-dimer concentrations with more favorable 90-day outcomes.²³ D-dimer reflects fibrin formation and degradation and may capture coagulation-fibrinolytic activation associated with tissue injury, hematoma burden, systemic stress, or thromboinflammatory responses. The present analysis extends these observations by demonstrating that ln(D-dimer) retained prognostic information after accounting for hematoma volume, IVH, neurological severity, and NLR. However, assay heterogeneity and differences in reporting units remain important considerations for future multicenter validation.

PLR was significantly higher in patients with poor outcome in the univariate analysis but was reduced to a zero coefficient during LASSO selection. Previous studies have reported associations between PLR and adverse ICH outcomes,^24,25 but its independent contribution appears less consistent. In the present cohort, PLR and NLR shared lymphocyte count as their denominator, whereas platelet counts alone did not differ significantly between outcome groups (P = 0.092). This pattern suggests that the univariate prognostic signal of PLR may have been driven more by relative lymphopenia than by platelet-related activation. Once NLR and D-dimer were considered simultaneously, PLR therefore provided limited additional information, supporting its exclusion from the final parsimonious model.

The principal contribution of the present study is therefore not the identification of NLR or D-dimer as novel prognostic biomarkers, because both have already been investigated extensively in ICH. Rather, the study evaluates their incremental contribution beyond a predefined clinical-imaging model in a relatively homogeneous cohort of deep hypertension-related ICH. The AUC improvement was accompanied by a lower Brier score, acceptable calibration, and stable bootstrap-corrected performance. Continuous NRI and IDI also favored the extended model, although these measures were considered supportive rather than definitive evidence of clinical superiority. Decision curve analysis demonstrated greater net benefit for the extended model over a clinically relevant threshold range,²⁶ but such benefit remains dependent on the clinical setting and intended decision. Sensitivity analyses, including complete-case analysis, alternative outcome definition, adjustment for acute interventions, and restriction to the 427 patients whose biomarker samples preceded major interventions, produced broadly consistent results. Collectively, these analyses support internal robustness but do not establish that the observed incremental gain will translate into clinically meaningful benefit in external populations.

Anatomical heterogeneity within deep ICH also deserves consideration. Emerging evidence indicates that prognostic information may depend not only on overall hematoma burden but also on the anatomical distribution of hemorrhage. Chen et al recently showed that topographical patterns of IVH provided prognostic information beyond conventional IVH scoring systems,²⁷ while another study in mild-to-moderate ICH identified specific hematoma locations, including thalamic and basal ganglia hemorrhage, among predictors of early functional dependence and demonstrated that performance of general ICH prognostic scores may vary across more specific patient subgroups.²⁸ In the present cohort, hematoma location was not statistically significant in univariate analysis, but the brainstem and cerebellar subgroups comprised only 29 (6.3%) and 37 (8.0%) patients, respectively. The absence of a statistically significant location effect should therefore not be interpreted as evidence that prognosis is equivalent across deep anatomical sites. These small subgroups provided limited statistical information for location-specific effects, and future multicenter studies should evaluate model performance separately in basal ganglia, thalamic, brainstem, and cerebellar ICH.

Post-admission interventions were deliberately excluded from the core admission-based model because treatment decisions occur after baseline assessment and are strongly determined by initial neurological and anatomical severity. Incorporating craniotomy, external ventricular drainage, or ICU admission into the primary model would therefore condition prediction on downstream management decisions along the pathway from initial severity to treatment and subsequent outcome, changing the model from an admission-based prognostic tool to a post-treatment conditional model. It could also obscure the prognostic contribution of baseline variables through confounding by indication. The exploratory intervention-adjusted analysis was retained to assess whether biomarker associations were sensitive to treatment intensity rather than to estimate causal treatment effects. This distinction is particularly relevant for EVD, because treatment allocation is strongly influenced by IVH severity and hydrocephalus; previous retrospective evidence in patients with mild IVH found no clear benefit of preventive EVD on 90-day functional outcome, illustrating the difficulty of interpreting treatment-outcome associations outside the context of indication.²⁹

Several limitations should be considered. First, the single-center retrospective design introduces potential selection and information bias and limits transportability. Second, although the cohort was restricted to deep hypertension-related ICH and alternative etiologies were reviewed clinically and radiologically, advanced vascular or magnetic resonance imaging was not performed uniformly; residual etiological misclassification is therefore possible. Third, the relatively small brainstem and cerebellar subgroups limited power to detect location-specific prognostic effects and precluded reliable stratified model assessment. Fourth, mRS was derived from clinical records or structured follow-up, and some outcome misclassification remains possible despite blinded assessment. Fifth, NLR and D-dimer may be influenced by unmeasured inflammatory, metabolic, renal, malignant, or medication-related factors, and laboratory assay variability may affect reproducibility across institutions. Sixth, post-admission interventions could modify subsequent outcomes, but their inclusion in the primary admission model would condition on downstream treatment decisions; the exploratory adjusted analysis therefore cannot establish treatment-independent or causal biomarker effects. Finally, only internal validation was performed. Bootstrap validation addresses model optimism but does not establish transportability or clinical impact. External multicenter validation should assess calibration across institutions and laboratory platforms, examine performance within specific deep hematoma locations, and determine whether the modest incremental information provided by NLR and D-dimer improves clinically relevant decisions. Until such validation is available, the model should be considered primarily a hypothesis-generating and internally validated risk-stratification tool for populations and clinical settings similar to the development cohort.

Conclusion

In patients with deep hypertension-related ICH, adding NLR and ln(D-dimer) to established admission clinical and neuroimaging predictors produced a statistically significant but modest improvement in discrimination, with an AUC increase from 0.765 to 0.828, together with improved overall prediction error and internally consistent performance. These findings suggest that routinely available inflammatory-coagulation biomarkers may provide complementary prognostic information, but their clinical value is likely to be context-dependent and cannot be established from internal validation alone. The model should currently be regarded as a hypothesis-generating and internal risk-stratification tool for similar clinical settings; multicenter external validation and prospective assessment of clinical utility are required before broader application or routine clinical implementation.

Funding Statement

This work was supported by the Natural Science Foundation of Hubei Province (2026AFC0451).

Data Sharing Statement

The experimental data used to support the findings of this study are available from the corresponding author upon request.

Ethics Approval

This retrospective chart review study was reviewed and approved by the Ethics Committee of Huanggang Central Hospital (approval No. HGYY-KY-2026-027). The study was conducted in accordance with the Declaration of Helsinki. All study-specific patient identification, eligibility screening, data extraction, database construction, and statistical analyses were initiated only after ethics approval had been obtained.

Consent to Participate

The need for informed consent was waived by the Ethics Committee of Huanggang Central Hospital in view of the retrospective nature of the study, and all the procedures being performed were part of routine care.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declared that they have no conflicts of interest regarding this work.