Association between the hemoglobin, albumin, lymphocyte, and platelet (HALP) score and 28-day mortality in neurocritical care patients: a retrospective cohort study based on the MIMIC-IV database Authors: Yajun Zou, Yu Xia, Hu Li, Long Teng, Zhipeng Wang, Min Shao, Yuxin Chen, Jianlin Zhang Published: 2025 ABSTRACT BackgroundIn recent years, reliable prognostic markers are increasingly important in neurocritical care. The hemoglobin, albumin, lymphocyte, and platelet (HALP) score, a composite biomarker reflecting systemic immune-nutritional status, has shown promise as a predictor. This study aims to evaluate the prognostic value of the HALP score for predicting 28-day mortality in neurocritical care unit (NCCU) patients.Materials and methodsWe conducted a retrospective cohort study by identifying neurocritical care patients from the MIMIC-IV database. The association between the HALP score and 28-day mortality was primarily assessed using multivariable Cox proportional hazards regression. Multiple analytical approaches were employed to comprehensively examine this relationship, including restricted cubic spline (RCS) regression to evaluate nonlinearity and Kaplan–Meier survival analysis. A risk prediction nomogram was developed based on a Cox regression model. The model’s discriminative performance at 28 days was evaluated using receiver operating characteristic (ROC) curve analysis, with the area under the curve (AUC) quantified using logistic regression.ResultFinally, 858 patients in the MIMIC database were analyzed, with 28-day ICU and in-hospital mortality rates of 21.0% and 20.4%, respectively. In the fully adjusted Cox model for ICU mortality, notably, the continuous HALP score was not independently predictive (HR = 0.99, 95% CI 0.98–1.00, p = 0.13). However, when analyzed categorically, which may better capture nonlinear risk relationships, higher HALP categories showed significant associations. Specifically, the 28-day ICU mortality rates progressively decreased across low, moderate, and high HALP score groups (29.7% vs 20.3% vs 12.9%, respectively; p < 0.001). After adjusting for all covariates using Cox regression, high HALP scores demonstrated significantly lower mortality risk compared to the lowest scores (HR = 0.56, 95% CI 0.36–0.87; p = 0.009). Restricted cubic spline analysis revealed a nonlinear threshold effect of the HALP score on mortality risk. The HALP score provided incremental predictive risk value for 28-day mortality when combined with conventional critical illness severity scores. The predictive model based on the HALP score achieved area under the curve (AUC) values of 0.71, 0.72, and 0.73 in the complete, Internal Validation Subset A, and Internal Validation Subset B, respectively. Our risk model provided additional identification of neurocritical patients when combined with traditional ICU scoring systems. This performance profile suggests that the HALP score may hold complementary predictive value to traditional ICU scoring systems.ConclusionOur findings are consistent with an inverse association between HALP scores and short-term survival in neurocritical care patients. With a modest predictive discrimination (AUC ≤ 0.73), this scoring system may have potential as a complementary clinical risk assessment tool. Future validation is warranted to confirm its utility in aiding the risk stratification of neurocritical care patients.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40001-025-03437-y. Introduction Neurocritical care encompasses life-threatening clinical conditions resulting from severe neurological disorders or injuries, including traumatic brain injury (TBI), massive cerebral infarction, stroke, intracerebral haemorrhage, subarachnoid haemorrhage (SAH), and central nervous system infections [1–3]. Patients admitted to the neurocritical care unit (NCCU) face grave prognoses, with in-hospital mortality rates ranging from 20 to 50% [4, 5]. This high mortality, often attributable to the primary neurological insult and its systemic complications, underscores the urgent need for robust prognostic tools. Beyond neurological severity, the patient’s systemic physiological reserve—encompassing inflammatory, immune, and nutritional status—is increasingly recognized as a critical determinant of recovery [5–8]. The systemic response to acute brain injury is characterized by a profound neuro-inflammatory cascade that disrupts immune homeostasis and precipitates a state of catabolism and nutritional depletion [9, 10]. It is within this context that the hemoglobin, albumin, lymphocyte, and platelet (HALP) score has been proposed as an integrative biomarker. HALP concurrently quantifies four key dimensions: oxygen-carrying capacity (hemoglobin), nutritional and anti-inflammatory status (albumin), immune competence (lymphocyte count), and thrombotic and inflammatory activity (platelet count) [11–14]. The prognostic utility of HALP is mechanistically grounded in the specific pathophysiology of neurocritical illness. First, anemia (low hemoglobin) is common in NCCU patients and can critically impair cerebral oxygen delivery, potentially exacerbating secondary neuronal injury in the ischemic penumbra surrounding the primary lesion [15, 16]. Studies on TBI and stroke have linked anemia to worse neurological outcomes. Second, hypoalbuminemia is not only a marker of malnutrition, but also reflects diminished anti-oxidant capacity and endothelial dysfunction, which can exacerbate cerebral edema and disrupt the blood–brain barrier [17]. Low albumin levels have been independently associated with mortality in mixed neurocritical care cohorts. Third, lymphopenia is a core feature of brain-induced immunodepression (CNS injury-induced immunodepression syndrome—CIDS), significantly heightening susceptibility to life-threatening infections (e.g., ventilator-associated pneumonia) that worsen outcomes [18]. The high incidence of infection in this population makes this an especially pertinent pathway. Finally, platelets are pivotal players in thrombo-inflammation; their consumption and dysfunction can influence the risk of secondary cerebral infarction or propagate microvascular thrombosis [19, 20]. Platelet activation and dysfunction are well documented in both ischemic and hemorrhagic brain injury. This strong pathophysiological rationale is supported by emerging clinical evidence linking HALP to outcomes in ischemic stroke and other conditions [21, 22]. While other composite indices such as the prognostic nutritional index (PNI), neutrophil-to-lymphocyte ratio (NLR), and platelet-to-lymphocyte ratio (PLR) have been investigated in critical care, HALP provides a more comprehensive integration of nutritional, immune, and thrombo-inflammatory pathways within a single metric. Despite the compelling pathophysiological rationale and promising preliminary evidence in specific conditions, the prognostic value of HALP remains unproven in the broader, etiologically heterogeneous neurocritical care population. To our knowledge, no prior study has comprehensively evaluated its association with mortality in a large, general cohort of mixed neurocritical care patients. Therefore, we hypothesize that a lower HALP score at ICU admission is independently associated with an increased risk of 28-day all-cause mortality, and that it provides incremental prognostic value beyond conventional severity of illness scores. Utilizing the MIMIC-IV database, this study aims to rigorously test this hypothesis. Materials and methods Data source We conducted a retrospective analysis using data from the Medical Information Mart for Intensive Care IV (MIMIC-IV v3.1) database [23]. This database contains comprehensive, de-identified clinical data from over 190,000 adult ICU patients admitted to a large US academic medical center between 2008 and 2019. The dataset includes a wide range of variables, such as patient demographics, vital signs, laboratory measurements, medications, and outcomes. Study population This study focused on neurocritical care patients during their first ICU admission. Diagnoses were classified using International Classification of Diseases (ICD-9/10) codes, including traumatic brain injury, ischemic stroke, intracerebral hemorrhage, subarachnoid hemorrhage, subdural hemorrhage, epidural hemorrhage, intracranial infections, and related disorders. Clinical laboratory data, including hemoglobin, albumin, lymphocyte count, and platelet count, were extracted from the MIMIC-IV database. To standardize the timing of exposure assessment across the cohort, we used the first available measurement for each component recorded within the first 24 h following ICU admission. This strategy was designed to capture the patient’s baseline physiological status upon ICU entry, before the influence of prolonged ICU interventions. Inclusion criteria were: (1) age ≥ 18 years; (2) survival and ICU stay duration ≥ 24 h; (3) availability of the required laboratory records (hemoglobin, albumin, lymphocyte, and platelet counts) within the first 24 h of ICU admission; and (4) first-time ICU admission. Clinical data were collected within the first 24 h of ICU admission. Patients were stratified into tertiles based on HALP scores. Variable extraction Data extraction was performed using PostgreSQL (version 13.7.2) and Navicat Premium (version 16) with structured query language (SQL). The extracted variables were categorized into six domains: (1) demographic characteristics: age, sex, weight, height, race, and body mass index (BMI); (2) comorbidities: hypertension, acute kidney injury (AKI), chronic kidney disease (CKD), heart failure (HF), liver cirrhosis (LIC), diabetes mellitus, and chronic obstructive pulmonary disease (COPD); (3) vital signs: respiratory rate (RR), heart rate (HR), non-invasive mean blood pressure (NBPM), systolic blood pressure (NBPS), and diastolic blood pressure (NBPD); (4) illness severity scores: Acute Physiology Score II (APS II), Simplified Acute Physiology Score II (SAPS-II), Oxford Acute Severity of Illness Score (OASIS), Sequential Organ Failure Assessment (SOFA), and Glasgow Coma Scale (GCS); (5) laboratory parameters: lymphocyte count, red blood cells (RBC), white blood cells (WBC), neutrophils (Neu), monocytes (Mon), hemoglobin (Hb), platelet count (PLT), red cell distribution width (RDW), monocyte percentage (Mon), hematocrit (HCT), albumin (ALB), anion gap (AG), lactate (Lac), sodium (Na), potassium (K), chloride (Cl), creatinine (CRE), blood urea nitrogen (URE), glucose (Glu), international normalized ratio (INR), prothrombin time (PT), partial thromboplastin time (PTT), alanine aminotransferase (ALT), aspartate aminotransferase (AST), and total bilirubin (TB); (6) treatment modalities: vasopressor use (VP), mechanical ventilation (MV), and continuous renal replacement therapy (CRRT). The extent of missing data was assessed for all candidate variables. Variables with more than 30% missing values were excluded from the analysis. In our dataset, missingness was exclusively present in continuous variables, all categorical variables, including the outcome variable (28-day mortality) and the HALP components (HB, ALB, lymphocyte count, and PLT), were complete (Supplementary Table 4). For continuous variables with ≤ 30% missing data, values were assumed to be missing at random (MAR). We performed multiple imputation using the R “mice” package (version 3.16). The imputation model included all variables intended for the primary analysis: the continuous variables requiring imputation, the complete primary exposure (the HALP its components), the complete outcome variable (28-day mortality), and all other complete categorical and continuous covariates. This approach strengthens the plausibility of the MAR assumption by using all available information to predict missing values. The HALP score and end points HALP = [hemoglobin (g/L) × albumin (g/L) × lymphocyte count (× 109/L)]/platelet count (× 109/L) [24]. For the purpose of survival analysis and visualizing mortality trends, patients were categorized into three groups ([low ≤ 14.9], [14.90 < moderate < 29.0], and [high ≥ 29.0]) based on HALP score tertiles. The primary end point of this study was 28-day ICU mortality, defined as death from any cause occurring during the initial ICU stay within 28 days following ICU admission. The secondary end point was 28-day in-hospital mortality, defined as death from any cause occurring during the entire hospitalization (including time spent in the ICU or after transfer to a general ward) within 28 days of ICU admission. Statistical analysis and risk prediction modeling and validation Continuous variables are presented as mean ± standard deviation (SD) or median with tertiles range, as appropriate. Group comparisons were performed using Student’s t-test or ANOVA for normally distributed data, and Mann–Whitney U or Kruskal–Wallis tests for non-parametric data. Categorical variables are reported as counts (percentages) and were compared using Pearson’s χ2 test or Fisher’s exact test, as appropriate. Univariate Cox regression analysis identified variables associated with 28-day ICU/HOSP mortality. The association between the HALP score and the time to event for each end point was evaluated using multivariable Cox proportional hazards regression, reported as hazard ratios (HRs) with 95% confidence intervals (CIs). Three sequential models were constructed: Model 1: unadjusted analysis, including only the HALP score. Model 2: adjusted for basic demographic variables considered fundamental confounders: age, sex, race, and body mass index (BMI). Model 3: additionally adjusted for an extensive set of potential confounders selected based on clinical relevance and univariate screening. These included: comorbidities (e.g., AKI, COPD, CKD), vital signs (heart rate, non-invasive systolic blood pressure), life-support interventions (vasopressor use, mechanical ventilation, continuous renal replacement therapy), and laboratory parameters that differed between survivors and non-survivors (Supplementary Table 3). After adjusting for all potential confounding covariates, we employed restricted cubic splines (RCS) with four knots (at the 5th, 35th, 65th, and 95th percentiles) to examine potential nonlinear relationships between HALP and clinical outcomes. Multicollinearity among the predictor variables was assessed using variance inflation factors (VIF) (Supplementary Table 5). All statistical analyses were performed using R software (version 4.2.2), with two-sided P values < 0.05 considered statistically significant. COX regression analysis and KM survival analysis were performed using the R package “survival (version 3.4-0)”, RCS analysis was performed using the R package “rms (version 6.3-0)”, and all visualizations were performed using the R package “ggplot2 (version 3.4.0)”. For the development of the predictive nomogram, we focused specifically on the primary end point of 28-day ICU mortality to create a tool directly applicable to the critical care setting. To identify predictors and prevent overfitting, we performed least absolute shrinkage and selection operator (LASSO) regression within the Cox regression framework using the R package ‘glmnet’. This penalized regression technique shrinks the coefficients of less important variables toward zero and can set them exactly to zero, thereby performing automatic variable selection. Using the complete cohort and the R package ‘glmnet’, we performed tenfold cross-validation to identify the optimal value of the tuning parameter (λ) that minimizes the prediction error. The predictors with non-zero coefficients at this optimal λ-value were retained as key predictors for our primary end point (28-day ICU death). We then applied multivariate Cox regression to these selected predictors to obtain final hazard ratios and construct our prognostic model, which was visualized as a nomogram. Discrimination was evaluated using the area under the receiver operating characteristic curve (AUC) for the binary outcome of 28-day ICU mortality. The predicted probabilities for this analysis were generated from the Cox model-based nomogram (Table 1). We report the AUC from this logistic regression analysis, as it is a standard and intuitive metric for performance at a fixed time point. To internally assess the model’s robustness and mitigate overfitting, we performed stratified random split validation. Using the createDataPartition function from the R package caret (version 6.0-94) with stratification by 28-day mortality status and random seed set to 1, the entire cohort was randomly divided into two internal validation subsets of equal size (1:1 ratio; Internal Validation Subset A and B) with balanced distribution of both clinical characteristics and the outcome variable. The model’s performance was then evaluated on both subsets to check for consistency. Table 1Summary descriptive table by groups of HALP score groupALLLowModerateHighP valueN = 858N = 286N = 286N = 286HALP20.8 [11.8;35.2]9.17 [5.55;11.8]20.8 [17.6;24.8]42.1 [35.2;55.5] < 0.001Age71.5 [61.0;81.0]76.5 [67.0;85.0]69.0 [60.0;80.0]68.0 [56.2;77.0] < 0.001Gender0.228 F341 (39.7%)123 (43.0%)115 (40.2%)103 (36.0%) M517 (60.3%)163 (57.0%)171 (59.8%)183 (64.0%)White464 (54.1%)155 (54.2%)161 (56.3%)148 (51.7%)0.551BMI27.4 [23.8;32.0]26.8 [22.7;32.4]27.4 [24.0;31.3]28.0 [24.7;32.3]0.149HYT330 (38.5%)88 (30.8%)113 (39.5%)129 (45.1%)0.002AKI421 (49.1%)174 (60.8%)128 (44.8%)119 (41.6%) < 0.001LIC46 (5.36%)11 (3.85%)13 (4.55%)22 (7.69%)0.094CKD240 (28.0%)104 (36.4%)76 (26.6%)60 (21.0%) < 0.001Diabetes315 (36.7%)109 (38.1%)103 (36.0%)103 (36.0%)0.835HF284 (33.1%)127 (44.4%)87 (30.4%)70 (24.5%) < 0.001COPD143 (16.7%)59 (20.6%)43 (15.0%)41 (14.3%)0.086SOFA5.00 [4.00;8.00]6.00 [4.00;8.00]5.00 [3.00;7.00]6.00 [4.00;8.00] < 0.001APSII46.0 [35.0;60.0]51.0 [39.0;66.0]44.0 [33.2;55.0]45.0 [32.2;59.0] < 0.001SAPSII40.0 [32.0;50.0]43.5 [36.0;54.0]38.0 [31.0;46.8]38.0 [30.0;47.8] < 0.001OASIS35.0 [30.0;40.0]36.0 [31.0;42.0]34.0 [28.0;40.0]34.0 [29.0;40.0]0.001GCS15.0 [12.0;15.0]15.0 [13.0;15.0]15.0 [12.0;15.0]15.0 [11.0;15.0]0.557APACHII19.0 [15.0;25.0]22.0 [16.0;26.0]18.0 [13.0;23.0]19.0 [15.0;24.0] < 0.001HR86.0 [74.0;99.0]86.5 [73.0;102]86.0 [73.2;98.0]85.0 [76.0;97.0]0.414NBPS122 [105;139]122 [103;137]125 [108;142]119 [104;139]0.104NBPD68.0 [57.2;80.0]67.0 [57.2;79.0]70.0 [58.0;83.0]67.0 [57.0;79.0]0.231RR19.0 [16.0;23.0]20.0 [17.0;23.0]19.0 [16.0;23.0]18.0 [14.2;22.0] < 0.001NBPM83.0 [71.0;95.0]83.0 [72.0;93.0]85.0 [72.0;96.0]81.0 [70.0;94.0]0.100Lymphocyte count1.13 [0.72;1.72]0.66 [0.42;0.92]1.12 [0.82;1.56]1.85 [1.34;2.44] < 0.001HCT32.5 [27.4;37.2]30.6 [26.9;35.9]34.0 [29.5;37.8]33.3 [26.7;38.2] < 0.001Hb10.5 [8.80;12.3]9.80 [8.43;11.4]11.1 [9.43;12.5]10.8 [8.72;12.8] < 0.001PLT178 [126;239]209 [151;291]192 [142;235]142 [100;192] < 0.001RDW14.4 [13.3;16.0]15.0 [13.9;17.0]14.1 [13.2;15.6]14.0 [13.2;15.4] < 0.001RBC3.53 [2.95;4.12]3.36 [2.86;3.97]3.64 [3.14;4.20]3.62 [2.89;4.26] < 0.001WBC12.3 [9.20;16.5]12.2 [8.93;17.2]11.6 [9.20;15.4]12.6 [9.33;16.7]0.223Neu9.55 [6.58;13.5]10.2 [6.86;15.8]8.98 [6.35;12.8]9.46 [6.60;12.4]0.018Mon0.23 [0.17;0.80]0.22 [0.16;0.75]0.23 [0.16;0.71]0.23 [0.19;0.99]0.009ALB3.00 [2.70;3.30]2.88 [2.50;3.18]3.10 [2.80;3.40]3.14 [2.89;3.50] < 0.001AG14.0 [12.0;17.0]15.0 [12.0;18.0]14.0 [12.0;17.0]14.0 [12.0;17.0]0.002Glucose137 [112;173]141 [113;175]138 [114;172]134 [107;170]0.406K4.10 [3.73;4.60]4.20 [3.80;4.80]4.10 [3.70;4.50]4.10 [3.70;4.50]0.029Na139 [136;142]138 [135;141]139 [136;142]140 [137;142]0.001LAC1.90 [1.30;2.80]1.90 [1.30;2.98]1.89 [1.30;2.60]1.90 [1.32;2.90]0.363INR1.30 [1.10;1.50]1.30 [1.10;1.50]1.20 [1.10;1.40]1.30 [1.20;1.60] < 0.001PT13.9 [12.2;16.5]14.1 [12.2;16.5]13.2 [12.1;15.6]14.4 [12.5;17.4]0.001APTT29.8 [26.5;35.0]30.6 [27.3;35.9]29.0 [25.9;34.1]30.1 [26.6;34.9]0.027ALT31.0 [17.0;66.0]31.0 [17.0;62.0]29.0 [16.2;59.8]35.9 [18.0;76.1]0.134AST44.0 [26.0;95.0]43.5 [24.0;85.8]40.0 [25.0;82.0]50.5 [31.0;120]0.005TB0.62 [0.40;1.00]0.60 [0.40;1.00]0.60 [0.40;0.90]0.70 [0.50;1.10]0.003CRE1.10 [0.80;1.60]1.30 [0.90;1.90]1.00 [0.70;1.50]0.90 [0.70;1.40] < 0.001UREA20.0 [14.0;34.0]27.0 [18.0;45.0]20.0 [14.0;29.0]17.0 [13.0;25.0] < 0.001VP617 (71.9%)209 (73.1%)190 (66.4%)218 (76.2%)0.029CRRT68 (7.93%)27 (9.44%)23 (8.04%)18 (6.29%)0.377MV544 (63.4%)164 (57.3%)177 (61.9%)203 (71.0%)0.003Hosp dead175 (20.4%)84 (29.4%)56 (19.6%)35 (12.2%) < 0.001Hosp time16.4 [9.73;27.5]15.4 [9.76;24.7]17.9 [9.73;28.0]16.7 [9.71;27.7]0.418ICU dead180 (21.0%)85 (29.7%)58 (20.3%)37 (12.9%) < 0.001ICU time6.66 [2.92;13.3]6.35 [2.70;12.2]7.74 [3.39;16.0]6.10 [2.72;12.6]0.063HYT, hypertension; AKI, acute kidney injury; LIC, liver cirrhosis; CKD, chronic kidney disease; HF, heart failure; COPD, chronic obstructive pulmonary disease; SOFA, Sequential Organ Failure Assessment; APSII, Acute Physiology and Chronic Health II Score; SAPSII, Simplified Acute Physiology Score II; OASIS, Oxford Acute Severity of Illness Score; GCS, Glasgow Coma Scale; APACHII, Acute Physiology and Chronic Health Evaluation II; HR, heart rate; NBPS, non-invasive blood pressure systolic; RR, respiratory rate; NBPM, non-invasive blood pressure Mean; HCT, hematocrit; Hb, hemoglobin; PLT, platelet; RDW, red blood cell distribution width; WBC, white blood cell; Neu, neutrophil; Mon, monocyte; ALB, albumin; AG, anion gap; K, potassium; Na, sodium; LAC, lactate; INR, international normalized ratio; PT, prothrombin time; APTT, activated partial thromboplastin time; ALT, alanine aminotransferase; AST, aspartate aminotransferase; TB, total bilirubin; CRE, creatinine; UREA, urea nitrogen; VP, vasoactive drugs; CRRT, continuous renal replacement therapy; MV, mechanical ventilation Table 1Summary descriptive table by groups of HALP score groupALLLowModerateHighP valueN = 858N = 286N = 286N = 286HALP20.8 [11.8;35.2]9.17 [5.55;11.8]20.8 [17.6;24.8]42.1 [35.2;55.5] < 0.001Age71.5 [61.0;81.0]76.5 [67.0;85.0]69.0 [60.0;80.0]68.0 [56.2;77.0] < 0.001Gender0.228 F341 (39.7%)123 (43.0%)115 (40.2%)103 (36.0%) M517 (60.3%)163 (57.0%)171 (59.8%)183 (64.0%)White464 (54.1%)155 (54.2%)161 (56.3%)148 (51.7%)0.551BMI27.4 [23.8;32.0]26.8 [22.7;32.4]27.4 [24.0;31.3]28.0 [24.7;32.3]0.149HYT330 (38.5%)88 (30.8%)113 (39.5%)129 (45.1%)0.002AKI421 (49.1%)174 (60.8%)128 (44.8%)119 (41.6%) < 0.001LIC46 (5.36%)11 (3.85%)13 (4.55%)22 (7.69%)0.094CKD240 (28.0%)104 (36.4%)76 (26.6%)60 (21.0%) < 0.001Diabetes315 (36.7%)109 (38.1%)103 (36.0%)103 (36.0%)0.835HF284 (33.1%)127 (44.4%)87 (30.4%)70 (24.5%) < 0.001COPD143 (16.7%)59 (20.6%)43 (15.0%)41 (14.3%)0.086SOFA5.00 [4.00;8.00]6.00 [4.00;8.00]5.00 [3.00;7.00]6.00 [4.00;8.00] < 0.001APSII46.0 [35.0;60.0]51.0 [39.0;66.0]44.0 [33.2;55.0]45.0 [32.2;59.0] < 0.001SAPSII40.0 [32.0;50.0]43.5 [36.0;54.0]38.0 [31.0;46.8]38.0 [30.0;47.8] < 0.001OASIS35.0 [30.0;40.0]36.0 [31.0;42.0]34.0 [28.0;40.0]34.0 [29.0;40.0]0.001GCS15.0 [12.0;15.0]15.0 [13.0;15.0]15.0 [12.0;15.0]15.0 [11.0;15.0]0.557APACHII19.0 [15.0;25.0]22.0 [16.0;26.0]18.0 [13.0;23.0]19.0 [15.0;24.0] < 0.001HR86.0 [74.0;99.0]86.5 [73.0;102]86.0 [73.2;98.0]85.0 [76.0;97.0]0.414NBPS122 [105;139]122 [103;137]125 [108;142]119 [104;139]0.104NBPD68.0 [57.2;80.0]67.0 [57.2;79.0]70.0 [58.0;83.0]67.0 [57.0;79.0]0.231RR19.0 [16.0;23.0]20.0 [17.0;23.0]19.0 [16.0;23.0]18.0 [14.2;22.0] < 0.001NBPM83.0 [71.0;95.0]83.0 [72.0;93.0]85.0 [72.0;96.0]81.0 [70.0;94.0]0.100Lymphocyte count1.13 [0.72;1.72]0.66 [0.42;0.92]1.12 [0.82;1.56]1.85 [1.34;2.44] < 0.001HCT32.5 [27.4;37.2]30.6 [26.9;35.9]34.0 [29.5;37.8]33.3 [26.7;38.2] < 0.001Hb10.5 [8.80;12.3]9.80 [8.43;11.4]11.1 [9.43;12.5]10.8 [8.72;12.8] < 0.001PLT178 [126;239]209 [151;291]192 [142;235]142 [100;192] < 0.001RDW14.4 [13.3;16.0]15.0 [13.9;17.0]14.1 [13.2;15.6]14.0 [13.2;15.4] < 0.001RBC3.53 [2.95;4.12]3.36 [2.86;3.97]3.64 [3.14;4.20]3.62 [2.89;4.26] < 0.001WBC12.3 [9.20;16.5]12.2 [8.93;17.2]11.6 [9.20;15.4]12.6 [9.33;16.7]0.223Neu9.55 [6.58;13.5]10.2 [6.86;15.8]8.98 [6.35;12.8]9.46 [6.60;12.4]0.018Mon0.23 [0.17;0.80]0.22 [0.16;0.75]0.23 [0.16;0.71]0.23 [0.19;0.99]0.009ALB3.00 [2.70;3.30]2.88 [2.50;3.18]3.10 [2.80;3.40]3.14 [2.89;3.50] < 0.001AG14.0 [12.0;17.0]15.0 [12.0;18.0]14.0 [12.0;17.0]14.0 [12.0;17.0]0.002Glucose137 [112;173]141 [113;175]138 [114;172]134 [107;170]0.406K4.10 [3.73;4.60]4.20 [3.80;4.80]4.10 [3.70;4.50]4.10 [3.70;4.50]0.029Na139 [136;142]138 [135;141]139 [136;142]140 [137;142]0.001LAC1.90 [1.30;2.80]1.90 [1.30;2.98]1.89 [1.30;2.60]1.90 [1.32;2.90]0.363INR1.30 [1.10;1.50]1.30 [1.10;1.50]1.20 [1.10;1.40]1.30 [1.20;1.60] < 0.001PT13.9 [12.2;16.5]14.1 [12.2;16.5]13.2 [12.1;15.6]14.4 [12.5;17.4]0.001APTT29.8 [26.5;35.0]30.6 [27.3;35.9]29.0 [25.9;34.1]30.1 [26.6;34.9]0.027ALT31.0 [17.0;66.0]31.0 [17.0;62.0]29.0 [16.2;59.8]35.9 [18.0;76.1]0.134AST44.0 [26.0;95.0]43.5 [24.0;85.8]40.0 [25.0;82.0]50.5 [31.0;120]0.005TB0.62 [0.40;1.00]0.60 [0.40;1.00]0.60 [0.40;0.90]0.70 [0.50;1.10]0.003CRE1.10 [0.80;1.60]1.30 [0.90;1.90]1.00 [0.70;1.50]0.90 [0.70;1.40] < 0.001UREA20.0 [14.0;34.0]27.0 [18.0;45.0]20.0 [14.0;29.0]17.0 [13.0;25.0] < 0.001VP617 (71.9%)209 (73.1%)190 (66.4%)218 (76.2%)0.029CRRT68 (7.93%)27 (9.44%)23 (8.04%)18 (6.29%)0.377MV544 (63.4%)164 (57.3%)177 (61.9%)203 (71.0%)0.003Hosp dead175 (20.4%)84 (29.4%)56 (19.6%)35 (12.2%) < 0.001Hosp time16.4 [9.73;27.5]15.4 [9.76;24.7]17.9 [9.73;28.0]16.7 [9.71;27.7]0.418ICU dead180 (21.0%)85 (29.7%)58 (20.3%)37 (12.9%) < 0.001ICU time6.66 [2.92;13.3]6.35 [2.70;12.2]7.74 [3.39;16.0]6.10 [2.72;12.6]0.063HYT, hypertension; AKI, acute kidney injury; LIC, liver cirrhosis; CKD, chronic kidney disease; HF, heart failure; COPD, chronic obstructive pulmonary disease; SOFA, Sequential Organ Failure Assessment; APSII, Acute Physiology and Chronic Health II Score; SAPSII, Simplified Acute Physiology Score II; OASIS, Oxford Acute Severity of Illness Score; GCS, Glasgow Coma Scale; APACHII, Acute Physiology and Chronic Health Evaluation II; HR, heart rate; NBPS, non-invasive blood pressure systolic; RR, respiratory rate; NBPM, non-invasive blood pressure Mean; HCT, hematocrit; Hb, hemoglobin; PLT, platelet; RDW, red blood cell distribution width; WBC, white blood cell; Neu, neutrophil; Mon, monocyte; ALB, albumin; AG, anion gap; K, potassium; Na, sodium; LAC, lactate; INR, international normalized ratio; PT, prothrombin time; APTT, activated partial thromboplastin time; ALT, alanine aminotransferase; AST, aspartate aminotransferase; TB, total bilirubin; CRE, creatinine; UREA, urea nitrogen; VP, vasoactive drugs; CRRT, continuous renal replacement therapy; MV, mechanical ventilation Summary descriptive table by groups of HALP score group HYT, hypertension; AKI, acute kidney injury; LIC, liver cirrhosis; CKD, chronic kidney disease; HF, heart failure; COPD, chronic obstructive pulmonary disease; SOFA, Sequential Organ Failure Assessment; APSII, Acute Physiology and Chronic Health II Score; SAPSII, Simplified Acute Physiology Score II; OASIS, Oxford Acute Severity of Illness Score; GCS, Glasgow Coma Scale; APACHII, Acute Physiology and Chronic Health Evaluation II; HR, heart rate; NBPS, non-invasive blood pressure systolic; RR, respiratory rate; NBPM, non-invasive blood pressure Mean; HCT, hematocrit; Hb, hemoglobin; PLT, platelet; RDW, red blood cell distribution width; WBC, white blood cell; Neu, neutrophil; Mon, monocyte; ALB, albumin; AG, anion gap; K, potassium; Na, sodium; LAC, lactate; INR, international normalized ratio; PT, prothrombin time; APTT, activated partial thromboplastin time; ALT, alanine aminotransferase; AST, aspartate aminotransferase; TB, total bilirubin; CRE, creatinine; UREA, urea nitrogen; VP, vasoactive drugs; CRRT, continuous renal replacement therapy; MV, mechanical ventilation Result Baseline characteristics of study participants This study analyzed data from 3118 patients in the MIMIC-IV database, of whom 858 neurocritical care patients met the inclusion criteria after exclusions (Fig. 1). Table 2 details the baseline characteristics of each group. The cohort had a median age of 71.5 years (tertiles range: 61–81), with 517 males (60.3%) and a median BMI of 27.4 kg/m2 (tertiles range: 23.8–32.0). Additionally, 38.5% had hypertension, 36.7% had diabetes, and 49.1% had acute kidney injury. The median HALP score in the cohort was 20.8 (tertiles range: 11.8–35.2). Patients were stratified into high-, moderate-, and low-HALP groups (n = 286 each). The median HALP scores were 9.17 (tertiles range: 5.55–11.8) in the low group, 20.8 (tertiles range: 17.6–24.8) in the moderate group, and 42.1 (tertiles range: 35.2–55.5) in the high group. Baseline data also indicated that patients with lower HALP scores tended to be older. Compared to the high-HALP group, the low-HALP group had higher RR, PLT, RDW, Neu, AG, creatinine, and urea levels. They also had lower Lym, hematocrit, hemoglobin, Red, Mon, albumin, sodium, and potassium levels. Clinical severity scores increased significantly as HALP scores decreased. Furthermore, absolute mortality rates decreased with higher HALP scores: 28-day in-hospital mortality occurred in 84 of 286 patients (29.4%) in the low-HALP group, 56 of 286 (19.6%) in the moderate group, and 35 of 286 (12.2%) in the high group (p < 0.001). Similarly, 28-day ICU mortality occurred in 85 (29.7%), 58 (20.3%), and 37 (12.9%) patients in the low-, moderate-, and high-HALP groups, respectively (p < 0.001).Fig. 1The research design and workflow of this studyTable 2The relationship between HALP score and 28-day ICU mortality rateCharacteristicModel 1Model 2Model 3HR195% CI1P valueHR195% CI1P valueHR195% CI1P valueHALP0.990.98, 0.990.0020.990.98, 1.000.0340.990.98, 1.000.13HALP groupLowRefRefRefRefRefRefModerate0.580.41, 0.810.0010.660.47, 0.930.0160.930.65, 1.330.7High0.430.29, 0.63 < 0.0010.530.35, 0.780.0020.560.36, 0.870.0091HR, hazard ratio; CI, confidence interval The research design and workflow of this study Table 2The relationship between HALP score and 28-day ICU mortality rateCharacteristicModel 1Model 2Model 3HR195% CI1P valueHR195% CI1P valueHR195% CI1P valueHALP0.990.98, 0.990.0020.990.98, 1.000.0340.990.98, 1.000.13HALP groupLowRefRefRefRefRefRefModerate0.580.41, 0.810.0010.660.47, 0.930.0160.930.65, 1.330.7High0.430.29, 0.63 < 0.0010.530.35, 0.780.0020.560.36, 0.870.0091HR, hazard ratio; CI, confidence interval The relationship between HALP score and 28-day ICU mortality rate 1HR, hazard ratio; CI, confidence interval Multivariable Cox regression analysis of the association between HALP score and short-term mortality in neurocritical care patients As shown in Table 3, after full adjustment for all covariates in the multivariable Model 3, no significant association was observed between continuous HALP scores and 28-day ICU mortality (HR = 0.99, 95% CI 0.98–1.00, p = 0.13). However, when analyzed categorically, the high HALP group demonstrated significantly lower mortality risk compared to the low HALP group in the fully adjusted model (HR = 0.56, 95% CI 0.36–0.87, p = 0.009). Similarly for 28-day in-hospital mortality (Table 4), higher HALP scores showed a modest inverse association when analyzed continuously in Model 3 (HR = 0.99, 95% CI 0.98–1.00, p = 0.02). The categorical analysis revealed more pronounced protection, with the high HALP group showing significantly reduced mortality versus the low group after multivariable adjustment in Model 3 (HR = 0.52, 95% CI 0.33–0.80, p = 0.003). Trend tests based on the adjusted models supported a significant dose–response relationship between decreasing HALP scores and increasing mortality risk (P < 0.05 for both ICU and hospital mortality). Restricted cubic spline analysis, adjusted for all covariates in Model 3, confirmed a significant nonlinear inverse relationship between HALP scores and mortality. The overall association was significant for both ICU (P-overall = 0.017) and hospital mortality (P-overall = 0.024), with the nonlinear component being highly significant (P-nonlinear = 0.007 for ICU mortality; P-nonlinear = 0.010 for hospital mortality). The RCS curves (Figs. 2A, B) visually demonstrate this reverse J-shaped relationship, characterized by a steep decrease in mortality risk as HALP increases from very low levels, followed by a plateau. This observed nonlinearity provides a statistical rationale for the loss of significance when HALP was modeled as a linear continuous variable and supports the use of tertile categories to capture its threshold-dependent association with mortality. Kaplan–Meier survival curves (which are unadjusted) visually supported the significant inverse association between higher HALP scores and improved survival probability. The accompanying univariable survival analysis quantified these findings, showing the high HALP group had lower 28-day ICU mortality (HR = 0.66, 95% CI 0.53–0.78) and hospital mortality (HR = 0.63, 95% CI 0.52–0.77) compared to the low group (Figs. 2C/D). Table 3The relationship between HALP score and 28-day hospital mortality rateModel 1Model 2Model 3CharacteristicHR195% CI1p-valueHR195% CI1p-valueHR195% CI1p-valueHALP0.980.97, 0.99 < 0.0010.990.98, 1.000.0040.990.98, 1.000.020HALP group––––––––– Low–––––––– Moderate0.630.45, 0.880.0070.720.51, 1.010.0570.960.67, 1.380.8 High0.400.27, 0.60 < 0.0010.480.32, 0.72 < 0.0010.520.33, 0.800.0031HR, hazard ratio; CI, confidence intervalTable 4Final multivariable Cox proportional hazards model for 28-day mortalityVariableCoefficient (β)Standard errorHazard ratio (HR)95% Confidence interval for HRP-valueAge0.02530.00591.026(1.014–1.037) < 0.001APSII0.01670.00361.017(1.010–1.024) < 0.001RDW0.10390.0251.11(1.057–1.165) < 0.001PT0.01680.00541.017(1.006–1.028)0.002APTT0.00790.00281.008(1.002–1.013)0.005HALP− 0.46680.20610.627(0.418–0.940)0.024HR, hazard ratio; CI, confidence interval; APS II, Acute Physiology Score II; RDW, red cell distribution width; PT, prothrombin time; APTT, activated partial thromboplastin time; HALP, hemoglobin, albumin, lymphocyte, platelet scoreFig. 2Nonlinear association and survival analysis between HALP score and mortality. A, B Restricted cubic spline (RCS) curves show the nonlinear relationship between continuous HALP score and the odds of A 28-day ICU mortality and B 28-day in-hospital mortality. The solid line represents the adjusted odds ratio (HR), with the dashed lines indicating the 95% confidence interval. The reference value (HR = 1) is set at the median HALP score. The P-values for overall association and nonlinearity are shown. C, D Kaplan–Meier survival curves illustrate the probability of survival from C ICU death and D in-hospital death 28 days, stratified by HALP score tertiles (low, blue; moderate, green; high, red). P-value and hazard ratios (HR) with 95% confidence intervals are derived from Cox proportional hazards models, using the low HALP group as reference Table 3The relationship between HALP score and 28-day hospital mortality rateModel 1Model 2Model 3CharacteristicHR195% CI1p-valueHR195% CI1p-valueHR195% CI1p-valueHALP0.980.97, 0.99 < 0.0010.990.98, 1.000.0040.990.98, 1.000.020HALP group––––––––– Low–––––––– Moderate0.630.45, 0.880.0070.720.51, 1.010.0570.960.67, 1.380.8 High0.400.27, 0.60 < 0.0010.480.32, 0.72 < 0.0010.520.33, 0.800.0031HR, hazard ratio; CI, confidence interval The relationship between HALP score and 28-day hospital mortality rate 1HR, hazard ratio; CI, confidence interval Table 4Final multivariable Cox proportional hazards model for 28-day mortalityVariableCoefficient (β)Standard errorHazard ratio (HR)95% Confidence interval for HRP-valueAge0.02530.00591.026(1.014–1.037) < 0.001APSII0.01670.00361.017(1.010–1.024) < 0.001RDW0.10390.0251.11(1.057–1.165) < 0.001PT0.01680.00541.017(1.006–1.028)0.002APTT0.00790.00281.008(1.002–1.013)0.005HALP− 0.46680.20610.627(0.418–0.940)0.024HR, hazard ratio; CI, confidence interval; APS II, Acute Physiology Score II; RDW, red cell distribution width; PT, prothrombin time; APTT, activated partial thromboplastin time; HALP, hemoglobin, albumin, lymphocyte, platelet score Final multivariable Cox proportional hazards model for 28-day mortality HR, hazard ratio; CI, confidence interval; APS II, Acute Physiology Score II; RDW, red cell distribution width; PT, prothrombin time; APTT, activated partial thromboplastin time; HALP, hemoglobin, albumin, lymphocyte, platelet score Nonlinear association and survival analysis between HALP score and mortality. A, B Restricted cubic spline (RCS) curves show the nonlinear relationship between continuous HALP score and the odds of A 28-day ICU mortality and B 28-day in-hospital mortality. The solid line represents the adjusted odds ratio (HR), with the dashed lines indicating the 95% confidence interval. The reference value (HR = 1) is set at the median HALP score. The P-values for overall association and nonlinearity are shown. C, D Kaplan–Meier survival curves illustrate the probability of survival from C ICU death and D in-hospital death 28 days, stratified by HALP score tertiles (low, blue; moderate, green; high, red). P-value and hazard ratios (HR) with 95% confidence intervals are derived from Cox proportional hazards models, using the low HALP group as reference Incremental prognostic value of the HALP score We quantitatively assessed the incremental prognostic value of the HALP score by measuring the absolute improvement in the area under the curve (ΔAUC) when added to conventional severity scores. The integration of HALP consistently and significantly enhanced the predictive accuracy for 28-day ICU mortality across all models, with AUC values increasing as follows: APACHE II (from 0.63 to 0.67; ΔAUC = + 0.04), SOFA (from 0.62 to 0.68; ΔAUC = + 0.06), APS II (from 0.67 to 0.70; ΔAUC = + 0.03), SAPS II (from 0.67 to 0.69; ΔAUC = + 0.02), and OASIS (from 0.63 to 0.67; ΔAUC = + 0.04) (Fig. 3).Fig. 3Incremental predictive value of the HALP score combined with conventional severity scores. Receiver operating characteristic (ROC) curves demonstrate the improvement in discriminative performance for predicting 28-day ICU mortality when the HALP score is integrated with established severity scores. Each panel compares the model with a severity score alone (red line) to the model combining the severity score with the HALP score (blue line). A APACHE II, B SOFA, C APS II, D SAPS II, E OASIS Incremental predictive value of the HALP score combined with conventional severity scores. Receiver operating characteristic (ROC) curves demonstrate the improvement in discriminative performance for predicting 28-day ICU mortality when the HALP score is integrated with established severity scores. Each panel compares the model with a severity score alone (red line) to the model combining the severity score with the HALP score (blue line). A APACHE II, B SOFA, C APS II, D SAPS II, E OASIS Construction of prediction models and nomogram We first evaluated 48 candidate variables in the complete cohort using LASSO regression (Figure S1A/B), followed by multivariate Cox proportional hazards analysis. LASSO regression identified eight key predictors (Figure S2), with subsequent multivariate Cox analysis supporting six independent risk factors for 28-day ICU mortality in neurocritical care patients: the HALP score, age, red cell distribution width (RDW), prothrombin time (PT), activated partial thromboplastin time (APTT), and Acute Physiology Score II (APS II) (Fig. 4). Compared with conventional severity scores (SOFA, APS II, SAPS II, OASIS), our nomogram demonstrated superior sensitivity and specificity in the complete cohort (Fig. 5A) and maintained robust performance in both internal validation subsets, with area under the receiver operating characteristic curve (AUROC) values of 0.71 (Internal Validation Subset A; Fig. 5B) and 0.73 (Internal Validation Subset B; Fig. 5C). Furthermore, the calibration performance of our Cox regression nomogram was rigorously assessed using bootstrap resampling with 1000 repetitions. The calibration curve (Fig. 6A) demonstrates excellent agreement between predicted probabilities and observed outcomes across the entire risk spectrum, with the fitting curve closely aligned with the ideal 45-degree diagonal line. Quantitative assessment further confirmed this strong calibration performance, with a calibration in the large (intercept) of − 0.023 (ideal = 0) and a calibration slope of 0.974 (ideal = 1). These metrics indicate minimal systematic over- or underprediction and well-maintained discriminative ordering of risk. The model demonstrates particular accuracy in the clinically relevant moderate-to-high risk ranges, supporting its potential utility for risk stratification in neurocritical care settings. Decision curve analysis confirmed the clinical utility of the nomogram (Fig. 6B). The model provided superior net benefit compared to default strategies across a wide range of threshold probabilities (approximately 5%–70%). Specifically, within the clinically most relevant threshold range of 20% to 50%—where most clinical decisions are made regarding treatment intensification or de-escalation—our model demonstrated moderate discrimination. For instance, at a threshold probability of 30%, the model yielded a net benefit of 0.12, indicating that for every 100 patients, using our nomogram could prevent 12 unnecessary interventions compared to a ‘treat-all’ strategy, without compromising patient outcomes. This robust performance confirms that using our nomogram to guide clinical decisions would be beneficial and could help avoid both overtreatment and undertreatment.Fig. 4Nomogram for predicting the risk of 28-day ICU mortality. This clinical tool integrates the HALP score and other independent predictors identified from multivariate analysis. To use the nomogram: for each patient variable, locate the value on the corresponding axis and draw a line upward to the ‘Points’ axis to determine the individual score. Sum the points for all variables, locate the total on the ‘Total Points’ axis, and then draw a line straight down to the ‘Risk of 28-day ICU mortality’ axis to obtain the individualized predicted probabilityFig. 5Discriminative performance of the nomogram versus conventional scores in the training and internal validation subset cohorts. Discriminative performance of the newly developed nomogram is compared against best-performing five conventional severity scores for predicting 28-day ICU mortality across different cohorts. A Complete cohort, B Internal Validation Subset A, C Internal Validation Subset B. The area under the curve (AUC) values with 95% confidence intervals are shown for the nomogram modelFig. 6Calibration and decision curve analysis of the nomogram. A The calibration curve demonstrates the agreement between predicted probabilities from our Cox regression model (x-axis) and observed outcomes (y-axis) after bootstrap validation with 1000 resamples. The solid line represents the bias-corrected calibration curve, while the dashed diagonal indicates perfect calibration. The close alignment across all risk strata confirms the model’s accuracy in predicting 28-day mortality risk, with quantitative metrics supporting excellent calibration performance (α = − 0.023, β = 0.974). B Decision curve analysis for the nomogram. Decision curve is an evaluation method used to help determine high-risk interventions, low-risk avoidance interventions, and assess the degree of benefit. The horizontal axis represents the threshold probability, and the vertical axis represents the net benefit. The green curve ‘None’ indicates no intervention and no benefit. All blue curve indicate that all samples were intervened. The red curve (model) is used to predict the net profit of decision-making Nomogram for predicting the risk of 28-day ICU mortality. This clinical tool integrates the HALP score and other independent predictors identified from multivariate analysis. To use the nomogram: for each patient variable, locate the value on the corresponding axis and draw a line upward to the ‘Points’ axis to determine the individual score. Sum the points for all variables, locate the total on the ‘Total Points’ axis, and then draw a line straight down to the ‘Risk of 28-day ICU mortality’ axis to obtain the individualized predicted probability Discriminative performance of the nomogram versus conventional scores in the training and internal validation subset cohorts. Discriminative performance of the newly developed nomogram is compared against best-performing five conventional severity scores for predicting 28-day ICU mortality across different cohorts. A Complete cohort, B Internal Validation Subset A, C Internal Validation Subset B. The area under the curve (AUC) values with 95% confidence intervals are shown for the nomogram model Calibration and decision curve analysis of the nomogram. A The calibration curve demonstrates the agreement between predicted probabilities from our Cox regression model (x-axis) and observed outcomes (y-axis) after bootstrap validation with 1000 resamples. The solid line represents the bias-corrected calibration curve, while the dashed diagonal indicates perfect calibration. The close alignment across all risk strata confirms the model’s accuracy in predicting 28-day mortality risk, with quantitative metrics supporting excellent calibration performance (α = − 0.023, β = 0.974). B Decision curve analysis for the nomogram. Decision curve is an evaluation method used to help determine high-risk interventions, low-risk avoidance interventions, and assess the degree of benefit. The horizontal axis represents the threshold probability, and the vertical axis represents the net benefit. The green curve ‘None’ indicates no intervention and no benefit. All blue curve indicate that all samples were intervened. The red curve (model) is used to predict the net profit of decision-making Discussion In this retrospective analysis of prospectively collected data from the MIMIC-IV database, we evaluated 858 neurocritical care patients to examine the association between HALP score and short-term (28-day) mortality. Our findings suggest that HALP could serve as a novel, readily assessable, exploratory biomarker associated with short-term outcomes in neurocritical patients. As a proof of concept, we also developed a nomogram (Fig. 4) that integrates multiple readily available parameters to illustrate how these factors might collectively inform mortality risk. This nomogram is presented solely as a research tool to visualize the associative relationships found in our dataset. Its potential clinical utility, for instance in possibly helping to identify patients at higher or lower risk, requires rigorous validation in future prospective studies before any clinical application can be considered. For example, after further validation, a clinician could use this nomogram upon ICU admission to help flag patients at unexpectedly high risk who could be candidates for more intensive monitoring or aggressive supportive care, even if they do not appear critically ill based on conventional single-organ assessments. Conversely, it could help identify patients with a very low predicted risk, potentially supporting decisions for less invasive management or earlier step-down unit transfer, thereby optimizing resource allocation. Cerebrovascular diseases pose a substantial global healthcare burden, creating an urgent need for reliable biomarkers to guide prognostication in neurocritical care [2, 3]. In this context, our preliminary study suggests that the hemoglobin, albumin, lymphocyte, and platelet (HALP) score may serve as a novel composite marker that is independently associated with 28-day mortality in neurocritical care patients. This finding extends the established utility of HALP from oncology and cardiology [25–27] into the realm of acute brain injury, suggesting its role may reflect a fundamental axis of physiological reserve pertinent to critical illness survival. To our knowledge, this is the first study to systematically evaluate HALP in a heterogeneous neurocritical care population. The prognostic value of HALP likely stems from its integrated measurement of both nutritional status and systemic inflammation—two domains intricately linked in a vicious cycle that exacerbates secondary brain injury [28, 29]. For neurocritical care patients, the HALP score provides a more comprehensive biological rationale than established markers like the Prognostic Nutritional Index (PNI), neutrophil-to-lymphocyte ratio (NLR), or platelet-to-lymphocyte ratio (PLR). While valuable, NLR and PLR primarily reflect the balance between pro-inflammatory (neutrophils) or pro-thrombotic (platelets) forces and immune regulation (lymphocytes) [30, 31]. Unlike PNI, HALP incorporates hemoglobin—critical for cerebral oxygen delivery—and platelets, which are integral to thromboinflammation, a key mechanism of secondary brain injury [32]. Furthermore, compared to the ratios of NLR and PLR, HALP encapsulates a broader synergy of oxygen supply, nutritional status, immune defense, and thrombotic tendency. This integrated assessment may offer a distinct advantage for risk stratification in neurocritical care patients, whose outcomes are driven by a complex interplay of ischemic, inflammatory, and catabolic processes [33, 34]. This comprehensive profile is particularly relevant in neurocritical care, where outcomes are dictated by a complex interplay of ischemia, inflammation, and catabolism. A low HALP score may thus represent a state of “neuro-inflammatory–metabolic exhaustion”, characterized by impaired oxygen delivery, diminished nutritional reserves, immune dysregulation, and pro-thrombotic tendencies—all contributing to reduced resilience to acute neurological insults. Beyond its stand-alone value, the HALP score provided modest but statistically significant incremental prognostic information when integrated with conventional severity scores (e.g., APACHE II, SOFA), suggesting it captures pathophysiological elements distinct from those assessed by existing tools. The RCS analysis revealed a statistically significant nonlinear relationship between the HALP score and mortality (p for nonlinearity < 0.01), characterized by a reverse J-shaped pattern. This suggests the presence of threshold effects, wherein mortality risk was highest at the lowest HALP scores and decreased steeply as the score increased, eventually transitioning into a wide plateau across the middle range of values. This pattern implies that the prognostic information provided by HALP is not linear and that the clinical implications of a severely low score may differ from variations within the mid-range. The subsequent loss of independent significance for HALP as a continuous variable in the multivariable model, contrasted with the prognostic utility of its categorical stratification, is consistent with this complex, nonlinear relationship. Therefore, the nomogram represents a pragmatic tool that incorporates this nonlinearity for early risk stratification. It is important to note that the precise inflection points of this curve and the potential slight variations at the extreme ends of the HALP distribution should be interpreted with caution, as they may be influenced by the sample size at these extremes and require confirmation in future studies with larger cohorts. In summary, our study proposes that the HALP score, as a composite marker reflecting both nutritional status and inflammatory response, has potential as an exploratory biomarker in neurocritical care. We emphasize that these findings are exploratory and must be validated externally in prospective studies before any clinical application can be considered. It is also important to note that by requiring a minimum 24-h ICU stay for inclusion, our study population, by design, does not include patients who died within the first 24 h of admission. Consequently, our findings are not generalizable to this specific subgroup of ultra-severe patients, and the reported associations between HALP and mortality apply to those who survive the initial acute phase of their critical illness. The generalizability (external validity) of our findings is limited. The data originated from a single US academic tertiary center, and the patient case mix, clinical protocols, and resource levels may not be representative of community hospitals, non-academic centers, or healthcare systems in other countries. Third, and most importantly, residual confounding remains a substantial concern despite our extensive adjustments. Our analysis lacked crucial data on detailed neurological severity (e.g., state of consciousness, pupillary reactivity), specific neuroimaging metrics (e.g., hematoma/ischemic core volume), sedation protocols, and the development of hospital-acquired infections. The absence of these potent prognostic factors means we cannot rule out that the HALP score is, in part, a proxy for this unmeasured disease severity. Fourth, despite the statistically significant incremental predictive value offered by the HALP score over traditional severity scores, we acknowledge that the absolute discriminative ability of our model (AUC ≤ 0.73) remains modest. This level of accuracy, while promising for a novel biomarker, is insufficient to definitively guide individual patient-level decisions in a complex clinical setting. The clinical utility of such a model may therefore lie not in definitive prediction, but rather in complementary risk stratification. Fifth, a key limitation of our study stems from the inherent heterogeneity of the neurocritical care population, which encompasses a wide spectrum of underlying diagnoses (e.g., ischemic stroke, intracerebral hemorrhage, traumatic brain injury). While our primary aim was to identify a prognostic marker with broad applicability across this spectrum—reflecting the real-world case mix of a neuro-ICU—we acknowledge that our reported pooled estimate represents an average effect. This approach risks obscuring important effect modification, whereby the association between HALP and mortality might differ in magnitude or even direction across specific patient subgroups. Sixth, and of critical importance, the generalizability of our findings is a primary limitation, as they are derived from a single-center, tertiary-care database (MIMIC-IV) with a specific patient case mix and clinical protocols. Consequently, our results may not be directly applicable to other settings, such as community hospitals or healthcare systems in different countries. Therefore, external validation in independent, multi-center, and geographically diverse cohorts is an essential prerequisite before any consideration of clinical application can be warranted. Seventh, our analysis utilized all laboratory values as extracted from the database without applying methods to adjust for statistical outliers (e.g., winsorizing or truncating). While this approach preserves the complete dataset and acknowledges that extreme values can reflect genuine, high-acuity pathophysiology in the ICU, it also carries the risk that rare measurement errors or data entry artifacts could influence the model estimates. Future prospective studies designed to validate the HALP score would benefit from implementing pre-defined protocols for the clinical verification of extreme laboratory values to further strengthen data quality. Eighth, the data for this study was collected over a period of 14 years (2008–2022) based on the MIMIC-IV (v3.1) database. During this time, advancements in neurocritical care management, changes in clinical protocols, and evolving guidelines (e.g., for blood product transfusion, glycemic control, or infection management) may have occurred. These temporal trends represent a potential source of unmeasured confounding, as patient outcomes could be influenced by the standard of care at the time of their admission, which we were unable to fully account for in our analysis. Finally, although we employed LASSO to reduce overfitting and demonstrated consistent performance across subsets, future studies with larger sample sizes would benefit from employing bootstrapping techniques to obtain a more precise optimism-corrected performance estimate. Moreover, it is crucial to formally evaluate the proportional hazards assumption of Cox regression models. This assumption is crucial for explaining the hazard ratio as constant over time. Therefore, the predictive model needs to be externally validated in future research, including rigorous testing of model assumptions. In addition, our study utilized Cox regression to estimate the average strength of association between HALP and mortality over the 28-day period. Although the primary aim of our study was exploratory—to identify a significant associative signal rather than to build a definitive predictive model—and our multi-model approach with consistently significant results supports the robustness of the association, the violation of the PH assumption cannot be ruled out. Therefore, the hazard ratios should be interpreted as average associations over the 28-day period rather than precise, time-invariant estimates. Besides, our study relies on a univariate screening approach for variable selection into the multivariable model. While this is a common practice, it is epidemiologically suboptimal, as it can contribute to model overfitting and, more critically, may lead to omitted variable. One more thing is that the prognostic nomogram we developed has not undergone internal validation using resampling methods such as bootstrap optimism correction or cross-validation. Consequently, the apparent model performance we report are likely subject to a degree of overoptimism and may not fully represent its performance when applied to new patient samples. This underscores the exploratory nature of our findings and highlights that the nomogram, in its current form, should be considered a preliminary tool for hypothesis generation rather than for direct clinical decision-making. Its utility must be confirmed through external validation in independent, multi-center cohorts. However, we acknowledge that this method has limitations compared to fully pre-specified models or techniques like penalized regression (e.g., LASSO), which can more efficiently handle multicollinearity and reduce overfitting. The application of such advanced statistical methods in future studies could help to validate and refine our predictive model. In short, as with any observational study that identifies novel associations, our findings are exploratory and must be considered hypothesis generating. The results we present require validation in independent, prospective cohorts to confirm their robustness and generalizability. Future studies may also investigate these associations within specific patient subgroups, which would necessitate careful statistical design to control for the risk of type I error due to multiple comparisons. Conclusion In summary, this exploratory analysis of a single-center database indicates an association between the admission HALP score and short-term mortality in a heterogeneous neurocritical care population. The findings generate the hypothesis that this composite biomarker, reflecting immune–nutritional status, could aid in risk stratification. However, given the study’s limitations—including its retrospective design, residual confounding, modest model discrimination, and limited generalizability—these results must be considered preliminary and hypothesis generating. The clinical nomogram we developed is a research tool that requires rigorous external validation in prospective, multi-center studies to confirm its value and general applicability before it can be considered for clinical use. Supplementary Information Supplementary material 1. Supplementary material 1.