Association between neutrophil percentage to albumin ratio and mortality in takotsubo syndrome patients: analysis of the MIMIC-IV database Authors: Chongzhe Pei, Kuan Yang, Xin Ding, Hongshi Li, Li Xue, Hongmei Zheng, Heng Cai Published: 2026 ABSTRACT AbstractBackgroundThis study investigates the association between the neutrophil percentage-to-albumin ratio (NPAR) and mortality in takotsubo syndrome (TS) patients.MethodsWe conducted a retrospective cohort study using data from the MIMIC-IV version 3.1 database. Multivariable Cox regression and restricted cubic spline (RCS) analyses were performed to evaluate the associations between NPAR and 28-day, 180-day, and 365-day mortality. Predictive performance for 28-day mortality was assessed using receiver operating characteristic (ROC) analysis and seven machine-learning models.ResultsA total of 201 patients with TS were included. Higher NPAR was independently associated with increased 28-day mortality (HR 5.494, 95% CI 1.939–12.190, P < 0.001), 180-day mortality (HR 5.141, 95% CI 2.367–11.166, P < 0.001), and 365-day mortality (HR 5.280, 95% CI 2.439–11.429, P < 0.001). RCS analysis showed a significant overall association between NPAR and 28-day mortality (P for overall = 0.041), without significant nonlinearity (P for nonlinear = 0.194). NPAR showed modest predictive ability for 28-day mortality (AUC 0.640). In machine-learning analyses, NPAR was identified as an important prognostic feature, and the random forest model achieved the highest AUC (0.771).ConclusionsHigher NPAR values are associated with increased mortality rates at 28-day, 180-day, and 365-day intervals in TS patients, suggesting NPAR as a potential predictive marker for TS outcomes.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12872-026-06024-z.Keywords: Neutrophil percentage to albumin ratio, Takotsubo syndrome, MIMIC-IV database, Mortality Introduction Takotsubo syndrome (TS), commonly referred to broken heart syndrome, is a complicated stress-triggered cardiomyopathy. It is usually featured by temporary left ventricular (LV) dysfunction, and no obvious obstruction is observed in the coronary arteries [1]. First described in 1990, TS has emerged as an increasingly recognized cause of acute heart failure and acute coronary syndrome (ACS)-like presentations, making up 1–3% of all patients worldwide who are admitted with suspected ACS [2]. Epidemiologically, TS exhibits a strong female predominance, with approximately 80–90% of cases occurring in postmenopausal women, though its incidence in men and younger populations is also rising [3]. The pathophysiology of TS remains incompletely elucidated, but it is widely accepted that intense emotional or physical stress triggers a surge in sympathetic activation, leading to excessive catecholamine release, myocardial stunning, and inflammatory cascade activation—key drivers of LV dysfunction [4]. Despite the transient nature of LV dysfunction in many TS patients, short and long-term mortality remain non-negligible. Recent meta-analyses report 30-day mortality rates of 4–5% and 5-year mortality rates of 15–20% in TS cohorts, with adverse outcomes including heart failure exacerbations, arrhythmias, and recurrent TS [5, 6]. A major clinical challenge in managing TS lies in identifying reliable prognostic markers to stratify patients with elevated mortality risk. The predictive accuracy of traditional risk factors is limited when they are used in isolation. Thus, there is an urgent need for accessible, cost-effective biomarkers that integrate key pathophysiological pathways to improve risk stratification in TS. Inflammation and nutritional derangement lie at the core of both the pathogenesis and prognosis of TS. During the acute phase of TS, sympathetic overactivation induces a robust inflammatory response, characterized by elevated circulating neutrophils (a primary effector of innate immunity) and proinflammatory cytokines (specifically interleukin-6, tumor necrosis factor-α) [7]. Neutrophil infiltration into the myocardium further exacerbates myocardial injury by releasing reactive oxygen species and proteases, impairing LV contractility and delaying recovery [8]. In contrast, hypoalbuminemia, a marker reflecting poor nutritional status and systemic inflammation, is prevalent among critically ill TS patients [9]. Albumin, a major plasma protein synthesized by the liver, serves vital functions in preserving vascular integrity, regulating inflammation, and transporting essential nutrients; its reduction in TS reflects not only malnutrition but also inflammation-mediated hepatic synthetic dysfunction and vascular leakage [9]. Given the complementary roles of neutrophil-driven inflammation and albumin-related nutritional status in TS, composite ratios integrating these two parameters have gained attention as prognostic tools in cardiovascular disease. For example, the neutrophil-to-lymphocyte ratio (NLR) has been linked to mortality in TS, but it focus solely on inflammatory cell counts and ignore nutritional status [10, 11]. The neutrophil percentage to albumin ratio (NPAR)—computed as (neutrophil percentage×100)/albumin (g/dL)—offers a unique advantage by combining an inflammatory marker (neutrophil percentage, a more stable measure than absolute neutrophil count) with a nutritional/inflammatory marker (albumin), thereby capturing both inflammatory burden and systemic reserve [12, 13]. In non-TS populations, NPAR has also been shown to predict mortality more effectively than individual markers or other ratios [14, 15]. Currently, research examining the connection between NPAR and mortality outcomes specifically in TS patients remains unexplored in the literature. To address this gap, a retrospective cohort analysis was executed. The investigation utilized patient records from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, which serves as a comprehensive, public repository of ICU patient information. Our primary objectives were to evaluate the link between baseline NPAR and short-term (28-day) and long-term (180-day, 365-day) mortality and assess the predictive capability of NPAR for 28-day mortality in TS patients using different models. Methods Data source The present retrospective observational research employed MIMIC-IV 3.1 as the primary data source. MIMIC-IV 3.1 represents an open-access clinical repository containing extensive, detailed information from 76,540 ICU patients at Beth Israel Deaconess Medical Center (BIDMC) in Boston between 2008 and 2019. The principal investigator completed required online training through the Collaborative Institutional Training Initiative (certification ID: 12486234; Record IDs: 57743676). Ethical clearance for utilizing PhysioNet Clinical Databases was obtained from both the Massachusetts Institute of Technology (MIT) and the BIDMC review boards. Patient privacy protection was ensured through data de-identification, eliminating the requirement for individual consent. Study population Subjects admitted to the ICU received TS diagnoses based on International Classification of Diseases codes (“429.83” in version 9 and “I51.81” in version 10) within the MIMIC-IV 3.1 database. For patients with multiple ICU admissions, only the first ICU admission was included. The study excluded individuals meeting these criteria: (I) Age below 18 years; (II) ICU length of stay shorter than 24 h; (III) Cases lacking complete data for serum albumin, absolute neutrophil count, and neutrophil percentage. After applying these criteria, 201 patients were included in the final analysis. Fig. 1 illustrates the detailed patient selection process. Flowchart of cohort selection for patients with Takotsubo syndrome from the MIMIC-IV 3.1 database Exposure and outcomes The neutrophil percentage represented the proportion of neutrophils among total white blood cells. The NPAR calculation employed the neutrophil percentage as numerator divided by albumin, obtained from identical admission blood samples, using this formula: Neutrophil percentage (%) ∗ 100/Albumin (g/dL). Albumin measurements were presented in grams per deciliter. For ensuring reliable and precise NPAR evaluation, the assessment incorporated patients experiencing their initial ICU admission with stays exceeding 24 h, with NPAR values determined within the initial 24 h post-ICU admission. The study population was categorized into two groups according to the ROC-derived cut-off value of NPAR (29.35): a low NPAR group (NPAR < 29.35, n = 119) and a high NPAR group (NPAR ≥ 29.35, n = 82). The primary endpoint examined was 28-day mortality, whereas 180-day mortality and 365-day mortality were defined as secondary endpoints. Study variables The dataset contained extensive patient information, encompassing demographics (age, sex, race), admission year, first care unit, laboratory measurements, and comorbidities (e.g., coronary heart disease, chronic pulmonary disease, liver disease, cerebrovascular disease, diabetes mellitus (DM), renal disease, congestive heart failure (CHF), and cancer). Additionally, it included SOFA scores, Glasgow Coma Scale (GCS) scores, and vital signs such as heart rate, respiration rate, mean arterial pressure (MAP), body temperature, and SpO2. The laboratory measurements consisted of white blood cells (WBC), platelet count, hemoglobin levels, mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), alanine aminotransferase (ALT), mean corpuscular volume (MCV), aspartate aminotransferase (AST), electrolytes (potassium, sodium), alkaline phosphatase (ALP), total bilirubin, creatinine, blood urea nitrogen (BUN), anion gap, bicarbonate, glucose, prothrombin time (PT), partial thromboplastin time (PTT), and international normalized ratio (INR). All SOFA scores, GCS scores, vital signs, and laboratory measurements were obtained from data gathered during the first 24 h post-ICU admission. Moreover, it incorporated the initial 24 h therapeutic management of patients following ICU admission, including antibiotics, glucocorticoid use, mechanical ventilation, and vasopressor. The average was computed for measurements taken multiple times within the first 24 h after ICU admission. Among these variables, the pre-specified covariates included in the fully adjusted Cox model were age, gender, acute kidney injury (AKI), acute respiratory failure (ARF), cerebrovascular disease, Charlson Comorbidity Index, COPD, ALT, ALP, potassium, PTT, sodium, glucocorticoid use, GCS score, ischemic heart disease (IHD), myocardial infarction (MI), and SOFA score. Statistical analysis Statistical analyses were executed utilizing R software (version 4.3.0). Statistical significance was established at P < 0.05. Baseline characteristics of patients with TS were compared between the low and high NPAR groups. For continuous variables, analysis employed Student’s t-test or Mann–Whitney U-test. Variable distribution determined their description method: normal distribution utilized mean with standard deviation, while non-normal distribution employed median with interquartile range. Numbers and percentages expressed categorical variables, analyzed through Chi-square or Fisher’s exact test. Missingness for all pre-specified covariates was assessed and is presented in Supplementary Figure S1A. Variables with missing values were handled using multiple imputation by chained equations (MICE) under the assumption that data were missing at random. Ten imputed datasets were generated, and the results were pooled using Rubin’s rules. The association between NPAR and mortality rates (28-day, 180-day, and 365-day) in TS patients was examined through multivariate Cox proportional hazards regression analyses, calculating hazard ratios (HRs) (95% CIs). Variables with p values < 0.05 in univariate analysis were initially selected as potential predictors (Table S1). To address potential multicollinearity, variance inflation factor (VIF) was calculated for candidate variables, and those with VIF > 5 were not included in the final multivariable model (Table S2). The remaining significant variables were subsequently included in the multivariate Cox regression model. Model1 excluded confounding factors. Model 2 incorporated age and gender adjustments. Model 3 expanded upon Model 2’s adjustments, incorporating acute kidney injury (AKI), acute respiratory failure (ARF), cerebrovascular disease, Charlson Comorbidity Index, SOFA score, chronic obstructive pulmonary disease (COPD), ALT, ALP, potassium, PTT, sodium, glucocorticoid use, GCS score, ischemic heart disease (IHD) and myocardial infarction (MI). The Kaplan-Meier method assessed 28-day survival probability, evaluated through log-rank testing. RCS with two knots illustrated NPAR’s correlation with 28-day mortality, incorporating the aforementioned multivariate adjustments. The proportional hazards assumption of the Cox regression models was evaluated using Schoenfeld residual tests. Subgroup analyses and interaction tests were performed to assess whether the association between NPAR and 28-day mortality differed across key clinical subgroups, and P values for interaction were reported. The Receiver Operating Characteristic (ROC) curve and seven machine learning models demonstrated NPAR’s predictive capacity for 28-day mortality. To identify the most prognostically relevant variables and reduce model complexity, we employed the Boruta algorithm prior to model construction. Features that consistently showed greater importance than the top-performing shadow feature were identified as significant predictors. The “Boruta” package within the R environment was used to carry out this analysis. The variables identified by this process were subsequently used for developing all machine learning models. Seven distinct machine learning (ML) models were developed, including eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Random Forest (RF), Adaptive Boosting (AdaBoost), Gradient Boosting Decision Trees (GBDT), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). Hyperparameters for each algorithm were optimized using grid search within a stratified 10-fold cross-validation framework, and the parameter combination with the best cross-validated performance was selected. Model performance was summarized by the mean ± standard deviation of AUC, accuracy, sensitivity, specificity, F1 score, and Cohen’s kappa across the 10 folds. The predictive performance of the finally tuned models was thoroughly evaluated using AUC, Precision-Recall (PR) curves, calibration curves, and Decision Curve Analysis (DCA). To formally compare the discriminative performance among the seven models, pairwise DeLong tests were employed to determine whether there were statistically significant differences in their AUCs. To assess the predictive performance, the area under the curve (AUC) for predicted survival—along with the corresponding 95% CIs—was calculated. Cross-validated AUC distributions and other classification metrics were further summarized in the Supplementary Table S5. Results Baseline characteristics The investigation examined 201 TS cases during initial ICU stays procured from the MIMIC-IV 3.1 database, following specified inclusion and exclusion parameters. Patients were stratified into a survivor group (n = 121, 60.2%) and a non-survivor group (n = 80, 39.8%) per their outcome. Table 1 provides a detailed overview of the baseline demographic, clinical, and laboratory characteristics for both the entire patient population and these two subgroups. A comparison between the groups revealed that non-survivors were significantly older, had a lower body weight, and presented with a higher burden of comorbidities, as indicated by a greater Charlson Comorbidity Index and a higher prevalence of cancer and COPD. Analysis of laboratory parameters showed that non-survivors exhibited significantly lower levels of hemoglobin, red blood cells, total calcium, and albumin. In contrast, the non-survivor group exhibited significantly higher INR and NPAR levels versus the survivor group (all P < 0.05). Moreover, statistical analysis revealed no meaningful distinctions between the two groups regarding ICU or hospital stay duration, therapeutic measures such as glucocorticoid administration, vasopressor use, antibiotic therapy, and mechanical ventilation, as well as other clinical characteristics specified in Table 1. Table 1Comparison of baseline characteristics between survivors and non-survivorsCharacteristicsOverall (N = 201)Survivor (N = 121)Non-survivor (N = 80)p-valueAge (years)65.00 (53.00–75.00)62.00 (50.00–73.00)68.00 (61.50–75.50)< 0.001Weight (kg)68.90 (58.70–85.00)70.10 (59.80–88.95)66.90 (55.15–78.95)0.026Gender0.764 Female151.00 (75.12%)90.00 (74.38%)61.00 (76.25%) Male50.00 (24.88%)31.00 (25.62%)19.00 (23.75%)Heart rate96.00 (84.00–113.00)93.00 (83.00–113.00)97.50 (85.00–113.50)0.146Respiratory rate22.00 (17.00–26.00)21.00 (18.00–26.00)22.00 (17.00–27.00)0.677Temperature98.30 (97.70–99.10)98.40 (97.70–99.10)98.20 (97.70–98.75)0.482SpO2(%)98.00 (94.00–100.00)98.00 (94.00–100.00)98.00 (94.50–100.00)0.418RBC (K/uL)3.51 (3.07–4.17)3.62 (3.09–4.26)3.43 (3.03–3.90)0.032Hemoglobin (g/dL)10.60 (9.30–12.20)10.80 (9.80–12.50)10.25 (8.75–11.85)0.004MCH (pg)30.20 (28.80–31.70)30.60 (28.80–32.00)29.80 (28.70–31.40)0.284MCHC (%)32.60 (31.40–33.80)32.70 (31.60–34.10)32.50 (31.30–33.60)0.087MCV (fL)91.00 (87.00–97.00)91.00 (87.00–97.00)92.00 (88.00–97.00)0.919WBC (K/uL)13.60 (8.40–19.60)14.10 (8.30–19.80)12.85 (8.65–19.05)0.614NEU (%)80.70 (73.00–87.80)80.00 (73.40–86.00)82.00 (72.95–89.65)0.603Monocytes (%)5.80 (3.60–8.20)5.80 (3.60–8.00)6.20 (3.50–8.90)0.336Anion Gap (mEq/L)15.00 (12.00–18.00)15.00 (12.00–17.00)15.00 (12.00–19.00)0.209Total calcium (mmol/L)8.10 (7.40–8.70)8.20 (7.60–8.70)8.00 (7.25–8.55)0.038Chloride (mEq/L)103.00 (98.00–107.00)103.00 (99.00–108.00)102.00 (97.00–107.00)0.183Glucose (mg/dL)136.00 (111.00–177.00)134.00 (110.00–171.00)139.00 (114.50–190.50)0.112Potassium (mEq/L)4.00 (3.70–4.50)4.00 (3.60–4.40)4.00 (3.70–4.60)0.359Sodium (mEq/L)138.00 (135.00–141.00)138.00 (136.00–141.00)137.50 (134.00–141.00)0.159Magnesium (mEq/L)1.90 (1.70–2.10)1.90 (1.70–2.20)1.90 (1.65–2.10)0.064Bicarbonate (mEq/L)21.00 (18.00–24.00)21.00 (18.00–24.00)21.00 (17.00–24.00)0.841Lactate (mmol/L)1.90 (1.20–3.00)1.90 (1.20–3.00)1.90 (1.30–3.55)0.388PT (sec)14.30 (12.60–16.80)14.10 (12.30–16.10)14.95 (13.00–18.00)0.057PTT (sec)30.30 (27.80–37.90)30.20 (27.40–36.70)30.65 (27.95–42.05)0.989INR1.30 (1.10–1.50)1.30 (1.10–1.40)1.40 (1.20–1.65)0.024NPAR27.24 (22.79–33.77)25.59 (22.46–29.69)31.30 (24.57–36.44)< 0.001ALT (IU/L)31.00 (17.00–72.00)32.00 (19.00–67.00)28.00 (14.00–95.50)0.704AST (IU/L)52.00 (29.00–122.00)53.00 (31.00–115.00)47.00 (27.00–137.00)0.728Total bilirubin (mg/dL)0.60 (0.40–1.00)0.60 (0.40–1.00)0.60 (0.40–1.00)0.056Creatinine (mg/dL)0.90 (0.70–1.40)0.90 (0.70–1.30)0.90 (0.70–1.50)0.550BUN (mg/dL)18.00 (13.00–29.00)17.00 (13.00–28.00)21.50 (12.50–29.50)0.632ALP (IU/L)85.00 (64.00–125.00)80.00 (61.00–119.00)91.50 (73.00–158.50)0.346Albumin (g/dL)2.80 (2.50–3.30)3.00 (2.60–3.40)2.60 (2.30–3.10)< 0.001Platelet (K/uL)222.00 (148.00–307.00)220.00 (145.00–297.00)229.50 (161.00–321.50)0.268SOFA6.00 (3.00–9.00)6.00 (3.00–8.00)7.00 (3.00–10.00)0.107GCS15.00 (14.00–15.00)15.00 (14.00–15.00)15.00 (13.00–15.00)0.092Charlson index5.00 (3.00–7.00)4.00 (3.00–6.00)7.00 (5.00–8.00)< 0.001Cardiomyopathy17.00 (8.46%)13.00 (10.74%)4.00 (5.00%)0.152Infective endocarditis6.00 (2.99%)4.00 (3.31%)2.00 (2.50%)0.742ARF132.00 (65.67%)76.00 (62.81%)56.00 (70.00%)0.293Obesity25.00 (12.44%)18.00 (14.88%)7.00 (8.75%)0.198Diabetes mellitus39.00 (19.40%)21.00 (17.36%)18.00 (22.50%)0.367Hypertension59.00 (29.35%)35.00 (28.93%)24.00 (30.00%)0.870AKI108.00 (53.73%)62.00 (51.24%)46.00 (57.50%)0.384Cerebrovascular disease6.00 (2.99%)3.00 (2.48%)3.00 (3.75%)0.604CKD25.00 (12.44%)11.00 (9.09%)14.00 (17.50%)0.077Cancer41.00 (20.40%)18.00 (14.88%)23.00 (28.75%)0.017Chronic bronchitis16.00 (7.96%)7.00 (5.79%)9.00 (11.25%)0.161Heart failure124.00 (61.69%)79.00 (65.29%)45.00 (56.25%)0.197MI34.00 (16.92%)23.00 (19.01%)11.00 (13.75%)0.330IHD82.00 (40.80%)48.00 (39.67%)34.00 (42.50%)0.689COPD27.00 (13.43%)11.00 (9.09%)16.00 (20.00%)0.026Glucocorticoid use29.00 (14.43%)14.00 (11.57%)15.00 (18.75%)0.156Vasopressor85.00 (42.29%)50.00 (41.32%)35.00 (43.75%)0.733Antibiotics136.00 (67.66%)77.00 (63.64%)59.00 (73.75%)0.134Mechanical ventilation182.00 (90.55%)106.00 (87.60%)76.00 (95.00%)0.079LOS in hospital (days)16.57 (9.80–27.91)17.90 (10.42–27.54)14.95 (9.52–31.84)0.819LOS in ICU (days)5.53 (2.87–11.18)5.10 (2.84–11.27)5.82 (2.96–11.17)0.743Abbreviations: RBC Red Blood Cells, MCH Mean Corpuscular Hemoglobin, MCHC Mean Corpuscular Hemoglobin Concentration, MCV Mean Corpuscular Volume, WBC White Blood Cells, NEU Neutrophil percentage, PT Prothrombin Time, PTT Partial Thromboplastin Time, INR International Normalized Ratio, NPAR Neutrophil Percentage-to-Albumin Ratio, ALT Alanine Aminotransferase, AST Aspartate Aminotransferase, BUN Blood Urea Nitrogen, ALP Alkaline Phosphatase, SOFA Sequential Organ Failure Assessment, GCS Glasgow Coma Scale, ARF Acute Respiratory Failure, AKI Acute Kidney Injury, CKD Chronic Kidney Disease, MI Myocardial Infarction, IHD Ischemic Heart Disease, COPD Chronic Obstructive Pulmonary Disease, LOS Length of Stay, ICU Intensive Care Unit Comparison of baseline characteristics between survivors and non-survivors Abbreviations: RBC Red Blood Cells, MCH Mean Corpuscular Hemoglobin, MCHC Mean Corpuscular Hemoglobin Concentration, MCV Mean Corpuscular Volume, WBC White Blood Cells, NEU Neutrophil percentage, PT Prothrombin Time, PTT Partial Thromboplastin Time, INR International Normalized Ratio, NPAR Neutrophil Percentage-to-Albumin Ratio, ALT Alanine Aminotransferase, AST Aspartate Aminotransferase, BUN Blood Urea Nitrogen, ALP Alkaline Phosphatase, SOFA Sequential Organ Failure Assessment, GCS Glasgow Coma Scale, ARF Acute Respiratory Failure, AKI Acute Kidney Injury, CKD Chronic Kidney Disease, MI Myocardial Infarction, IHD Ischemic Heart Disease, COPD Chronic Obstructive Pulmonary Disease, LOS Length of Stay, ICU Intensive Care Unit NPAR and 28-day, 180-day, and 365-day mortality A ROC curve analysis was conducted to assess NPAR’s standalone ability to predict 28-day mortality, resulting in an AUC of 0.640 (95% CI: 0.534–0.746) (Fig. 2). The optimal NPAR threshold was determined to be 29.35, with this value yielding a sensitivity of 63.9% and a specificity of 65.6%. Using this cutoff, the patient cohort was divided into two groups: a low NPAR group (NPAR < 29.35, n = 119) and a high NPAR group (NPAR ≥ 29.35, n = 82). Kaplan-Meier survival analysis revealed that over time, patients in the high NPAR group had a markedly lower survival probability than those in the low NPAR group (log-rank P = 0.00037, Fig. 3). To further quantify this association, Cox proportional hazards models were developed to examine mortality risk at multiple time points (Table 2). For the primary outcome of 28-day mortality, the high NPAR group had a 241.5% elevated mortality risk (HR: 3.415, 95% CI: 1.664–7.007, P < 0.001) in the unadjusted model. This link was maintained in the partially adjusted model, with the risk increasing to 269% (HR: 3.69, 95% CI: 1.79–7.605, P < 0.001), and the risk further rose to 449.4% in the fully adjusted model (HR: 5.494, 95% CI: 1.939–12.19, P < 0.001). The marked link between higher NPAR and elevated mortality risk persisted for the long-term outcomes of 180 and 365 days in both unadjusted and multivariable-adjusted models. ROC curve analysis of NPAR for the prediction of 28-day mortality. The ROC curve was used to assess the discriminative performance of NPAR for 28-day mortality in patients with TS and to determine the optimal cut-off value of 29.35 KM Curve between patient groups stratified by the optimal NPAR cut-off (29.35). Patients were divided into low-NPAR and high-NPAR groups according to the ROC-derived cut-off value of 29.35. Survival differences between the two groups were assessed using the log-rank test Table 2The association between NPAR group and 28-day, 180-day, and 365-day mortalityExposureN (%)aModel 1Model 2Model 3HR95% CIP valueHR95% CIP valueHR95% CIP value28-day mortality Low NPAR group119 (59.2)refrefref High NPAR group82 (40.8)3.4151.664, 7.0070.0013.691.79, 7.605P < 0.00015.4941.939, 12.190.001180-day mortality Low NPAR group119 (59.2)refrefref High NPAR group82 (40.8)3.6041.866, 6.961P < 0.00013.861.99, 7.486P < 0.00015.1412.367, 11.166P < 0.0001365-day mortality Low NPAR group119 (59.2)refrefref High NPAR group82 (40.8)3.7481.947, 7.212P < 0.00014.0062.073, 7.741P < 0.00015.282.439, 11.429P < 0.0001arepresented the number and percentage of patients in each NPAR groupModel 1: UnadjustedModel 2: Adjusted for age and genderModel 3: Adjusted for age, gender, acute kidney injury (AKI), acute respiratory failure (ARF), Cerebrovascular disease, Charlson Index, SOFA, chronic obstructive pulmonary disease (COPD), ALT, ALP, potassium, PTT, sodium, glucocorticoid use, GCS scores, ischemic heart disease (IHD) and myocardial infarction (MI)HR Hazard Ratio, CI Confidence Interval The association between NPAR group and 28-day, 180-day, and 365-day mortality arepresented the number and percentage of patients in each NPAR group Model 1: Unadjusted Model 2: Adjusted for age and gender Model 3: Adjusted for age, gender, acute kidney injury (AKI), acute respiratory failure (ARF), Cerebrovascular disease, Charlson Index, SOFA, chronic obstructive pulmonary disease (COPD), ALT, ALP, potassium, PTT, sodium, glucocorticoid use, GCS scores, ischemic heart disease (IHD) and myocardial infarction (MI) HR Hazard Ratio, CI Confidence Interval To further explore the dose-response relationship between NPAR (considered as a continuous variable) and mortality risk, a restricted cubic spline (RCS) analysis was conducted. This analysis demonstrated a significant overall association between NPAR levels and the risk of 28-day mortality (P for Overall = 0.041). The test for non-linearity was not significant (P for Nonlinear = 0.194), indicating that the mortality risk increases in a linear relationship with elevated NPAR levels (Fig. 4). Additionally, prognostic models were developed in which NPAR was considered as a continuous variable. The findings confirmed its marked link to mortality at multiple time points (Table S4). The predictive accuracy and clinical utility of this continuous variable model were further evaluated by ROC and DCA curves (Figure S2, S3). Schoenfeld residual tests showed no significant violation of the proportional hazards assumption at the global model level (global P = 0.075), although the test for PTT was statistically significant, suggesting a possible time-varying effect of this covariate (Supplementary Figure S1B). RCS curve for the association between NPAR and the risk of 28-day mortality. The curve shows the hazard ratio (HR) for 28-day mortality across NPAR values, with NPAR = 29.35 as the reference. The shaded area represents the 95% confidence interval. The overall association was significant (P for overall = 0.041), whereas the nonlinear component was not statistically significant (P for nonlinear = 0.194) Subgroup and interaction analyses Subgroup analyses were further performed to evaluate whether the association between NPAR and 28-day mortality differed across key clinical subgroups. As shown in Supplementary Figure S1C, the association between higher NPAR and increased 28-day mortality remained generally consistent across subgroups defined by sex, AKI, ARF, IHD, and MI. No significant interactions were detected (P for interaction = 0.420 for sex, 0.615 for AKI, 0.525 for ARF, 0.407 for IHD, and 0.702 for MI), suggesting that the prognostic effect of NPAR was relatively stable across these clinical strata. Developing and comparing prognostic models based on machine learning The Boruta algorithm identified six key features as significant predictors of 28-day mortality: age, weight, Charlson Comorbidity Index, initial sodium, initial alkaline phosphatase, and NPAR (Fig. 5). Based on these features, all seven machine learning models demonstrated strong discriminative ability, good calibration, and positive clinical net benefit (Fig. 6, Figure S4). In stratified 10-fold cross-validation, the RF model achieved the highest AUC (0.771 ± 0.057) and also showed favorable overall classification performance, with an accuracy of 0.738 ± 0.066, sensitivity of 0.475 ± 0.075, specificity of 0.894 ± 0.033, F1-score of 0.574 ± 0.074, and Cohen’s kappa of 0.397 ± 0.119 (Fig. 6, Table S5). For the purpose of validating this result, pairwise DeLong tests were performed to make comparisons between the AUCs of the models. The results confirmed that while the RF model was numerically superior, its performance was not statistically different from the other six classifiers (P > 0.05 for all comparisons, Table S6, S7). Variable importance plot derived from the Boruta feature selection analysis. The boxplots present the distribution of variable importance scores across Boruta iterations. Confirmed important variables are shown in green, tentative variables in yellow, rejected variables in red, and shadow features in blue Internal validation summary of the seven machine-learning models based on stratified 10-fold cross-validation. A Mean cross-validated AUC values ± SD for XGBoost, LightGBM, Random Forest, AdaBoost, GBDT, SVM, and KNN. B Summary of the corresponding mean cross-validated classification metrics, including accuracy, sensitivity, specificity, and F1 score Discussion This retrospective cohort analysis examining 201 TS patients from the MIMIC-IV database represents the initial exploration of NPAR’s relationship with mortality in this specific group. Our key findings are threefold: (1) higher baseline NPAR is independently linked to a markedly elevated risk of 28-day, 180-day, and 365-day mortality in TS patients, with a nearly 5.5-fold higher risk of 28-day mortality in patients with NPAR above the exploratory ROC-derived cut-off value of 29.35 (compared with NPAR < 29.35) after full multivariable adjustment; (2) The association between NPAR (when treated as a continuous variable) and 28-day mortality exhibits a linear pattern: for every one-unit rise in NPAR, the risk of 28-day mortality increases by 7.7%; and (3) NPAR showed moderate discriminative ability for predicting 28-day mortality (AUC = 0.640) and emerges as one of six key features in machine learning models (with Random Forest model achieving the highest internally validated AUC of 0.771). These results collectively support NPAR as a simple, accessible, and clinically relevant prognostic marker for TS. The strong link between NPAR and mortality in TS is consistent with the pathophysiological roles of inflammation and nutritional imbalance in this condition. NPAR combines two complementary biomarkers: neutrophil percentage, which indicates acute inflammation, and albumin, a marker of nutritional status and systemic inflammatory response [16]. In the acute phase of TS, sympathetic stress triggers a surge in catecholamines, which activates neutrophils and promotes their migration to the myocardium [17]. Neutrophils release cytotoxic mediators (such as myeloperoxidase, matrix metalloproteinases) that damage myocardial tissue, impair LV contractility, and delay functional recovery [18]. Concurrently, inflammation suppresses hepatic albumin synthesis and increases vascular permeability, leading to hypoalbuminemia [19]. Albumin deficiency further exacerbates myocardial injury by reducing antioxidant capacity, disrupting vascular integrity, and impairing nutrient delivery to the heart [20]. Thus, a high NPAR reflects a “double blow” of unchecked inflammation and compromised systemic reserve—two factors that directly contribute to adverse outcomes in TS. The approximately linear dose-response relationship between NPAR and 28-day mortality (RCS: P for Overall = 0.041, P for Nonlinear = 0.194) is particularly notable, as it suggests that the association between NPAR and mortality is not solely dependent on categorization by a single threshold. Unlike binary markers (such as the presence or absence of hypoalbuminemia), our analyses of NPAR as both a categorical and a continuous variable showed consistent results, and the continuous-variable approach provided additional support for the robustness of the main Cox regression findings across a spectrum of values. In this study, our Cox regression results show that each 1-unit increase in NPAR is linked to a 7.7% higher 28-day mortality risk (fully adjusted model), further supporting the prognostic relevance of NPAR in TS. Accordingly, the cut-off value of 29.35 should be interpreted as an exploratory threshold for risk stratification rather than a definitive clinical boundary, while the continuous-variable analyses may better reflect the underlying gradient of mortality risk. ROC analyses and machine learning analyses provided complementary evidence regarding the prognostic role of NPAR. With an AUC of 0.640, NPAR alone performs at a level that is either comparable to, or even slightly better than, that of established multi‑parameter prognostic scoring systems evaluated in similar Takotsubo syndrome cohorts—such as the SOFA score and the Charlson Comorbidity Index [21–24]. When integrated into machine learning models (alongside age, weight, Charlson index, sodium, and alkaline phosphatase), NPAR contributed to the Random Forest model’s superior performance (AUC = 0.771), confirming its incremental prognostic value beyond clinical and laboratory variables. However, this finding should be interpreted cautiously, as the machine-learning analyses were exploratory and pairwise DeLong tests showed no statistically significant differences among the candidate models. This lack of statistical separation may reflect the relatively small sample size (n = 201) and underscores the need for external validation before any model can be considered clinically applicable. Our findings are consistent with studies investigating NPAR in other cardiovascular and critical care populations. For example, in a cohort of 3106 patients with coronary artery disease (CAD), NPAR ≥ 27.9 was independently linked to short- and long-term mortality [25]. Similarly, a recent MIMIC-IV-based study in critically ill patients with ischemic stroke reported that elevated NPAR was independently associated with increased 30-day, 90-day, and 1-year all-cause mortality, and that NPAR showed modestly better discriminatory ability than neutrophil percentage or albumin alone [26]. These studies, together with ours, suggest that NPAR is a broadly applicable prognostic marker across diseases characterized by inflammation-nutrition imbalance. In TS-specific research, most prognostic studies have focused on single biomarkers (such as troponin, BNP) or clinical factors (such as age, heart failure). A recent study of 231 TS patients found that NLR > 12 was associated with 1-year mortality, but NLR does not account for nutritional status [10]. Our study extends this work by showing that NPAR, which integrates inflammation and nutrition, highlights its added value in capturing the multifaceted pathophysiology of TS. The clinical utility of NPAR lies in its simplicity and accessibility. Unlike novel biomarkers (such as microRNAs, cytokines) that require specialized testing, NPAR is calculated from routine admission laboratory data and therefore may be readily incorporated into early risk assessment in patients with TS. However, the cut-off value of 29.35 identified in this study should be regarded as exploratory rather than a definitive clinical decision threshold, particularly in the absence of external validation. At present, NPAR may be most appropriately viewed as a convenient adjunct for identifying patients who may warrant closer clinical attention or intensified monitoring during the acute phase. More broadly, these findings raise the possibility that NPAR could contribute to future risk-guided management strategies; for example, high-NPAR patients may be candidates for enhanced surveillance (such as telemetry or more frequent echocardiographic assessment) and specific therapeutic interventions. For example, high-NPAR patients might benefit from anti-inflammatory therapies (low-dose corticosteroids, if appropriate) or nutritional support (albumin supplementation in severely hypoalbuminemic patients), though future interventional studies are needed to test these strategies. The limitations of our study should be acknowledged. First, this was a retrospective analysis based on the MIMIC-IV database, and the diagnosis of TS was identified using ICD-9 and ICD-10 codes rather than formal clinical adjudication, which may introduce misclassification bias. Second, the cut-off value of 29.35 was derived and evaluated within the same cohort and should therefore be considered exploratory rather than definitive. In addition, the lack of external validation may limit the generalizability of our findings. Future large-scale, multicenter prospective studies are needed to validate the association between NPAR and mortality in TS across diverse populations. Secondly, investigations into dynamic NPAR monitoring—such as daily measurements during hospitalization—should be conducted to ascertain whether changes in NPAR are correlated with clinical outcomes, including left ventricular (LV) function recovery and mortality. Thirdly, mechanistic research should explore the connections between NPAR and the pathophysiology of TS, such as myocardial inflammation, catecholamine levels, and LV strain. Fourthly, researchers need to assess the value of combining NPAR with other biomarkers like BNP and troponin to develop a comprehensive prognostic model for TS. Lastly, interventional trials should examine whether targeting the inflammation-nutrition imbalance—through anti-inflammatory or nutritional interventions—can reduce mortality in TS patients with high NPAR levels. Conclusions This investigation demonstrates a significant positive link between NPAR and mortality rates among TS patients. Individuals presenting elevated NPAR values showed heightened mortality across 28-day, 180-day, and 365-day periods, indicating a direct link between elevated NPAR and adverse outcomes in TS. These results suggest that NPAR measurements may function as essential prognostic markers in TS management, potentially informing therapeutic strategies and enhancing patient survival rates. Supplementary Information Supplementary Material 1. Figure S1. Methodological robustness checks for the fully adjusted Cox regression model. (A) Pre-specified covariates included in the fully adjusted model, grouped by clinical domain; percentages indicate variable-level missingness in the raw dataset before multiple imputation. (B) Schoenfeld residual test results for the proportional hazards assumption, including the global test. (C) Subgroup-specific adjusted hazard ratios (HRs) and 95% confidence intervals for the association between NPAR and 28-day mortality, together with the corresponding P values for interaction. Supplementary Material 2. Figure S2. ROC curve of the multivariable Cox regression model, using NPAR as a continuous variable, for predicting 28-day mortality. The ROC curve illustrates the discriminative performance of the multivariable Cox regression model for 28-day mortality in patients with TS. The area under the curve (AUC) was 0.869 (95% CI: 0.808–0.929). Supplementary Material 3. Figure S3. DCA curve for a multivariable Cox regression model predicting 28-day mortality using NPAR as a continuous variable. The DCA curve illustrates the net benefit of the multivariable Cox regression model across different threshold probabilities, compared with the treat-all and treat-none strategies. Supplementary Material 4. Figure S4. Performance comparison of the seven machine learning models. (A) Precision–recall (PR) curves for XGBoost, LightGBM, Random Forest, AdaBoost, GBDT, SVM, and KNN in predicting 28-day mortality. The average precision (AP) for each model is provided in the figure. (B) Calibration curves showing the agreement between mean predicted probabilities and observed event rates for the seven models. The diagonal dotted line denotes perfect calibration. (C) Decision curve analysis (DCA) comparing the mean net benefit of the seven models across different threshold probabilities. The red dotted line represents the treat-none strategy, and the black dashed line represents the treat-all strategy. Supplementary Material 5. Acknowledgements We express appreciation to the participants in this study for their support and cooperation. We additionally value the valuable insights and suggestions from the editors and reviewers. Clinical trial number Not applicable. Abbreviations Alkaline phosphatase Alanine aminotransferase Aspartate aminotransferase Area under the curve Beth Israel Deaconess Medical Center Body mass index Blood urea nitrogen C-reactive protein Glasgow Coma Scale Intensive care unit International Normalized Ratio Mean arterial pressure Mean corpuscular hemoglobin concentration Medical Information Mart for Intensive Care IV Neutrophil percentage to albumin ratio Prognostic nutritional index Prothrombin time Restricted cubic spline Takotsubo syndrome Sequential Organ Failure Assessment White blood cell Authors’ contributions CZP, KY and HC designed the study, evaluated data, and made contributions to the writing of this manuscript. YK, XD, LX and HSL evaluated data. CZP and HC availed supervision to the study and played a central function in drafting the manuscript. All authors have read and approved the final manuscript. Funding The National Natural Science Foundation of China (Grant No. 82100352), Tianjin Medical University General Hospital “New Century Talents” Cultivation Project (Grant No. XSJ-2021-02), Tianjin Medical University General Hospital Young and Middle-Aged Scientific Research Cultivation Fund (Grant No. 303078100501), and Tianjin Major BioTech Special Projects (Grant No.25ZXSWSY00530) funded this research. Data availability The datasets can be obtained upon request from the corresponding author. For our analyses, we utilized data screening codes supplied by the authors of the MIMIC-IV databases. Declarations Ethics approval and consent to participate The dataset utilized in this study was sourced from MIMIC-IV. To gain access to the database, we completed the CITI Program course on Human Research and Data or Specimens Only Research and obtained permission (Record ID: 57743676). Patient information within this database was anonymized, and ethical review and informed consent requirements were waived. All procedures were conducted in compliance with applicable guidelines and regulations. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Chongzhe Pei and Kuan Yang contributed equally to this work. Contributor Information Chongzhe Pei, Email: chongzhepei@163.com. Heng Cai, Email: caihengch@163.com. References Associated Data Supplementary Materials Supplementary Material 1. Figure S1. Methodological robustness checks for the fully adjusted Cox regression model. (A) Pre-specified covariates included in the fully adjusted model, grouped by clinical domain; percentages indicate variable-level missingness in the raw dataset before multiple imputation. (B) Schoenfeld residual test results for the proportional hazards assumption, including the global test. (C) Subgroup-specific adjusted hazard ratios (HRs) and 95% confidence intervals for the association between NPAR and 28-day mortality, together with the corresponding P values for interaction. Supplementary Material 2. Figure S2. ROC curve of the multivariable Cox regression model, using NPAR as a continuous variable, for predicting 28-day mortality. The ROC curve illustrates the discriminative performance of the multivariable Cox regression model for 28-day mortality in patients with TS. The area under the curve (AUC) was 0.869 (95% CI: 0.808–0.929). Supplementary Material 3. Figure S3. DCA curve for a multivariable Cox regression model predicting 28-day mortality using NPAR as a continuous variable. The DCA curve illustrates the net benefit of the multivariable Cox regression model across different threshold probabilities, compared with the treat-all and treat-none strategies. Supplementary Material 4. Figure S4. Performance comparison of the seven machine learning models. (A) Precision–recall (PR) curves for XGBoost, LightGBM, Random Forest, AdaBoost, GBDT, SVM, and KNN in predicting 28-day mortality. The average precision (AP) for each model is provided in the figure. (B) Calibration curves showing the agreement between mean predicted probabilities and observed event rates for the seven models. The diagonal dotted line denotes perfect calibration. (C) Decision curve analysis (DCA) comparing the mean net benefit of the seven models across different threshold probabilities. The red dotted line represents the treat-none strategy, and the black dashed line represents the treat-all strategy. Supplementary Material 5. Data Availability Statement The datasets can be obtained upon request from the corresponding author. For our analyses, we utilized data screening codes supplied by the authors of the MIMIC-IV databases.