The relationship between prognostic nutritional index and in-hospital mortality in patients with severe coronary heart disease: A multicenter retrospective cohort study and predictive model based on machine learning

Abstract

Background

The prognostic value of the Prognostic Nutritional Index (PNI) in critically ill patients with coronary heart disease (CHD) remains unclear. This multicenter retrospective cohort study aimed to investigate the association between PNI and in-hospital mortality in severe CHD patients and develop a machine learning-based predictive model.

Methods

This study is a multicenter retrospective cohort study, extracting data of adult critically ill patients with CHD who met the inclusion criteria from the MIMIC-IV database and the eICU-CRD database. Multivariate logistic regression and restricted cubic spline (RCS) analyses were used to evaluate the relationship between PNI and in-hospital mortality and 365-day mortality. The Boruta algorithm and LASSO regression were used to screen predictive features, six machine learning models were constructed, and SHapley Additive explanation (SHAP) was used to interpret the importance of model features.

Results

Low PNI was significantly associated with higher in-hospital mortality (OR=0.95, 95% CI 0.94–0.97). RCS revealed a linear negative relationship between PNI and mortality. The Random Forest model demonstrated superior predictive performance (AUC=0.846), significantly outperforming traditional scoring systems. SHAP analysis identified CRRT, Sepsis, Vasopressor, PNI, and BUN as key predictors. An online clinical calculator was developed based on these findings.

Conclusions

This study developed a machine learning-based calculator to predict in-hospital mortality in patients with CHD in the ICU (https://leongxj.shinyapps.io/chd-mortality-prediction/). This calculator may guide clinical decision-making and improve the treatment of patients with CHD by identifying those at higher risk of in-hospital death.

1. Introduction

Coronary heart disease (CHD) as one of the leading cardiovascular diseases worldwide, remains the primary cause of death and disability. According to the World Health Organization, CHD accounts for nearly one-third of total deaths globally, and with population aging and changes in lifestyle, its incidence and mortality rates are showing an increasing trend year by year.^1–3 Patients with severe CHD often have complex pathophysiological mechanisms and multiple comorbidities, with in-hospital mortality significantly higher than that of patients with ordinary CHD.^4,5 Traditional scoring systems and single biomarkers have limited predictive accuracy and are insufficient to meet the demands of personalized treatment.^6,7 Therefore, exploring new prognostic indicators and applying advanced data analysis methods are of great significance for improving outcomes in critically ill patients with CHD. In recent years, particularly ensemble learning and boosting-type machine learning methods have demonstrated tremendous potential in the early prediction and risk stratification of coronary heart disease, often surpassing traditional single-classifier models.^8,9

The Prognostic Nutritional Index (PNI) is a composite nutritional-immune indicator calculated based on serum albumin and lymphocyte count, and it has been widely used in the prognostic assessment of various diseases.^10,11 PNI not only reflects the nutritional status of patients but can also indirectly indicate immune function and systemic inflammation levels. Recent studies have shown that PNI is closely related to the prognosis of various malignant tumors and also demonstrates strong predictive value in cardiovascular diseases.^12,13 In patients with acute coronary syndrome (ACS) and heart failure, low PNI values are independently associated with higher all-cause mortality, rehospitalization rates, and adverse cardiovascular event incidences.^14,15 The mechanism may be related to malnutrition leading to decreased immune function, exacerbated inflammatory response, and reduced tissue repair capacity.^16,17 However, current research on the relationship between PNI and in-hospital mortality in critically ill CHD patients is still limited, and there is a lack of systematic analysis based on large-scale multicenter data.

Therefore, this study aims to systematically explore the association between PNI and in-hospital mortality in patients with severe CHD through a multicenter retrospective cohort study, and to develop and validate a practical predictive model based on machine learning algorithms for clinical application, to assist in accurately identifying high-risk patients and provide a reliable tool for clinical decision-making, ultimately improving the clinical outcomes of patients with severe CHD.

2. Method

2.1. Data sources

The data used in this study were obtained from the Medical Information Mart for Intensive Care IV (MIMIC-IV version 3.1) and the eICU Collaborative Research Database (eICU-CRD version 2.0). The MIMIC-IV database contains clinical information for all patients admitted to the ICUs at Beth Israel Deaconess Medical Center from 2008 to 2022, including patient demographics, vital signs, laboratory test results, medication records, and survival data. The eICU-CRD is a multicenter database covering over 200,000 ICU admissions in the United States between 2014 and 2015, including demographic records, bedside monitored physiological parameters, and other laboratory data. The authors of this study have completed the collaborative training project examination and obtained the corresponding certification (Record ID: 65401013). All data were anonymized to protect patient privacy, and the requirement for informed consent was waived. Data reporting followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.

2.2. Study design and population

This is a multicenter retrospective cohort study. The main study population included 3,672 patients from the MIMIC-IV database, with 6,919 patients from the eICU-CRD database serving as an external validation cohort. Inclusion criteria were: (1) adult patients during their first ICU admission, with only the first ICU admission data retained for patients with multiple ICU admissions; (2) patients diagnosed with CHD according to ICD-9 and ICD-10 codes. Exclusion criteria were: (1) patients under 18 years of age; (2) patients with an ICU stay shorter than 24 hours; (3) patients with missing lymphocyte count or serum albumin data. The study flowchart is shown in Figure 1.

Figure 1.

Flowchart of patient selection from the MIMIC-IV databases and the eICU-CRD. MIMIC-IV, Medical Information Mart for Intensive Care-IV; eICU-CRD, eICU Collaborative Research Database.

2.3. Data collection

PostgreSQL and Navicat Premium software were used to extract patient data that met the inclusion criteria. The extracted data included: 1. Demographic characteristics: age, gender, ethnicity, height, and weight; 2. Comorbidities: hypertension, diabetes, hyperlipidemia, chronic kidney disease, myocardial infarction, heart failure, chronic obstructive pulmonary disease (COPD), malignant tumors, AKI and sepsis. 3. Vital signs and disease scores at admission: heart rate (HR), respiratory rate (RR), body temperature, non-invasive systolic blood pressure (SBP), non-invasive diastolic blood pressure (DBP), non-invasive mean arterial pressure (MBP), oxygen saturation (SpO2), sequential organ failure score (SOFA), acute physiology score III (APSIII), SAPSII, OASIS; 4. First laboratory measurements within 24 hours of admission: lymphocyte count, albumin, hemoglobin, red blood cell count (RBC), white blood cell count (WBC), platelet count (PLT), red blood cell distribution width (RDW), hematocrit, glucose, lactate, potassium, sodium, calcium, chlorine, anion gap, prothrombin time (PT), partial thromboplastin time (APTT), international normalized ratio (INR), creatinine, urea nitrogen (BUN), total bilirubin, Aspartate aminotransferase (AST), alanine aminotransferase (ALT), creatine kinase isoenzyme (CK-MB), lactate dehydrogenase (LDH); 5. Medications and treatments during hospitalization: ACEIs/ARBs, B-blockers, P2Y12 receptor antagonists, statins, aspirin, vasoactive drugs, coronary artery bypass grafting (CABG), percutaneous coronary intervention (PTCA), mechanical ventilation, continuous renal replacement therapy (CRRT). 6. Prognostic data: ICU length of stay, mortality during hospitalization and 365-day mortality. In addition, variables with missing values ≥25% were excluded. For variables with missing values <25%, multiple imputation was performed using the ‘mice’ package in R. The missing data are shown in Supplementary Table 1.

2.4. Calculation of GV and endpoint events

The formula for PNI is: PNI = 10 * serum albumin (g/dL) + 0.005 * total lymphocyte count (μL). ¹⁸ The primary outcome of this study was all-cause in-hospital mortality, and the secondary outcome was 365-day mortality after admission.

2.5. Statistical analysis

The normality of continuous variables was assessed by the Kolmogorov-Smirnov test in this study. The normally distributed data were expressed as mean ± standard deviation, and the t-test was used for comparison between groups. The data for skewed distribution were expressed as median (interquartile range), and the Mann-Whitney U test was used for comparison between groups. Categorical variables were assessed by Pearson chi-square test or Fisher’s exact test and expressed as frequencies (%). The study population was divided into T1 (25.00 [22.00 - 27.00]), T2 (31.00 [29.01 - 32.00]), and T3 (36.01 [35.00 - 39.00]) according to the PNI triple score, and Kaplan-Meier curves were used to compare the differences in outcomes between the three groups. Taking the T1 group as a reference, multivariate logistic regression was used to analyze the relationship between PNI and in-hospital mortality and 365-day all-cause mortality, and the results were expressed as odds ratio (OR) and 95% confidence interval (CI). Multicollinearity of variables was assessed by variance expansion factor (VIF) and variables with VIF > 5 were excluded (Supplementary Table 2). Among them, Model 1 was not adjusted, Model 2 adjusted for age, gender and weight, Model 3 added adjustments for laboratory results and vital signs, and Model 4 made comprehensive adjustments, including age, gender, weight, laboratory results, vital signs, comorbidities and treatment measures. Secondly, restrictive cubic splines were used to evaluate the nonlinear relationship between PNI and in-hospital mortality and 365-day all-cause mortality. In addition, we performed subgroup analyses based on age (<65 years or ≥65 years), sex, and major comorbidities to assess consistency across strata. Statistical significance was defined as two-sided P<0.05. Data analysis was performed using R (version 4.5.1) and DecisionLinnc (version 2.0) software. DecisionLinnc provides a unified platform that supports multiple programming environments and enables data processing, advanced statistical modeling, and machine learning in an intuitive interface.

2.6. Construction and performance evaluation of machine learning models

The MIMIC-IV data were randomly divided into training and validation sets at a 7:3 ratio. ¹⁹ To preserve the incidence of natural events in the real population and ensure the calibration performance of the predicted absolute risk values, oversampling or weighting methods were not adopted in this study. In the training set, feature selection was performed using the Boruta algorithm and Least Absolute Shrinkage and Selection Operator (LASSO) regression. ²⁰ The Boruta algorithm ranks variables based on Z-scores derived from random forest importance, displays them as boxplots, and classifies them as confirmed, tentative, or unimportant. ²¹ Meanwhile, LASSO regression performs penalized selection by shrinking coefficients and setting non-informative ones to zero, removing redundant covariates while retaining the most influential predictors ²² ; this approach provides greater stability than univariate screening or stepwise selection methods. Finally, based on references and expert opinions, the intersection of features selected by LASSO regression and the Boruta algorithm was chosen as the feature variables for the primary outcomes of this study. Based on the above feature variables, prediction models were constructed using six machine learning algorithms, including Logistic Regression (LR), Random Forest (RF), Multi-Layer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (Light GBM), and Support Vector Machine (SVM). Hyperparameter tuning was performed using grid search with cross-validation, and performance was summarized using the area under the receiver operating characteristic curve (AUROC), accuracy, recall, F1 score, precision, and specificity. Precision-recall (PR) curves, calibration curves, and decision curve analysis were used to evaluate predictive behavior, calibration, and clinical utility. Then, the best-performing model underwent 10-fold cross-validation and external validation to assess its robustness and generalizability. Shapley additive explanations (SHAPs) were used to test model interpretability, quantifying each predictor’s contribution at both the overall and individual levels to enhance transparency in the prediction process. ²³ In addition, the data from eICU-CRD database was used to externally validate the optimal model. Ultimately, based on the core predictive factors and the best model, a web-based calculator was developed to facilitate the practical application of the model. The above workflow is shown in Supplementary Figure 1.

3. Results

3.1. Study population and baseline features

According to the inclusion and exclusion criteria, 3672 patients were finally included from the MIMIC-IV database, with a mean age of 73.00 (range 64.00 - 80.00) and 64.92% male. According to the tertiles of PNI at ICU admission, the investigators were divided into 3 groups: T1 (25.00 [22.00 - 27.00]), T2 (31.00 [29.01 - 32.00]), and T3 (36.01 [35.00 - 39.00]). Table 1 shows the baseline characteristics of these groups. Compared with the T3 group, the patients in the T1 group were older at admission, had higher disease severity scores, and had higher heart rate, white blood cells, blood urea nitrogen, creatinine, lactate levels, and coagulation indicators (APTT, PT, INR), while the patients in the T1 group had lower systolic blood pressure, diastolic blood pressure, mean blood pressure, albumin, red blood cells, hemoglobin and hematocrit, and calcium, while SpO2, body temperature, AST, ALT, sodium, glucose, and anion gap were no different between the two groups. Among the comorbidities, there were significant differences between groups in hypertension (34.10% in group T3), hyperlipidemia (60.51% in group T3), sepsis (74.84% in group T1), and acute kidney injury (AKI, 81.37% in group T1), while there were no significant differences between groups in diabetes, myocardial infarction, chronic obstructive pulmonary disease (COPD) and malignant tumors. In terms of treatment, the use of vasopressors and continuous renal replacement therapy (CRRT) was higher in the T1 group (75.82%/15.93%), and the use of chronic disease-related drugs (ACEI/ARB, β blockers, aspirin) and revascularization (CAGB/PTCA) was higher in the T3 group. Outcome measures showed that the T3 group had the upper hand, with the lowest in-hospital mortality (13.08%) and 365-day mortality (15.62%), while the T1 group had the highest (32.35% and 36.11%, respectively, both p<0.001).

Table 1.

Baseline characteristics according to PNI groups in MIMIC-IV databases.

Variable	N = 3,672	T1 = 1,224	T2 = 1,225	T3 = 1,223	p-value
Age(year)	73.00 (64.00 - 80.00)	73.00 (65.00 - 81.00)	74.00 (65.00 - 81.00)	71.00 (62.00 - 79.00)	<0.001
Weight(Kg)	80.00 (67.80 - 95.50)	77.30 (65.40 - 92.25)	80.30 (67.20 - 96.70)	82.40 (70.00 - 98.50)	<0.001
Gender, n (%)	​	​	​	​	0.010
Female	1,288.00 (35.08%)	466.00 (38.07%)	428.00 (34.94%)	394.00 (32.22%)	​
Male	2,384.00 (64.92%)	758.00 (61.93%)	797.00 (65.06%)	829.00 (67.78%)	​
SOFA score	6.00 (3.00 - 9.00)	7.00 (4.00 - 10.00)	6.00 (4.00 - 9.00)	5.00 (2.00 - 7.00)	<0.001
APSIII score	48.00 (36.00 - 62.00)	56.00 (43.50 - 72.00)	47.00 (36.00 - 61.00)	41.00 (30.00 - 53.00)	<0.001
SAPSII score	41.00 (33.00 - 51.00)	46.00 (37.00 - 55.00)	41.00 (34.00 - 51.00)	37.00 (29.00 - 45.00)	<0.001
OASIS score	33.00 (28.00 - 40.00)	36.00 (30.00 - 43.00)	34.00 (28.00 - 39.00)	31.00 (26.00 - 37.00)	<0.001
Vital Signs
HR (beats/min)	86.00 (75.00 - 101.00)	90.00 (77.00 - 106.00)	85.00 (74.00 - 100.00)	83.00 (74.00 - 96.00)	<0.001
SBP (mmHg)	117.00 (102.00 - 134.50)	114.00 (99.00 - 131.00)	118.00 (104.00 - 134.00)	121.00 (105.00 - 138.00)	<0.001
DBP (mmHg)	67.00 (57.00 - 79.00)	65.00 (55.00 - 78.00)	67.00 (57.00 - 78.00)	70.00 (60.00 - 82.00)	<0.001
MBP (mmHg)	81.00 (70.00 - 93.00)	78.00 (68.00 - 90.00)	81.00 (71.00 - 91.00)	83.00 (73.00 - 96.00)	<0.001
RR (beats/min)	19.00 (16.00 - 24.00)	20.00 (16.00 - 24.00)	19.00 (16.00 - 23.00)	19.00 (15.00 - 23.00)	<0.001
SPO2(%)	98.00 (95.00 - 100.00)	97.00 (94.00 - 100.00)	97.00 (95.00 - 100.00)	98.00 (95.00 - 100.00)	0.310
Temperature (°C)	98.20 (97.70 - 98.70)	98.20 (97.70 - 98.80)	98.10 (97.70 - 98.70)	98.20 (97.70 - 98.70)	0.235
Laboratory Indicators
ALC (10^9/L)	1.09 (0.64 - 1.74)	0.91 (0.50 - 1.42)	1.10 (0.63 - 1.73)	1.30 (0.81 - 2.03)	<0.001
Albumin (g/dL)	3.10 (2.70 - 3.50)	2.50 (2.20 - 2.70)	3.10 (2.90 - 3.20)	3.60 (3.50 - 3.90)	<0.001
WBC (10^9/L)	12.00 (8.60 - 16.80)	12.70 (8.75 - 18.35)	11.80 (8.50 - 16.60)	11.60 (8.50 - 15.50)	<0.001
RBC (10^9/L)	3.40 (2.85 - 4.04)	3.14 (2.71 - 3.70)	3.35 (2.81 - 3.95)	3.76 (3.17 - 4.39)	<0.001
RDW (%)	14.80 (13.50 - 16.60)	15.50 (14.30 - 17.20)	14.70 (13.60 - 16.50)	14.00 (13.20 - 15.70)	<0.001
PLT (10^9/L)	182.00 (128.00 - 245.00)	178.00 (118.00 - 254.00)	174.00 (123.00 - 239.00)	192.00 (142.00 - 244.00)	<0.001
Hemoglobin (mmol/L)	10.10 (8.40 - 11.90)	9.30 (7.90 - 10.90)	9.80 (8.40 - 11.50)	11.30 (9.30 - 13.00)	<0.001
Hematocrit (%)	31.30 (26.40 - 36.90)	29.00 (25.00 - 34.10)	30.60 (26.20 - 36.00)	34.70 (29.20 - 39.80)	<0.001
LDH (mg/dL)	317.00 (236.00 - 481.00)	327.00 (234.50 - 494.00)	319.00 (242.00 - 495.00)	304.00 (233.00 - 451.00)	0.016
BUN (mg/dL)	26.00 (17.00 - 44.00)	31.00 (18.00 - 51.00)	26.00 (18.00 - 45.00)	21.00 (15.00 - 35.00)	<0.001
Creatinine(mmol/L)	1.30 (0.90 - 2.10)	1.40 (0.90 - 2.40)	1.30 (0.90 - 2.20)	1.10 (0.90 - 1.70)	<0.001
AST (u/L)	40.50 (24.00 - 96.00)	43.00 (24.00 - 100.00)	42.00 (24.00 - 98.00)	38.00 (24.00 - 91.00)	0.192
ALT (u/L)	25.00 (15.00 - 57.00)	27.00 (14.00 - 62.50)	25.00 (15.00 - 65.00)	24.00 (15.00 - 50.00)	0.294
APTT(s)	31.90 (27.60 - 41.80)	32.25 (27.80 - 40.30)	31.30 (27.30 - 40.30)	32.00 (27.70 - 46.40)	0.020
PT(s)	14.50 (12.70 - 17.50)	15.10 (13.30 - 18.70)	14.90 (12.80 - 18.10)	13.70 (12.20 - 16.00)	<0.001
INR	1.30 (1.20 - 1.60)	1.40 (1.20 - 1.70)	1.40 (1.20 - 1.70)	1.30 (1.10 - 1.50)	<0.001
Lactate (mmol/L)	1.80 (1.20 - 2.70)	1.90 (1.30 - 3.00)	1.80 (1.30 - 2.70)	1.60 (1.20 - 2.40)	<0.001
Sodium (mEq/l)	138.00 (135.00 - 141.00)	138.00 (135.00 - 141.00)	138.00 (135.00 - 141.00)	138.00 (135.00 - 141.00)	0.774
Potassium (mEq/l)	4.30 (3.90 - 4.80)	4.25 (3.80 - 4.80)	4.30 (3.90 - 4.80)	4.30 (3.90 - 4.70)	0.010
Glucose (mg/dL)	134.00 (108.00 - 180.00)	134.50 (106.00 - 185.00)	135.00 (109.00 - 179.00)	133.00 (109.00 - 176.00)	0.773
Chloride (mEq/l)	103.00 (99.00 - 107.00)	103.00 (99.00 - 107.00)	103.00 (98.00 - 107.00)	103.00 (99.00 - 106.00)	0.009
Calcium (mEq/l)	8.40 (7.90 - 8.90)	8.05 (7.60 - 8.60)	8.30 (7.90 - 8.80)	8.70 (8.30 - 9.10)	<0.001
AG (mEq/l)	15.00 (12.00 - 18.00)	15.00 (12.00 - 18.00)	15.00 (12.00 - 18.00)	15.00 (12.00 - 18.00)	0.444
Comorbidities n (%)
Hypertension	1,052.00 (28.65%)	327.00 (26.72%)	308.00 (25.14%)	417.00 (34.10%)	<0.001
Diabetes	1,496.00 (40.74%)	490.00 (40.03%)	525.00 (42.86%)	481.00 (39.33%)	0.171
Hyperlipidemia	1,952.00 (53.16%)	547.00 (44.69%)	665.00 (54.29%)	740.00 (60.51%)	<0.001
CKD	1,218.00 (33.17%)	415.00 (33.91%)	432.00 (35.27%)	371.00 (30.34%)	0.028
Myocardial Infarction	1,571.00 (42.78%)	502.00 (41.01%)	512.00 (41.80%)	557.00 (45.54%)	0.053
Heart Failure	1,874.00 (51.03%)	590.00 (48.20%)	669.00 (54.61%)	615.00 (50.29%)	0.005
COPD	753.00 (20.51%)	273.00 (22.30%)	253.00 (20.65%)	227.00 (18.56%)	0.071
Malignant Tumor	594.00 (16.18%)	200.00 (16.34%)	214.00 (17.47%)	180.00 (14.72%)	0.178
Sepsis	2,262.00 (61.60%)	916.00 (74.84%)	746.00 (60.90%)	600.00 (49.06%)	<0.001
AKI	2,854.00 (77.72%)	996.00 (81.37%)	963.00 (78.61%)	895.00 (73.18%)	<0.001
Treatment (p%)
ACEI/ARB	1,395.00 (37.99%)	352.00 (28.76%)	464.00 (37.88%)	579.00 (47.34%)	<0.001
Vasopressor	2,583.00 (70.34%)	928.00 (75.82%)	879.00 (71.76%)	776.00 (63.45%)	<0.001
Beta Blockers	2,581.00 (70.29%)	793.00 (64.79%)	854.00 (69.71%)	934.00 (76.37%)	<0.001
P2Y12 Receptor Antagonist	267.00 (7.27%)	46.00 (3.76%)	84.00 (6.86%)	137.00 (11.20%)	<0.001
Statin	746.00 (20.32%)	216.00 (17.65%)	258.00 (21.06%)	272.00 (22.24%)	0.014
Aspirin	1,147.00 (31.24%)	235.00 (19.20%)	410.00 (33.47%)	502.00 (41.05%)	<0.001
CRRT	404.00 (11.00%)	195.00 (15.93%)	135.00 (11.02%)	74.00 (6.05%)	<0.001
Ventilation	3,037.00 (82.71%)	997.00 (81.45%)	1,042.00 (85.06%)	998.00 (81.60%)	0.028
CABG	653.00 (17.78%)	67.00 (5.47%)	247.00 (20.16%)	339.00 (27.72%)	<0.001
PTCA	212.00 (5.77%)	33.00 (2.70%)	63.00 (5.14%)	116.00 (9.48%)	<0.001
Outcomes
In-hospital mortality	805.00 (21.92%)	396.00 (32.35%)	249.00 (20.33%)	160.00 (13.08%)	<0.001
365-day mortality	925.00 (25.19%)	442.00 (36.11%)	292.00 (23.84%)	191.00 (15.62%)	<0.001

Abbreviations: MIMIC, Medical Information Mart for Intensive Care; SOFA, Sequential Organ Failure Assessment; APSIII, Acute Physiology Score System III; SAPSII, Simplified Acute Physiology Score II;OASIS, oxford acute severity of illness score; HR, heart rate; SBP, systolic blood pressure; DBP, diastolic blood pressure; MBP, Mean arterial pressure; RR, Respiratory rate;SPO2, Peripheral oxygen saturation; ALC, Absolute lymphocyte count; WBC, White blood cell count; RBC, red blood cell; RDW, red blood cell distribution width; PLT, platelet; LDH, Lactate dehydrogenase; BUN, blood urea nitrogen; AST, aspartate aminotransferase; ALT, alanine aminotransferase; APTT, Activated partial thromboplastin time; PT, Prothrombin time; INR, International normalized ratio; AG, Anion gap; CKD, Chronic kidney disease; COPD, Chronic obstructive pulmonary disease; AKI, Acute kidney injury; ACEI/ARB, angiotensin-converting enzyme inhibitors/Angiotensin II Receptor Blocker; CRRT, continuous renal replacement therapy; CABG, Coronary Artery Bypass Grafting; PTCA, Percutaneous Transluminal Coronary Angioplasty.

3.2. Association and dose–response relationships between PNI and in-hospital mortality

As shown in Figure 2, we compared in-hospital mortality and 356-day mortality in the three groups using Kaplan-Meier survival curves. The results showed that there were significant differences in mortality between the low, moderate, and high PNI groups. Specifically, the mortality rate in the low PNI group was significantly higher than that in the middle and high PNI group (P<0.001).

Figure 2.

Survival analysis curves for in-hospital and 365-day mortality of patients with CHD from MIMIC-IV. (A) In-hospital mortality. (B) 365-day mortality.

Subsequently, we constructed a multivariate logistic regression model using the T1 group as a reference, controlled for potential confounders through multi-model adjustment strategies, and analyzed the correlation between PNI and outcomes (Table 2). The results showed that PNI was significantly associated with lower in-hospital mortality (OR= 0.91∼95, P < 0.001) and lower 365-day mortality (OR = 0.92 ∼ 0.96, P < 0.001) when PNI was used as a continuous variable. In addition, the results of terrific analysis were consistent: T2 and T3 showed a lower risk of in-hospital death in all models compared with T1 (T2: OR=0.52∼0.73, P < 0.001 or < 0.01; T3: OR=0.31∼0.64, P < 0.001) and lower risk of death at 365 days (T2: OR=0.31∼0.44, P < 0.001 or < 0.05; T3: OR=0.33-0.69, P < 0.001 or < 0.01). In all models, the trend test across tertiles was positive (P < 0.001 or < 0.01), further underscoring the significant potential of PNI as a prognostic marker for patients with severe CHD. To delve into the relationship between mortality and PNI, we conducted an RCS analysis. As shown in Figure 3, there is a significant and nearly linear relationship between PNI and mortality risk: in-hospital mortality (overall p<0.001, nonlinear p-value = 0.045) and 365-day mortality (overall p-value <0.001, nonlinear p-value = 0.108). There was an L-shaped association between PNI and all-cause mortality.

Table 2.

Association of PNI with in-hospital and 365-day mortality in patients with CHD from MIMI-IV databases.

Variables	Model 1		Model 2		Model 3		Model 4
	OR (95%CI)	P-value	OR (95%CI)	P-value	OR (95%CI)	P-value	OR (95%CI)	P-value
In-hospital mortality
PNI Continuous	0.92 (0.90-0.93)	<0.001	0.91 (0.90-0.93)	<0.001	0.92 (0.91-0.94)	<0.001	0.95 (0.94-0.97)	<0.001
PNI Tertiles
T1	1(Ref)		1(Ref)		1(Ref)		1(Ref)
T2	0.53 (0.44-0.64)	<0.001	0.52 (0.43-0.63)	<0.001	0.60 (0.49-0.72)	<0.001	0.73 (0.59-0.91)	<0.01
T3	0.31 (0.26-0.39)	<0.001	0.32 (0.26-0.39)	<0.001	0.40 (0.32-0.51)	<0.001	0.64 (0.49-0.83)	<0.001
P for trend		<0.001		<0.001		<0.001		<0.001
365-day mortality
PNI Continuous	0.92 (0.91-0.93)	<0.001	0.92 (0.90-0.93)	<0.001	0.93 (0.91-0.94)	<0.001	0.96 (0.94-0.97)	<0.001
PNI Tertiles
T1	1(Ref)		1(Ref)		1(Ref)		1(Ref)
T2	0.55 (0.46-0.66)	<0.001	0.55 (0.46-0.65)	<0.001	0.63 (0.52-0.76)	<0.001	0.79 (0.64-0.97)	<0.05
T3	0.33 (0.27-0.40)	<0.001	0.33 (0.27-0.41)	<0.001	0.43 (0.35-0.54)	<0.001	0.69 (0.54-0.88)	<0.01
P for trend		<0.001		<0.001		<0.001		<0.01

Model 1: No adjustment.

Model 2: Adjusted for Age, Gender, and Weight.

Model 3: Adjusted for Age, Gender, Weight, LDH, BUN, Creatinine, INR, Lactate, Sodium, Potassium, Chloride, Calcium, WBC, RBC, RDW, PLT, Glucose, HR, MBP.

Model 4: Adjusted for Age, Gender, Weight, LDH, BUN, Creatinine, INR, Lactate, Sodium, Potassium, Chloride, Calcium, WBC, RBC, RDW, PLT, Glucose, HR, MBP, Hypertension, Chronic kidney disease, Hyperlipidemia, Heart Failure, CRRT, Ventilation, Sepsis, AKI, ACEI/ARB, Vasopressor, Beta Blockers, P2Y12 Receptor Antagonist, Statin, Aspirin, CABG, PTCA.

Figure 3.

Restricted cubic spline of the relationship between PNI and prognosis in patients with CHD in MIMIC-IV. (A) In-hospital mortality (B) 365-day mortality. Covariate adjustment was performed according to Model 4, and analysis was conducted using the likelihood ratio test.

3.3. Subgroup analysis

We conducted subgroup analyses based on age (<65 years and ≥65 years), gender, hypertension, chronic kidney disease (CKD), Tumor, diabetes, hyperlipidemia, heart failure (HF), myocardial infarction (MI), and chronic obstructive pulmonary disease (COPD). The results showed that none of these stratification factors significantly affected the relationship between PNI and in-hospital or 365-day mortality in patients with severe CHD, confirming the robustness of our findings (Figure 4).

Figure 4.

Forest plots of stratified analyses of PNI and prognosis of patients with CHD from MIMIC-IV. (A) In-hospital mortality (B) 365-day mortality. CKD, Chronic kidney disease; HF, Heart Failure; COPD, Chronic obstructive pulmonary disease.

3.4. Sensitivity analysis

We defined the lowest tertile of PNI (T1 group) as the low PNI group and the highest tertile of PNI (T3 group) as the high PNI group. Using the MatchIt package in R, 1:1 nearest neighbor matching with replacement was performed, with a caliper of 0.2 standard deviations of the log-transformed propensity score for propensity score matching. After matching, 609 matched patient pairs were retained, and the two matched cohorts were balanced in demographic characteristics, vital signs, laboratory results, major comorbidities, and treatment regimens (most SMDs < 0.1, supplementary table 3). The results showed that the in-hospital mortality rate (24.96% vs 17.57%) and 365-day mortality rate (28.24% vs 21.35%) in the low PNI group remained significantly higher than those in the high PNI group (p < 0.01).

3.5. Mediation analyses

Mediation analysis results showed that RDW and BUN partially mediated the relationship between PNI and mortality risk in patients with severe CHD. Specifically, RDW mediated 12.99% of in-hospital mortality and 15.3% of 365-day mortality; BUN mediated 8.03% of in-hospital mortality and 8.68% of 365-day mortality (Supplementary Figure 2).

4. Machine learning—PNI and in-hospital mortality

4.1. Selection of characteristic variable

Using the Boruta algorithm and LASSO regression to screen for predictors of in-hospital mortality in the study population, the Boruta algorithm identified 27 acceptable variables, including “CRRT,” “Vasopressor,” “PNI,” and “Beta Blockers” (Figure 5(a)). The LASSO regression path plot and cross-validation curve show the included variables (Figures 5(b) and 5(c)). The visualization of feature selection is shown in Figure 5(d), where variables shared by the red and blue lines represent the intersection of the two methods and the variables ultimately used in this study. Finally, by taking the intersection of the Boruta algorithm and LASSO regression results, 14 predictors were identified—CRRT, Vasopressor, PNI, Beta Blockers, BUN, CABG, Sepsis, Aspirin, AG, Age, Chloride, RDW, AKI, and ACEI/ARB.

Figure 5.

Feature selection. (A)The regression coefficient paths of each variable in Lasso regression; (B) The regression cross-validation curve of Lasso (C) Feature selection based on Boruta principle. (D)Venn diagram of the two screening results.

4.2. Model performance comparisons

We compared six machine learning models—Logistic Regression (LR), Random Forest (RF), Multilayer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (Light GBM), and Support Vector Machine (SVM) for predicting in-hospital mortality in patients with severe CHD. Table 3 shows the predictive performance of each model in the test set. Among them, RF achieved the highest AUC value (AUC=0.846) and the highest F1 score (F1=0.5943). Figure 6 shows the ROC curves, DCA curves, PR curves, and calibration curves of each model. A comprehensive comparison of the performance metrics of each model confirms that RF performed the best. In addition, we used ten-fold cross-validation to evaluate the predictive performance of the RF model. Its performance metrics again confirmed that the model we constructed has good robustness and generalization ability, with no signs of overfitting or underfitting (Supplementary Figure 3). Based on the above results, the RF model was identified as the optimal model in this study for further research.

Table 3.

Predictive performance of each model in the test datasets.

Model	AUC	Accuracy	Sensitivity	Recall	Specificity	F1-score
RF	0.8455	0.7659	0.7810	0.7810	0.7616	0.5943
MLP	0.8060	0.7250	0.7397	0.7397	0.7209	0.5416
XGBoost	0.8163	0.7296	0.7810	0.7810	0.7151	0.5592
LightGBM	0.8099	0.7132	0.7727	0.7727	0.6965	0.5420
SVM	0.8024	0.7232	0.8182	0.8182	0.6965	0.5649
LR	0.7607	0.7414	0.5826	0.5826	0.7860	0.4974

RF: Random Forest; MLP: Multilayer Perceptron; XGBoost: Extreme Gradient Boosting; LightGBM: Light Gradient Boosting Machine; SVM: Support Vector Machine; LR: Logistic Regression.

Figure 6.

Machine learning model construction and evaluation on the test set. (A) ROC curve plot. (B) PR curve plot. (C) DCA curve plot. (D) Calibration plot.

4.3. Comparison of the optimal model with traditional scoring systems

We compared the RF model with traditional scoring systems using the same dataset. Our results show that, compared with traditional scores (SOFA, OASIS, SAPS II, APS III), the RF model performs best in prediction, with an AUC value much higher than the other models (Table 4).

Table 4.

Comparison of the optimal model with traditional scoring systems.

Model	AUC	Accuracy	Sensitivity	Recall	Specificity	F1-score
SOFA	0.667	0.782	0.467	0.089	0.972	0.149
APSIII	0.721	0.775	0.432	0.148	0.947	0.220
SAPSII	0.693	0.784	0.491	0.110	0.969	0.179
OASIS	0.655	0.786	0.514	0.080	0.979	0.139
RF	0.846	0.766	0.781	0.781	0.762	0.594

RF: Random Forest.

4.4. external validation

We selected 6,919 patients with coronary heart disease from the eICU-CRD for external validation. In the eICU-CRD, the in-hospital mortality rate for these patients was 12.7%, lower than the 21.9% observed in the MIMIC database. Before conducting external validation, we applied the same data processing methods to the eICU-CRD data as we did for the MIMIC data. The validation results showed an AUC of 0.774 and a Brier score of 0.101 (Supplementary Figure 4). Additionally, the accuracy, sensitivity, specificity, and F1 score were 0.793, 0.490, 0.837 and 0.377. The difference in AUC between external validation and the validation/test sets was less than 0.1, leading us to conclude that the RF model demonstrates good stability.

We further evaluated the calibration performance of the best model in the internal test set and the external eICU-CRD cohort using the calibration slope, calibration intercept, and Hosmer-Lemeshow test. In the internal cohort, the calibration slope of the RF model was 1.246, the intercept was 0.195, and the Brier score was 0.125; in the external eICU-CRD cohort, the model’s calibration slope was 1.080 and the intercept was -0.206. The Hosmer-Lemeshow test showed no significant differences in either cohort (internal: χ²=8.359, p=0.399; external: χ²=4.983, p=0.759). The above calibration results further validate the RF model’s excellent predictive performance and good generalization ability.

4.5. Model visualization based on SHAP principle

We utilized the SHAP (Shapley Additive Explanations) principle to visualize and explain the contributions of each predictive factor in the RF model. The SHAP feature importance ranking and summary plot show the relative contribution of each feature to the overall model predictions, while the SHAP force plot demonstrates the contribution of these factors to specific individuals. The SHAP-based feature importance ranking plot (Figure 7(a)) and summary plot (Figure 7(b)) display the global contributions of feature variables in the RF model, with the x-axis representing SHAP values, indicating the contribution of each feature to the model’s predictions. The y-axis ranks the features based on the cumulative impact of their SHAP values. Our results indicate that CRRT, Sepsis, Vasopressor, PNI, and BUN are the five most important features. The SHAP force plot highlights the direction and magnitude of each feature’s effect on the prediction for specific critical coronary heart disease patients based on the RF model, with red indicating a positive impact and blue indicating a negative impact. Figure 7(c) presents a high-risk individual sample: the use of CRRT, lower PNI (23.0695), and higher RDW (24.1) are the main factors driving increased risk; Figure 7(d) presents a low-risk individual sample: lower RDW (12.9), no use of CRRT, and use of Aspirin are the main factors driving reduced risk. Finally, we developed a web-based calculator using the selected feature variables to facilitate individualized prognostic risk assessment for critical CHD, improving accessibility and convenience (https://leongxj.shinyapps.io/chd-mortality-prediction/).

Figure 7.

Model interpretation and feature importance. (A) SHAP Importance Plot; (B) SHAP Bees Plot; (C and D) SHAP Heat Force P.lot.

6. Discussion

This study is based on large intensive care databases (MIMIC-IV and eICU-CRD), systematically explored for the first time the association between PNI and in-hospital and 365-day mortality in critically ill patients with CHD, and successfully developed and validated a machine learning-based clinical prediction model. Our main findings are: (1) PNI is an independent protective factor against mortality in critically ill patients with CHD, showing a significant negative correlation and an approximately linear L-shaped dose-response relationship with mortality; (2) this association remains robust across different clinical subgroups and is confirmed through PSM; (3) mediation analysis revealed that RDW and BUN partially mediate the relationship between PNI and mortality risk; (4) among various machine learning models, the RF model demonstrated the best predictive performance, significantly outperforming traditional critical illness scoring systems and was effectively validated in the eICU-CRD cohort; (5) SHAP-based interpretability analysis confirmed that PNI is the fourth most important feature influencing model predictions, highlighting its significant clinical value. Collectively, these findings indicate that PNI is not only a strong prognostic biomarker but can also be effectively integrated into modern clinical prediction tools, offering new insights for risk stratification and individualized management of critically ill patients with CHD.

Malnutrition and systemic inflammation are key drivers of the occurrence, progression, and poor prognosis of coronary heart disease.^24,25 PNI was originally proposed by Onodera and colleagues to assess the nutritional status and surgical risk of patients undergoing gastrointestinal surgery. ²⁶ Importantly, PNI is not merely a nutritional variable, because it incorporates albumin and lymphocyte count, it reflects the integrated effects of malnutrition, systemic inflammation, immune dysfunction, catabolic stress, and impaired physiological reserve. These mechanisms are relevant in both acute and chronic cardiovascular disease. In the short term, low PNI may identify patients with greater perioperative vulnerability, inflammatory activation, infection susceptibility, impaired tissue repair, and organ dysfunction.^27,28 In the longer term, persistent nutritional–inflammatory imbalance may contribute to myocardial remodeling, worsening heart failure, arrhythmic vulnerability, recurrent hospitalization, and mortality.^29,30 Furthermore, the relevance of PNI may extend to acute infectious and inflammatory conditions such as COVID-19, where inflammation, lymphopenia, hypoalbuminemia, endothelial injury, and cardiovascular risk converge. ³¹ The results of this study indicate that low PNI is an independent prognostic risk factor for patients with severe coronary heart disease, and this association remains stable after multivariable adjustment, subgroup analysis, and PSM matching. This is consistent with previous findings in cardiovascular and other critical illness populations.^32,33

The pathophysiological mechanism of PNI affecting poor prognosis may involve multiple levels. First of all, serum albumin is not only a marker of nutritional status, but also an important negative reactive protein in the acute phase. ³⁴ Hypoalbuminemia tends to indicate a more severe systemic inflammatory response and endothelial dysfunction. ³⁵ In severe coronary heart disease, the body is at high metabolic and oxidative stress levels, and albumin protects the cardiovascular system through its antioxidant, antithrombotic, and colloidal osmotic maintenance effects.^36,37 Second, lymphocyte counts reflect the body’s immune response, and lymphopenia suggests immune exhaustion, which is associated with stress hormone activation, increased plaque instability, and increased susceptibility to infection.^38,39 In acute coronary syndrome or heart failure, elevated cortisol levels and activation of the neuroendocrine system lead to increased lymphocyte apoptosis and the production of stress lymphocytopenia, which is closely related to the degree of myocardial injury and poor remodeling.^40–42 PNI combines the two, enabling a comprehensive assessment of a patient’s physiological reserves and compensatory capacity under severe stress. Low PNI indicates an imbalance in inflammatory and immune function, a condition closely associated with sarcopenia, impaired immune function, endothelial dysfunction, increased plaque instability, and reduced tissue repair capacity, thereby leading to poor prognosis.^43,44 The mediated analysis of this study further found that RDW and BUN partially mediated the relationship between PNI and mortality risk. RDW is a measure of red blood cell volume variability, and its elevation is often associated with chronic inflammation, oxidative stress, and impaired bone marrow function. ⁴⁵ As an indicator of renal function and protein catabolism, elevated levels of BUN indicate renal hypoperfusion or high catabolic status. ⁴⁶ This suggests that the nutritional inflammatory state represented by low PNI further leads to a worsening of patient outcomes in part by exacerbating erythropoiesis disorders and promoting high catabolism/impaired renal function.^47,48 The discovery of mediating effects provides a preliminary mechanistic explanation for the role of the “nutrition-inflammation-organ function” axis in the prognosis of severe coronary heart disease.

Traditional clinical prediction models are mostly based on logistic regression, making it difficult to capture nonlinear relationships and complex interactions between variables. In recent years, machine learning has been successfully applied multiple times to cardiovascular risk prediction and other clinical risk stratification tasks, demonstrating its unique value in capturing complex nonlinear relationships in heterogeneous clinical data.^8,9,49 We robustly identified 14 core predictive factors from numerous variables using the Boruta and LASSO feature selection methods, including PNI. Among the six machine learning models compared, the RF model demonstrated the best and most stable performance on both the test set and the external validation set (test set AUC = 0.846, external validation set AUC = 0.774), with predictive performance significantly surpassing traditional critical illness scoring systems such as SOFA and SAPS II. This finding aligns with the recent trend of using machine learning to enhance predictive accuracy in the field of critical care.^50,51

Through interpretable analysis of the model using the SHAP framework, the key features driving the predictions were further revealed. Among them, CRRT, Sepsis, Vasopressor, PNI, and BUN were identified as the five most important predictive factors, highly consistent with clinical knowledge. The application of CRRT indicates that acute kidney injury has reached a severity requiring renal replacement therapy. Such patients often have volume overload, metabolic acidosis, and accumulation of uremic toxins, which can worsen cardiac load, suppress myocardial function, and thereby trigger a vicious cycle of cardiorenal syndrome. ⁵² Sepsis induces vascular endothelial dysfunction, myocardial depression, and microcirculatory disorders through uncontrolled inflammatory responses, and when coexisting with severe coronary heart disease, it exerts a synergistic damaging effect, significantly increasing the risk of death. ⁵³ The degree of dependence on vasoactive drugs directly reflects the severity of circulatory failure. Although high doses can barely maintain perfusion, they may adversely affect prognosis by increasing myocardial oxygen consumption and afterload. ⁵⁴ PNI as a core indicator integrating nutritional and immune status, its decrease suggests a nutritional-inflammatory imbalance and is closely related to sarcopenia, impaired immune function, and reduced tissue repair ability.^43,55 BUN as a comprehensive prognostic indicator that integrates hemodynamic status with systemic metabolic load, reflects the severity of the condition when elevated, from the two aspects of renal hypoperfusion and high catabolic metabolism. BUN as a prognostic indicator integrating hemodynamic status and systemic metabolic load, reflects the severity of the condition along two dimensions when elevated: inadequate renal perfusion and increased catabolic metabolism. ⁵⁶ The identification of these five core factors not only validates the reliability of the model but also provides clear targets for early clinical identification of high-risk patients and implementation of targeted interventions. Finally, we developed an online calculator, which can predict the risk of critically ill CHD patients in the ICU by entering the corresponding clinical data.

However, this study also has several limitations. First, although the retrospective cohort study used various statistical methods to control bias, it cannot entirely rule out the impact of residual confounding factors. Second, our data comes from a US ICU database, and the patient population and treatment protocols may differ from those in other regions of the world, so the applicability of the model to a broader population needs further validation. Third, this study mainly focused on data within the first 24 hours after ICU admission, and did not dynamically track how changes in PNI or other indicators over time affect prognosis. Finally, while machine learning models have certain advantages in predictive accuracy, challenges such as usability, integration into workflows, and prospective validation still need to be addressed for clinical implementation and dissemination. Future research should aim to conduct prospective cohort studies to verify our findings, explore the prognostic value of dynamic PNI monitoring, and carry out implementation studies of this predictive model in real clinical settings, with the ultimate goal of improving clinical outcomes in this high-risk population of critically ill patients with CHD.

7. Conclusion

In summary, this study conclusively confirmed through a large-scale multicenter retrospective cohort study that a low PNI is an independent risk factor for both in-hospital and long-term mortality in patients with severe CHD. We further developed and validated an RF machine learning prediction model that incorporates PNI, and used the SHAP interpretability method to explain the model. Finally, based on the RF prediction model, a web-based application was developed to provide strong support for the assessment and treatment by clinical healthcare professionals.

ORCID iDs

Caiping Tang https://orcid.org/0009-0004-7063-4647

Yulai Wu https://orcid.org/0009-0007-8782-0667

Lin Xie https://orcid.org/0009-0009-3438-7163

Ziyu Lin https://orcid.org/0009-0009-1713-0593

Lihua Wang https://orcid.org/0009-0009-6757-4926

Zhenyi Yan https://orcid.org/0000-0001-7152-9428

Xiaoyan Liu https://orcid.org/0009-0009-3624-0878

Xijian Liang https://orcid.org/0000-0002-4730-6280

Zhiyi Li https://orcid.org/0009-0002-0159-1896

Ethical considerations

This retrospective study analyzed data from the MIMIC-IV and eICU-CRD databases. Both datasets contain fully de-identified clinical records that comply with HIPAA standards. MIMIC-IV was approved by the Institutional Review Boards(IBR) of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center; eICU-CRD was approved by the MIT IRB. The authors have completed CITI training and signed data use agreements.

Consent to participate

This study is a secondary analysis of publicly available de-identified data and does not require obtaining individual informed consent.

Data Availability Statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found at: https://mimic.mit.edu/; https://physionet.org/content/eicu-crd/2.0/.*