Albumin and Machine Learning Model for Predicting 28-day All-cause Mortality in ICU Patients with Lung Cancer

Summary

This protocol details the extraction of ICU clinical data from the MIMIC-IV database and the stepwise application of survival analysis and interpretable machine-learning models to evaluate serum albumin as a predictor of 28-day mortality in lung cancer patients.

Abstract

Serum albumin reflects nutritional status, systemic inflammation, and disease burden. However, its prognostic significance in critically ill patients with lung cancer remains unclear. This study aimed to investigate the association between serum albumin and 28-day all-cause mortality in intensive care unit (ICU) patients with lung cancer and to assess its predictive value in machine learning (ML) models. This retrospective cohort study included 1,274 adult ICU patients with lung cancer from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. Multivariable Cox proportional hazards models evaluated the association between serum albumin and 28-day mortality. Additional analyses included restricted cubic splines, Kaplan–Meier survival curves, and subgroup analyses. ML classifiers were developed using the least absolute shrinkage and selection operator (LASSO) and Boruta for feature selection. Logistic regression was selected as the optimal model, with SHapley Additive Explanations (SHAP) values used for interpretation. Among 1,274 patients, the 28-day mortality rate was 26.9%. Lower serum albumin levels were independently associated with higher 28-day mortality. The hazard ratio for each 1 g/dL increase in serum albumin was 0.729 (95% CI, 0.597–0.891; p = 0.002). An inverse association was observed across serum albumin quartiles (p for trend <0.001), and Kaplan–Meier analysis showed lower mortality among patients with albumin ≥2.9 g/dL (log-rank p < 0.0001). The logistic regression model achieved an AUC of 0.767 in the validation cohort, indicating good discrimination and calibration. SHAP analysis identified serum albumin as a key predictor of 28-day mortality. Lower serum albumin was independently associated with increased short-term mortality in ICU patients with lung cancer. Serum albumin may serve as a practical biomarker for early risk stratification in this population. ML models incorporating albumin demonstrated strong predictive performance and interpretability.

Introduction

Lung cancer continues to rank among the most common and deadliest forms of malignancy worldwide. As reported by the Global Cancer Observatory (GLOBOCAN 2020), lung cancer represents around 11.4% of all newly identified cancer cases and remains the primary cause of cancer-related deaths, contributing to roughly 18% of cancer fatalities worldwide¹. Although significant progress has been made in early diagnostic techniques, chemotherapeutic regimens, and targeted treatment approaches, the overall 5-year survival rate for lung cancer persists below 20%, predominantly owing to its late-stage presentation and highly aggressive nature².

A significant number of individuals diagnosed with lung cancer necessitate intensive care unit (ICU) admission throughout their disease trajectory, most often as a result of complications like acute respiratory distress, severe infections, or hemorrhage, which may arise from either cancer progression or adverse effects associated with treatment. Among this patient group, ICU admission is strongly linked to poor short-term prognoses, with reported 28-day mortality rates ranging between 30% and 60%, substantially surpassing those observed in the broader ICU population³. In prior ICU cohorts of patients with lung cancer, comorbid conditions are highly prevalent, with hypertension and diabetes reported in approximately 46.1% and 37.9% of patients, respectively. In addition, nearly half of these patients present with moderate-to-severe baseline functional disability (45.6%), which may complicate ICU management and contribute to poor short-term outcomes⁴. The care of critically ill lung cancer patients within the ICU environment presents significant therapeutic complexities. Clinical decision-making for these patients frequently requires intricate evaluations regarding the initiation of mechanical ventilation, administration of vasopressors, and deployment of broad-spectrum antibiotics, all of which must be considered within the framework of an underlying cancer that typically exhibits limited potential for reversal. Several variables underlie their heightened vulnerability to early mortality, such as extensive tumor progression, persistent systemic inflammatory responses, compromised organ functionality, and nutritional deficiencies⁵. Hypoalbuminemia is frequently identified as one of the key clinical factors associated with early mortality in this patient population. Serum albumin serves as an indicator that reflects both chronic malnutrition resulting from cancer-associated cachexia and acute systemic inflammatory responses during critical illness, potentially providing important prognostic information⁶. Although serum albumin is routinely measured in clinical practice, its prognostic value has not been thoroughly investigated in lung cancer patients admitted to the ICU, and it is largely absent from current risk prediction tools⁷. Considering the elevated short-term mortality and the complexity of management in this clinical context, there is a pressing need for easily obtainable and practical biomarkers to facilitate early risk assessment and inform treatment decisions. Against this backdrop, serum albumin merits additional research to clarify its prognostic significance in critically ill individuals with lung cancer^8,9,10.

Due to the complex and interrelated characteristics of cancer-associated risk factors, conventional statistical approaches frequently face challenges in effectively modeling the intrinsic nonlinear associations embedded within high-dimensional clinical datasets. The advent of machine learning (ML) methodologies has introduced novel avenues for improving cancer risk prediction. Techniques including random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM) have gained widespread application in biomedical investigations, exhibiting strong capabilities in managing intricate data patterns and delivering superior predictive accuracy^11,12. Nevertheless, the opaque or black-box characteristics of such models restrict their interpretability and hinder their integration into clinical settings. To overcome this drawback, SHapley Additive Explanations (SHAP), a method grounded in game theory, has been introduced as a powerful tool to improve model transparency¹³. Through the computation of Shapley values for each feature, SHAP effectively quantifies the impact of specific variables on model outputs, thereby enhancing both interpretability and reliability¹⁴. This approach has seen growing adoption within biomedical research as a means to promote the clinical applicability of ML-based models.

Considering these factors, this study was designed to thoroughly explore the prognostic significance of serum albumin among critically ill individuals diagnosed with lung cancer, utilizing clinical information derived from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. To achieve this objective, we utilized both conventional statistical analyses and ML methods. To our knowledge, few studies have examined serum albumin as both an independent prognostic factor and a clinically interpretable contributor within machine-learning models, specifically in ICU patients with lung cancer¹⁵. Most existing studies have focused on either general ICU populations or non–ICU oncology cohorts, while machine-learning-based–based prognostic models in critically ill cancer patients have largely emphasized predictive performance without explicitly assessing the clinical contribution of individual biomarkers¹⁶.

In this context, the present study integrates conventional survival analysis with machine-learning approaches and SHAP-based interpretation to clarify the association between serum albumin and 28-day all-cause mortality and to quantify its relative importance within an interpretable prediction framework tailored to this high-risk population. Initially, multivariable Cox proportional hazards models were applied to evaluate the independent relationship between serum albumin levels and 28-day all-cause mortality, with adjustments for relevant confounders. Concurrently, predictive models were constructed using ML algorithms to identify major prognostic determinants and to examine the specific contribution of serum albumin to the overall predictive capacity. Particular focus was directed toward improving model interpretability by employing SHAP, allowing for the quantification of serum albumin’s relative impact compared with other clinical features. By incorporating serum albumin alongside routinely collected clinical indicators, this investigation seeks to establish a robust and interpretable risk stratification tool to enhance early clinical decision-making and enable individualized care for this high-risk cohort.

Protocol

Database access and software environment
Obtain authorized access to the Medical Information Mart for Intensive Care IV database (MIMIC-IV, v3.1) after completing the required training on ethical data use (Certification ID: 69002811). According to the Measures for Ethical Review of Life Science and Medical Research Involving Human Subjects (February 18, 2023, China), Article 32 stipulates that research using legally obtained public data, data generated without interfering with public behavior, or anonymized information is exempt from ethics review. Consistent with these provisions, this study was exempt from institutional ethical approval. The data use protocol and ethical guidelines were strictly followed, and the study was conducted solely for scientific research purposes. Set up the database environment using PostgreSQL (v17.1.11). Use Navicat Premium (v17) as the database management and query interface to run SQL queries and export extracted datasets for analysis.

Patient selection
Identify patients with lung cancer in MIMIC-IV using International Classification of Diseases, Ninth and Tenth Revision (ICD-9 and ICD-10) codes¹⁷. Apply the following exclusion criteria sequentially: Age <18 years or >90 years; Missing serum albumin measurements within the first 24 h of ICU admission; Multiple ICU or hospital admissions; ICU length of stay <24 h.

Variable extraction
Extract variables from MIMIC-IV (v3.1) using PostgreSQL (v17.1.11) via Navicat Premium (v17), including: demographics, vital signs, laboratory tests, comorbidities, therapeutic interventions, severity scores, and outcomes. Demographics: age, gender. Vital signs within the first 24 h after ICU admission: heart rate (HR), respiratory rate (RR), oxygen saturation (SpO₂), and temperature. Laboratory variables within the first 24 h: serum albumin, glucose, creatinine, urea nitrogen, total bilirubin, hemoglobin, red blood cell (RBC) count, white blood cell (WBC) count, platelet count, red cell distribution width (RDW), alanine aminotransferase (ALT), aspartate aminotransferase (AST), prothrombin time (PT), calcium, potassium, sodium, chloride. Comorbidities from diagnostic records: hypertension, hepatic cirrhosis, chronic kidney disease, diabetes mellitus, chronic bronchitis, congestive heart failure, and chronic obstructive pulmonary disease (COPD). Severity scores at ICU admission: Sequential Organ Failure Assessment (SOFA), Simplified Acute Physiology Score II (SAPS II), Charlson Comorbidity Index, Acute Physiology and Chronic Health Evaluation II (APACHE II). Therapeutic interventions: systemic corticosteroids, antibiotics, vasopressors. Lifestyle factor: smoking history. For all laboratory variables and severity scores within the first 24 h, if multiple measurements are available, compute the average value across the 1^st day and use this value for analysis. Exclude variables with more than 20% missing data. For variables with missingness ≤20%, perform imputation using the missForest algorithm, a random forest–based machine-learning method for handling mixed-type clinical data. The proportion of missing data for key variables is provided in Supplementary Table 1.

Outcome definition
Define the primary outcome as all-cause mortality within 28 days after ICU admission. Ascertain mortality status using death records available in the MIMIC-IV database. Classify patients as survivors or non-survivors based on 28-day mortality status.

Statistical analysis
Assess the distribution of continuous variables using the Shapiro–Wilk test. Compare normally distributed variables using Student’s t-test and non-normally distributed variables using the Wilcoxon rank-sum test. Compare categorical variables using the chi-square test. Perform least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation to select candidate variables and reduce multicollinearity. Construct multivariable Cox proportional hazards models to evaluate the independent association between serum albumin and 28-day mortality. Serum albumin was entered as both a continuous and categorical variable, and covariates selected by LASSO regression were included for adjustment. Hazard ratios (HRs) and 95% confidence intervals (CIs) were estimated. Explore nonlinear associations using restricted cubic spline (RCS) analysis. Knots were placed at predefined percentiles of the albumin distribution, and nonlinearity was assessed by testing the spline terms. Generate Kaplan–Meier survival curves and compare survival distributions using the log-rank test.

Machine learning model development and validation
Randomly split the dataset at the patient level into a training set (70%), a validation set (15%), and an independent test set (15%). Perform feature selection using the training set only by identifying the intersection of variables selected by LASSO regression and the Boruta algorithm, ensuring that only stable and relevant predictors are retained. Train eight machine-learning classifiers on the training set, including logistic regression, random forest, decision tree, support vector machine, XGBoost, LightGBM, complement naïve Bayes, and support vector classifier. Model performance was first assessed in the validation set and subsequently confirmed in the independent test set using the AUC-ROC, AUC-PR, accuracy, sensitivity, specificity, calibration curves, decision curve analysis, and learning curves.

Model interpretation and endpoint
Identify the best-performing machine-learning model based on validation performance metrics. Apply SHapley Additive Explanations (SHAP) to quantify the contribution of each feature to mortality prediction. Generate SHAP summary and dependence plots to visualize the relative importance and directionality of serum albumin and other predictors. Complete the analysis by integrating statistical and machine-learning results to establish an interpretable risk stratification framework.

Results

Baseline characteristics
A total of 1,274 patients with lung cancer were identified and extracted from the MIMIC-IV database (Figure 1). The median age of the cohort was 69 years (IQR: 61–77), and 51.18% were male. Overall, 343 patients (26.92%) died within 28 days of ICU admission. Baseline characteristics stratified by 28-day survival status are presented in Table 1. Compared with survivors, non-survivors exhibited significantly higher severity of illn...

Discussion

In this study, we examined the association between serum albumin levels and 28-day all-cause mortality among ICU patients with lung cancer, using data from the MIMIC-IV database. Lower serum albumin levels were independently associated with a higher risk of 28-day mortality, even after adjustment for a wide range of clinical covariates. This association remained consistent across various analytical approaches, including Cox proportional hazards models and ML methods. Subgroup analyses further demonstrated the stability o...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Medical Information Mart for Intensive Care IV (MIMIC-IV) Database	Massachusetts Institute of Technology		http: //mimic.physionet.org/
Navicat Premium (v17)	PremiumSoft CyberTech Ltd.
PostgreSQL (v17.1.11)	PostgreSQL Global Development Group
Python (v3.10)	Python Software Foundation
R Statistical Software (v4.2.3)	R Foundation for Statistical Computing
SHapley Additive exPlanations (SHAP)	Open-source software