Relationship between atherogenic index of plasma and length of stay in critically ill patients with atherosclerotic cardiovascular disease: a retrospective cohort study and predictive modeling based on machine learning

Abstract

Background

The atherogenic index of plasma (AIP) is considered an important marker of atherosclerosis and cardiovascular risk. However, its potential role in predicting length of stay (LOS), especially in patients with atherosclerotic cardiovascular disease (ASCVD), remains to be explored. We investigated the effect of AIP on hospital LOS in critically ill ASCVD patients and explored the risk factors affecting LOS in conjunction with machine learning.

Methods

Using data from the Medical Information Mart for Intensive Care (MIMIC)-IV. AIP was calculated as the logarithmic ratio of TG to HDL-C, and patients were stratified into four groups based on AIP values. We investigated the association between AIP and two key clinical outcomes: ICU LOS and total hospital LOS. Multivariate logistic regression models were used to evaluate these associations, while restricted cubic spline (RCS) regressions assessed potential nonlinear relationships. Additionally, machine learning (ML) techniques, including logistic regression (LR), decision tree (DT), random forest (RF), extreme gradient boosting (XGB), and light gradient boosting machine (LGB), were applied, with the Shapley additive explanation (SHAP) method used to determine feature importance.

Results

The study enrolled a total of 2423 patients with critically ill ASCVD, predominantly male (54.91%), and revealed that higher AIP values were independently associated with longer ICU and hospital stays. Specifically, for each unit increase in AIP, the odds of prolonged ICU and hospital stays were significantly higher, with adjusted odds ratios (OR) of 1.42 (95% CI, 1.11–1.81; P = 0.006) and 1.73 (95% CI, 1.34–2.24; P < 0.001), respectively. The RCS regression demonstrated a linear relationship between increasing AIP and both ICU LOS and hospital LOS. ML models, specifically LGB (ROC:0.740) and LR (ROC:0.832) demonstrated superior predictive accuracy for these endpoints, identifying AIP as a vital component of hospitalization duration.

Conclusion

AIP is a significant predictor of ICU and hospital LOS in patients with critically ill ASCVD. AIP could serve as an early prognostic tool for guiding clinical decision-making and managing patient outcomes.

Introduction

Cardiovascular diseases (CVDs) are one of the major threats to human survival and quality of life in today’s world, claiming approximately 17.7 million lives annually [1, 2]. Atherosclerotic cardiovascular disease (ASCVD), encompassing ischaemic heart disease and ischemic stroke, now accounts for over 60% of global cardiovascular deaths, and its burden continues to escalate at an alarming rate [3, 4, 5, 6]. Controlling the growth of the burden of CVD, especially ASCVD, is, therefore, a serious challenge. Addressing this challenge demands innovative and precise strategies to mitigate its progression. However, research on hospital length of stay (LOS), particularly in intensive care units (ICU), in patients with critically ill ASCVD remains relatively limited. While CVD mortality rates have declined over the years, there remains a critical need for further research on the clinical course and prognosis of hospitalized patients, particularly in terms of ICU stays.

Among various indicators, the Atherogenic Index of Plasma (AIP), introduced by Dobiásová and Frohlich, has emerged as a valuable biomarker. Calculated as the log-transformed TG/HDL-C ratio, AIP reflects circulating lipid imbalances and serves as an independent predictor of rapid plaque progression [7, 8, 9]. Mounting evidence links elevated AIP with an increased risk of myocardial infarction, stroke, diabetes, and other metabolic disorders [7, 10, 11, 12]. AIP may influence clinical outcomes in ASCVD patients through mechanisms such as endothelial dysfunction and increased inflammation, which can affect hospitalization duration [13, 14].

In patients with severe ASCVD, the length of hospitalization is a critical metric, reflecting not only disease severity but also treatment efficacy and recovery outcomes. While existing research has demonstrated that factors such as lipid profiles and metabolic characteristics influence hospital stays [15, 16, 17], the role of AIP in this context remains underexplored. Could AIP, as a marker of lipid dysregulation, provide insights into hospitalization dynamics for severe ASCVD patients?

In our study, we sought to bridge this knowledge gap by investigating the relationship between AIP and the length of hospital stay in patients with severe ASCVD. Utilizing the comprehensive MIMIC-IV database and employing machine learning techniques, we aimed to identify key predictors of prolonged hospitalization, shedding light on the intricate interplay between metabolic risk factors and patient outcomes.

Methods

Data selection

This study is a retrospective observational cohort analysis with longitudinal follow-up of patients. It utilizes the Medical Information Marketplace for Intensive Care-IV (MIMIC-IV-3.1), a publicly available database that includes data from over 70,000 intensive care unit admissions at Beth Israel Deaconess Medical Center in Boston, Massachusetts, spanning 2008 to 2019 [18]. The MIMIC-IV database provides comprehensive patient information, including demographics, vital signs, examination results, and diagnoses coded using the International Classification of Diseases, Ninth (ICD-9) and Tenth (ICD-10) Editions. Yu Guo was granted access to the database after obtaining certification (Record ID: 66829613) and extracted the necessary variables for this study. As the database anonymizes patient health information, individual consent was not required.

Diagnoses were determined by manual review of ICD-9 and ICD-10 codes (the ICD-9 and ICD-10 codes for all diseases are shown in supplementary material Table S1) [19]. Exclusion criteria were as follows: (1) minors (< 18 years of age), (2) patients with an ICU stay of less than 24 h, and (3) patients with advanced renal impairment, severe liver disease, or malignancy. For repeat admissions, we only collected indicators for the first admission. The flow chart of the study is shown in Fig. 1.

Fig. 1

Flow chart of the study design. LR: logistic regression; DT: decision tree; RF: random forest; XG: extreme gradient boosting; LGB: light gradient boosting machine

Measurement and calculation of AIP

AIP was chosen as the primary study variable. Serum triglycerides (TG) and serum hyperlipidemia (HLD) cholesterol were measured for the first time after admission to minimize interference with TG and HDL cholesterol values by subsequent treatment [20, 21, 22]. The AIP calculation formula is as follows:

$$ {\text{AIP}} = \log {\text{ }}({\text{TG }}({\text{mmol/L}}){\text{ }}/{\text{ HDL - C }}({\text{mmol/L}}))\quad [23] $$

Endpoint events

The study endpoints were hospital LOS and ICU LOS.

Data collection

The data extraction tool uses PostgreSQL software (v13.7.1) and Navicate Premium software (version 15) to extract data through the running Structured Query Language (SQL). Potential confounders were extracted including (1) demographics: age, gender, race; (2) Vital signs: heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, respiratory rate; (3) Clinical management: use of aspirin, clopidogrel, warfarin, vasopressin, statins; mechanical ventilation; continuous renal replacement therapy; (4) Comorbidities: acute kidney injury (AKI), chronic kidney disease (CKD), sepsis, chronic obstructive pulmonary disease (COPD), HLD, respiratory failure (RF), heart failure (HF), atrial fibrillation (AF), hypertension, diabetes mellitus; (5) Laboratory indices: red blood cells (RBD), white blood cells (WBC), red blood cell width of distribution (RDW), platelets (PLT), hemoglobin (Hb), hematocrit (HCT), total bilirubin (TB), alanine transferase (ALT ), TG, alanine transferase (AST), high-density lipoprotein cholesterol (HDL), serum glucose, prothrombin time (PT), serum creatinine (Cr), blood urea nitrogen (BUN), anion gap (AG), international normalized ratio (INR), serum potassium, serum sodium, serum calcium, and serum magnesium; and (6) severity of disease scores at the time of admission: sequential Organ Failure Estimate (SOFA) score, Acute Physiology Score III (APS III), Systemic Inflammatory Response Syndrome Score (SIRS), Simplified Acute Physiology Score II (SAPS II), Oxford Acute Severity of Illness Score (OASIS), Glasgow Coma Score (GCS), Charlson Comorbidity Index (CCI).

In the model, variables with more than 25% missing values are deleted, and missing values are multiple imputations using multiple imputations by chained equations (MICE) [24].

Statistical analysis

Subjects were stratified into quartiles based on their AIP values (Q1-Q4) [12]. Continuous variables were expressed as median (interquartile spacing (IQR)), and comparisons between groups were made using the student’s t-test or the Kruskal-Wallis H test, and categorical variables were expressed as frequencies and percentages (%), and comparisons between groups were made using the Pearson’s chi-square test or Fisher’s exact test. Kaplan-Meyer(K-M) survival analysis and Restricted mean survival time (RMST) were used to estimate ICU length of stay and length of hospitalization grouped according to the AIP index. Multivariate logistic regression analyses examined the relationship between AIP and the two outcomes. Age, weight, gender, RDW, Sodium, PTT, HLD, AKI, Diabetes, and COPD were included as confounders in the Q1 group as the reference group and trend tests were performed. We applied restricted cubic spline (RCS) curves to investigate the potential nonlinear relationship between AIP and outcome events and created a threshold effects model to identify the inflection point of AIP. Furthermore, we performed subgroup analyses to verify the reliability of the findings. To control the risk of false positives due to multiple hypothesis testing, the Bonferroni correction method was used in this study. The statistical analyses for this research were carried out using Python (version 3.9.12), SPSS (version 26.0), and DecisionLnnc1.0 software. The p-values were set according to the Bonferroni multiple correction criterion. A value of 0.025 (0.05/2) was considered statistically significant in the main analysis and 0.0023 (0.05/22) in the subgroup interaction analysis.

Construction and performance evaluation of machine learning models

The choice of variables is determined by the intersection of lasso regression and Boruta algorithm results together. The patients were randomly assigned to two groups, with 80% allocated to the training set and the remaining 20% assigned to the test set. Using the chosen predictors, five machine learning models were developed, including Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGB) [25, 26]. LR is highly interpretable and suitable for linear relationships [26]; DT captures nonlinear relationships and interactions; RF is suitable for high-dimensional data and robust to noise [27]; XGB and LGB are optimized versions of the gradient boosting framework that support parallel computation and automatic handling of missing values, and perform well in clinical prediction tasks [28]. The models were evaluated based on the area under the receiver operating characteristic (AUROC) curve, specificity, sensitivity, accuracy, and F1 score, with AUROC serving as the primary metric. The model demonstrating the best predictive performance was selected as the main model for this study. Calibration curves were utilized to evaluate the alignment between observed and predicted outcomes, while decision curve analysis (DCA) helped assess the net clinical benefit. The interpretability of the final prediction model was explored using the Shapley summation interpretation (SHAP) method.

Result

We screened and finally included 2374 patients with a mean age of 71 years, of which 54.91% were male. Missing values have been added using multiple imputations (Supplementary materials Table 2).

Baseline characteristics of study individuals

The patients were grouped into quartiles based on the AIP values calculated at the time of admission as Q1 (n = 594, -0.813-1.168), Q2 (n = 594, 1.168–1.521), Q3 (n = 592, 1.521–1.875), and Q4 (n = 594, 1.875–2.228). The analysis showed that the patient population with the highest AIP values had the youngest age of 62 (56–75), the highest body weight of 85.5 kg (73-100.57) the highest scores on all scales (SOFA, ASPIII, SIRS, GCS, CCI), higher NBPm, the highest heart rate, respiratory rate, RDW, and BUN, and also demonstrated the highest ICU hospitalization as well as longer length of stay (Supplementary materials Table 3).

Clinical outcomes.

The results of the Kaplan-Meier survival analyses for the two outcomes are shown in Fig. 2, where the horizontal coordinate is the length of hospital stay and the vertical coordinate represents the discharge rate, i.e., as the length of hospital stay increases, the discharge rate decreases progressively. The results for both quartile groupings show significant differences (both p < 0.001), and both graphs show that the largest quartile group in each group had the longest length of stay for the same discharge rate. In Fig. 3, the difference between the two groups grouped by median AIP was also significant in the different endings.

Fig. 2

Kaplan–Meier survival analysis curves of the two cohorts, ICU length of stay (A), length of hospitalization (B). The horizontal coordinate is the length of hospital stay, the vertical coordinate represents the discharge rate

Fig. 3

Restricted mean survival time graph of A ICU length of stay and B length of hospitalization

As shown in Table 1, when AIP was analyzed as a categorical variable, three multivariate logistic regression analyses were used, with model 1 unadjusted, model 2 adjusted for age, weight, gender, and model 3 adjusted for age, weight, gender, RDW, Sodium, PTT, HLD, AKI, Diabetes, and COPD. Using model 1 as a reference, the odds ratio (OR) and 95% confidence interval (CI) for the highest quartile subgroup of AIP were 1.69 (1.33–2.15) and 1.42 (1.11–1.81) for models 2 and 3, respectively, suggesting that elevated levels of AIP were significantly associated with ICU LOS, and that, after further adjusting for a wide range of confounders, AIP remained an independent risk factor for prolonged ICU stay. The results of the trend test further suggest a dose-effect relationship between AIP and prolonged ICU stay.

Table 1 Relationship between AIP and ICU LOS

Similarly, we used three multivariate logistic regression analysis models to investigate the relationship between AIP and length of hospitalization. As shown in Table 2, compared with model 1, the OR (95% CI) was 1.56 (1.23–1.96) and 2.33 (1.83–2.97) for Q3 and Q4 in model 2, and 1.41 (1.10–1.80) and 1.73 (1.34–2.24) for Q3 and Q4 in model 3, respectively. All findings were statistically significant with a P-value of less than 0.025. This indicates that as AIP increases, the risk of a longer hospital stay also rises, and this association remained robust after adjusting for various potential confounders. This suggests that AIP could serve as an independent risk factor for the length of hospital stay. Patients in the fourth quartile faced a significantly greater risk of prolonged hospitalization compared to those in the first quartile, highlighting the importance of focusing on this high-risk group.

Table 2 Relationship between AIP and hospital LOS

Detection of linear relationships

RCS curve analysis revealed a linear correlation (p-value < 0.05, non-linear p > 0.05) between AIP and ICU LOS as well as the length of hospitalization (Fig. 4A, B). The linear correlation scatterplot shows that AIP was positively correlated with ICU LOS and length of hospitalization (p < 0.001, Correlation = 0.124; p < 0.001, Correlation = 0.184) (Fig. 4C, D).

Fig. 4

Restricted cubic spline analysis illustrating AIP with ICU length of stay (A), and length of hospitalization (B). Linear correlation scatterplot for AIP with ICU length of stay (C), and length of hospitalization (D)

Subgroup analysis

We performed subgroup analyses of ICU LOS and hospital LOS, with P-values corrected using Bonferroni (Fig. 5). In the subgroup analysis with ICU LOS as the endpoint, in the age < 65 years, age > 65 years, male, female, with sepsis, with or without AF, with or without mechanical ventilation, with or without hypertension, with or without AKI, without CKD, with or without diabetes, with or without HLD, and with or without HF groups, interquartile group 4 had longer hospital stays (HR > 1, P < 0.0023 in each subgroup) (Fig. 5A). In subgroup analyses with hospital LOS as the endpoint, all fourth quartile subgroups had longer hospital stays in the age < 65 years, male, with or without AF, with hypertension, without CKD, without diabetes mellitus, without HLD, and without HF groups (HR > 1, P < 0.0023 in each subgroup) (Fig. 5B).

Fig. 5

Forest plots illustrating stratified analyses of the association of AIP and ICU LOS (A), and hospital LOS (B)

AIP vs. existing clinical scoring systems

Furthermore, we performed a comparison between AIP and other traditional scores (Fig. 6), constructing logistic regression models for the comparison [29, 30, 31]. The traditional scores incorporated by the items were SOFA, ASPIII, SIRS, SAPSII, OASIS and GCS scores [32]. For the prediction of ICU LOS(Fig. 6A), the ROC of AIP with the above five scores were 0.578; 0.683; 0.684; 0.624; 0.670; 0.658; and 0.461, respectively. Similarly, for the prediction of hospitalization time (Fig. 6B), they are 0.624; 0.683; 0.684; 0.670; 0.658; and 0.461, respectively.

Fig. 6

Comparison of AIP and existing clinical scoring systems A ICU LOS B Hospital LOS

Machine learning results

ICU LOS

Selection of characteristic variable

The variables characterizing AIP and ICU LOS were selected based on the intersection of the LASSO regression and Boruta algorithm results. After calculation, the Boruta algorithm identified 24 admissible variables, including “sepsis,” “mechanical ventilation,” “vasopressin,” and “OASIS” (Fig. 7A). The path diagrams and crossover curves of the LASSO regression illustrate the included variables (Fig. 7B and C, Supplementary materials Table 4). The combined results of both methods are visualized, with purple feature variables in the feature selection network diagram representing the intersection of the two selection approaches and the variables ultimately used in this study (Fig. 7D).

Fig. 7

AIP and ICU LOS Feature Selection Chart (A) Feature selection based on the Boruta algorithm. The horizontal axis is the name of each variable, and the vertical axis is the Z value of each variable. The box plot shows the Z value of each variable during model calculation. (B, C) Path diagrams and cross-validation plots of lasso regression analysis results. (D) Feature selection network diagram. The yellow section shows the results of the LASSO regression analysis, the red section shows the results of the Boruta algorithm, and the purple section shows the overlapping variables of the results of the two algorithms

Model performance comparisons

We constructed five models to predict ICU length of stay for ASCVD: LR, DT, RF, XGB, and LGB. As shown in Fig. 8, the LGB could provide relatively better model fitting performance with an AUC of 0.740 (sensitivity:0.508, specificity:0.879, accuracy:0.678, F1:0.676) compared to the other models, which were DT:0.632; RF:0.737; LR:0.733; and XGBoost:0.739. We performed parameter tuning and prevented model overfitting [33], such as using 5-fold cross-validation, using an independent validation set in cross-validation (not involved in model training and tuning), and using multiple metrics to evaluate the model to ensure that it has good generalization ability (Supplementary Material Table S6, S7). Fig.S1 shows the DCA curves, decision curves, and error histograms for the five models. The sensitivity, specificity, and accuracy of the other models are in the appendix material Table S8. In this study, we chose the LGB model for further investigation.

Fig. 8

ROC curves for the machine learning models. LR: logistic regression; DT: decision tree; RF: random forest; XG: extreme gradient boosting; LGB: light gradient boosting machine ROC: receiver operating characteristic; AUC: area

under the curve

Model explanation

The SHAP method, a comprehensive interpretation approach suitable for both global and individual sample analysis, was used to interpret the model. For global interpretation, the SHAP means to assess the contribution of features to the model, with the five most important features being sepsis, mechanical ventilation, vasopressin, Charlson score, and age are listed in descending order of importance (Fig. 9A). The direction and role of these five features in the predictive model are visualized, where red indicates high feature values and blue indicates low values (Fig. 9B).

Fig. 9

Global and local model explanation by the SHAP method. A SHAP summary bar plot. This plot evaluates the contribution of each feature to the model using mean SHAP values, displayed in descending order. B SHAP summary dot plot. The probability of the length of stay in ICU increases with the SHAP values of the features. Each dot represents a patient’s SHAP value for a given feature, with red indicating higher feature values and blue indicating lower values. Dots are stacked vertically to show density. C SHAP waterfall plot. This plot shows the contribution of each feature to the prediction result of one patient using the LGB(LightGBM) model. Red bars indicate features that contribute positively to the prediction, while blue bars indicate negative contributions. D, E SHAP force plot. Force diagrams for two different ending patients. RDW: Red blood cell distribution width; Nbps: Noninvasive Blood Pressure; AST: Aspartate transaminase; WBC: White blood cell; AIP: atherogenic index of plasma; AKI: Acute kidney injury; CRRT: Continuous Renal Replacement Therapy

To interpret the prediction results of individual samples, the first sample’s data was analyzed, revealing that mechanical ventilation significantly and positively influenced ICU length of stay (+ 1.28), followed by sepsis (+ 1.14) (Fig. 8C). Additionally, the model’s predictions for two patients—one with a long ICU stay and another with a short ICU stay—are visualized, with red indicating positive contributions to the prediction and blue indicating negative effects. The f(x) values represent the actual SHAP contributions for each factor (Fig. 9D and E).

The LGB model variables predicting ICU LOS for patients with AIP are visualized in a heat map, highlighting the significant impact of sepsis on the model output. Sepsis displays the widest range of colors across all samples and the longest importance bar on the right, indicating its prominent influence (Fig. S2A). Mechanical ventilation shows a notable positive effect in certain samples (darker red areas) while negatively impacting others (darker blue areas). Although vasopressor exhibits a strong influence on specific samples, its overall effect is limited, as reflected by the sporadic distribution of red and blue colors.

To illustrate the influence of AIP features on ICU LOS, a SHAP force diagram is presented, showing the direction and magnitude of AIP’s contribution to the model predictions. A predominance of red areas with significant magnitude suggests a greater impact of AIP in prolonging LOS, whereas a dominance of blue indicates a tendency to shorten LOS (Fig. S2B).

Hospital LOS

Selection of characteristic variables

The selection of variables characterizing AIP and LOS was based on the overlapping results from the LASSO regression and Boruta’s algorithm. The Boruta algorithm identified 38 admissible variables, including “sepsis,” “mechanical ventilation,” “vasopressin,” and “ASPIII” (Fig. 10A). The paths of the LASSO regression and the crossover curves illustrate the included variables, as detailed in previous sections (Fig. 10B and C, Supplementary materials Table 5). The combined results of the two methods are visualized in the feature selection network diagram, where the purple feature variables represent the intersection of both methods and the variables selected by ML in this study, totaling 29 variables (Fig. 10D).

Fig. 10

AIP and hospital LOS Feature Selection Chart A Feature selection based on the Boruta algorithm. The horizontal axis is the name of each variable, and the vertical axis is the Z value of each variable. The box plot shows the Z value of each variable during model calculation. B, C Path diagrams and cross-validation plots of lasso regression analysis results. D Feature selection network diagram. The yellow section shows the results of the LASSO regression analysis, the red section shows the results of the Boruta algorithm, and the purple section shows the overlapping variables of the results of the two algorithms

Building predictive models

As in the previous steps, we constructed five models to predict the hospital LOS for ASCVD patients, as shown in Fig. 11. The LR model can provide relatively better model fitting performance with an AUC of 0.832 compared to the other models (sensitivity:0.802, specificity:0.727, accuracy:0.741, F1:0.74). The AUCs of the other models were DT:0.791; RF:0.813; XGBoost:0.808; and LGB:0.812. We also performed parameter tuning and prevention of model overfitting, etc. Fig. S3 shows the DCA curves, decision curves, and error histograms for the five models. The sensitivity, specificity, and accuracy of the other models are in the appendix material Table S6. In this study, the LR model was chosen for further investigation.

Fig. 11

ROC curves for the machine learning models. LR: logistic regression; DT: decision tree; RF: random forest; XG: extreme gradient boosting; LGB: light gradient boosting machine ROC: receiver operating characteristic; AUC: area under the curve

Model explanation

The SHAP means were used to assess the contribution of features to the model, identifying sepsis, vasopressin, mechanical ventilation, SOFA score, and SAPS II as the five most important features in descending order of importance (Fig. 12A). The direction and role of these features in the predictive model are visualized, with red indicating a positive contribution and blue indicating a negative contribution (Fig. 12B).

Fig. 12

Global and local model explanation by the SHAP method. A SHAP summary bar plot. This plot evaluates the contribution of each feature to the model using mean SHAP values, displayed in descending order. B SHAP summary dot plot. The probability of the length of hospitalization increases with the SHAP values of the features. Each dot represents a patient’s SHAP value for a given feature, with red indicating higher feature values and blue indicating lower values. Dots are stacked vertically to show density. C SHAP waterfall plot. This plot shows the contribution of each feature to the prediction result of one patient using the LR (Logistic Regression) model. Red bars indicate features that contribute positively to the prediction, while blue bars indicate negative contributions. D, E SHAP force plot. Force diagrams for two different ending patients. RDW: Red blood cell distribution width; RBC: Red blood cell; AST: Aspartate transaminase

For individual samples, predictive results are interpreted using graphs that reveal the significant influence of sepsis on hospitalization time, contributing positively (+ 0.11), while mechanical ventilation was the second most impactful feature, contributing negatively (-1.14) (Fig. 12C). Further, SHAP plots illustrate the in-hospital prognoses for two patients: one with a long hospital stay and another with a short hospital stay, highlighting the contributions of various factors (Fig. 12D and E).

The heat map of the LR model variables predicting hospital LOS for patients with AIP highlights sepsis as the most influential feature in the model output (Fig. S4A). Vasopressin demonstrates a significant positive effect on certain samples (darker red areas) while negatively affecting others (darker blue areas). Mechanical ventilation exerts a strong influence on specific samples, though its overall range of effects remains limited, as reflected by the sporadic distribution of red and blue colors.

A SHAP force diagram illustrates the direction and magnitude of AIP’s influence on the length of hospitalization, offering insights into how AIP contributes to model predictions. A predominance of red areas with significant magnitude indicates a greater impact of AIP in lengthening the hospital stay, while a predominance of blue areas suggests a tendency to shorten it (Fig.S4B).

Discussion

In this study, we analyzed data from 2,374 adults and observed a compelling association between the AIP and LOS in both the ICU and overall hospitalization among patients with critically ill ASCVD. Patients in higher quartiles of AIP experienced significantly prolonged ICU and hospital stays compared to those in the lowest quartile. After adjusting for various confounders, the relationship between AIP and LOS remained linear, reinforcing the notion that higher AIP levels are consistently linked to longer stays. For the comparison of AIP with traditional scores, although AIP cannot completely replace traditional scores, it is still relevant in predicting the length of hospitalization. While the traditional score is based on a composite of the patient’s multidimensional status, the AIP depends on two blood indicators, but we still hope that the AIP will be a complementary predictor of the length of hospitalization.

To ensure rigorous variable selection, we utilized both LASSO regression and Boruta’s algorithm, two widely recognized feature selection methods. AIP was identified as a significant predictor in both methods, a finding that aligns with our logistic regression results. These observations suggest that AIP holds promise as a reliable predictor of hospital and ICU LOS in patients with critically ill ASCVD.

The rapid evolution of artificial intelligence in recent years has led to the widespread adoption of ML algorithms in medical research, particularly for predicting treatment outcomes and patient prognosis. LGB and RF were prioritized for their efficiency in handling high-dimensional data and capturing non-linear interactions, as demonstrated in prior critical care studies [34, 35, 36]. LR, one of the foundational tools in ML, offers not only categorical predictions but also the probability of each outcome, enabling nuanced insights into model confidence. Its interpretable coefficients allow a clear understanding of how individual features influence predictions [37]. Meanwhile, advanced algorithms such as LGB offer significant advantages, including rapid training speeds, low memory consumption, robust handling of high-dimensional datasets, and efficient parallel processing. These attributes make LGB particularly suitable for analyzing large-scale clinical data with intricate feature interactions [27, 34].

From a biological perspective, the role of AIP in atherosclerosis and cardiovascular outcomes is well-supported by prior research. Cholesterol in TG-rich lipoproteins has been identified as a key contributor to atherosclerosis, facilitating plaque formation in arterial walls. Atherogenic Apo B-containing lipoproteins, primarily LDL-C, drive the progression of atherosclerosis by promoting foam cell formation, stimulating immune responses, and producing reactive inflammatory mediators [38, 39, 40, 41]. Observational studies have consistently highlighted a negative correlation between HDL-C and atherosclerotic events, suggesting that HDL-C plays a protective role in vascular health [14, 42].

AIP, which encapsulates the balance between TG and HDL-C, has been linked to adverse outcomes across a spectrum of conditions, including diabetes, stroke, cardiovascular disease, and kidney disease [43]. For example, You et al. reported a significant linear association between elevated AIP and increased mortality from diabetes in a healthy population [44]. A cross-sectional study examining blood samples from Mexican women found that AIP could be used as a potential biomarker for the early diagnosis of cardiovascular disease in developing countries [10]. Zheng et al. found that high levels of cumulative AIP were associated with a higher risk of ischemic stroke in the Kailuan Study of 54,123 participants [45]. Furthermore, elevated AIP levels have been associated with worse clinical outcomes and longer hospital stays in patients with acute myocardial infarction [46]. Building on this foundation, our study is the first to specifically investigate the relationship between AIP and ICU LOS as well as overall hospitalization duration in critically ill ASCVD, highlighting the potential value of AIP as a marker of disease severity and prognosis.

The mechanisms by which elevated AIP prolongs hospitalization are multifaceted, with inflammation likely playing a central role. Elevated AIP reflects not only lipid metabolism disturbances but also heightened systemic inflammation, both of which are intimately linked to the pathophysiology of ASCVD [47, 48, 49]. Inflammatory responses drive atherosclerotic plaque formation, rupture, and thrombosis, with HDL-C playing a crucial protective role by facilitating reverse cholesterol transport, reducing vascular lipid deposits, and modulating the inflammatory response [50, 51, 52]. However, elevated AIP is often accompanied by reduced HDL-C levels, which diminishes its anti-inflammatory and antioxidant functions. This reduction hampers the clearance of oxidized LDL and weakens the regulation of inflammatory pathways, potentially exacerbating vascular damage [53, 54].

AIP, therefore, serves as a biomarker that integrates lipid metabolism and inflammatory status, reflecting the dual pathways contributing to ASCVD progression. The interplay between chronic inflammation and lipid abnormalities leads to endothelial damage, increased plaque instability, and heightened cardiovascular risk [55, 56]. AIP, as a crossroads of lipid metabolism and inflammatory responses may be a valuable biomarker for predicting length of hospitalization and clinical outcomes in patients with ASCVD.

This study also has some limitations. Firstly, we only used a single-center database for the study, which may result in the regional, healthcare setting, and patient population bias, and future studies speak to considering the use of multicenter data to improve external validity. Secondly, there were a large number of missing values in the database, such as height and imaging data, all of which we did not include in the metrics because we considered the effect of missing values. We look forward to further studies as the database is improved. Thirdly, only the baseline AIP index was analyzed in this study. The dynamics of the AIP index were not explored during hospitalization and ICU. Fourthly, despite the inclusion of many potential confounders, due to the limitations of the database, there may still be some unconsidered factors such as patient lifestyle (diet, exercise), treatment regimen, family history, and psychological factors. This will be explored in greater depth as the database is refined.

Conclusion

AIP is an important influence on the length of ICU stay and hospitalization in patients with critically ill ASCVD. Early detection and intervention of AIP can be clinically used as a prognostic assessment tool and play an important role in future treatment strategies. Future studies, particularly those employing prospective designs and external validation cohorts, are needed to confirm and extend these findings.