Development and validation of an explainable machine learning model for predicting acute kidney injury in critically ill patients with primary peritonitis: a multicenter cohort study

Abstract

Primary peritonitis-associated acute kidney injury (PPA-AKI) represents a complex complication that substantially elevates mortality risk. Currently, no effective machine learning (ML) models exist for precise detection. This investigation aims to construct an interpretable ML model for forecasting PPA-AKI risk while determining modifiable risk factors. This investigation used two cohorts: a derivation cohort (n = 329) from Hengyang Central Hospital (HYCH) and a validation cohort (n = 367) from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. Using the Boruta algorithm selection, 15 characteristics were determined and incorporated into 12 ML approaches, yielding 113 combinations, from which we identified the Stepglm[both] + gradient boosting machine as a superior algorithm for predicting PPA-AKI. The PPA-AKI index, comprising 9 variables, exhibited excellent diagnostic capabilities, attaining an area under the curve (AUC) of 0.976 (95%CI: 0.963–0.989) within the HYCH and 0.901 (95%CI: 0.870–0.931) within the MIMIC-IV, along with notable fitting performance and robustness. The Shapley Additive Explanations ranks the importance of features (urine output, prognostic nutritional index, blood urea nitrogen (BUN) and phosphate), and visualizes individual and global PPA-AKI risk prediction. Restricted cubic spline regression alongside threshold effect analysis revealed a nonlinear association between magnesium, BUN, creatinine and PPA-AKI, whilst generating inflection points for these features. To offer a more flexible predictive instrument, the PPA-AKI model was developed utilizing a free, publicly accessible web-based calculator (https://lglcz.shinyapps.io/DynMod/). This study developed and validated a low-cost, accurate, and readily available diagnostic tool for PPA-AKI, offering potential utility in PPA-AKI preventive management.

Introduction

Primary peritonitis (PP), also known as spontaneous bacterial peritonitis (SBP), represents an acute or subacute form of diffuse bacterial peritonitis that develops without any identifiable intra-abdominal infectious source [1]. It frequently occurs in advanced hepatic disease, impacting approximately 7% to 30% of inpatient cases presenting with cirrhosis and ascites [2]. The PP links to an unfavorable prognosis, demonstrating mortality rates spanning from 20% to exceeding 50% at 6 and 12 months, respectively [3]. Despite advances in PP management (such as prompt antibiotic treatment), the mortality rate observed among hospitalized PP patients remains high [1,2]. This phenomenon may be explained by multiple underlying factors, including elevated incidence of acute kidney injury (AKI), multidrug-resistant bacterial infections, and multi-organ dysfunction, which represent critical areas and priorities for subsequent research [4].

AKI represents one of the serious complications of PP, distinguished by a sudden decline in renal function that includes both structural damage and functional deterioration [5]. PP-related AKI (PPA-AKI) develops in up to 54% of patients [6]. Significantly, multiple investigations have demonstrated that AKI development functions as an independent risk factor for unfavorable outcomes in PP [7,8]. Upon AKI onset in PP patients, serum creatinine concentrations increase whilst glomerular filtration rate declines, impeding the body’s capacity to clear metabolic waste products efficiently. This leads to fluid dysregulation, electrolyte imbalances, and metabolic acidosis, thereby elevating mortality rates amongst PP patients [8]. Numerous AKI instances are identified belatedly. As no therapeutic interventions for established AKI can enhance outcomes, precise and prompt identification of high-risk patients remains vital for AKI prevention through providing essential opportunities for proactive management [9]. Therefore, there is an urgent necessity to identify pivotal and alterable risk determinants and develop a noninvasive, dependable, and accessible diagnostic framework for PPA-AKI identification.

In clinical settings, elevated serum creatinine levels and reduced urine output serve as widely adopted diagnostic markers for AKI per the Kidney Disease: Improving Global Outcome (KDIGO) criteria [10]; nevertheless, they lack sufficient sensitivity and specificity for kidney injury [11]. Rather than identifying AKI prior to its occurrence, creatinine levels only commence elevation following the establishment of kidney injury, which may constrain the intervention timing for AKI. The machine learning (ML) methodologies derived from electronic medical records (EMR) have attracted considerable clinician interest and recognition in recent years [12,13]. The widespread adoption of EMR across healthcare settings has facilitated more accurate and readily available collection of clinical data for patients. Numerous ML approaches are currently being employed in developing AKI prediction models for a range of critical conditions [14–16], with most demonstrating favorable predictive capabilities. At present, no dependable ML model exists for PPA-AKI accurate identification. Therefore, establishing and confirming a robust, generalizable ML model for the individualized prediction of PPA-AKI is imperative, thus diminishing the occurrence of PPA-AKI and enhancing survival outcomes in PP patients.

In this large-scale multicenter cohort study, a novel multi-criteria decision-making framework was developed, utilizing the leave-one-out cross-validation (LOOCV) approach for predictive analysis by transforming 12 conventional ML algorithms into 113 combinations. Routine clinical characteristics at admission derived from EMR were systematically employed to predict AKI in PP patients, and a robust and generalizable PPA-AKI model was subsequently established through the LOOCV framework. Subsequently, we conducted various comprehensive statistical approaches for model assessment and model interpretation. Finally, we developed a web calculator for the model’s application.

Materials and methods

This investigation was approved by the Ethics Review Committee of the Affiliated Hengyang Hospital of Hunan Normal University & Hengyang Central Hospital (No. 2025(204)), adhering to the ethical standards established in the World Medical Association Declaration of Helsinki. The retrospective design using anonymized data negated informed consent requirements.

Study design

The design of this investigation encompasses six phases: Data acquisition, feature identification, model development and validation, model assessment, model interpretation, and model online calculator. This investigation adhered to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) + artificial intelligence (AI) guidelines [17].

Study population

This multicenter retrospective study drew upon data derived from two distinct cohorts. The derivation cohort was included from the Hengyang Central Hospital (HYCH) cohort (January 2020 to April 2025), which is a private database and derived from the EMR of Affiliated Hengyang Hospital of Hunan Normal University & Hengyang Central Hospital. The validation cohort was collected from the MIMIC-IV database (certification number: 75757419), which represents a longitudinal, single-center database encompassing 247,366 individuals and 196,527 adults who received admission to the Beth Israel Deaconess Medical Center between 2008 and 2019. The MIMIC-IV database is a large-scale database of the intensive care unit (ICU). In order to ensure homogeneity with the MIMIC-IV cohort, for the HYCH cohort, we selected the patients and data from the gastroenterology ICU.

The inclusion criteria for this investigation comprised the following: 1) adults (≥18 years old) diagnosed with PP; 2) ICU length of stay (LOS) ≥ 24 h. The exclusion criteria encompassed: 1) prior diagnosis of chronic kidney disease (CKD) stage 5 or AKI; 2) previous receipt of renal replacement therapy; 3) more than 20% missing data overall; 4) death within 24 h of ICU admission. A detailed study population flowchart is displayed in Figure 1. Notably, the HYCH cohort guarantees no missing data as far as possible. PP was diagnosed in accordance with international guidelines, defined by a polymorphonuclear (PMN) cell count in the ascitic fluid of ≥ 250/m³ alongside a positive ascitic fluid culture (culture-positive PP), or by a PMN count of ≥ 250/m³ with a negative ascitic fluid culture (culture-negative neutrocytic ascites), provided that alternative causes of peritonitis and hemorrhagic ascites were excluded [18].

Figure 1.

A detailed flowchart of the study population.

ICU: intensive care unite; AKI: acute kidney injury; PP: primary peritonitis; CKD: chronic kidney disease; HYCH: Hengyang Central Hospital; MIMIC-IV: Medical Information Mart for Intensive Care IV.

Definition of outcome

The outcome of this study is that PP patients develop AKI. AKI is diagnosed according to the 2012 KDIGO criterion [10]: a serum creatinine elevation of no less than 0.3 mg/dL within 48 h, or a rise to a minimum of 1.5-fold the baseline concentration within the preceding 7 days, where baseline creatinine represents the lowest creatinine concentration recorded within the 7-day period before each AKI episode, or urine production below 0.5 mL/kg/h for a minimum of 6 successive hours.

Sample size estimation

We determined the sample size by applying the events per variable (EPV) metric [17], which represents a broadly recognized approach in statistical analyses. Within the derivation cohort, the occurrence rate of PPA-AKI was 0.50. Considering our objective to incorporate nine predictor variables whilst setting the EPV to 10, we computed the necessary sample size through the subsequent formula: Sample size = Number of variables × EPV/(Incidence rate) = 9 × 10/(0.5) = 180.

Data extraction and preprocessing

For each participant, we obtained records during ICU admission. Comprehensive details include: (1) demographic information; (2) past medical history; (3) vital signs; (4) first-recorded laboratory values; (5) illness severity indices encompassing sequential organ failure assessment (SOFA), simplified acute physiology score II (SAPS II), acute physiology and chronic health evaluation II (APACHE II), and glasgow coma scale (GCS); (6) interventions and medications during hospitalization; (7) ICU/hospital LOS and 28-day ICU/hospital mortality.

Additionally, the prognostic nutritional index (PNI) is computed by serum albumin and lymphocyte count [19] employing the following formula: PNI = 10 × serum albumin (g/dL) + 5 × lymphocyte count (10⁹/L).

To address missing values, we utilized multiple imputation with chained equations (MICE), working on the premise that data absence occurred at random. Variables suitable for the multivariate model were included in the imputation procedure, with Rubin’s rule utilized to consolidate findings across the 10 imputed datasets.

Feature selection

Selecting appropriate variables is crucial for ensuring diagnostic model efficacy. Therefore, we utilized the Boruta algorithm with the Boruta R package, which utilizes a random forest classifier to execute feature selection across multiple datasets exhibiting high-dimensional and multivariate properties [20]. The parameter settings for the Boruta feature selection algorithm are as follows: Iterations = 100, p-value threshold = 0.01, Bonferroni’s method = TRUE.

Model construction and validation

Employing feature variables determined through the Boruta algorithm, a consensus diagnostic framework for PPA-AKI was constructed and evaluated. The approach proceeded as follows:

1. ML algorithm combinations: Initially, 12 conventional ML algorithms were incorporated: Stepglm (MASS R package), Least Absolute Shrinkage and Selection Operator (LASSO) (glmnet R package), Linear Discriminant Analysis (LDA) (caret R package), gradient boosting machine (GBM) (gbm R package), extreme Gradient Boosting (XGBoost) (xgboost R package), Naive Bayes (e1071 R package), SVM (e1071 R package), random forest (RF) (RF) (randomForest R package), partial least squares regression for generalized linear models (plsRglm) (plsRglm R package), glmBoost (mboost R package), elastic network (Enet) (glmnet R package) and ridge regression (glmnet R package). These 12 ML methods underwent hyperparameter tuning via grid search. Through the integration of 12 ML algorithms with parameter optimization, numerous distinct combinations were established as predictive frameworks employing the LOOCV structure.

2. Model development: Among these algorithms, RF, glmBoost, Stepglm, and LASSO demonstrate feature-selection functionalities with 10-fold cross-validation. Initially, feature variables are selected, followed by model training with 10-fold cross-validation. In a combined approach, the initial method conducts feature selection whilst the subsequent method executes modeling.

3. Model evaluation: Afterwards, 113 ML combinations underwent assessment through the area under the curve (AUC) within the derivation cohort. AUC computations were performed for every model generated from the derivation cohort upon application to the validation cohort.

4. Model selection: The ultimate consensus diagnostic framework for PPA-AKI (PPA-AKI index) was developed employing the mean AUC from both cohorts, taking into account model simplicity together with goodness of fit.

Model evaluation

To assess multicollinearity amongst variables in the PPA-AKI model, variance inflation factor (VIF) and Spearman correlation analyses were executed. Multicollinearity was indicated when VIF values exceeded 5, tolerance was below 0.2, or correlation coefficients between variables surpassed 0.5.

The PPA-AKI index’s predictive capability was assessed using receiver operating characteristic (ROC) and precision recall (PR) analyses of area under the curve (AUC) values with 95% confidence interval (CI). The model’s performance underwent comprehensive assessment employing various metrics encompassing sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), Matthews correlation coefficient (MCC), precision, accuracy, F1-score, prevalence, recall, false negative rate (FNR) and false positive rate (FPR). The calibration of the PPA-AKI index was assessed through calibration curves, which illustrated the correspondence between predicted and observed probabilities. Model agreement was assessed using the Brier score (optimal = 0; values > 0.3 indicate poor calibration) and calibration slope (optimal = 1). The clinical utility of the PPA-AKI index was assessed using decision curve analysis (DCA), computing net benefits across threshold probabilities to evaluate performance against baseline metrics.

Multivariable logistic regression modeling was utilized to demonstrate that the PPA-AKI index functions as an independent predictor of AKI amongst PP patients, subsequent to covariate adjustment. To explore potential heterogeneity within specific populations, subgroup evaluations were conducted through patient stratification according to discrete clinical characteristics. The link between these parameters and the PPA-AKI index employed for stratification in subgroup assessments was investigated employing likelihood ratio testing.

Sensitivity analysis

To justify this choice that the HYCH cohort as the derivation set and the MIMIC-IV cohort as the validation set, we performed a sensitivity analysis, namely MIMIC-IV cohort as the derivation set and HYCH cohort as the validation set. PPA-AKI was developed and assessed using an innovative LOOCV framework.

Considering that urine output and serum creatinine are the diagnostic criteria for AKI and may therefore artificially enhance model performance, we excluded these variables in a sensitivity analysis to further assess the model’s true predictive power.

Given that urine output is calculated based on the first 24 h following admission, if a patient meets the KDIGO diagnostic criteria for AKI at hour 12, the model will incorporate post-event data into the prediction. Thus, we will conduct a sensitivity analysis by excluding patients diagnosed with AKI within 12 h of admission, in order to further evaluate the model’s predictive performance.

Model interpretability

To enhance model interpretability, we utilized the SHAP approach with shapviz R package [21]. SHAP values were computed to establish the importance hierarchy of individual clinical characteristics within the PPA-AKI model. Furthermore, the SHAP approach offers both localized and globalized explanations for model comprehension. Localized interpretation enables the demonstration of specific AKI risk predictions for individual PP patients through the input of particular data. Globalized interpretation provides consistent and dependable attribution values for individual characteristics within the model, thereby illustrating the relationship between PPA-AKI and input parameters. The SHAP dependency plot illustrates the nonlinear interactive association between PPA-AKI and predictor variables.

Logistic regression, RCS regression and threshold effect analysis

To further investigate feature significance in the PPA-AKI model, we assessed linear and nonlinear links between PPA-AKI and predictive variables. Linear associations were examined via univariable logistic regression with Wald tests examining coefficient linearity. Nonlinear associations were analyzed employing RCS regression (rms package) with chi-square tests. Subsequently, threshold effect analysis identified inflection points for predictive variables.

Visualized online calculator

To offer a more adaptable predictive instrument for the PPA-AKI model, a web-based calculator featuring a visual interface was created to enable straightforward entry of clinical parameters and produce clear, interpretable results showing PPA-AKI risk.

Statistical analysis

Continuous data exhibiting normal distribution were denoted as mean ± standard deviation (SD), whilst comparisons between groups were executed employing independent samples t-tests. Data not following normal distribution were denoted as median (interquartile range [IQR], p25-p75) and analyzed between groups through Mann-Whitney U tests. Categorical variables are denoted as frequencies (percentages, %) and assessed through chi-square tests or Fisher’s exact tests, as appropriate. R software (R version 4.3.1), SPSS statistics 22.0 and DecisionLinnc 1.0 software (Python) were utilized to perform the statistical analysis. A two-sided p < 0.05 was deemed significant except when specific P values were specified.

Results

Study population and patient characteristics

The incidence of PPA-AKI was 165 (165/329 = 50.15%) in the HYCH cohort and 184 (184/367 = 50.13%) in the MIMIC-IV cohort. We conducted a comparison of 66 standard clinical features, encompassing pathophysiological characteristics upon admission, therapeutic interventions, and clinical outcomes, across the AKI and non-AKI cohorts within two cohorts, demonstrating notable differences between these two cohorts (Table 1). Relative to the non-AKI group, it was found that pneumonia incidence, furosemide/vasopressor treatment, hospital/ICU LOS, 28-day hospital/ICU mortality were substantially increased in the AKI cohort.

Table 1.

Comparison of baseline characteristics and biochemical indicators between AKI and non-AKI groups in PP patients in HYCH and MIMIC-IV cohorts.

Cohorts			HYCH				MIMIC-IV
Variable	Levels	Overall	Non-AKI	AKI	p value	Overall	Non-AKI	AKI	p value
		N = 329	N = 164	N = 165		N = 367	N = 183	N = 184
Age (y), mean (SD)		61.84 ± 14.87	60.50 ± 15.66	63.18 ± 13.96	0.103	59.71 ± 13.89	61.58 ± 14.89	57.85 ± 12.59	0.010
Sex, n (%)					0.338				0.024
	Female	139 (42.25%)	65 (39.63%)	74 (44.85%)		161 (43.87%)	91 (49.73%)	70 (38.04%)
	Male	190 (57.75%)	99 (60.37%)	91 (55.15%)		206 (56.13%)	92 (50.27%)	114 (61.96%)
Height (cm), mean (SD)		169.46 ± 10.32	170.47 ± 9.99	168.45 ± 10.57	0.075	168.28 ± 10.88	167.19 ± 10.58	169.36 ± 11.09	0.057
Weight (kg), mean (SD)		81.31 ± 23.21	78.49 ± 22.15	84.11 ± 23.95	0.028	82.95 ± 23.21	78.13 ± 20.64	87.74 ± 24.65	<0.001
BMI (kg/m2), mean (SD)		28.30 ± 7.91	26.94 ± 7.20	29.65 ± 8.36	0.002	29.17 ± 7.17	27.84 ± 6.59	30.49 ± 7.49	<0.001
Pneumonia, n (%)					0.006				0.002
	No	291 (88.45%)	153 (93.29%)	138 (83.64%)		257 (70.03%)	142 (77.60%)	115 (62.50%)
	Yes	38 (11.55%)	11 (6.71%)	27 (16.36%)		110 (29.97%)	41 (22.40%)	69 (37.50%)
COPD, n (%)					0.066				0.123
	No	314 (95.44%)	160 (97.56%)	154 (93.33%)		328 (89.37%)	159 (86.89%)	169 (91.85%)
	Yes	15 (4.56%)	4 (2.44%)	11 (6.67%)		39 (10.63%)	24 (13.11%)	15 (8.15%)
Stroke, n (%)					0.179				0.753
	No	324 (98.48%)	163 (99.39%)	161 (97.58%)		356 (97%)	177 (96.72%)	179 (97.28%)
	Yes	5 (1.52%)	1 (0.61%)	4 (2.42%)		11 (3%)	6 (3.28%)	5 (2.72%)
Cancer, n (%)					0.503				0.012
	No	291 (88.45%)	147 (89.63%)	144 (87.27%)		312 (85.01%)	147 (80.33%)	165 (89.67%)
	Yes	38 (11.55%)	17 (10.37%)	21 (12.73%)		55 (14.99%)	36 (19.67%)	19 (10.33%)
T2DM, n (%)					0.218				0.017
	No	294 (89.36%)	150 (91.46%)	144 (87.27%)		277 (75.48%)	148 (80.87%)	129 (70.11%)
	Yes	35 (10.64%)	14 (8.54%)	21 (12.73%)		90 (24.52%)	35 (19.13%)	55 (29.89%)
HF, n (%)					0.508				0.457
	No	308 (93.62%)	155 (94.51%)	153 (92.73%)		310 (84.47%)	152 (83.06%)	158 (85.87%)
	Yes	21 (6.38%)	9 (5.49%)	12 (7.27%)		57 (15.53%)	31 (16.94%)	26 (14.13%)
IHD, n (%)					0.006				0.552
	No	312 (94.83%)	161 (98.17%)	151 (91.52%)		309 (84.20%)	152 (83.06%)	157 (85.33%)
	Yes	17 (5.17%)	3 (1.83%)	14 (8.48%)		58 (15.80%)	31 (16.94%)	27 (14.67%)
HR (bpm), mean (SD)		98.73 ± 22.66	97.95 ± 22.81	99.52 ± 22.55	0.531	97.26 ± 20.85	101.38 ± 21.03	93.17 ± 19.91	<0.001
RR (insp/min), mean (SD)		20.48 ± 7.07	19.88 ± 7.51	21.08 ± 6.58	0.122	20.82 ± 6.83	19.84 ± 6.63	21.80 ± 6.90	0.006
Temperature (◦C), mean (SD)		36.82 ± 0.76	36.90 ± 0.79	36.74 ± 0.73	0.060	36.58 ± 2.95	36.81 ± 0.73	36.36 ± 4.10	0.140
SpO2 (%), mean (SD)		96.55 ± 3.44	96.68 ± 3.32	96.42 ± 3.56	0.496	96.82 ± 3.73	96.73 ± 4.16	96.91 ± 3.26	0.633
SBP (mmHg), mean (SD)		114.75 ± 26.18	115.10 ± 25.70	114.41 ± 26.72	0.811	116.35 ± 24.39	118.20 ± 24.24	114.52 ± 24.47	0.149
DBP (mmHg), mean (SD)		64.03 ± 17.15	65.19 ± 17.09	62.87 ± 17.18	0.221	66.02 ± 17.13	68.02 ± 17.57	64.04 ± 16.49	0.026
MAP (mmHg), mean (SD)		77.79 ± 17.88	78.60 ± 17.87	76.99 ± 17.91	0.416	78.91 ± 17.53	80.54 ± 18.11	77.29 ± 16.83	0.076
APACHEII, mean (SD)		21.13 ± 7.79	18.56 ± 6.98	23.68 ± 7.73	<0.001	21.67 ± 7.27	18.26 ± 5.80	25.06 ± 7.00	<0.001
SOFA, median (p25 - p75)		7 (5 - 10)	6 (4 - 8)	8 (6 - 12)	<0.001	8 (5 - 11)	6 (3 - 8)	10 (7 - 13)	<0.001
APSIII, mean (SD)						62.70 ± 24.36	51.38 ± 19.22	73.97 ± 23.74	<0.001
SIRS, mean (SD)						2.86 ± 0.85	2.92 ± 0.82	2.79 ± 0.87	0.160
SAPSII, mean (SD)		42.12 ± 16.94	36.07 ± 15.19	48.15 ± 16.47	<0.001	43.22 ± 14.50	38.21 ± 13.20	48.21 ± 14.03	<0.001
OASIS, mean (SD)						34.26 ± 8.79	33.15 ± 8.66	35.36 ± 8.80	0.016
GCS, mean (SD)		13.05 ± 2.89	13.66 ± 2.43	12.43 ± 3.18	<0.001	13.55 ± 2.75	13.90 ± 2.27	13.20 ± 3.12	0.014
Charlson, mean (SD)						5.82 ± 3.05	5.26 ± 3.19	6.37 ± 2.81	<0.001
Urine output (mL), mean (SD)		1,620.62 ± 902.30	2,050.50 ± 976.30	1,193.35 ± 558.99	<0.001	1,618.16 ± 940.57	2,095.23 ± 985.25	1,143.68 ± 591.59	<0.001
Balance (mL), median (p25 - p75)		3,220 (-450 − 10,205)	1,703.67 (-940 − 7,307.59)	4,180 (0 - 12,044.82)	0.111	2,207.30 (246.83 − 5,098.26)	2,540.57 (517.59 − 5,815.63)	1,759.25 (-79.13 − 3,988.05)	0.003
WBC (×109/L), mean (SD)		20.22 ± 11.12	18.34 ± 10.88	22.09 ± 11.07	0.002	18.98 ± 8.20	18.79 ± 8.41	19.16 ± 8.00	0.664
RBC (×109/L), mean (SD)		3.76 ± 0.69	3.73 ± 0.67	3.80 ± 0.71	0.331	3.12 ± 0.83	3.44 ± 0.82	2.80 ± 0.70	<0.001
RDW (%), mean (SD)		17.85 ± 3.64	17.36 ± 3.53	18.35 ± 3.69	0.013	17.32 ± 3.19	16.22 ± 2.55	18.41 ± 3.38	<0.001
PLT (×109/L), median (p25 - p75)		168 (109 - 250)	164 (111 − 220.50)	175 (105 - 286)	0.126	136 (77 - 227)	198 (109 - 287)	103 (64 - 162)	<0.001
Lymphocytes (×109/L), median (p25 - p75)		1.86 (1.06 − 3.24)	1.88 (1.15 − 3.29)	1.86 (1.01 − 3.23)	0.583	2.00 (1.13 − 3.18)	2.52 (1.68 − 4.37)	1.38 (0.66 − 2.46)	<0.001
Hb (g/dL), mean (SD)		11.37 ± 2.00	11.30 ± 1.95	11.44 ± 2.05	0.543	9.12 ± 1.92	9.58 ± 1.97	8.67 ± 1.76	<0.001
Lactate (mmol/L), median (p25 - p75)		2.40 (1.40 − 6.00)	2.10 (1.30 − 3.80)	2.90 (1.70 − 7.08)	0.001	2.50 (1.60 − 4.00)	2.10 (1.50 − 3.60)	2.80 (1.80 − 4.40)	<0.001
PaCO2 (mmHg), mean (SD)		44.13 ± 14.19	42.16 ± 12.57	46.08 ± 15.42	0.012	38.10 ± 9.98	38.67 ± 9.42	37.52 ± 10.51	0.270
PH, mean (SD)		7.44 ± 0.09	7.44 ± 0.08	7.44 ± 0.10	0.626	7.35 ± 0.10	7.36 ± 0.09	7.34 ± 0.11	0.014
PaO2 (mmHg), median (p25 - p75)		108 (75 - 166)	113.50 (81.50 - 160)	94 (55 - 167)	0.190	85 (52 - 151)	99 (59 - 181)	75.50 (46.50 − 126.50)	0.003
Base excess (mEq/L), median (p25 - p75)		0.10 (-4.40 − 3.60)	0.10 (-3.85 − 2.90)	0.80 (-4.90 − 4.60)	0.832	−3 (-7 − 0)	−2 (-6 − 0)	−4 (-8 - −1)	<0.001
Glucose (mg/dL), median (p25 - p75)		139 (94 - 196)	131.50 (88 - 183)	147 (104 - 229)	0.005	123 (101 - 160)	129 (104 - 158)	118.50 (97.50 - 169)	0.681
Potassium (mmol/L), mean (SD)		4.33 ± 0.86	4.16 ± 0.85	4.51 ± 0.83	<0.001	4.29 ± 0.79	4.21 ± 0.69	4.37 ± 0.87	0.045
Sodium (mmol/L), mean (SD)		143.29 ± 5.94	142.35 ± 5.56	144.22 ± 6.18	0.004	135.60 ± 6.42	135.62 ± 5.84	135.57 ± 6.96	0.938
Chloride (mmol/L), mean (SD)		110.78 ± 6.88	109.93 ± 6.33	111.61 ± 7.31	0.027	101.81 ± 7.80	102.52 ± 7.20	101.10 ± 8.31	0.081
Bicarbonate (mmol/L), mean (SD)		27.39 ± 5.27	27.31 ± 4.97	27.47 ± 5.56	0.783	20.16 ± 4.77	21.34 ± 4.06	18.98 ± 5.13	<0.001
Phosphate (mg/dL), mean (SD)		4.69 ± 1.85	4.13 ± 1.39	5.24 ± 2.07	<0.001	4.32 ± 1.93	3.90 ± 1.48	4.73 ± 2.22	<0.001
Calcium (mg/dL), mean (SD)		8.62 ± 0.93	8.60 ± 0.78	8.65 ± 1.06	0.617	8.30 ± 0.97	8.11 ± 0.89	8.50 ± 1.00	<0.001
Magnesium (mg/dL), mean (SD)		2.30 ± 0.45	2.23 ± 0.43	2.38 ± 0.46	0.003	2.18 ± 0.47	2.08 ± 0.46	2.28 ± 0.45	<0.001
Anion gap (mEq/L), mean (SD)		15.43 ± 6.32	14.09 ± 5.29	16.75 ± 6.96	<0.001	16.04 ± 5.28	14.47 ± 4.38	17.61 ± 5.64	<0.001
BUN (mg/dL), median (p25 - p75)		26 (15 − 51)	17 (11 − 33)	38 (24 - 66)	<0.001	27 (15 - 49)	18 (11 - 29)	38 (26 − 65.50)	<0.001
Creatinine (mg/dL), median (p25 - p75)		1.50 (0.75 − 2.81)	0.87 (0.63 − 1.74)	2.19 (1.35 − 3.20)	0.005	1.30 (0.70 − 2.30)	0.80 (0.60 − 1.20)	1.95 (1.30 − 3.30)	<0.001
Albumin (g/dL), mean (SD)		2.62 ± 0.69	2.81 ± 0.72	2.43 ± 0.61	<0.001	2.97 ± 0.76	3.15 ± 0.75	2.79 ± 0.72	<0.001
Bilirubin (mg/dL), median (p25 - p75)		1.70 (0.70 − 4.70)	1.40 (0.60 − 4.70)	2.20 (0.90 − 5.10)	0.117	2.70 (0.90 − 8.40)	1.20 (0.60 − 3.50)	5.85 (2.40 − 13.55)	<0.001
AST (IU/L), median (p25 - p75)		64 (33 - 174)	54 (30 − 110.50)	76 (37 − 300)	0.041	52 (27 − 102)	41 (21 − 72)	66.50 (35.50 − 140.50)	0.090
ALT (IU/L), median (p25 - p75)		36 (22 - 93)	31 (18.50 - 63)	46 (24 - 145)	0.039	25 (16 − 51)	21 (14 - 49)	28 (18 - 66)	0.177
PNI, mean (SD)		38.26 ± 11.92	40.42 ± 12.30	36.11 ± 11.15	<0.001	42.01 ± 13.85	48.12 ± 15.21	35.93 ± 8.91	<0.001
INR, median (p25 - p75)		1.68 (1.30 − 2.30)	1.50 (1.30 − 2.10)	1.74 (1.40 − 2.30)	0.055	1.80 (1.40 − 2.40)	1.50 (1.30 − 2.00)	2.10 (1.60 − 2.70)	<0.001
PTT (s), median (p25 - p75)		39.60 (31.60 − 53.00)	37.00 (31.60 − 48.05)	42.00 (33.00 − 57.90)	0.053	38.20 (31.50 − 49.10)	34.80 (29.80 − 43.50)	43.45 (34.40 − 57.15)	<0.001
AKI stage, n (%)									<0.001
	1	16 (4.86%)		16 (9.70%)		15 (4.09%)		15 (8.15%)
	2	29 (8.81%)		29 (17.58%)		37 (10.08%)		37 (20.11%)
	3	120 (36.47%)		120 (72.73%)		132 (35.97%)		132 (71.74%)
Ventilation, n (%)					<0.001				0.262
	No	199 (60.49%)	117 (71.34%)	82 (49.70%)		58 (15.80%)	25 (13.66%)	33 (17.93%)
	Yes	130 (39.51%)	47 (28.66%)	83 (50.30%)		309 (84.20%)	158 (86.34%)	151.00 (82.07%)
Furosemide, n (%)					0.017				0.870
	No	233 (70.82%)	126 (76.83%)	107 (64.85%)		306 (83.38%)	152 (83.06%)	154 (83.70%)
	Yes	96 (29.18%)	38 (23.17%)	58 (35.15%)		61 (16.62%)	31 (16.94%)	30 (16.30%)
Vasopressor, n (%)					0.002				<0.001
	No	231 (70.21%)	128 (78.05%)	103 (62.42%)		170 (46.32%)	105 (57.38%)	65 (35.33%)
	Yes	98 (29.79%)	36 (21.95%)	62 (37.58%)		197 (53.68%)	78 (42.62%)	119 (64.67%)
Hospital LOS, median (p25 - p75)		11.24 (6.18 − 20.14)	9.47 (5.17 − 16.90)	13.91 (7.42 − 24.93)	<0.001	17.28 (8.88 − 32.42)	15.18 (7.77 − 24.46)	20.53 (11.64 − 36.36)	0.001
ICU LOS, median (p25 - p75)		3.92 (2.08 − 9.88)	2.83 (1.60 − 6.85)	5.25 (2.96 − 12.71)	<0.001	3.51 (2.01 − 7.48)	3.16 (1.95 − 5.89)	4.03 (2.14 − 9.04)	0.032
28 days hospital mortality, n (%)					0.006				<0.001
	No	247 (75.08%)	134 (81.71%)	113 (68.48%)		273 (74.39%)	158 (86.34%)	115 (62.50%)
	Yes	82 (24.92%)	30 (18.29%)	52 (31.52%)		94 (25.61%)	25 (13.66%)	69 (37.50%)
28 days ICU mortality, n (%)					0.005				<0.001
	No	267 (81.16%)	143 (87.20%)	124 (75.15%)		265 (72.21%)	156 (85.25%)	109 (59.24%)
	Yes	62 (18.84%)	21 (12.80%)	41 (24.85%)		102 (27.79%)	27 (14.75%)	75 (40.76%)

Abbreviations: MIMIC-IV: Medical Information Mart for Intensive Care IV; ICU: intensive care unite; AKI: acute kidney injury; PP: primary peritonitis; SD: standard deviation; BMI: body mass index; HR: heart rate; SBP: systolic blood pressure; DBP: diastolic blood pressure; MAP: mean arterial pressure; RR: respiratory rate; T2DM: type 2 diabetes mellitus; HF: heart failure; IHD: ischemic heart disease; COPD: chronic obstructive pulmonary disease; APACHEII: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; APSIII: acute physiological score III; SIRS: systemic inflammatory response syndrome; SAPSII: simplified acute physiology score II; OASIS: oxford acute severity of illness score; GCS: glasgow coma scale; RBC: red blood cell; WBC: white blood cell; PLT: platelet; Hb: hemoglobin; RDW: red blood cell volume distribution width; PTT: partial thromboplastin time; BUN: blood urea nitrogen; AST: aspartate aminotransferase; ALT: alanine aminotransferase; PNI: prognostic nutritional index; INR: international normalized ratio; LOS: length of stay; p25 - p75: interquartile range percentage.

Additionally, patients with PPA-AKI in pathophysiological characteristics had a higher weight, body mass index (BMI), SOFA, SAPS II, APACHE II, lactate, red blood cell volume distribution width (RDW), potassium, phosphate, magnesium, anion gap, blood urea nitrogen (BUN), creatinine, and lower GCS, urine output, albumin, and PNI. These observations demonstrate a marked connection between these conventional admission parameters and AKI development in PP patients.

Variable selection

To mitigate confounding factors and pinpoint dependable variables, the Boruta algorithm was utilized to conduct feature variable screening across 54 pathophysiological variables at admission. Using the Boruta algorithm, we identified 15 variables in the HYCH cohort (Figure 2A) and 20 variables in the MIMIC-IV cohort (Figure 2B). Variable selection details are depicted in Supplementary material 1. Thereafter, we obtained the intersection to identify 15 indicators (BMI, APACHE II, SOFA, SAPS II, GCS, Urine output, RDW, lactate, phosphate, magnesium, anion gap, albumin, PNI, BUN, and creatinine) (Figure 2C).

Figure 2.

Important characteristic variables identified by the Boruta algorithm in the Hengyang Central Hospital (HYCH) and Medical Information Mart for Intensive Care IV (MIMIC-IV) cohorts. The horizontal axis shows the names of variables, and the vertical axis shows the Z-score of variables. The boxplot shows the Z-score of variables during the model calculation process. (A) HYCH cohort. (B) MIMIC-IV cohort. (C) Venn diagram displays the shared feature variables in HYCH and MIMIC-IV cohorts.

MIMIC-IV: Medical Information Mart for Intensive Care IV; HYCH: Hengyang Central Hospital; BMI: body mass index; APACHEII: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; SAPSII: simplified acute physiology score II; GCS: glasgow coma scale; RBC: red blood cell; PLT: platelet; RDW: red blood cell volume distribution width; PTT: partial thromboplastin time; BUN: blood urea nitrogen; PNI: prognostic nutritional index; INR: international normalized ratio.]

Optimized model development and verification

To establish a dependable and consensus-based diagnostic model for PPA-AKI, we implemented an innovative LOOCV framework through transforming 12 classical ML algorithms into 113 combinations within the derivation cohort. Our analytical framework incorporated 15 feature variables. Among these ML, RF, glmBoost, Stepglm, and LASSO further performed feature selection, selecting 6, 9, 10 and 11 variables respectively. For the derivation and validation cohorts, AUC values were computed for each individual algorithm, along with the mean AUC across both cohorts. As a result, among these predictive models, “RF”, “Lasso + RF”, “Stepglm[backward]+RF”, “Stepglm[both]+RF”, “glmBoost + RF”, “glmBoost + GBM”, and “Stepglm[both]+GBM” demonstrated superior predictive performance (Figure 3A and Figure S1). However, “RF”, “Lasso + RF”, “Stepglm[backward]+RF”, “Stepglm[both]+RF”, “glmBoost + RF” displayed elevated AUC values (AUC ≥ 0.996) within the derivation cohort whilst showing diminished performance (AUC ≤ 0.894) in the validation cohort, suggesting the presence of overfitting. Moreover, the efficacy of “glmBoost + GBM” (AUC = 0.980 in derivation cohort, AUC = 0.898 in validation cohort, average AUC = 0.939) and “Stepglm[both]+GBM” (AUC = 0.976 in derivation cohort, AUC = 0.901 in validation cohort, average AUC = 0.939) are acceptable in both the derivation and validation cohorts, with their predictive capabilities being comparable. Given the simplicity of the models, namely that “glmBoost + GBM” and “Stepglm [both] + GBM” incorporated 10 (BMI, APACHE II, SOFA, SAPS II, urine output, magnesium, phosphate, BUN, creatinine, and albumin) and 9 (urine output, PNI, creatinine, BUN, BMI, APACHE II, phosphate, magnesium, and SOFA) variables respectively. Regarding the model calibration ability: for the “glmBoost + GBM” model, the Brier score was 0.061 and the calibration slope was 1.010 in the HYCH cohort, while the corresponding values were 0.210 and 0.898 in the MIMIC-IV cohort; for the “Stepglm[both]+GBM” model, the Brier score was 0.062 and the calibration slope was 1.012 in the HYCH cohort, with values of 0.130 and 0.945 in the MIMIC-IV cohort, respectively. These results indicate that the “Stepglm[both]+GBM” model exhibits superior calibration performance compared with the “glmBoost + GBM” model in both cohorts. The ultimate model materialized as a combination of Stepglm[both] and GBM. The “Stepglm[both]+GBM” model serves as a consensus diagnostic model for PPA-AKI designated as the PPA-AKI index, which yields a satisfactory AUC, utilizes fewer variables, better calibration performance and demonstrates greater stability across both cohorts. Detailed information regarding models’ performance, feature selection methodologies for individual models, patient-specific risk scoring, and predictive classification outcomes are depicted in Supplementary material 2.

Figure 3.

Establishment and validation of a consensus diagnostic model for primary peritonitis-associated acute kidney injury (PPA-AKI) via 12 machine learning (ML)-based integrative procedures. (A) ML algorithm combinations of prediction models using the LOOCV framework and further calculated the area under curve (AUC) of each model in HYCH and MIMIC-IV cohorts. The “Stepglm[both]+GBM” model, including 9 feature variables (urine output, PNI, creatinine, BUN, BMI, APACHE II, phosphate, magnesium, and SOFA), serves as a consensus diagnostic model for PPA-AKI designated as the PPA-AKI index, which yields a satisfactory AUC, utilizes fewer variables, and demonstrates greater stability across both cohorts. (B, C) Receiver operating characteristic (ROC) curves with AUC values to evaluate predictive efficacy of PPA-AKI index in HYCH cohort (B) and MIMIC-IV cohort (C).

AUC: area under curve; Enet: elastic network; LASSO: least absolute shrinkage and selection operator; RF: random forest; GBM: gradient boosting machine; SuperPC: supervised principal components; plsRglm: partial least squares regression for generalized linear models, SVM: support vector machine; LOOCV: leave-one-out cross-validation; LDA: linear discriminant analysis; XGBoost: eXtreme Gradient Boosting; BMI: body mass index; APACHEII: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; BUN: blood urea nitrogen; PNI: prognostic nutritional index.

Model evaluation

Firstly, the VIF scores were below 2, tolerance exceeded 0.5, and the correlation coefficients among variables were below 0.5 across both the HYCH and MIMIC-IV cohorts. This indicates no substantial multicollinearity issues among variables within the PPA-AKI model (Supplementary material 4).

Secondly, ROC and PR curve analyses revealed that the PPA-AKI index demonstrated excellent predictive capability, attaining an AUC of 0.976 (95%CI: 0.963–0.989) (Figure 3B) and 0.974 (95%CI: 0.967–0.981) (Figure 4A) in the HYCH cohort and an AUC of 0.901 (95%CI: 0.870–0.931) (Figure 3C) and 0.903 (95%CI: 0.887–0.919) (Figure 4B) in the MIMIC-IV cohort. The sensitivity, specificity, PPV, NPV were 0.958 (95% CI: 0.927–0.988), 0.884 (95% CI: 0.835–0.933), 0.893 (95% CI: 0.847–0.938), and 0.954 (95% CI: 0.921–0.987) in the HYCH cohorts, respectively. Those in MIMIC-IV cohort were 0.886 (95%CI: 0.840–0.932), 0.765 (95%CI: 0.704–0.826), 0.791(95%CI: 0.736–0.847), and 0.870(95%CI: 0.818–0.922), respectively. The FNR, FPR, MCC, precision, accuracy, F1-score, prevalence, and recall were computed across both cohorts and are detailed in Supplementary material 5, demonstrating high predictive performance. The calibration curves exhibited excellent concordance with the reference line (y = x) across both cohorts (Figure 4C and 4D).

Figure 4.

Evaluation of diagnostic value, fitting ability, clinical usefulness and nonlinear relationship of PPA-AKI index. (A, B) precision recall (PR) curves with AUC values to evaluate predictive efficacy of PPA-AKI index in HYCH cohort (A) and MIMIC-IV cohort (B). (C, D) Calibration curves for PPA-AKI index in HYCH cohort (C) and MIMIC-IV cohort (D). X-axis is predicted probability of PPA-AKI. Y-axis is observed probability of PPA-AKI. (E, F) Decision curve analysis was applied to evaluate the clinical usefulness of PPA-AKI index in HYCH cohort (E) and MIMIC-IV cohort (F). The Y-axis represents the net benefit. The black line represents the hypothesis that no patients’ treatment. The X-axis represents the threshold probability. The threshold probability is where the expected benefit of treatment is equal to the expected benefit of avoiding treatment. (G, H) Potential nonlinear for the levels of PPA-AKI index with PPA-AKI risk measured by restricted cubic spline regression with optimal knots in HYCH cohort (G) and MIMIC-IV cohort (H). The brown line and shadow area represent the estimated OR and the 95% CI.

PPA-AKI: primary peritonitis-associated acute kidney injury; AUC: area under curve.

Finally, DCA analysis indicated that when the threshold probabilities ranged between 0.01 and 1.0 in the HYCH cohort, and between 0.05 and 0.94 in the MIMIC-IV cohort, utilizing the PPA-AKI index (predictive model) demonstrated greater net benefits in contrast to “intervention all” or “non-intervention” strategies, which indicates the clinical usefulness of the PPA-AKI model (Figure 4E and 4F). RCS regression demonstrated no significant non-linear relationship between the PPA-AKI index and PPA-AKI, with evidence of a significant linear relationship (p for Overall < 0.001 and P for nonlinearity = 0.188 in HYCH cohort; p for Overall < 0.001 and P for nonlinearity = 0.071 in MIMIC-IV cohort) (Figure 4G and 4H). Inflection point analysis revealed that the PPA-AKI index demonstrated threshold values of 0.47 and 0.45 within the HYCH and MIMIC-IV cohorts, respectively.

In the unadjusted model (without covariate adjustment), univariate logistic regression analysis demonstrated a markedly positive association between PPA-AKI index and PPA-AKI, with an OR and 95%CI of 2.88 (2.69–3.61) in HYCH cohort, and 1.76 (1.59–1.94) in MIMIC-IV cohort (Table 2) (all p < 0.001). After adjusting for all covariates, multivariable logistic regression analysis suggested that PPA-AKI index was an independent predictor of PPA-AKI risk, with an OR and 95%CI of 24.19 (7.71–75.86) in HYCH cohort, and 1.85 (1.31–2.61) in MIMIC-IV cohort (all p < 0.001).

Table 2.

Univariable and multivariable logistic regression analysis for prediction of AKI in PP patients.

	HYCH cohort		MIMIC-IV cohort
	OR (95 % CI)	p-value	OR (95 % CI)	p-value
Coarse model	2.88 (2.69-3.61)	< 0.001*	1.76 (1.59-1.94)	< 0.001*
Model 1	3.15 (2.44-4.08)	< 0.001*	1.85 (1.65-2.08)	< 0.001*
Model 2	3.21 (2.46-4.20)	< 0.001*	1.91 (1.62-2.18)	< 0.001*
Model 3	24.19 (7.71-75.86)	< 0.001*	1.85 (1.31-2.61)	< 0.001*

Coarse model, no covariate was adjusted.

Model 1, Age, sex, weight, height, and BMI were adjusted.

Model 2, Age, sex, weight, height, BMI, pneumonia, COPD, stroke, cancer, T2DM, HF, IHD, HR, RR, Temperature, SpO2, SBP, DBP, and MAP were adjusted.

Model 3, Age, sex, weight, height, BMI, pneumonia, COPD, stroke, cancer, T2DM, HF, IHD, HR, RR, Temperature, SpO2, SBP, DBP, MAP, APACHEII, SOFA, SAPSII, GCS, urine output, balance, WBC, RBC, RDW, PLT, lymphocytes, Hb, lactate, PaCO2, PH, PaO2, base excess, glucose, potassium, sodium, chloride, bicarbonate, phosphate, calcium, magnesium, anion gap, BUN, creatinine, albumin.

Bilirubin, AST, ALT, PNI, INR, and PTT were adjusted.

Abbreviations: MIMIC-IV: Medical Information Mart for Intensive Care IV; ICU: intensive care unite; AKI: acute kidney injury; PP: primary peritonitis; OR: odds ratio; BMI: body mass index; HR: heart rate; SBP: systolic blood pressure; DBP: diastolic blood pressure; MAP: mean arterial pressure; RR: respiratory rate; T2DM: type 2 diabetes mellitus; HF: heart failure; IHD: ischemic heart disease; COPD: chronic obstructive pulmonary disease; APACHEII: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; SAPSII: simplified acute physiology score II; GCS: glasgow coma scale; RBC: red blood cell; WBC: white blood cell; PLT: platelet; Hb: hemoglobin; RDW: red blood cell volume distribution width; PTT: partial thromboplastin time; BUN: blood urea nitrogen; AST: aspartate aminotransferase; ALT: alanine aminotransferase; PNI: prognostic nutritional index; INR: international normalized ratio.

*p < 0.05.

To investigate potential differences across distinct populations, logistic regression analyses were executed within various subgroups. These analyses demonstrated a marked positive link between the PPA-AKI index and PPA-AKI throughout all subgroups (p < 0.05) in both HYCH and MIMIC-IV cohorts, excluding certain subgroups characterized by restricted sample sizes (chronic obstructive pulmonary disease (COPD), n = 15; stroke, n = 5/11; ischemic heart disease (IHD), n = 17) (Table 3 and Supplementary material 6), indicating that the PPA-AKI model demonstrates general robustness. Interaction tests identified several significant interactions between the PPA-AKI index and fluid balance (p = 0.038 for interaction) in the HYCH cohort, along with interactions between the PPA-AKI index and sex (p = 0.004 for interaction), as well as lactate (p = 0.039 for interaction) in the MIMIC-IV cohort.

Table 3.

Subgroup analysis for PPA-AKI index in predicting the risk of AKI after PP in HYCH and MIMIC-IV cohorts.

	HYCH							MIMIC-IV
Variable	Count	Percent	OR	Lower	Upper	p value	P for interaction	Count	Percent	OR	Lower	Upper	p value	P for interaction
Overall	329	100	2.87	2.29	3.61	<0.001*		367	100	1.76	1.59	1.94	<0.001*
Sex							0.973							0.004*
Female	139	42.2	2.92	2.03	4.2	<0.001*		161	43.9	2.28	1.81	2.87	<0.001*
Male	190	57.8	2.9	2.15	3.91	<0.001*		206	56.1	1.56	1.39	1.75	<0.001*
COPD							0.725							0.992
No	314	95.4	2.85	2.27	3.58	<0.001*		328	89.4	1.75	1.58	1.95	<0.001*
Yes	15	4.6	4.05	0.58	28.44	0.159		39	10.6	1.76	1.26	2.45	0.001*
Stroke							0.992							0.383
No	324	98.5	2.85	2.27	3.58	<0.001*		356	97	1.75	1.58	1.94	<0.001*
Yes	5	1.5	NA	NA	NA	1		11	3	7.37	0.29	186.6	0.226
IHD							0.991							0.924
No	312	94.8	2.82	2.25	3.54	<0.001*		309	84.2	1.81	1.61	2.03	<0.001*
Yes	17	5.2	NA	NA	NA	1		58	15.8	1.78	1.33	2.39	<0.001*
DBP (mmHg)							0.263							0.172
< 80	275	83.6	2.81	2.21	3.58	<0.001*		297	80.9	1.82	1.62	2.05	<0.001*
≥ 80	54	16.4	17.81	0.71	448.63	0.08		70	19.1	1.56	1.28	1.89	<0.001*
Balance (mL)							0.038*							0.509
≤ 0	114	34.7	2.26	1.72	2.99	<0.001*		84	22.9	1.9	1.5	2.4	<0.001*
> 0	215	65.3	4	2.52	6.36	<0.001*		283	77.1	1.74	1.55	1.95	<0.001*
Lactate (mmol/L)							0.661							0.039*
< 2	123	37.4	3.12	2.03	4.8	<0.001*		135	36.8	2.14	1.69	2.71	<0.001*
≥ 2	206	62.6	2.79	2.13	3.65	<0.001*		232	63.2	1.63	1.45	1.82	<0.001*
Albumin (g/dL)							0.993							0.931
< 3.5	294	89.4	2.78	2.21	3.5	<0.001*		264	71.9	1.76	1.56	1.99	<0.001*
≥ 3.5	35	10.6	NA	NA	NA	0.999		103	28.1	1.78	1.47	2.15	<0.001*
INR							0.274							0.939
≤ 1.2	57	17.3	20.41	6.62	736.05	0.099		49	13.4	1.76	1.27	2.45	0.001*
> 1.2	272	82.7	2.75	2.52	3.46	<0.001*		318	86.6	1.74	1.56	1.93	<0.001*

Abbreviations: MIMIC-IV: Medical Information Mart for Intensive Care IV; ICU: intensive care unite; AKI: acute kidney injury; PP: primary peritonitis; OR: odds ratio; PPA-AKI: primary peritonitis-associated acute kidney injury; NA: not avaliable; IHD: ischemic heart disease; COPD: chronic obstructive pulmonary disease; DBP: diastolic blood pressure; INR: international normalized ratio.

*p < 0.05.

Sensitivity analysis

PPA-AKI was developed and evaluated using a derivation cohort (MIMIC-IV cohort) and a validation cohort (HYCH cohort) through an innovative LOOCV framework. The AUC values were calculated for individual algorithms and the average AUC across both derivation and validation cohorts (Figure S2). We found that all predictive models exhibited substantial AUC values in the derivation cohort but demonstrated lower performance in the validation cohort, indicating overfitting. Detailed information regarding feature selection, predictive classification, and individual risk scores is provided in Supplementary material 3.

Sensitivity analysis through excluding urine output and creatinine found that the PPA-AKI index exhibited acceptable predictive performance, with an AUC of 0.846 (95%CI: 0.808–0.885) in HYCH cohort, 0.793 (95%CI: 0.745–0.841) in MIMIC-IV cohort, respectively (Figure S3A and S3B). The sensitivity, specificity, PPV, NPV were 0.729 (95%CI: 0.662–0.797), 0.858 (95%CI: 0.807–0.909), 0.828 (95%CI: 0.768–0.888), and 0.727 (95%CI: 0.667–0.786) in HYCH cohort, respectively. Those in MIMIC-IV cohort were 0.666 (95%CI: 0.602–0.730), 0.824 (95%CI: 0.770–0.878), 0.806 (95%CI: 0.737–0.876), and 0.683 (95%CI: 0.619–0.747), respectively. This suggests that the PPA-AKI model demonstrates satisfactory predictive performance and can serve as an effective diagnostic tool for PPA-AKI.

Sensitivity analysis through excluding patients diagnosed with AKI within 12 h of admission indicated that the PPA-AKI index exhibited excellent predictive performance, with an AUC of 0.955 (95%CI: 0.932–0.979) in HYCH cohort, 0.881 (95%CI: 0.845–0.917) in MIMIC-IV cohort, respectively (Figure S3C and S3D). The sensitivity, specificity, PPV, NPV were 0.938 (95% CI: 0.917–0.949), 0.852 (95% CI: 0.826–0.878), 0.843 (95% CI: 0.807–0.879), and 0.941 (95% CI: 0.909–0.973) in HYCH cohort, respectively. Those in MIMIC-IV cohort were 0.876 (95%CI: 0.826–0.925), 0.741 (95%CI: 0.690–0.792), 0.775(95%CI: 0.716–0.834), and 0.854(95%CI: 0.797–0.912), respectively.

Model interpretation

We employed the SHAP algorithm to highlight the predictive importance of chosen variables in the optimal PPA-AKI model. The relative importance and contribution of the 9 features into the PPA-AKI model in HYCH and MIMIC-IV cohorts are shown in Figure 5A and 5B, which was obtained by the SHAP algorithm interpretation. From the two cohorts, it can be noted that the top 5 variables are urine output, PNI, BUN, phosphate and creatinine.

Figure 5.

Global and local explanation by the Shapley Additive Explanations (SHAP) method for PPA-AKI model in HYCH and MIMIC-IV cohorts. (A, B) Summary plot showed the 9 features ranking by mean absolute SHAP values in HYCH cohort (A) and MIMIC-IV cohort (B). (C, D) Global model explanation in HYCH cohort (C) and MIMIC-IV cohort (D). Each variable name is shown on the left-hand side with the variable with the greatest contribution listed at the top. To the right of the variables, there are colored lines, which are individual points that correspond to observations in the population. A higher value for the variable is represented in yellow, while a lower value for the variable will be shown in purple. A value farther to the right (i.e. a higher SHAP value) indicates that the variable contributes to a prediction of a positive target, such as PPA-AKI. (E, F) An example of risk factor analysis for a patient with PPA-AKI which represented the individual PP toward the “AKI” class in HYCH cohort (E) and MIMIC-IV cohort (F).

PPA-AKI: primary peritonitis-associated acute kidney injury; BMI: body mass index; APACHEII: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; BUN: blood urea nitrogen; PNI: prognostic nutritional index.

Global explanations: Figure 5C and 5D provide visual representations of the range and distribution of the 9 characteristics in relation to model performance. Features demonstrating positive correlations encompassed: BUN, phosphate and creatinine (the higher the value of these features, the higher the probability of developing PPA-AKI); while the negatively correlated features are urine output and PNI. It is notable that features can drive predictions in different directions (increasing or decreasing PPA-AKI) for patients with different features, where the influence of a given feature value on the prediction is invariant.

Local explanations: To better interpret the individualized PPA-AKI risk, we analyzed the predictive variables and the risk sources specified by the SHAP values. For the PPA-AKI patient in HYCH cohort with the higher predicted SHAP value (1.68), PNI (18.9, SHAP value = 2.07), urine output (825 mL, SHAP value = 1.09), and APACHE II (19, SHAP value = 0.202) were the sources of risk (Figure 5E). A low PNI value and reduced urine output resulted in an elevated SHAP value. Similarly, another PPA-AKI patient exhibiting a higher predicted SHAP value (2.28) within the MIMIC-IV cohort was displayed in Figure 5F.

Finally, the partial dependence plots offered a visual depiction of the global association between features and risk distribution. via these plots, we distinctly recognized the linear or nonlinear correlations and variational patterns between BMI, APACHE II, SOFA, phosphate, magnesium, PNI, urine output, BUN, creatinine and SHAP values throughout both cohorts (Figure 6 and Figure S4).

Figure 6.

One-way SHAP dependence plot of the 9 feature predictors in HYCH cohort. (A) BMI; (B) APACHE II; (C) SOFA; (D) phosphate; (E) magnesium; (F) PNI; (G) urine output; (H) BUN; (I) creatinine. Each dependence plot shows how a single feature affects the output of the prediction model, and each dot represents a single patient. Specifically, the values of the predictor are represented by the x-axis, and its SHAP values are represented by the y-axis. To interpret these plots, for example, in (A), patients with higher BMI (as x-axis increased) were associated with a higher SHAP value, which indicated a higher likelihood of PPA-AKI (y-axis also increased). BMI: body mass index; APACHE II: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; PNI: prognostic nutritional index; BUN: blood urea nitrogen.

SHAP: Shapley Additive Explanations; BMI: body mass index; APACHEII: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; BUN: blood urea nitrogen; PNI: prognostic nutritional index.

Logistic regression, RCS regression and threshold effect analysis

To further explore the significance of features within the PPA-AKI model, we conducted univariable logistic and RCS regression analyses to examine the linear and nonlinear relationships across two cohorts. Supplementary material 7 offered the optimized knot for RCS analysis. Consequently, three and nine predictors were found to exhibit non-linear and linear associations with the risk of PPA-AKI respectively (Table 4), with these findings illustrated in Figure 7 and Figure S5. Magnesium (p for nonlinearity = 0.035 or 0.048), BUN (P for nonlinearity ≤ 0.002), and creatinine (p for nonlinearity < 0.001) displayed marked nonlinear links to PPA-AKI index in HYCH and MIMIC-IV cohorts (Figure 7). Nevertheless, BMI (P for nonlinearity = 0.234 or 0.275), APACHE II (P for nonlinearity = 0.528 or 0.711), SOFA (p for nonlinearity = 0.689 or 0.425), phosphate (p for nonlinearity = 0.133 or 0.032), PNI (p for nonlinearity = 0.142 or 0.050), urine output (p for nonlinearity = 0.005 or 0.958) demonstrated significant linear yet no significant nonlinear correlations with PPA-AKI index. Comprehensive details regarding RCS regression analysis were presented in Supplementary material 8.

Table 4.

Nonlinear and linear associations between clinical variables and PPA-AKI.

Cohorts	HYCH				MIMIC
Factors	Linear association		Nonlinear association		Linear association		Nonlinear association
	OR (95%CI)	p	Chi-square	p	OR (95%CI)	p	Chi-square	p
BMI (kg/m2)	1.047 (1.016-1.079)	0.002*	1.417	0.234	1.057 (1.024-1.091)	0.001*	1.190	0.275
APACHEII	1.099 (1.064-1.135)	< 0.001*	0.397	0.528	1.178 (1.132-1.226)	< 0.001*	0.137	0.711
SOFA	1.170 (1.101-1.244)	0.003*	0.160	0.689	1.363 (1.271-1.463)	< 0.001*	0.625	0.429
Phosphate (mmol/L)	1.473 (1.271-1.708)	< 0.001*	5.591	0.133	1.290 (1.136-1.465)	< 0.001*	6.903	0.032*
Magnesium (mmol/L)	2.197 (1.291-3.740)	0.004*	10.392	0.035*	2.715 (1.663-4.430)	< 0.001*	6.076	0.048*
PNI	0.969 (0.950-0.988)	0.001*	8.193	0.146	0.899 (0.876-0.923)	< 0.001*	9.517	0.050
Urine output (mL)	0.995 (0.990-0.999)	< 0.001*	16.847	0.005*	0.996 (0.991-0.999)	< 0.001*	0.003	0.958
BUN (mg/dL)	1.030 (1.020-1.040)	< 0.001*	23.837	< 0.001*	1.042 (1.030-1.054)	< 0.001*	19.395	0.002*
Creatinine (mg/dL)	1.171 (1.045-1.312)	0.007*	53.736	< 0.001*	1.299 (1.133-1.490)	< 0.001*	86.188	< 0.001*

Abbreviations: MIMIC-IV: Medical Information Mart for Intensive Care IV; PPA-AKI: primary peritonitis-associated acute kidney injury; OR: odds ratio; BMI: body mass index; APACHEII: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; BUN: blood urea nitrogen; PNI: prognostic nutritional index.

*p < 0.05.

Figure 7.

Potential nonlinear association for the levels of 9 predictors with PPA-AKI risk measured by restricted cubic spline regression with optimal knots in HYCH cohorts. (A) BMI; (B) APACHE II; (C) SOFA; (D) phosphate; (E) magnesium; (F) PNI; (G) urine output; (H) BUN; (I) creatinine.

BMI: body mass index; APACHEII: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; BUN: blood urea nitrogen; PNI: prognostic nutritional index.

Moreover, threshold effect analysis was executed to ascertain the inflection point for BMI (28.16 or 28.10 kg/m²), APACHE II (18), SOFA (4), phosphate (3 mg/dL), magnesium (2 mg/dL), PNI (36.32 or 36.76), urine output (615 or 622 mL), BUN (23 mg/dL), and creatinine (1.3 mg/dL) in HYCH and MIMIC-IV cohorts. Specifically, when the values of BMI, APACHE II, SOFA, phosphate, magnesium, BUN, and creatinine were greater than certain thresholds or the values of PNI and urine output were smaller than certain thresholds, the likelihood of having PPA-AKI risk was markedly higher. Comprehensive details regarding threshold effect analysis were presented in Supplementary material 9.

Clinical utility of the PPA-AKI model

Utilizing the 9 predictive features identified, we developed a web-based computational platform designed to support clinical decision-making at the point of care for PPA-AKI management. This online calculator is freely accessible via https://lglcz.shinyapps.io/DynMod/.

Discussion

PPA-AKI constitutes a common and serious complication, leading to adverse outcomes in PP patients. At present, PP lacks an efficient ML model for precise identification of AKI risk and provision of a vital opportunity for AKI prevention, thus enhancing PP survival. This study included two large-scale cohorts, HYCH (n = 329) and MIMIC-IV (n = 367), to develop and validate a consensus and optimized PPA-AKI model based on Stepglm[both]+GBM algorithms using routine clinical indicators (BMI, APACHE II, SOFA, phosphate, magnesium, PNI, urine output, BUN, and creatinine) at admission. The PPA-AKI model provides a cost-efficient, dependable, interpretable and readily accessible diagnostic tool for PPA-AKI prevention, showing promising clinical application value. Targeted monitoring and early intervention of these modifiable risk factors may prevent AKI in PP patients.

A pivotal component that bolstered our predictive model’s generalizability and robustness is our recently developed LOOCV framework. This study utilized two comprehensive PP cohorts to enhance statistical power, ensuring model dependability and precision. Moreover, in the existing extensive research on AKI prediction models, researchers predominantly select modeling algorithms based on their own preferences and level of expertise [22,23]. To remedy this shortcoming, we integrated 12 classic ML algorithms into 113 algorithm combinations through the LOOCV framework, with variable selection capabilities and data dimensionality reduction. This facilitates the generation of consensus and optimal diagnostic models based on model performance, fitting and simplicity. It is worth noting that overfitting represents one of the formidable challenges encountered by AI and ML during biomedical model development, whereby multiple models exhibit satisfactory performance within the derivation cohort whilst displaying suboptimal outcomes in different external validation cohorts. In our research, we observed that the top five models based on average AUC of two cohorts fitted the derivation dataset well, yet failed to do so when validated using the validation dataset. The subsequent two excellent models (“glmBoost + GBM” and “Stepglm[both]+GBM”) did not exhibit significant overfitting issues. Given the simplicity of the model, the final approach materialized as a combination of Stepglm[both] and GBM utilizing nine predictor variables while avoiding substantial multi-collinearity issues. Promisingly, via ROC, PR, calibration curve, and DCA evaluation, the PPA-AKI index demonstrated superior predictive capability, model adequacy and clinical utility value. After covariate adjustment, multivariable logistic regression demonstrated that the PPA-AKI index remained an independent AKI risk predictor in PP patients. Our study evaluated model efficacy within different subgroups via interaction testing to detect possible dataset biases. Consistent findings across subgroups confirmed the PPA-AKI model’s generalizability and robustness.

The ML technique represents a “black-box” approach with insufficient transparency in prediction generation [24]. This may cause clinicians to avoid its use, as they hesitate to base medical decisions on unclear information. Thus, interpretability of medical diagnostic frameworks remains crucial for physician acceptance. We initially applied the SHAP methodology [,25] to clarify the “black-box” nature of PPA-AKI frameworks in forecasting AKI amongst PP patients. The SHAP methodology facilitates model interpretation through comprehensive explanations that describe the overall functionality of the model and specific explanations that detail how individual predictions are formed for particular PPA-AKI risk utilizing personalized data inputs. In our study, SHAP analysis ranks the feature importance that the top 5 variables are urine output, PNI, BUN, phosphate and creatinine for the PPA-AKI model. This finding aligns with the observation that urine output, PNI, BUN, and creatinine emerge as the most significant variables during feature selection via Boruta’s algorithm. Furthermore, utilizing 1-way SHAP dependence plots in conjunction with RCS regression, we clearly observed that typically, the SHAP value (i.e. the risk of PPA-AKI) increases as the BMI, APACHE II, SOFA, phosphate, magnesium, BUN, and creatinine increase, as well as PNI and urine output decrease. Most significantly, a web-based computational framework was established employing the optimal PPA-AKI model to facilitate accurate identification and risk assessment of AKI in PP patients. This patient-specific interpretable framework continues to be freely accessible to clinicians, allowing them to integrate it with their clinical expertise to enhance decision-making processes.

To enhance the clinical utility of these predictive factors within the PPA-AKI model, we employed RCS regression alongside threshold effect analysis to elucidate the non-linear associations between PPA-AKI and its predictors, whilst identifying critical inflection points for these factors. Previous studies have identified these extrarenal indicators, such as BMI [26], APACHE II [27], SOFA [28], phosphate [29] magnesium [30] and malnutrition [31] as important predictors of AKI. Meanwhile, the pathophysiological roles of these markers in AKI have also been elucidated, such as obesity in PP patients may lead to kidney dysfunction via a combination of venous congestion and compromised arterial perfusion; hyperphosphatemia promotes the deposition of calcium-phosphorus crystals, resulting in direct nephrotoxic effects [32]; hypermagnesemia exacerbates renal injury by affecting creatinine clearance and magnesium excretion [33]. Our research demonstrates for the first time that PNI served as a simple and objective measure for quantifying and assessing patients’ nutritional status, and represented one of the most important predictive factors for PPA-AKI. Additionally, there is a significant non-linear association between magnesium, blood urea nitrogen and creatinine, and the PPA-AKI, while BMI, APACHE II, SOFA, and PNI showed linear correlations with PPA-AKI. Notably, a non-linear association between urine output and PPA-AKI in the HYCH cohort, but a linear association in the MIMIC-IV cohort. Similarly, phosphate showed a linear relationship with PPA-AKI in the HYCH cohort, whereas the relationship was non-linear in the MIMIC-IV cohort. This may be attributed to: 1) Differences in study populations: the HYCH cohort is composed of the Chinese population, while the MIMIC-IV cohort is made up of the American population; 2) Differences in fluid management: patients in the HYCH cohort had a greater positive fluid balance (mean 3220 mL), whereas that in the MIMIC-IV cohort was relatively lower (mean 2207 mL). Most importantly, threshold effect analysis was performed to determine the corresponding inflection points for each predictor: BMI, APACHE II, SOFA, phosphate, magnesium, urine output, BUN, and creatinine, with values of 28.16 or 28.10 kg/m², 18, 4, 3 mg/dL, 2 mg/dL, 36.32 or 36.76, 615 or 622 mL, 23 mg/dL, 1.3 mg/dL, respectively. Inflection point of these modifiable predictors can combine with clinicians’ empirical knowledge to facilitate decision-making to reduce PPA-AKI risk, such as when phosphate exceeds 3 mg/dL or magnesium exceeds 2 mg/dL, clinicians should correct electrolyte disturbances by identifying the underlying cause, discontinuing magnesium- or phosphate-containing medications, and administering diuretics, among other measures; when the PNI is below 36.72, clinicians should consider nutritional supplementation (albumin, energy intake, vitamins, etc.); when BUN exceeds 23 mg/dL or creatinine 1.3 mg/dL, clinicians should review nephrotoxic medications, optimize fluid status, consider nephrology consultation; when urine output falls below 622 mL, clinicians should administer intravenous fluids, improve renal perfusion, etc. As these findings have not yet been implemented in clinical practice, we are unable to draw a clear utility. It is necessary to validate these findings in future clinical studies.

Relative to previously published studies, our research exhibits several unique features. (1) This represents the first development of PPA-AKI prediction models using standard clinical parameters at admission across two large cohorts. Our predictive framework, incorporating routine clinical markers, can efficiently guide individualized interventions. (2) Through implementing Boruta’s feature selection methodology and 12 ML techniques, 113 combinations were created to construct consensus-based and robust predictive models, enabling identification of significant variables, preventing overfitting, and selecting optimal models. (3) To tackle issues related to the “black-box” nature and restricted interpretability of ML approaches, the SHAP methodology was employed. SHAP clarifies feature significance rankings whilst visualizing both global and individual risk assessments. (4) RCS regression combined with threshold effect evaluation demonstrates nonlinear associations between PPA-AKI and predictors, whilst establishing inflection points for each predictive factor. Based on predictor thresholds, clinicians may implement timely interventions to diminish PPA-AKI incidence. (5) To offer a more adaptable predictive instrument, the PPA-AKI model was developed utilizing a freely available, publicly accessible web-based calculator for clinical implementation.

Despite our efforts to execute this research with utmost rigor and thoroughness, certain limitations require recognition. First, being a retrospective investigation, it lacks prospective verification, rendering our conclusions potentially susceptible to inherent biases. Notwithstanding encouraging external validation outcomes, the proposed model remains unsuitable for widespread application until validation through prospective, multicenter, large-scale cohort studies. Second, this study included only critically ill patients with primary peritonitis; the generalizability of this model still needs to be further validated in a population of patients in general medical wards. Third, the predictive model fundamentally represents a correlational instrument rather than a causal mechanism, potentially compromising the precision and applicability of clinical decision-making processes. Causal ML models (e.g. causal forest) are introduced to fully elucidate the causal relationships between predictor variables and the development of PPA-AKI. Fifth, prediction models constructed upon predetermined temporal frameworks rather than real-time clinical modifications may compromise the practicality of implementing these AKI models within clinical settings. It will be necessary in the future to use time-series data to build a dynamic prediction model based on different time windows. Fourth, the clinical effectiveness of the model, that is, whether it can truly improve patient outcomes after being integrated into the workflow, still needs to be further confirmed by prospective intervention studies. Sixth, novel biomarkers including NGAL, KIM-1 and TIMP-2 × IGFBP7 were not incorporated owing to limited availability in routine clinical data, yet our framework can easily facilitate their future integration. Seventh, the correlations between these predictor variables (such as BMI, phosphate, magnesium and PNI) and PPA-AKI have been elaborately clarified. Yet, the fundamental physiological mechanisms by which these variables cause PPA-AKI remain unclear. Future investigation of PPA-AKI mechanisms will necessitate further experiments concerning physiological pathways and supplementary randomized controlled trials.

Conclusion

This investigation developed and confirmed a consensus diagnostic framework utilizing Stepglm[both]+GBM algorithms for predicting PPA-AKI incorporating multiple adjustable parameters. SHAP analysis elucidates feature significance rankings whilst providing visualized global and individualized interpretations for the PPA-AKI framework. The PPA-AKI framework represents a cost-effective, precise, interpretable, and readily available instrument for PPA-AKI diagnosis with promising clinical utility. Timely identification and management of these aberrant and modifiable risk factors shall assist in preventing PPA-AKI occurrence.

Abbreviation

PPA-AKI: primary peritonitis-associated acute kidney injury; PP: primary peritonitis; SBP: spontaneous bacterial peritonitis; AKI: acute kidney injury; KDIGO: Kidney Disease: Improving Global Outcome; MIMIC-IV: Medical Information Mart for Intensive Care IV; EMR: electronic medical records; PNI: prognostic nutritional index; ML: machine learning; GCS: Glasgow Coma Scale; AI: artificial intelligence; RCS: restricted cubic spline; SHAP: Shapley Additive Explanations; CI: confidence interval; OR: odds ratio; SE: standard error; ROC: receiver operating characteristic; AUC: area under curve; Enet: elastic network; LASSO: least absolute shrinkage and selection operator; RF: random forest; GBM: gradient boosting machine; SuperPC: supervised principal components; plsRglm: partial least squares regression for generalized linear models, SVM: support vector machine; LOOCV: leave-one-out cross-validation; DCA: decision curve analysis; LDA: linear discriminant analysis; XGBoost: eXtreme Gradient Boosting; GBM: gradient boosting machine; TRIPOD: Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis; MCC: Matthews correlation coefficient; FNR: false negative rate; FPR: false positive rate; SD: standard deviation; BMI: body mass index; HR: heart rate; SBP: systolic blood pressure; DBP: diastolic blood pressure; MAP: mean arterial pressure; RR: respiratory rate; T2DM: type 2 diabetes mellitus; HF: heart failure; IHD: ischemic heart disease; COPD: chronic obstructive pulmonary disease; APACHE II: acute physiology and chronic health evaluation II; SOFA: sequential organ failure assessment; APS III: acute physiological score III; SIRS: systemic inflammatory response syndrome; SAPS II: simplified acute physiology score II; OASIS: oxford acute severity of illness score; GCS: glasgow coma scale; RBC: red blood cell; WBC: white blood cell; PLT: platelet; Hb: hemoglobin; RDW: red blood cell volume distribution width; PTT: partial thromboplastin time; BUN: blood urea nitrogen; AST: aspartate aminotransferase; ALT: alanine aminotransferase; INR: international normalized ratio; LOS: length of stay

Funding Statement

This work was supported by the Natural Science Foundation of Hunan Province (No. 2026JJ81630).

Ethical approval and consent to participate

The studies involving human participants from MIMIC-IV database were reviewed and approved by the Institutional Review Boards of Beth Israel Deaconess Medical Center (Boston, MA) and the Massachusetts Institute of Technology (Cambridge, MA). The studies involving human participants from HYCH database were reviewed and approved by the Institutional Review Boards of the Affiliated Hengyang Hospital of Hunan Normal University & Hengyang Central Hospital (No. 2025(204)), adhering to the ethical standards established in the World Medical Association Declaration of Helsinki. The need for informed consent was waived based on the study’s retrospective design and anonymized data usage.

Consent for publication

Not applicable.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The original data presented in the study are included in the article/Supplementary material. Supplementary materials and figures used are stored at the Zenodo repository https://doi.org/10.5281/zenodo.19449559. MIMIC-IV database is freely available at https://mimic-iv.mit.edu/. Further inquiries can be directed to the corresponding author.