Utilization of interpretable machine learning model to forecast the risk of major adverse kidney events in elderly patients in critical care

Abstract

Major adverse kidney events within 30 d (MAKE30) implicates poor outcomes for elderly patients in the intensive care unit (ICU). This study aimed to predict the occurrence of MAKE30 in elderly ICU patients using machine learning. The study cohort comprised 2366 elderly ICU patients admitted to the Second Xiangya Hospital of Central South University between January 2020 and December 2021. Variables including demographic information, laboratory values, physiological parameters, and medical interventions were used to construct an extreme gradient boosting (XGBoost) -based prediction model. Out of the 2366 patients, 1656 were used for model derivation and 710 for testing. The incidence of MAKE30 was 13.8% in the derivation cohort and 13.2% in the test cohort. The average area under the receiver operating characteristic curve of the XGBoost model was 0.930 (95% CI: 0.912–0.946) in the training set and 0.851 (95% CI: 0.810–0.890) in the test set. The top 8 predictors of MAKE30 tentatively identified by the Shapley additive explanations method were Acute Physiology and Chronic Health Evaluation II score, serum creatinine, blood urea nitrogen, Simplified Acute Physiology Score II score, Sequential Organ Failure Assessment score, aspartate aminotransferase, arterial blood bicarbonate, and albumin. The XGBoost model accurately predicted the occurrence of MAKE30 in elderly ICU patients, and the findings of this study provide valuable information to clinicians for making informed clinical decisions.

Keywords:

• Critical illness
• major adverse kidney events
• the elderly
• XGBoost
• machine learning

Introduction

Medical advancements have significantly prolonged human life expectancy [1], leading to a growth of critically ill patients over the age of 65 [2]. Such patients often experience a high rate of renal adverse events, and to better reflect overall renal outcomes, the use of a composite and objective clinical outcome indicator called major adverse kidney events (MAKE30) is recommended [3]. Studies have shown that the incidence of MAKE30 in intensive care unit (ICU) patients ranges from 13.6% to 38.9% [4–10]. The incidence might be higher in the elderly population due to their increased comorbidities and higher severity of illness [11,12].

Early identification of patients at high risk of adverse kidney outcomes is essential for clinicians to adjust clinical decisions and reduce the risk of MAKE30 [5]. Several factors have been found to correlate with the occurrence of MAKE30 in critically ill patients, including serum chloride [8], serum renin [13], plasma Galectin-3 [10], positive liquid balance [14], and abnormal hepatic or portal venous Doppler [7]. Logistic regression models have been developed for the prediction of MAKE30 in critically ill patients with acute kidney injury (AKI) [15], acute pancreatitis [16], and those receiving veno-arterial extracorporeal membrane oxygenation [17]. However, most of these models were based on a logistic regression approach and were not been established in critically ill elderly patients.

Machine learning has emerged as a promising approach for clinical prognosis evaluation due to its ability to handle complex nonlinear relationships [18–20]. One such algorithm, extreme gradient boosting (XGBoost), has recently dominated applied machine learning and structured data competitions. Its interpretability has made it a popular choice in the medical field. Previous studies have shown that XGBoost performed well in predicting prognosis of elderly ICU patients [20–22]. Thus, the aim of this study is to establish a prediction model for MAKE30 in elderly patients in the ICU using the XGBoost algorithm.

Methods

Ethics approval

This study followed the Declaration of Helsinki and the transparent reporting statement for individual prognostic or diagnostic multivariate prediction models [23]. Central South University’s Medical Ethics Committee (2022-K031) approved the study protocol. Due to the retrospective nature of the study, informed consent was waived.

Data source

This cohort study included all elderly patients admitted to the ICUs of the Second Xiangya Hospital of Central South University from January 1, 2020, to December 31, 2021. Medical information of the ICU patients was extracted from electronic medical record system.

Patient population

Patients with age ≥ 65 years who were admitted to the ICU were included in the study. The following patients were excluded: (1) length of ICU stay < 48 h; (2) patients with chronic kidney diseases (CKD) stage 5 or uremia; (3) patients on maintenance hemodialysis; (4) patients with creatinine values less than once during hospitalization; (5) kidney transplant recipients. For patients who were admitted to the ICU multiple times within one year, only the first admission information was recorded.

Feature selection

We collected patients’ demographic information, comorbidities, physiological parameters, invasive and noninvasive laboratory indicators, medical interventions, and clinical scores as potential predictive variables. The first values of physiological parameters and laboratory indicators within 24 h since admission to the ICU were used, and medical interventions were identified within 48 h since admission to the ICU. Sequential Organ Failure Assessment (SOFA) score, Simplified Acute Physiology Score II (SAPS II), and Acute Physiology and Chronic Health Evaluation II (APACHE II) score were calculated based on the first measurement of the patient within 24 h after admission to the ICU. Baseline creatinine was defined as the lowest creatinine value in the 12 months before admission and was estimated with the use of the following formula if no value was available. [creatinine = 0.74 − 0.2 (if female) + 0.08 (if African American) + 0.003 × age (in years)]. If a patient had documented CKD, it was defined as the lowest creatinine value during hospitalization [6,9,24,25]. Baseline glomerular filtration rates were calculated using the Modification of Diet in takes diseases equation estimation [26]. AKI was determined following the 2012 Kidney Disease: Improving Global Outcomes Clinical Practice Guidelines [27].

Outcome assessment

The main outcome is the occurrence of MAKE30, which is a compound outcome, including (1) emergent dialysis; (2) the last creatinine at hospitalization was 200% of baseline creatinine (persistent kidney impairment); (3) death [7,14]. Outcomes were obtained at discharge (if discharged within 30 d) or on the 30th day since admission to the ICU. The length of hospital stay and ICU stay were secondary outcomes.

Data set partitioning and statistical analysis

The patients were randomly divided into a training group and a test group in a 7:3 ratio by caret package in R. Patients’ characteristics are described as numbers (percentages) for categorical variables and medians (interquartile ranges) for continuous variables. We compared the distributions of continuous variables using the Mann–Whitney U test and categorical variables using the chi-square test. Categorical variables were transformed in advance to factor variables. The missForest package in R was used to impute missing data using the random forest method [28]. Each variable’s data type and proportion of missing values are listed in Table S1.

Feature selection

Variables with missing proportions greater than 30% are eliminated. The nearZeroVar function was applied to identify variables with variance close to zero and exclude them from further calculations. For the final model, the variables that were chosen by at least three methods, including the Least Absolute Shrinkage and Selection Operator, Boruta algorithm, random forest-recursive feature elimination, and random forest-filtering, were selected. The variance inflation factor (VIF) was calculated to identify multicollinearity.

Modeling and testing

We trained the XGBoost model using 5 random shuffles of 5-fold cross-validation and evaluated the model performance based on the average area under the receiver operating characteristic curve (AUROC). The performance of the final prediction model was further evaluated in the test set using AUROC, the area under the precision-recall curve (AUPRC), and calibration. The modeling framework is shown in Figure S1. Table S2 describes the functions, packages, and tuning parameters used.

Interpretability of the model

We use the Shapley additive explanations (SHAP) method to try to open the black box of machine learning [29]. It can preliminarily provide interpretability for the model, and quantify the influence of each feature on the prediction made by the model [30,31].

Sensitivity analysis

Model performance was assessed in different subgroups. We focused on patients with different age groups (≤ 79 years old and > 79 years old), different types of hospitalization (medical hospitalization and surgical hospitalization), different severity of illness (APACHE II score ≤ 16 and APACHE II score > 16), and different time to endpoint (time to endpoint ≤ 13 d and > 13 d). For clinical translation, low and high-risk thresholds were delineated and the corresponding statistical indicators were calculated. To evaluate the impact of dataset imbalance on model construction, upsampling and downsampling models were established using traincontral function in R.

Results

Characteristics of the study population

A total of 2366 patients were identified in this study, with 1656 patients divided into the derivation cohort and the remaining 710 patients into the test cohort (Figure 1). The characteristics of the derivation and test cohorts are shown in Table 1. In the derivation cohort, 229 patients developed MAKE30, of which 53 patients died, 128 patients had persistent kidney impairment, and 141 patients received new dialysis. In the test cohort, 94 patients occurred MAKE30, of whom 22 patients died, 59 patients had persistent kidney impairment, and 57 patients received new dialysis. Comparisons of baseline characteristics in patients with and without MAKE30 in the derivation and test cohorts were shown in Tables S3 and S4 respectively.

Figure 1. Flow diagram of patient selection. CKD: chronic kidney diseases; ICU: intensive care unit; SCr: serum creatinine.

Table 1. Baseline characteristics and outcomes of the patients in the derivation and test cohorts.

Variables	Derivation cohort (n = 1656)	Test cohort (n = 710)	p value
Demographics
Age, year	72 (68–77)	72 (68–78)	0.98
Gender (Female), n (%)	545 (32.9)	248 (34.9)	0.37
Days before ICU admission, day	0 (0–4)	0 (0–4)	0.77
Complication, n (%)
Acute kidney injury at ICU admission	140 (8.5)	80 (11.3)	0.037
Congestive heart failure	210 (12.7)	113 (15.9)	0.042
Chronic heart failure (NYHA class IV)	89 (5.4)	51 (7.2)	0.11
Peripheral vascular disease	19 (1.1)	7 (1.0)	0.90
Cerebrovascular disease	513 (31.0)	204 (28.7)	0.30
Diabetes	440 (26.6)	192 (27.0)	0.85
Chronic lung disease	389 (23.5)	143 (20.1)	0.08
Mild liver disease	385 (23.2)	169 (23.8)	0.81
Moderate or severe liver disease	47 (2.8)	22 (3.1)	0.83
Chronic kidney disease	238 (14.4)	99 (13.9)	0.83
Chronic kidney disease stage 3/4	169 (10.2)	75 (10.6)	0.85
Malignancy	443 (26.8)	186 (26.2)	0.82
Leukemia	14 (0.8)	1 (0.1)	0.09
Metastatic tumor	64 (3.9)	24 (3.4)	0.65
Sepsis	163 (9.8)	82 (11.5)	0.24
Hypertension	890 (53.7)	406 (57.2)	0.14
Dyslipidemia	147 (8.9)	59 (8.3)	0.71
Gastrointestinal bleeding	162 (9.8)	69 (9.7)	1.00
Clinical intervention, n (%)
Vancomycin	30 (1.8)	19 (2.7)	0.23
Aminoglycosides	66 (4.0)	28 (3.9)	1.00
Hydroxyethyl starch	209 (12.6)	93 (13.1)	0.80
Furosemide	887 (53.6)	381 (53.7)	1.00
Infusion of red blood cells	176 (10.6)	79 (11.1)	0.78
Mechanical ventilation	962 (58.1)	392 (55.2)	0.21
Contrast agent	743 (44.9)	330 (46.5)	0.50
Infusion of vasopressors	663 (40.0)	318 (44.8)	0.035
Laboratory tests
Baseline SCr (µmol/L), median [IQR]	83.5 (68.4–85.8)	83.7 (68.2–85.8)	0.70
Baseline eGFR (mL/min/1.73 m^2), median [IQR]	87.2 (82.9–89.5)	87.0 (81. 9–90.1)	0.91
SCr at admission (µmol/L), median [IQR]	82.7 (65.1–111.0)	85.0 (66.1–115.5)	0.33
Neutrophil-lymphocyte ratio, median [IQR]	5.84 (3.00–11.93)	5.80 (3.11–12.46)	0.42
Serum sodium (mmol/L), median [IQR]	139.6 (136.8–142.3)	139.8 (136.9–142.4)	0.76
Serum potassium (mmol/L), median [IQR]	4.04 (3.67–4.40)	3.98 (3.65–4.36)	0.11
Serum chlorine (mmol/L), median [IQR]	104.6 (100.9–107.9)	104.4 (100.8–107.7)	0.38
Alanine aminotransferase (U/L) median [IQR]	19.2 (12.3–34.8)	19.2 (12.5–33.5)	0.66
Aspartate aminotransferase (U/L) median [IQR]	27.0 (19.6–46.2)	27.2 (19.7–44.4)	0.78
Albumin (g/L) median [IQR]	33.1 (28.9–37.2)	33.4 (29.3–37.0)	0.45
Total bilirubin (µmol/L), median [IQR]	11.6 (8.1–17.3)	11.4 (7.9–17.2)	0.46
Blood urea nitrogen (mmol/L) median [IQR]	7.26 (5.40–10.36)	7.40 (5.33–11.10)	0.53
SCr (µmol/L), median [IQR]	82.0 (64.6–115.0)	85.6 (65.9–119.0)	0.26
Serum uric acid (µmol/L), median [IQR]	286 (199–395)	305 (199–401)	0.21
White blood cells (×10 9/L), median [IQR]	9.6 (7.1–13.2)	9.4 (6.7–13.0)	0.33
Hemoglobin (g/L), median [IQR]	115 (99–129)	112 (96–128)	0.049
Red blood cells (×10 9/L), median [IQR]	3.79 (3.27–4.28)	3.69 (3.16–4.20)	0.019
Platelets (×10 9/L), median [IQR]	181 (136–233)	174.00 (132–230)	0.16
Hematocrit (%), median [IQR]	35.3 (30.3–39.4)	34.6 (29.6–38.6)	0.046
Red blood cell distribution width coefficient of variation (%), median [IQR]	13.6 (12.9–14.7)	13.6 (12.9–14.9)	0.74
Prothrombin time (sec), median [IQR]	14.4 (13.6–15.6)	14.4 (13.5–15.8)	0.63
Activated partial thromboplastin time ratio, median [IQR]	1.18 (1.04–1.34)	1.17 (1.04–1.38)	0.18
Proteinuria grade, n (%)			0.31
−	1044 (65.2)	457 (66.4)
+ −	215 (13.4)	90 (13.1)
+	170 (10.6)	62 (9.0)
++	133 (8.3)	56 (8.1)
+++	37 (2.3)	20 (2.9)
++++	1 (0.1)	3 (0.4)
Arterial blood pH median [IQR]	7.42 (7.38–7.46)	7.42 (7.38–7.46)	0.82
Arterial blood HCO3- (mmol/L), median [IQR]	23.1 (20.2–26.2)	22.9 (20.2–26.1)	0.52
Lactic acid (mmol/L), median [IQR]	1.4 (1.0-2.1)	1.4 (1.0–2.0)	0.61
Arterial partial pressure of oxygen (mmHg), median [IQR]	102 (77–138)	101 (76–139)	0.69
Oxygenation index median (mmHg), [IQR]	328 (230–440)	328 (218–445)	0.98
Physiological parameter, median [IQR]
Systolic blood pressure at ICU admission, mmHg	134 (115–153)	130 (112–151)	0.14
Diastolic blood pressure at ICU admission, mmHg	73 (63–83)	72 (62–82)	0.28
Body temperature at ICU admission, °C	36.6 (36.4–36.9)	36.6 (36.4–37.0)	0.25
Heart rate at ICU admission, bpm	83 (70–100)	86 (70–101)	0.204
Respiratory rate at ICU admission, bpm	20 (16–23)	20 (18–22)	0.51
Pathophysiologic conditions, n (%)
Metabolic acidosis	138 (8.4)	55 (7.8)	0.70
Respiratory failure on admission to ICU	659 (40.1)	293 (41.7)	0.51
Shock at ICU admission	197 (11.9)	103 (14.5)	0.10
Clinical scores, median [IQR]
GCS	15 (14–15)	15 (13–15)	0.89
SOFA score	3 (1–4)	3 (2–5)	0.40
APACHE II score	11 (8–16)	12 (8–16)	0.54
SAPS II	32 (27–40)	32 (26–41)	0.80
Outcomes
MAKE30, n (%)	229 (13.8)	94 (13.2)	0.75
Dialysis, n (%)	141 (8.5)	57 (8.0)	0.76
Death, n (%)	53 (3.2)	22 (3.1)	1.00
Persistent renal dysfunction, n (%)	128 (7.7)	59 (8.3)	0.69
Hospital length of stay (day), median [IQR]	14 (9–21)	14 (9–21)	0.81
Intensive care unit length of stay (day), median [IQR])	6 (3–11)	6 (3–12)	0.29

ICU: intensive care unit; NYHA: New York Heart Association; SCr: serum creatinine; eGFR: Estimate glomerular filtration rate; GCS: Glass coma scale; SOFA: Sequential Organ Failure assessment; APACHE II: Acute Physiology and Chronic Health Evaluation II; SAPS II: Simplified Acute Physiology Score II; MAKE30: Major adverse kidney events within 30 d.

Predictive variable and model performance

Variables exhibiting multicollinearity (VIF >10) were removed [32]. The final prediction model included 38 variables (Table S5). The XGBoost model established by the 38 variables has good discrimination with the AUROC in the derivation cohort and test cohort being 0.930 (95% CI: 0.912–0.946) and 0.851 (95% CI: 0.810–0.890), respectively (Figure 2). The AUPRC of the XGBoost model in the derivation cohort and test cohort were 0.76 and 0.51, respectively (Figure S2). The Brier scores of the XGBoost model in the derivation cohort and test cohort were 0.066 and 0.087, respectively. Figures S2 and S3 provide details about the performance metrics, including AUPRC and calibration plots. The inference time of the XGBoost classifier for one case is 6.1489 (lower quartile - upper quartile:5.9230–6.5812) milliseconds. The hardware of our computer for modeling is listed in Table S5.

Figure 2. Receiver operating characteristic curves of the extreme gradient boosting model.

Interpretability of the model

By using the SHAP algorithm, each factor’s importance for the MAKE30 was initially explored. Figure 3 shows the feature importance of the 38 variables most relevant to MAKE30 (in descending order). Among them, APACHE II score, serum creatinine, blood urea nitrogen (BUN), SAPS II, SOFA score, aspartate aminotransferase (AST), arterial blood bicarbonate and albumin (ALB) were the top eight predictor variables associated with MAKE30 in the XGBoost model. How individual variable values affect SHAP values is shown in Figure 4. To facilitate the clinical utility of the XGBoost model, low and high-risk thresholds were determined. Based on low-risk and high-risk cutoffs, Table 2 presents diagnostic test characteristics. When 0.154 was used as the low-risk critical value, the sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio of the model on the training set were 85.6%, 85.1%, 5.73 and 0.17, respectively; the sensitivity, specificity, positive likelihood ratio and negative likelihood ratio of the model in the test set were 77.7%, 79.5%, 3.80 and 0.28, respectively. When 0.186 was used as the high-risk critical value, the sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio of the model on the training set were 81.7%, 87.5%, 3.80 and 0.28, respectively; the sensitivity, specificity, positive likelihood ratio and negative likelihood ratio of the model in the test set were 74.5%, 84.1%, 4.68, 0.30, respectively.

Figure 3. Shapley additive explanations summary plots of the extreme gradient boosting model for MAKE30. AKI: acute kidney injury; ALB: albumin; ALT: alanine aminotransferase; APACHE II: Acute Physiology and Chronic Health Evaluation II; AST: aspartate aminotransferase; BUN: blood urea nitrogen; CKD: Chronic kidney disease; eGFR: Estimated Glomerular Filtration Rate; GCS: Glasgow Coma Scale; HB: hemoglobin; RBC: Red blood cells; HCT: hematocrit; CHF: chronic heart failure; NYHA: New York Heart Association; HR: heart rate at ICU admission; PaO₂: arterial partial pressure of oxygen; PLT: platelet; PT: prothrombin time; RDW-CV: Red blood cell distribution width coefficient of variation; SAPS II: Simplified Acute Physiology Score II; SCr: serum creatinine; SOFA: sequential organ failure assessment score; T: body temperature at ICU admission; WBC: white blood cell.

Figure 4. Shapley additive explanations dependence plots for the association between the predictors and MAKE30 in the extreme gradient boosting model. ALB: albumin; APACHE II: Acute Physiology and Chronic Health Evaluation II; AST: aspartate aminotransferase; BUN: blood urea nitrogen; SAPS II: Simplified Acute Physiology Score II; SCr: serum creatinine; SOFA: Sequential Organ Failure Assessment.

Table 2. Diagnostic test characteristics of the extreme gradient boosting model at the low- and high-risk cutoff points.

	Cutoff value	Sensitivity (%)	Specificity (%)	Positive predic-tive value (%)	Negative predic-tive value (%)	Positive likeli-hood ratio	Negative likeli-hood ratio
Derivation cohort
Low-risk cutoff	0.154	85.6	85.1	47.9	97.4	5.73	0.17
High-risk cutoff	0.185	81.7	87.5	51.1	96.7	6.51	0.21
Test cohort
Low-risk cutoff	0.154	77.7	79.5	36.7	95.9	3.80	0.28
High-risk cutoff	0.185	74.5	84.1	41.7	95.6	4.68	0.30

Sensitivity analysis

The XGBoost model achieved good prediction performance in different age groups, different types of hospitalization, and different severity of illness, different time to endpoint (Figure S4–S11). The AUROC in subgroup with time to endpoint ≤ 13 d was 0.947 (95% CI: 0.931–0.961) in the derivation cohort and 0.882 (95% CI: 0.842–0.916) in the test cohort, and the AUROC in subgroup with time to endpoint >13 d was 0.871 (95% CI:0.820–0.910) in the derivation cohort and 0.748 (95% CI:0.637–0.849). The Delong test showed p = .016, suggesting the XGBoost model performs better in early-on cohort patients than in late-on cohort patients. In addition, when we used upsampling or downsampling to balance the derived cohorts to train the model, the model still provided good predictive performance (Figure S12 and S13).

Discussion

This cohort study aimed to establish a prediction model that accurately predicts the risk of MAKE30 in critically ill elderly patients using XGBoost. The AUROC for the model was 0.930 (95% CI: 0.912–0.946) in the derivation cohort and 0.851 (95% CI: 0.810–0.890) in the test cohort, demonstrating its effectiveness. We attempted to analyze the interpretability of XGBoost model using SHAP, and the results were consistent with clinical practice and other studies. This novel machine-learning model has potential applications in clinical decision-making and risk stratification.

Combining artificial intelligence and the medical field has shown great promise, with numerous studies demonstrating its efficacy [33–36]. For example, Huang et al. trained the XGBoost model to predict in-hospital mortality for elderly patients with ischemic stroke, achieving an AUROC of 0.994 in the derivation cohort and 0.733 in the test cohort [37]. Similarly, the XGBoost model developed by Li et al. to predict in-hospital mortality for patients with heart failure reached an AUROC of 0.824 in the test cohort [38]. Given the good performance of XGBoost in predicting clinical prognosis and the lack of machine learning models for the prediction of MAKE30, this study presents a novel and valuable XGBoost model for the occurrence of MAKE30 in critically ill elderly patients.

In this study, we utilized the SHAP method to enhance the interpretability of the XGBoost model and identified several key factors that contribute to the risk of MAKE 30 in elderly patients in the ICU. Our analysis identified low platelet count, high baseline estimated glomerular filtration rate, and SOFA score as high-risk factors for MAKE30, which have been previously implicated in the construction of prognostic models for adverse renal outcomes in other populations [3,8,16,17]. Our findings revealed that various demographic and physiological factors, including the APACHE II score, serum creatinine and BUN levels, and SOFA score, are positively associated with the risk of MAKE30. These results suggest that the physiological health status of patients at ICU admission can predict adverse renal outcomes within 30 d. Additionally, our analysis showed that low levels of ‘Arterial blood HCO3’ were linked to a high risk of MAKE30. This may be due to renal compensatory mechanisms for maintaining normal acid-base homeostasis, leading to further renal injury in patients. For example, up-regulating renal endocrine hormones such as angiotensin II, aldosterone, and endothelin-1 by kidney promote renal injury, inflammation, and fibrosis [39,40]. The presence of metabolic acidosis, as indicated by low HCO3- levels, may reflect tubular interstitial disease and poor renal function, leading to an increased risk of AKI [41]. Previous retrospective cohort studies have confirmed this association, and the use of sodium bicarbonate infusion has been explored as a potential intervention to improve kidney and survival outcomes [25,42].

Furthermore, we found that a significant increase in non-hepatic AST was associated with mortality, which may reflect skeletal muscle damage, myocardial damage, or hematological diseases [43]. Our analysis also showed that low levels of albumin related to a high risk of MAKE30. Albumin is a protein that is mainly produced by the liver and circulates in the blood plasma. It has various functions, such as maintaining fluid balance, transporting hormones and drugs, and modulating immune responses. Low albumin levels might indicate liver dysfunction, kidney dysfunction, inflammation, or infections, which are all potential risk factors for adverse kidney events [14,44,45]. Therefore, albumin levels can reflect the overall health status of elderly ICU patients and their susceptibility to MAKE30. Albumin levels might also be used as a biomarker to monitor the response to interventions and the prognosis of elderly ICU patients with AKI. These findings emphasize the importance of considering various physiological factors in predicting adverse renal outcomes in critically ill elderly patients and may have implications for clinical decision-making and risk stratification.

The XGBoost model has achieved exceptional prediction performance and has won multiple machine learning competitions. This is due to its ability to provide a relatively transparent solution to the problem of nonlinear relationships, as well as its avoidance of black-box machine learning models [18,46]. The impact of sample imbalance can be assessed by changing the sampling method. In our study, we confirmed that the XGBoost model performed well in predicting the occurrence of MAKE30 in elderly ICU patients.

This cohort study establishes the efficacy of XGBoost in predicting MAKE30 in critically ill elderly patients, with robustness and applicability in different populations. The model employed variables that are readily available in clinical practice, enhancing its accessibility and affordability. Furthermore, the ranking of predictors can aid healthcare providers in identifying high-risk patients and optimizing medical management.

Nevertheless, this study has several limitations. Firstly, the data were retrospectively collected from a single center. Single-center studies often have the deficiencies of insufficient sample size and data bias. The model may not generalize well when extended to different clinical practice settings and different patient populations. External validation and prospective studies may help to test the accurate prediction ability of the model and accelerate the clinical universality of the model. Therefore, external validation and prospective validation are necessary in the future. Secondly, intravenous fluids were not included as a potential variable due to data collection challenges. However, previous studies suggest that high-risk patients should avoid a higher cumulative fluid balance and receive lactated Ringer solution instead of normal saline. Further studies are required to investigate the impact of intravenous fluids on MAKE30 [6,47]. Thirdly, our study did not exclude patients with antineutrophil cytoplasmic antibody (ANCA)-associated vasculitis or anti-glomerular basement membrane disease who might have different renal events incidence with other patients, and further studies are needed to compare the performance of our model in different subgroups of patients. Fourth, the model was established in patients staying in ICU ≥48 h during the hospital admission, and the feasibility for patients staying in ICU <48 h is unknown. Finally, we preliminarily explored the interpretability of the model using the SHAP algorithm, but this method applied the assumption that the features were independent of each other, and it only provides correlation, not causation, between features and predictions, and it might be affected by spurious correlations or omitted variables. More general methods are to be explored.

Conclusion

Our study successfully developed an XGBoost machine learning model that exhibited good performance in both the derivation and test groups. This model has the potential to provide valuable information for clinical decision-making in the management of critically ill elderly patients in the ICU.

Disclosure statement

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

Additional information

Funding

The author(s) reported there is no funding associated with the work featured in this article.