A machine learning predictive model for acute kidney injury among aneurysmal subarachnoid hemorrhage patients

Abstract

Background

Acute kidney injury (AKI) has been confirmed to be related to the prognosis of aSAH patients. Evaluating the risk of AKI in the early stage is important to avoid the unfavorable outcome of aSAH patients. However, no study has explored the predictive value of machine learning algorithms for AKI after aSAH. This study was designed to develop a machine learning algorithm-based predictive model for AKI among aSAH patients.

Methods

The outcome of this study was the AKI confirmed using the KDIGO criteria. The predictive value of seven machine learning algorithms for the AKI among aSAH patients was explored and verified using the 5-fold cross-validation. The predictive efficiency of machine learning algorithms-based predictive models was evaluated by the area under the receiver operating characteristics curve (AUC). The Shapley Additive explanation method was performed to visualize the importance of features incorporated in machine learning algorithms-based predictive models.

Results

711 aSAH patients were enrolled with an AKI incidence of 7.7%. The AKI group had higher WFNS (p = 0.011), Hunt Hess (p = 0.006), and lower Glasgow Coma Scale (GCS) (p = 0.004). The multiple aneurysm was more frequently observed in the AKI group (p = 0.027). The AKI group had longer length of ICU stay (p < 0.001), length of hospital stay (p < 0.001), and higher mortality (p < 0.001). Three algorithms performed well in predicting the AKI in the training dataset including the random forest (AUC = 1.000), AdaBoost (AUC = 0.954), and XGBoost (AUC = 0.947). The random forest performed the best in the validation dataset with an AUC of 0.724. The top ten features in the random forest algorithm were GCS, mean blood pressure, initial serum creatinine, cystatin C level, albumin, neutrophil, lactate dehydrogenase, glucose, white blood cell, and sodium.

Conclusions

The random forest model demonstrated superior performance in predicting AKI in aSAH patients, achieving a high AUC value, predictive accuracy, and remarkable stability. This model could help clinicians evaluate the risk of AKI in the early stage and guide therapeutic options among aSAH patients.

Introduction

It has been estimated that 6.9-9.0 of every 100,000 people would suffer aneurysmal subarachnoid hemorrhage (aSAH) around the world each year [1]. Although accounting for only 5% of all stroke cases, aSAH causes poor prognosis with mortality as high as 45% [2]. The poor prognosis of aSAH is attributable to acute brain injury in the early stage, secondary brain injury, and extracranial dysfunction during hospitalizations [3–5]. As a common complication developed in critically ill patients, the AKI has been investigated widely developed among aSAH patients with the incidence ranging from 1.9 to 38% [6, 7]. And the correlation between AKI and unfavorable outcomes of aSAH patients has been discovered [7–9]. Evaluating the risk of AKI in the early stage may help clinicians avoid nephrotoxic medications and measure the mortality risk among aSAH patients. Some risk factors for AKI among aSAH patients have been identified including male gender, hypertension, diabetes, aneurysm size, HIV, coagulopathy, serum levels of sodium, chloride and albumin, nephrotoxic drugs [6, 7, 10, 11]. Machine learning has been extensively utilized in medical research due to its exceptional ability to handle high-dimensional datasets and complex nonlinear relationships. In some clinical contexts, machine learning algorithms may outperform traditional logistic regression, delivering superior predictive accuracy and stability. Furthermore, the application of machine learning may facilitate the discovery of novel potential risk factors that conventional logistic regression fails to identify. Many studies have been conducted verifying the predictive value of machine learning algorithms on AKI among various patients including those diagnosed with acute pancreatitis, active infective endocarditis, cerebral infarction, or those who underwent cardiac surgery [12–15]. Some complications after aSAH have also been effectively predicted using machine learning algorithms including pneumonia, delayed cerebral ischemia, and shunt-dependent hydrocephalus [16–18]. There is still no study exploring the value of machine learning algorithms in predicting the AKI after aSAH. We performed this study to develop machine learning algorithms based predive models to evaluate the AKI risk among aSAH patients incorporating variables collected in the early stage after ruptures.

Methods

Patients

Patients hospitalized and received treatments due to the aSAH between January 2017 and June 2019 in the West China Hospital, a medical center in the southwestern China region, were included in this study. The diagnosis of aSAH was identified based on symptomatic signs of ruptured intracranial aneurysm and radiological signs from computed tomography angiography (CTA) or digital subtraction angiography (DSA). Some patients were excluded from this study due to these criteria: [1] admission to our hospital 48 h after the onset of initial symptoms; [2] transferring from other hospitals; [3] incomplete records of laboratory examinations on the first day of admission; [4] serum creatinine on the first day of admission ≥ 353.6umpl/L. 711 aSAH patients were enrolled after the screening. Informed consent forms of enrolled patients were signed by patients themselves or their legal representatives based on the study policy of our hospital. This study was performed complying with the Helsinki Declaration and obtained ethical approval from the West China Hospital Review Board (2021 − 1684).

Variables collection

The following variables were collected into this study: (1) patient demographics and vital signs on admission; (2) comorbidities (diabetes mellitus, hypertension, chronic renal disease); (3) clinical scores (GCS, WFNS, Hunt Hess, mFisher score) on admission; (4) laboratory examinations analyzed from the first blood sample on the first day of admission (white blood cell, neutrophil, lymphocyte, monocyte, platelet, hemoglobin, red blood cell, glucose, red cell distribution width, serum creatinine, blood urea nitrogen, cystatin C, serum sodium, serum potassium, serum chloride, serum calcium, serum magnesium, serum phosphorus, lactate dehydrogenase, alkaline phosphatase); (5) aneurysm location, occurrence of multiple aneurysm and Intraventricular hemorrhage; (6) delayed cerebral ischemia; (7) medical operations including angiogram, coiling, clipping. The outcome of this study was the AKI since admission, which was identified using the Kidney Disease Improving Global Outcome (KDIGO) criteria [19].

Statistical analysis

The Kolmogorov-Smirnov test was used to verify the normality of collected variables. The difference of normally distributed variables (expressed as mean ± standard deviation) and non-normally distributed variables [expressed as median (interquartile range)] between the AKI group and non-AKI group was testified by Student’s t-test and Mann-Whitney U test, respectively. The difference of categorical variables expressed as counts (percentage) between the AKI group and non-AKI group were testified by the Chi-square test or Fisher exact test. The p-value < 0.05 was considered statistically significant.

Machine learning algorithms

To predict the AKI risk at the early stage after aSAH, variables collected on the first day were included in the training process of machine learning algorithms-based models, including patients demographics and vital signs on admission, comorbidities, clinical scores, laboratory examinations on the first day of admission, aneurysm location, occurrence of multiple aneurysm, intraventricular hemorrhage, and medical operations on the first day. Variables with missing rates exceeding 70% were removed to prevent potential distortion caused by data absence. Delayed cerebral ischemia was not included because it commonly occurred 4 days after aneurysm rupture, which limited its role in the early assessment of AKI risk. The predictive value of seven machine learning algorithms including extreme gradient boosting (XGBoost), logistic regression, light gradient boosting machine (light GBM), Random Forest, adaptive boosting (AdaBoost), complement naïve Bayes (CNB), and support vector machine (SVM) was explored and verified using the 5-fold cross-validation, which randomly divided the dataset into 5 subsets of similar size, and subsequently conducted 5 rounds of training and validation. Each round used one subset as the validation set and the remaining four subsets as the training set. The optimal parameters of these machine learning algorithms for predicting the AKI were automatically searched and set during the cross-validation process. The predictive efficiency of these machine learning algorithms-based predictive models in both the training dataset and validation dataset was evaluated using a variety of indexes including area under the receiver operating characteristics curve (AUC), accuracy, sensitivity, specificity, PPV, NPV, and F1 score. The calibration curve and decision curve of predictive models were drawn. Based on Shapley values from cooperative game theory, Shapley Additive Explanations (SHAP) values are a powerful tool for interpreting machine learning models, providing a fair way to distribute the “contribution” of each feature to the machine learning algorithms-based model’s prediction. The importance of a feature is determined by the average absolute value of its SHAP values across all instances in the dataset. Features with higher average absolute SHAP values are ranked higher. The sign of the SHAP value indicates whether a feature increases or decreases the prediction. For example, a positive SHAP value for a feature means it increases the predicted value, while a negative value means it decreases the predicted value. The overall modeling process was shown as Fig. 1. All statistical analyses and figures drawings were performed using an online statistical analysis platform-Extreme smart analysis (https://www.xsmartanalysis.com/) [20], which executed instructions utilizing the R software (version 4.2.3) and Python (version 3.11.4).

Fig. 1.

Modeling process of this study

Results

Baseline characteristics of included aSAH patients

711 aSAH patients were enrolled with an AKI incidence of 7.7% (Table 1). Compared with the non-AKI group, the AKI group had higher heart rate (p = 0.010), WFNS (p = 0.011), Hunt Hess (p = 0.006), and lower GCS (p = 0.004). The incidence of comorbidities including diabetes mellitus, hypertension, and chronic renal disease did not differ between the AKI group and the non-AKI group. Laboratory examinations presented the AKI group had higher levels of white blood cell (p = 0.012), neutrophil (p = 0.002), glucose (p < 0.001), serum creatinine (p = 0.025), blood urea nitrogen (p < 0.001), cystatin C (p < 0.001), serum sodium (p = 0.047), lactate dehydrogenase (p = 0.009) and lower lymphocyte (p = 0.010). The aneurysm location did not show statistical significance between the AKI group and the non-AKI group. The incidence of delayed cerebral ischemia and the usage incidence of coiling and clipping did not differ between two groups. While the multiple aneurysm was more frequently observed in the AKI group (p = 0.027). Furthermore, the AKI group had longer length of ICU stay (p < 0.001), length of hospital stay (p < 0.001) and higher mortality (p < 0.001).

Table 1.

Characteristics of included aSAH patients

	Total (n = 711)	Non-AKI group (n = 656, 92.3%)	AKI group (n = 55, 7.7%)	p
Age (year)	55 (49–65)	55 (49–65)	59 (50–68)	0.130
Male gender, n (%)	244 (34.3%)	222 (33.8%)	22 (40.0%)	0.355
Smoking, n (%)	125 (17.6%)	112 (17.1%)	13 (23.6%)	0.219
Alcoholism, n (%)	106 (14.9%)	98 (14.9%)	8 (14.5%)	0.937
Comorbidities
Diabetes mellitus, n (%)	35 (4.9%)	31 (4.7%)	4 (7.3%)	0.402
Hypertension, n (%)	304 (42.8%)	276 (42.1%)	28 (50.9%)	0.203
Chronic renal disease, n (%)	19 (2.7%)	16 (2.4%)	3 (5.5%)	0.183
Vital signs on admission
Mean blood pressure (mmHg)	105 (95–118)	105 (94–118)	114 (96–122)	0.087
Heart rate (s− 1)	78 (69–88)	78 (69–88)	84 (74–92)	0.010
Clinical scores
GCS	13 (11–14)	13 (12–14)	12 (7–14)	0.004
WFNS	2 (2–4)	2 (2–4)	4 (2–4)	0.011
Hunt Hess	2 (2–3)	2 (2–3)	3 (2–4)	0.006
mFisher	4 (2–4)	4 (2–4)	3 (2–4)	0.217
Laboratory tests
White blood cell (109/L)	10.34 (7.99–13.54)	10.23 (7.93–13.48)	11.57 (9.13–15.85)	0.012
Neutrophil (109/L)	8.84 (6.63–12.10)	8.76 (6.53–12.01)	10.46 (8.39–14.19)	0.002
Lymphocyte (109/L)	0.99 (0.72–1.36)	1.01 (0.74–1.37)	0.83 (0.62–1.16)	0.010
Monocyte (109/L)	0.52 (0.36–0.69)	0.51 (0.36–0.69)	0.55 (0.36–0.68)	0.571
Platelet (109/L)	165 (128–207)	165 (128–207)	174 (131–207)	0.950
Hemoglobin (g/dL)	125 (110–136)	125 (110–136)	127 (112–136)	0.861
RBC (109/L)	39.7 (35.0-42.8)	39.7 (35.2–42.8)	39.2 (34.8–43.1)	0.927
Glucose (mmol/L)	6.58 (5.58–8.09)	6.47 (5.52-8.00)	7.59 (6.81–8.64)	< 0.001
Red cell distribution width (%)	13.3 (12.8–14.2)	13.3 (12.8–14.2)	13.4 (12.9–14.2)	0.692
Serum creatinine (umol/L)	57 (49–70)	57 (49–68)	64 (48–100)	0.025
Blood urea nitrogen (g/L)	4.4 (3.5–5.5)	4.3 (3.4–5.4)	5.4 (4.2–6.9)	< 0.001
Cystatin C (mg/L)	0.72 (0.64–0.84)	0.72 (0.64–0.82)	0.83 (0.74–1.07)	< 0.001
Serum sodium (mmol/L)	139.7 (137.3-142.9)	139.6 (137.3-142.5)	141.7 (138.6-145.5)	0.047
Serum potassium (mmol/L)	3.76 (3.50–4.07)	3.78 (3.51–4.08)	3.69 (3.41–3.93)	0.107
Serum chloride (mmol/L)	101.9 (98.5-105.6)	101.8 (98.5-105.3)	104.3 (99.3-108.4)	0.067
Serum calcium (mmol/L)	2.16 ± 0.13	2.16 ± 0.13	2.13 ± 0.14	0.080
Serum magnesium (mmol/L)	0.87 (0.80–0.95)	0.87 (0.81–0.94)	0.87 (0.80–0.96)	0.795
Serum phosphorus (mmol/L)	0.91 (0.74–1.06)	0.92 (0.75–1.07)	0.89 (0.69–1.03)	0.161
Lactate dehydrogenase (U/L)	195 (165–241)	194 (164–237)	222 (171–268)	0.009
Alkaline phosphatase (U/L)	73 (59–91)	74 (59–91)	73 (56–93)	0.787
Aneurysm of posterior circulation, n (%)	55 (7.7%)	52 (7.9%)	3 (5.5%)	0.510
Multiple aneurysm, n (%)	69 (9.7%)	59 (9.0%)	10 (18.2%)	0.027
Intraventricular hemorrhage, n (%)	308 (43.3%)	290 (44.2%)	18 (32.7%)	0.099
Angiogram, n (%)	290 (40.8%)	265 (40.4%)	25 (45.5%)	0.463
Coiling, n (%)	75 (10.5%)	69 (10.5%)	6 (10.9%)	0.928
Clipping, n (%)	524 (73.7%)	484 (73.8%)	40 (72.7%)	0.865
Delayed cerebral ischemia, n (%)	119 (16.7%)	108 (16.5%)	11 (20.0%)	0.500
Length of ICU stay (day)	3 (0–9)	3 (0–8)	9 (5–20)	< 0.001
Length of hospital stay (day)	12 (9–19)	12 (9–17)	19 (11–26)	< 0.001
Mortality, n (%)	146 (20.534%)	124 (18.902%)	22 (40%)	< 0.001

GCS, Glasgow Coma Scale; WFNS, World Federation of Neurosurgical Societies; mFisher, modified Fisher

Value of different machine learning algorithms for predicting AKI among aSAH patients

Three algorithms performed well in predicting the AKI in the training dataset including the random forest (AUC = 1.000), AdaBoost (AUC = 0.954), and XGBoost (AUC = 0.947) (Table 2) (Fig. 2A). These three algorithms still performed well in predicting the AKI in the validation dataset, especially the random forest, which performed the best with an AUC of 0.724 (Table 3) (Fig. 2B). Additionally, the random forest had the highest accuracy both in the training dataset (0.998) and validation dataset (0.924). The sensitivity of the random forest was also the highest both in the training dataset (1.000) and validation dataset (0.691), which meant the random forest is more beneficial to identify aSAH patients with a high AKI risk at the early stage. The calibration plot (Fig. 2C) indicated the predicted probability of the random forest algorithm achieved a closer consistency with the actual probability. Evaluating the clinical utility of models at different thresholds, the decision curve showed most machine learning algorithms have high net benefits, in addition to the CNB (Fig. 2D). Based mainly on the AUC value, accuracy, and calibration plot, the random forest was considered the optimal predictive model for the AKI among aSAH. The learning curve of our random forest model showed continuous improvement in the performance of the validation set after increasing the data volume (Fig. 2E). The top ten features in the random forest algorithm were GCS, mean blood pressure, initial serum creatinine, cystatin C level, albumin, neutrophil, lactate dehydrogenase, glucose, white blood cell, and sodium (Fig. 3A and B). The SHAP output value of one patient was presented as Fig. 3C.

Table 2.

Performance of machine learning algorithms for predicting the AKI in the training cohort of aSAH patients

Classification models	AUC (95%CI)	Accuracy	Sensitivity	Specificity	PPV	NPV	F1 score
XGBoost	0.947 (0.917–0.977)	0.886	0.927	0.882	0.431	0.991	0.578
Logistic regression	0.698 (0.606–0.791)	0.807	0.532	0.832	0.238	0.953	0.303
Light GBM	0.809 (0.751–0.868)	0.855	0.736	0.809	Na	0.961	Na
Random Forest	1.000	0.998	1.000	1.000	1.000	0.998	1.000
AdaBoost	0.954 (0.931–0.976)	0.868	0.936	0.864	0.418	0.992	0.562
CNB	0.679 (0.584–0.774)	0.727	0.605	0.740	0.161	0.955	0.253
SVM	0.707 (0.614-0.800)	0.839	0.505	0.869	0.252	0.952	0.328

XGBoost, extreme gradient boosting; Light GBM, light gradient boosting machine; AdaBoost, adaptive boosting; CNB, complement naïve Bayes; SVM, support vector machine; AUC, area under the receiver operating characteristic curve; PPV, Positive Predictive Value; NPV, Negative Predictive Value

Fig. 2.

(A) Receiver operating characteristics curve of machine learning based models for predicting the AKI in the training dataset of aSAH patients; (B) Receiver operating characteristics curve of machine learning based models for predicting the AKI in the validation dataset of aSAH patients; (C) Calibration plot of machine learning based models for predicting the AKI among aSAH patients; (D) Validation decision plot of machine learning based models for predicting the AKI among aSAH patients; (E) Learning curve of random forest model showed AUC discrepancy between the training dataset and the validation set with increasing training samples

Table 3.

Performance of machine learning algorithms for predicting the AKI in the validation cohort of aSAH patients

Classification models	AUC (95%CI)	Accuracy	Sensitivity	Specificity	PPV	NPV	F1 score
XGBoost	0.674 (0.500-0.849)	0.797	0.600	0.775	0.157	0.942	0.244
Logistic regression	0.644 (0.453–0.835)	0.801	0.618	0.730	0.224	0.942	0.311
Light GBM	0.629 (0.459–0.799)	0.805	0.455	0.830	Na	0.938	Na
Random Forest	0.724 (0.557–0.891)	0.924	0.691	0.715	Na	0.924	Na
AdaBoost	0.678 (0.502–0.854)	0.790	0.564	0.789	0.168	0.942	0.254
CNB	0.651 (0.460–0.842)	0.716	0.618	0.738	0.137	0.945	0.221
SVM	0.585 (0.383–0.788)	0.809	0.382	0.890	0.173	0.939	0.231

XGBoost, extreme gradient boosting; Light GBM, light gradient boosting machine; AdaBoost, adaptive boosting; CNB, complement naïve Bayes; SVM, support vector machine; AUC, area under the receiver operating characteristic curve; PPV, Positive Predictive Value; NPV, Negative Predictive Value

Fig. 3.

(A) SHAP value of all patients output in the random forest model; (B) Feature importance derived from the random forest model;C. SHAP value of one patient output in the random forest model

Discussion

Our study showed that AKI developed less commonly in aSAH patients than the most of previously reported incidence exceeding 10% [7, 8, 10, 21, 22]. Though one study enrolling patients receiving cerebral artery aneurysm clipping showed an AKI incidence of 1.9%, which was mainly attributable to the fact that most of the participants did not experience rupture of the intracranial aneurysm [6]. The variation of AKI incidence cross studies may be caused by the difference of AKI definition, disease severity, medical level, and therapeutic options. aSAH patients with the AKI had significantly higher mortality, longer length of ICU stay, and hospital stay than those without in our study. This finding was consistent with conclusions of previous studies confirming that the AKI occurrence and AKI stage were both related to the poor prognosis of aSAH patients [7, 9, 11]. The impaired renal function may exacerbate electrolyte disorders and disturb the effect of dehydration therapy with subsequent brain edema and intracranial hypertension. Therefore, it is important to explore risk factors for AKI and evaluate the risk of AKI in the early stage after the rupture of the aneurysm. The random forest algorithm-based predictive model performed the best in predicting the risk of AKI in our enrolled aSAH patients. Previous studies exploring risk factors for AKI after aSAH mainly used the conventional logistic regression [6, 7, 10, 11]. While our study indicated that the random forest performed better than the logistic regression, with a higher AUC value and predictive accuracy.

The random forest is a classifier using multiple trees to train and predict samples. The output category of the random forest is determined by the mode of the categories output by individual trees. Therefore, the random forest has higher accuracy than many individual algorithms. Furthermore, the random forest can be built in parallel, accelerating the training process and efficiently managing overfitting. The relatively higher AUC of our developed random forest-based predictive model in the validation dataset indicated this model had higher stability than other models we explored. Although the XGBoost and AdaBoost also performed well in the training dataset, they did not show sustained stability and strong predictive value in the validation dataset with the low AUC, which indicated they had overfitting during the training process for predicting AKI. Compared with XGBoost and AdaBoost, the random forest demonstrates stronger robustness, capable of managing abnormal data, missing values, and more insensitive to the scaling of input features. Additionally, the random forest effectively withstands data noise and generally exhibits strong generalization capabilities, leveraging ensemble learning and random sampling techniques, making it versatile in handling diverse datasets, including high-dimensional and imbalanced datasets.

The top ten features in the random forest algorithm-based predictive model were GCS, mean blood pressure, initial serum creatinine, cystatin C level, albumin, neutrophil, lactate dehydrogenase, glucose, white blood cell, and sodium. The lower GCS indicated more severe intracranial injury after aSAH. Similarly, one previous study found that WFNS was significantly correlated with early AKI in patients with non-traumatic subarachnoid hemorrhage [23]. The mean blood pressure in the AKI group was higher than the non-AKI group, which implied that extremely abnormal hypertension might increase the risk of AKI after aSAH. A meta-analysis found hypertension was an independent risk factor for contrast-associated AKI [24]. Another study showed that intraoperative hypertension was positively related to postoperative AKI in patients undergoing laparoscopic surgery [25]. Certainly, the hypotension indicating insufficient perfusion to the renal tissue could also promote the development of AKI after aSAH. One recent study showed the association between systolic blood pressure and in-hospital mortality of critically ill patients with AKI was U-shaped [26]. Therefore, the relationship between blood pressure and AKI after aSAH may be nonlinear. Considering the cystatin C, many studies have validated its value in predicting the AKI due to its higher sensitivity and stability than the serum creatinine [27–29]. Finally, the influence of hypoalbuminemia on the risk of AKI has also been verified in other studies [30–32]. And one study found that preoperative albumin ≤ 3.9 g/dL was independently associated with the AKI after cerebral artery aneurysm clipping [6]. Both the number of white blood cell and neutrophil could reflect the inflammatory and immune status. One Korean study found a low preoperative platelet to white blood cell ratio was related to AKI following cerebral aneurysm treatment [33]. Another study showed that increased preprocedural white blood cell count was associated with an increased risk of developing contrast-induced acute kidney injury in patients receiving percutaneous coronary intervention [34]. Furthermore, neutrophil percentage has been confirmed as a biomarker of AKI in elderly patients with acute myocardial infarction [35]. As a common complication among critically ill patients, hyperglycemia would directly or indirectly lead to the development of AKI through various mechanisms such as oxidative stress, inflammatory response, and endothelial dysfunction [36]. Studies found elevated glucose was an independent risk factor for the AKI among patients undergoing coronary intervention, and those undergoing cardiac surgery [37–39]. As a glycolysis-related enzyme catalyzing the interconversion between pyruvate and lactate, lactate dehydrogenase would increase under the condition of cell damage or death, caused by ischemia, hypoxia, infection, or physical damage [40]. Lactate dehydrogenase has been verified efficient in predicting AKI among severe heatstroke patients with rhabdomyolysis, severe burn patients, and septic patients [41–43]. The adverse effect of increased serum sodium level on the AKI risk also was verified among critically ill patients [44]. Another study confirmed that hypernatremia was a significant risk factor for AKI after SAH [45].

Among above mentioned features, the mean blood pressure, albumin, glucose, and sodium are modifiable factors. Managing these factors with appropriate levels may decrease the risk of AKI after aSAH. Certainly, the optimal targets level of these factors should be further explored in future studies. The random forest algorithm-based predictive model incorporating those features can be integrated into electronic health records (EHR) systems to provide real-time risk scores for AKI in the future. Clinicians could use these risk scores to prioritize high-risk patients for closer monitoring and early intervention, such as adjusting medication dosages, optimizing fluid management, and maintaining adequate blood pressure and perfusion.

Several limitations should be noted in this study. Firstly, this study developed a machine learning based predictive model for AKI utilizing data from a single medical center. There are differences in patient characteristics and therapeutic strategies among different medical centers, which may limit the generalizability and stability of our developed predictive model. The number of AKI events was also limited due to the low prevalence of AKI among our included aSAH patients. Therefore, the accuracy of the trained predictive model should be externally verified in the future on other medical centers with larger sample sizes. And parameters of the trained predictive model should be further optimized after incorporating more potentially relevant features. Secondly, some risk factors for AKI were not collected such as nephrotoxic antibiotics and options of blood transfusion. These medications were rarely used at the early stage in our included aSAH patients. Some nephrotoxic drugs may cause damage to the kidney only when used for a long duration or excessively such as furosemide, steroids. While the main aim of our study is to develop a predictive model for AKI using features at the early stage, future studies with larger sample sizes could be designed to develop models incorporating dynamic features during the overall hospitalizations to continuously evaluate the risk of AKI in hospitalized aSAH patients. Thirdly, the learning curve of our random forest model indicated the discrepancy in AUC of our random forest model between training and validation datasets was mainly caused by the limited sample size. Though the 5-fold cross validation has been used to narrow the gap of AUC between training and validation datasets, it is still necessary to perform the research with a larger sample size to further optimize the parameters and stabilize the performance of the random forest model. Finally, the random forest-based predictive model for AKI should be transformed and integrated into a readily available software program on electronic devices for convenient use in clinical settings.

Conclusion

The random forest algorithm-based model performed well in predicting the AKI among aSAH patients. In the future, by seamlessly integrating with EHR systems, this predictive model could offer real-time AKI risk scores, enabling clinicians to identify high-risk patients for enhanced monitoring and avoid administering nephrotoxic medications during the early stages following aSAH. This proactive approach would mitigate the adverse impact of AKI on aSAH prognosis.

Author contributions

R.W. created the study protocol, performed the statistical analyses, trained machine learning based-predictive models, and wrote the first manuscript raft. L.Q. assisted with manuscript revision. Y.Z. assisted the analysis and explained statistical methods. L.C. assisted with manuscript revision and data confirmation. H.M. assisted with data collection and manuscript revision. J.X. contributed to data interpretation and manuscript revision. Y.Z. contributed to manuscript revision. All authors read and approved the final manuscript.

Funding

This study was funded by 1·3·5 project for disciplines of excellence–Clinical Research Incubation Project, West China Hospital, Sichuan University (2020HXFH036), Knowledge Innovation Program of the Chinese Academy of Sciences (JH2022007), General Program of the National Natural Science Foundation of China (82173175).

Data availability

The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

Ethics declarations

Ethics approval and consent to participate

This study was performed complying with the Helsinki Declaration. This study has been approved by the Ethics Committee of the West China Hospital (2021 − 1684). Informed consent forms of each patient were legally signed by themselves or their authorized families according to the research policy of our hospital.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.