Machine learning models for predicting severe acute kidney injury in patients with sepsis-induced myocardial injury

Abstract

Severe acute kidney injury (sAKI) is a prevalent and serious complication among patients with sepsis-induced myocardial injury (SIMI). Prompt and early prediction of sAKI has an important role in timely intervention, ultimately improving the patients’ survival rate. This study aimed to establish machine learning models to predict sAKI via thorough analysis of data derived from electronic medical records. The data of eligible patients were retrospectively collected from the Medical Information Mart for Intensive Care IV database (MIMIC-IV database) from 2008 to 2019. A total of 1,467 patients with SIMI were included and the primary outcome was the development of sAKI within 7 days after intensive care unit admission. Nine predictive variables were selected and further used to establish the machine learning models. Five different machine learning models were established. The random forest model yielded the most accurate predictions with the highest area under receiver operating characteristic curve (AUC = 0.81) and accuracy (0.79), while the AUC of the traditional SOFA model is 0.66, with an accuracy of 0.71. The machine learning models can be effective tools for predicting the risk of sAKI in patients with SIMI and the RF model performed best.

Sepsis is a life-threatening condition that can lead to dysfunctions of various organs, including the lung, kidney, liver and heart. These organ dysfunctions are not only associated with poor prognosis, but are a major cause of death¹. Among septic patients, sepsis-induced myocardial injury (SIMI) is particularly prevalent, with an incidence of approximately 40–50%². Moreover, SIMI is often accompanied by the dysfunction of other organs, most commonly acute kidney injury (AKI)³. AKI is known to increase the morbidity and mortality of critically ill patients in the intensive care unit setting⁴ and can lead to chronic kidney disease^5–7. In fact, the death rate due to AKI is about 40%⁸. Research by Wald et al. has shown that for patients with sepsis-associated AKI, there is almost a 10% chance of requiring chronic dialysis in the next few years, even under the best circumstances⁹. However, cumulative evidence has revealed that early-stage AKI could be reversed, and approximately 11% of fatalities occurring in hospitals can be attributed to untimely identification and intervention when the conditions are deteriorating¹⁰. Severity staging for AKI drives prognostic and management recommendations, with more severe stages portending worse outcomes¹¹. Additionally, a retrospective observational study demonstrated that acute heart failure (AHF) and AKI had a strong synergic impact on in-hospital mortality¹². Therefore, early detection of severe AKI (sAKI) in patients with SIMI and effective intervention measures are crucial, as they may prevent sAKI from worsening into chronic kidney dysfunction and reduce mortality and the risk of cardiovascular complications.

In recent years, machine learning (ML) has gained recognition from clinicians due to advancements in statistical theory and computer technology^13,14. Novel ML techniques have been widely used in predictive models of various diseases. However, there is currently no evidence on the feasibility of machine learning approaches for predicting sAKI in adult patients with SIMI. Given the importance of early detection and intervention, this study aimed to develop different machine learning models to predict sAKI and identify the most precise and reliable model, thereby providing guidance for clinical interventions.

Methods

Data source

We conducted a retrospective cohort study based on the MIMIC-IV database, containing all patients admitted to the ICUs of Beth Israel Deaconess Medical Center (BIDMC) in Boston between 2008 and 2019¹⁵. MIMIC-IV is the result of a collaboration between BIDMC and Massachusetts Institute of Technology (MIT). Data from this database is identified, transformed, and made available to researchers who have completed training in human research and signed a data use agreement. To apply for access to the database, An author (T Mi) passed the protection of human research participants examination and obtained the certificate (No. 9983480).

Study population

The inclusion criteria for our study were as follows: age ≥ 18 years, non pregnant state, meeting the Sepsis-3 criteria¹⁶ and the SIMI criteria within 24 h after first admission to ICU, and ICU time exceeding 24 h. SIMI was defined as cTnT > 0.01 ng/mL^17,18. Direct or indirect causes leading to abnormal release of cardiac troponin T (cTnT) were excluded from the study, including acute coronary syndrome (ACS), cardiomyopathy, myocarditis, valvular disease, endocarditis, pericarditis, chronic obstructive pulmonary disease (COPD), chronic heart failure (CHF), prior cardiac surgery, cardiac arrest before ICU admission and history of severe tachyarrhythmia. According to the Kidney Disease: Improving Global Outcomes (KDIGO) criteria¹⁹stage 3 in the staging of AKI was defined as sAKI. Patients with chronic kidney diseases were excluded. Besides, patients who had missing data records > 20% were not included in the study.

Data extraction

The general information of included patients was extracted, including age, sex, body mass index (BMI), and previous comorbidities (coronary artary disease (CAD), cerebrovascular disease (CVD), diabetes mellitus (DM), liver disease). Other clinical information within the first 24 h after their admission to the ICU was extracted, which encompassed vital signs (temperature, respiratory rate (RR), heart rate (HR), systolic blood pressure (SBP), diastolic blood pressure (DBP), pulse oxygen saturation (SpO2)), laboratory test results (e.g. anion gap, arterial oxygen saturation (SaO2), creatinine, hemoglobin, glucose, platelet, lactate), Sequential Organ Failure Assessment (SOFA) score and urinie output. For some laboratory test results, the maximum and minimum values were included in the analysis. In this study, missing data were inputted using multiple imputation methods based on the random forest algorithm, and the data set was randomly divided into training and validation sets in a ratio of 7:3.

Statistical analysis

Continuous variables were summarized as the median with interquartile range and were compared using the Wilcoxon rank-sum test. Categorial variables were expressed as number and percentage and were compared using the Chi-square tests or Fisher’s exact probability method. We used the Least Absolute Shrinkage and Selection Operator (LASSO) analysis and the Boruta algorithm analysis respectively to filter feature variables, and then used the selected common feature variables as predictive factors to construct machine learning models.

After feature selection, five ML algorythms, including k-nearest neighbors (KNN), support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost) and multilayer perceptron (MLP), were employed for model construction. We developed the five machine learning models in the training cohort and used the most appropriate hyper-parameters for the final model construction. The performance of each model was validated and compared in the validation cohort. The receiver operating characteristic (ROC) curve, the confusion matrix, the calibration plot and the precision-recall curve of each model were depicted to evaluate the predictive performance. As SOFA was the traditional and common tool for predicting the disease severity, we also compared the performance of ML-based predictive models with that of SOFA model. A shapley additive explanation (SHAP) summary plot was depicted to illustrate the strength and the direction of the correlation between feature variables and the risk of sAKI. All statistical analyses were performed using Python version 3.12. A P-value less than 0.05 was considered statistically significant.

Results

Baseline characteristics

A total of 31,911 adult patients were diagnosed with sepsis in the MIMIC-IV database. Among the patients, those who met the exclusion criteria were excluded (Fig. 1), and ultimately 1467 patients were included in the study. In addition, 35.7% of the 1467 patients (n = 524) developed sAKI within 7days after ICU admission. Table 1 showed a comparison of differences on baseline characteristics between the training cohort and the validation cohort. There was no difference in comparison between the training cohort and the validation cohort, except for white blood cell (WBC) and potassium. The proportion of patients who developed sAKI in the training cohort and validation cohort was 37.1% and 32.4%, respectively.

Fig. 1.

The flowchart of patients’ selection.

Table 1.

Baseline characteristics of included patients.

Characteristics	Total (n = 1467)	Train cohort (n = 1027)	Validation cohort (n = 440)	P-Value
Demographic variables
Age, y	70 (57–82)	71 (58–83)	69 (56–81)	0.098
Sex, n (%)				0.368
Male	783 (53.4%)	556 (54.2%)	227 (51.5%)
Female	684 (46.6%)	470 (45.8%)	214 (48.5%)
BMI, kg/m2	23.1 (20.7–25.9)	23.1 (20.9–25.9)	23.0 (20.5–25.9)	0.662
Comorbidities
CAD, n (%)	255 (17.4%)	173 (16.9%)	82 (18.6%)	0.467
Hypertension, n (%)		445 (43.4%)	197 (44.7%)	0.687
CVD, n (%)	287 (19.6%)	198 (19.3%)	89 (20.2%)	0.750
DM, n (%)	411 (28%)	281 (27.4%)	130 (29.5%)	0.450
Liver disease, n (%)	249 (17%)	169 (16.5%)		0.481
Vital sign
SpO2_min, %	92 (89–95)	92 (89–95)	92 (90–95)	0.187
SpO2_max, %	100	100	100	0.233
Temperature_min, °C	36.4 (35.7–36.7)	36.3 (35.7–36.7)	36.4 (35.7–36.7)	0.777
Temperature_max, °C	37.4 (37-38.1)	37.4 (37–38)	37.5 (37.1–38.1)	0.128
RR_min, times/min	13 (10–16)	13 (10–16)	13 (11–15)	0.602
RR_max, times/min	29 (25–33)	29 (25–34)	28 (25–33)	0.755
HR_min, times/min	73 (61.5–85)	73 (61–84)	73 (63–87)	0.185
HR_max, times/min	109 (95–124)	109 (95–124)	109 (94–123)	0.979
SBP_min, mmHg	84 (75–94)	84 (75–94)	86 (76–95)	0.175
SBP_max, mmHg	146 (132–164)	146 (132–163)	147 (134–166)	0.254
DBP_min, mmHg	43 (37–50)	43 (37–50)	43 (37–50)	0.684
DBP_max, mmHg	88 (75.5–102)	88 (75–102)	87 (76–101)	0.664
Laboratory test
Anion gap_min, mmol/L	13 (11–15)	13 (11–15)	13 (11–16)	0.707
Anion gap_max, mmol/L	18 (15–22)	18 (14–22)	18 (15–22)	0.410
PH_min	7.29 (7.20–7.35)	7.29 (7.2–7.39)	7.29 (7.21–7.36)	0.646
PH_max	7.40 (7.37–7.44)	7.40 (7.36–7.44)	7.41 (7.37–7.45)	0.374
Lactate_min, mmol/L	1.6 (1.1–2.4)	1.7 (1.1–2.4)	1.6 (1.1–2.4)	0.469
Lactate_max, mmol/L	2.6 (1.6–4.6)	2.6 (1.6–4.6)	2.6 (1.6–4.6)	0.644
SaO2_min, %	85 (73–95)	85 (73–94)	85 (74–95)	0.831
SaO2_max, %	91 (85–97)	92 (85–97)	92 (85–97)	0.772
PaCO2_min, mmHg	35 (30–40)	35 (30–40)	34 (30–39)	0.179
PaCO2_max, mmHg	46 (40–55)	47 (40–55)	46 (40–53)	0.156
PaO2_min, mmHg	61 (40–83)	62 (40–82)	52 (40–77)	0.713
PaO2_max, mmHg	153 (98–237)	153 (97–236)	149 (101–236)	0.943
Albumin_min, g/dL	3.1 (2.6–3.5)	3.0 (2.6–3.5)	3.1 (2.7–3.4)	0.912
Albumin_max, g/dL	3.2 (2.8–3.6)	3.2 (2.8–3.6)	3.2 (2.8–3.6)	0.694
Bilirubin_min, mg/dL	0.6 (0.4–1.2)	0.6 (0.4–1.3)	0.6 (0.4–1.2)	0.769
Bilirubin_max, mg/dL	0.8 (0.4–1.6)	0.8 (0.4–1.6)	0.8 (0.4–1.6)	0.946
BUN_min, mg/dL	23 (14.5–38)	24 (15–38)	22 (14–38)	0.319
BUN_max, mg/dL	29 (19–48)	29 (20–47)	28 (18–49)	0.451
Creatinine_min, mg/dL	1.1 (0.8–1.7)	1.1 (0.8–1.7)	1.1 (0.8–1.7)	0.577
Creatinine_max, mg/dL	1.4 (1.0-2.3)	1.4 (1.0-2.4)	1.4 (1.0-2.2)	0.574
Glucose_min, mg/dL	104 (85–127)	105 (86-127.8)	103 (83–126)	0.154
Glucose_max, mg/dL	174 (134-245.5)	175 (133–247)	172 (136–243)	0.988
WBC_min, K/UL	10.3 (7.2–14.6)	10.4 (7.5–14.7)	9.9 (6.9–14.1)	0.056
WBC_max, K/UL	14.9 (10.6–21)	15.2 (10.9–21.4)	14.2 (10.1–19.9)	0.037
Platelet_min, K/UL	166 (108-233.5)	166 (108–232)	166 (107–240)	0.820
Platelet_max, K/UL	217 (149.5–295)	219 (151–296)	212 (146–293)	0.337
Hematocrit_min, %	30.5 (25.7–35.1)	30.6 (26-35.4)	30.5 (25.2–34.7)	0.179
Hematocrit_max, %	35.7 (31.2–40.7)	35.6 (31.3–41.1)	35.8 (30.8–39.9)	0.240
Hemoglobin_min, g/dL	10.1 (8.5–11.6)	10.1 (8.5–11.7)	10 (8.4–11.6)	0.385
Hemoglobin_max, g/dL	11.7 (10.1–13.4)	11.7 (10.2–13.4)	11.8 (10-13.1)	0.502
ALT_min, IU/L	42 (20–126)	42 (19–119)	50 (21–206)	0.507
ALT_max, IU/L	52 (22–176)	53 (22–167)	30 (22–96)	0.610
AST_min, IU/L	68 (31–184)	67 (31–173)	70 (30–209)	0.417
AST_max, IU/L	87 (35–315)	85 (34–301)	101 (36–355)	0.381
Potassium_min, mEq/L	3.8 (3.4–4.2)	3.8 (3.4–4.2)	3.7 (3.4–4.1)	0.013
Potassium_max, mEq/L	4.5 (4.1–5.2)	4.6 (4.1–5.2)	4.5 (4.1–5.2)	0.059
Sodium_min, mEq/L	137 (134–141)	137 (134–140)	137 (135–141)	0.430
Sodium_max, mEq/L	141 (138–144)	141 (138–144)	141 (138–144)	0.255
Calcium_min, mEq/L	7.8 (7.2–8.4)	7.8 (7.2–8.4)	7.7 (7.1–8.3)	0.050
Calcium_max, mEq/L	8.5 (7.9-9.0)	8.5 (8.0-9.1)	8.5 (7.9-9.0)	0.203
Chloride_min, mEq/L	103 (98–107)	102 (98–107)	103 (99–107)	0.299
Chloride_max, mEq/L	107 (104–112)	107 (103–112)	108 (104–112)	0.373
INR_min	1.2 (1.1–1.4)	1.2 (1.1–1.4)	1.2 (1.1–1.5)	0.625
INR_max	1.4 (1.2–1.7)	1.4 (1.2–1.7)	1.4 (1.2–1.8)	0.736
PT_min, s	13.8 (12.3–16.0)	13.8 (12.3–15.9)	13.8 (12.4–16.2)	0.746
PT_max, s	15 (13.2–19)	14.9 (13.1–18.9)	15.2 (13.2–19.3)	0.803
PTT_min, s	29.0 (25.6–33.8)	28.9 (25.5–33.9)	29.0 (25.9–33.7)	0.743
PTT_max, s	34.3 (28.8–51.0)	34.0 (28.7–51.6)	34.9 (28.8–49.5)	0.803
cTnT, ng/ml	0.09 (0.04–0.27)	0.08 (0.04–0.27)	0.09 (0.04–0.23)	0.568
SOFA	3 (2–5)	3 (2–5)	4 (2–5)	0.374
Urine output, ml/h	50 (28–90)	50 (28–89)	55 (30–94)	0.074
sAKI, n (%)	524 (35.7%)	381 (37.1%)	143 (32.4%)	0.100

sAKI severe acute kidney injury, BMI body mass index, CAD coronary artary disease, CVD cerebrovascular disease, DM diabetes mellitus, SpO₂ pulse oxygen saturation, RR respiratory rate, HR heart rate, SBP systolic blood pressure, DBP diastolic blood prssure, SaO₂ arterial oxygen saturation, PaCO₂ partial pressure of carbon dioxide in artery, PaO₂ partial pressure of oxygen in artery, BUN serum urea nitrogen, WBC white blood cells, ALT alanine aminotransferase, AST aspartate transaminase, INR international normalized ratio, PT prothrombin time, PTT partial thromboplastin time, cTnT cardiac troponin T, SOFA Sequential Organ Failure Assessment.

Feature selection

We used the LASSO regression analysis and the Boruta algorithm analysis respectively to screen feature variables among the seventy-one clinical features. Figure 2 presented the 32 feature variables obtained from the LASSO analysis in the form of a path diagram, as well as their changing trends with increasing regularization parameter α intensity. Figure 3 was a boruta feature importance plot, which showed the importance of the 18 feature variables selected by the boruta algorithm analysis. At last, the common feature variables selected based on those two analyses were used as the final predictors for the development of ML models. We ultimately screened out nine feature variables, including urine output, serum creatinine, BMI, PH, platelets, SBP, PTT, PT and sodium.

Fig. 2.

LASSO path for selected features.

Fig. 3.

Boruta feature importance plot.

Model performance comparison on validation cohort

We generated five ML models and the SOFA model to predict the development of sAKI in patients with SIMI. Table 2 summarized the predictive performance of all constructed ML models and the SOFA model on both the validation cohort and the training cohort. Figure 4 presented the confusion matrix of each prediction model. Figure 5 showed the discriminative performance of six models in terms of ROC curves. Among the six models, the area under ROC curve of the RF model was 0.81, performing better than other models, including the KNN model (AUC = 0.71), the SVM model (AUC = 0.69), the XGBoost model (AUC = 0.77), the MLP model (AUC = 0.70) and the SOFA model (AUC = 0.67). The precision-recall curve of each model was described in Fig. 6. Although the recall value of the SVM model was higher than that of other models (Table 2), Fig. 6 illustrated that the predictive performance of the RF model was still the best in terms of average precision. The calibration curve plot (Fig. 7) indicated that there was a good consistency between the predicted probabilities of sAKI developed from the RF model and the actual results. Figure 8 exhibited the strength and direction of the correlation between selected features and the research outcome based on the random forest predictive model. As shown in the Fig. 8, the lower output of urine, the higher risk of sAKI occurrence. A higher level of creatinine or a larger value of BMI were associated with a higher risk of sAKI, while a smaller value of pH or a lower value of platelet count were correlated with the risk of sAKI.

Table 2.

Predictive performance of five models in training and validation cohorts.

Models	Accuracy validation (training)	Precision validation (training)	Sensitivity (recall) validation (training)	Specificity validation (training)	AUC validation (training)
KNN	0.73 (0.73)	0.62 (0.68)	0.57 (0.53)	0.87 (0.84)	0.71 (0.75)
SVM	0.48 (0.51)	0.37 (0.42)	0.86 (0.86)	0.30 (0.30)	0.69 (0.72)
XGBoost	0.75 (0.75)	0.63 (0.70)	0.57 (0.58)	0.83 (0.85)	0.77 (0.78)
RF	0.79 (0.77)	0.70 (0.76)	0.58 (0.55)	0.88 (0.90)	0.81 (0.81)
MLP	0.73 (0.74)	0.56 (0.74)	0.43 (0.47)	0.84 (0.90)	0.75 (0.78)
SOFA	0.71 (0.67)	0.64 (0.64)	0.20 (0.17)	0.95 (0.94)	0.66 (0.61)

AUC area under the curve, KNN k-nearest neighbors, SVM support vector machine, XGBoost extreme gradient boosting, RF random forest, MLP multilayer perceptron, SOFA Sequential Organ Failure Assessment.

Fig. 4.

The confusion matrix of the six models. (0) non severe acute kidney injury, (1) severe acute kidney injury, (A) k-nearest neighbors model, (B) multilayer perceptron model, (C) random forest model, (D) support vector machine model, (E) extreme gradient boosting, (F) Sequential Organ Failure Assessment.

Fig. 5.

Receiver operating characteristic curve of the six models.

Fig. 6.

The precision-recall curve of the six models.

Fig. 7.

The calibration curve plot of the six models.

Fig. 8.

The shapley additive explanations summary plot of the selected variables based on the RF model. The x-axis represented the degree of influence of each feature on the model output, the y-axis displayed the feature variables used in the model development.

Discussion

Sepsis can lead to myocardial depression^19,20 and renal insufficiency^21,22. Notably, myocardial dysfunction often exacerbates AKI, further increasing mortality risk^23,24which rises progressively with declining kidney function^25,26. A study by Levy et al. involving 1,036 patients with severe sepsis demonstrated that early improvement (within 24 h) in cardiovascular, renal, or respiratory function was significantly associated with survival, highlighting the importance of early AKI identification and intervention for improving outcomes²⁷. Clinically, most septic patients with mild to moderate AKI can recover to baseline renal function without any specific renal interventions. However severe AKI patients often experience prolonged renal recovery and are at higher risk of progressing to chronic kidney disease²¹. Therefore, early identification of sAKI and timely renal support are crucial for improving prognosis and survival in patients with SIMI. To our knowledge, this is the first study specifically focused on predicting sAKI in patients with SIMI.

In this study, five machine learning models for predicting sAKI in patients with SIMI were established and validated. Among the five machine learning models and the SOFA model, the RF model had the best discrimination and calibration ability. Random forest (RF), also known as random trees, is a joint prediction model composed of multiple decision trees, which can naturally serve as a fast and effective multi-class classification model^28,29. The RF model is known for its high accuracy, strong robustness, and the ability to handle high-dimensional data. Although the RF model is less prone to overfitting compared to individual decision trees due to their ensemble nature, it is important to note that they can still overfit, particularly in noisy datasets³⁰. Therefore, in this study. we carefully considered the characteristics of the training set and adjusted the hyperparameters appropriately to mitigate the risk.

To enhance the interpretability of the model, the SHAP summary plot derived from the RF model identified the impact on model output of the 9 selected factors, including urine output, creatinine, BMI, PH, platelets, blood pressure, PTT, PT, and sodium. The three most important predictive factors for the best-performing model in this study were urine output, serum creatinine, and BMI. In this study, we performed AKI staging diagnosis according to the KDIGO criteria³¹. It must be remembered that although urine output and creatinine are used to define AKI, they can also be used as predictive biomarkers. As BMI was a novel risk factor for AKI, the pathogenesis was not fully understood. Current research studies have found that obesity is associated with increased levels of pro-inflammatory cytokines³²endothelial dysfunction³³ and oxidative stress³⁴as well as with elevated intraabdominal pressure leading to venous congestion and poor arterial organ perfusion^35,36. Numerous studies had concluded that there was an association between obesity and the development of AKI in critically ill patients with sepsis^37–39. Each 5-kg/m² increase in body mass index was related to a 10% risk of more severe AKI³⁹which ultimately induced multi-system complications and poor patient outcomes. Similarly, this study suggested the predictive role of BMI in sAKI among patients with SIMI.

In addition to the factors mentioned above, PH, blood pressure, and coagulation indicators were also correlated with the occurrence of sAKI. Acidemia negatively impacts the cardiovascular system^40–42and severe metabolic acidosis is associated with moderate to severe kidney injury in 50% of critically ill patients⁴³. Continuous acidosis can directly damage renal tubular cells, leading to apoptosis, necrosis, and loss of renal function^44,45 while low serum PH may induce inflammation and impair immune cell function^46,47. Reduced SBP can cause inadequate renal perfusion, further aggravating kidney injury. Studies have shown that hypotension is an independent risk factor for AKI-related mortality^48,49as impaired renal autoregulation in AKI patients exacerbates renal perfusion instability.

In the RF model, coagulation indicators such as platelets, PT, and PTT also predicted sAKI risk. Platelets and coagulation play critical roles in inflammation and immunity^50,51and thrombocytopenia is an independent predictor of ICU mortality^52,53. Prolonged PT and decreased platelets have been identified as significant predictors of sepsis-induced AKI⁵⁴ and disseminated intravascular coagulation (DIC) is strongly associated with sAKI risk in septic shock patients⁵⁵. This highlights the interplay between AKI and coagulation dysfunction.

Although hyponatremia showed a correlation with sAKI in this study, the relationship between blood sodium and AKI remains controversial. Some studies suggest hypernatremia predicts AKI in septic patients⁵⁶ while others highlight an underrecognized association between hyponatremia and AKI⁵⁷. Further research is needed to clarify the underlying mechanisms.

Risk predictive models aid in making clinical decision regarding the use of resources, invasive procedures, and experimental therapies. According to the KIDGO summary, It is advisable to tailor management to AKI stage. For patients with SIMI, early prediction of sAKI can help remind us whether continuous blood purification is necessary, and urge us to take prompt interventions. Furthermore, we hypothesize that early identification of sAKI may also facilitate to explore novel anti-inflammatory and antioxidant therapies, thereby facilitating renal recovery and prevent short or long-term complications.

There are some limitations in the present research. Firstly, this was a retrospective study based on a publicly available database, which inevitably led to some biases, such as inaccuracy and incompleteness of partial data. Secondly, the data analyzed in this study were from a single center in the United States, which weakened the generalization ability of the developed model. Thirdly, the present study lacked verification of external validity. Lastly, the AUC of the RF model in this study was 0.81, which was not very high. This might be related to the limited features we used in the current model, such as first day feature indicators rather than continuity indicators. Thus, we imagine that adding some continuity indicators or medication information may improve the predictive performance of the model. Further more, future multi-center prospective studies are warranted to verify the results and validate the reliability of ML models in predicting clinical outcomes.

In conclusion, ML models are effective tools for predicting the risk of sAKI in patients with SIMI, with the RF model demonstrating superior performance. By leveraging the RF model, clinicians can tailor precise management strategies and implement early interventions for patients with SIMI at risk of sAKI, thereby preventing renal function deterioration and reducing mortality.

Author contributions

TM contributed to acquisition of data. XL, QF and MF contributed to data processing, analysis of data, and model design and training. TM and XL drafted the article, and all authors provided critical revision. All authors read and approved the final manuscript.

Funding

None.

Data availability

All data in the article can be obtained from MIMIC-IV database (https://physionet.org/content/mimiciv/3.0/).

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

The health information obtained from the MIMIC-IV database was unidentifed, so informed consent of patients was not required. TM passed the protection of human research participants examination and obtained the certificate (No. 9983480). This study was reviewed and approved by the Ethics Committee of Jining No.1 People’s Hospital (2024-IIT-077).