Machine Learning for Dynamic and Early Prediction of Acute Kidney Injury after Cardiac Surgery

Abstract

Objective:

Acute kidney injury (AKI) after cardiac surgery increases morbidity and mortality. Diagnosis relies on oliguria or increased serum creatinine, which develop 48–72 hours after injury. We hypothesized machine learning (ML) incorporating preoperative, operative, and intensive care unit (ICU) data could dynamically predict AKI before conventional identification.

Methods:

Cardiac surgery patients at a tertiary hospital (2008–2019) were identified using electronic medical records (EMR) in the MIMIC-IV database. Pre-/intraoperative parameters included demographics, Charlson Comorbidity subcategories, and operative details. ICU data included hemodynamics, medications, fluid intake/output, and laboratory results. KDIGO creatinine criteria were used for AKI diagnosis. An ensemble ML model was trained for hourly predictions of future AKI within 48 hours. Performance was evaluated by area under the receiver operating characteristic curve (ROC-AUC) and balanced accuracy.

Results:

Within the cohort (n=4,267) there were ~7 million data points. Median baseline creatinine was 1.0 g/dL [IQR 0.8–1.2], with 17% (735/4,267) of patients having chronic kidney disease. Postoperative Stage 1 AKI occurred in 50% (2,129/4,267), Stage 2 in 8% (324/4,267), and Stage 3 in 4% (183/4,267). For hourly prediction of any AKI over the next 48 hours, ROC-AUC was 0.82 and balanced accuracy 75%. For hourly prediction of stage 2 or greater AKI over the next 48 hours, ROC-AUC was 0.95 and balanced accuracy 86%. The model predicted AKI before clinical detection in 89% of cases.

Conclusion:

Ensemble ML models using EMR data can dynamically predict AKI risk after cardiac surgery. Continuous postoperative risk assessment could facilitate interventions to limit or prevent renal injury.

Ultramini-Abstract

Current clinical indicators of acute kidney injury are delayed 48–72 hours after onset of injury. Utilizing electronic medical record data, machine learning models can provide accurate, dynamic predictions of future acute kidney injury risk during postoperative care of cardiac surgery patients to facilitate interventions to limit or prevent injury.

INTRODUCTION

Acute kidney injury (AKI) occurs in 20% of patients after cardiac surgery and is associated with a 5-fold increase in operative mortality and ~20% reduction in 10-year survival.^1–4 Postoperative renal failure is also tracked as a major quality metric for cardiac surgery, including publicly reported scores from the Society of Thoracic Surgeons.⁵ Presently, AKI is identified clinically by a decrease in urine output or an increase in serum creatinine.⁶ These findings develop 48 to 72 hours after the onset of injury, potentially missing the therapeutic intervention window. Early detection may allow therapeutic interventions, such as volume resuscitation and hemodynamic optimization to reduce the extent of AKI.⁶ Previous attempts at early detection of AKI using urine biomarkers or risk calculators have produced mixed results. Novel urine biomarkers, such as g-glutamyl transpeptidase (GGT) and alkaline phosphatase (AP), have yielded poor predictive performance in clinical trials and are not cost-effective. ^7–9 Newer risk calculators are in development, but static calculation of AKI risk is of limited clinical value in the dynamic post-cardiac surgery critical care environment.^10–12

For improved clinical utility, predictive models should ideally provide a dynamic, updated estimation of AKI risk over a patient’s clinical course to inform the management of modifiable risk factors to prevent or mitigate renal damage. During postoperative recovery, a given patient’s baseline AKI risk is modified by their hemodynamics, medication administration, fluid balance, blood product transfusion, and laboratory values. We hypothesized that this data contained within the electronic medical record (EMR) could be leveraged to dynamically predict future AKI risk based on changes in clinical condition after cardiac surgery.

METHODS

Dataset and Study Population

Patients undergoing non-transplant cardiac surgery at a tertiary care center in the United States between 2008–2019 were identified using ICD-9/10 procedure codes (see Supplemental Table 1) in the publicly available MIMIC-IV database, which contains comprehensive electronic medical record data from patient hospital encounters.¹³ Patients with end-stage renal disease (preoperative dialysis or baseline creatinine ≥4.0mg/dl) or without baseline creatinine values were excluded. Patient comorbidities were extracted by mapping ICD-9/10 diagnosis codes to Charlson Comorbidity Index categories.¹⁴ Cardiac procedures were designated to non-exclusive categories of aortic valve, mitral valve, coronary revascularization, or other cardiac procedure. Some intraoperative details, such as cardiopulmonary bypass times, were not available. Medication names, routes of administration, and dosages were extracted from free text pharmacy orders using a python implementation of MedEx, an established natural language processing system for medication data.¹⁵

Postoperative clinical data, including physiologic measurements, fluid intake and output, and medication administration events, were transformed from timestamped measurements/events into time series format with time zero defined as time of intensive care unit (ICU) admission. Vital signs and fluid intake and output data are available only in the ICU setting, while the remaining variables are available throughout the hospital stay. Frequency of data recording in the MIMICIV dataset varies by data components. For the study cohort, routine vital signs were recorded on average every 53 minutes, intake and output measurements every 13 minutes, and medication administration every 23 minutes. Data preprocessing was performed by removing extreme outliers (>99th and <1st percentile), as extreme outliers have been shown to adversely affect validity and applicability of clinical data including laboratory measurements.¹⁶ Physiologically infeasible measurements (e.g., systolic blood pressure greater than 300mmHg) were removed as presumed signal artifact or recording errors in the EMR data.^{17, 18} Time series were resampled into hourly observations using arithmetic mean for linear measurements, harmonic mean for rates, and sum for volumes and medication dosage.¹⁹ For dimension reduction, the minimum and maximum value for each parameter over the prior 24 hours was included for prediction.²⁰

This study was approved by the Baylor College of Medicine Institutional Review Board (H-44702, 12/17/2019) and informed consent was waived for analysis of deidentified data.

Acute Kidney Injury Definition

Creatinine-based criteria from the Kidney Disease: Improving Global Outcomes (KDIGO) definition of AKI was used (Table 1). Baseline creatinine was defined as the most recent preoperative serum creatinine value within the 48 hours preceding surgery. Presence of AKI was determined on a rolling basis. For each hour a given patient’s ICU stay, their most recent creatinine value was compared to their baseline creatinine and their lowest creatinine over the past 48 hours. If a patient’s creatinine value met the KDIGO criteria for AKI in table 1, they were designated as having an AKI condition present for that hour. The urine output criteria of the KDIGO definition were not used as they proved to be overly sensitive and non-specific for the post-cardiac surgery population, as reported previously.²¹

Table 1.

Kidney Disease: Improving Global Outcomes (KDIGO) Acute Kidney Injury Creatinine Criteria

AKI Stage	Serum Creatinine
Stage 1	1.5–1.9x baseline or ≥0.3mg/dL increase over 48 hours
Stage 2	2.0–2.9x baseline
Stage 3	≥3.0x baseline or ≥4.0mg/dL or New renal replacement therapy

Model Development and Testing

The dataset was split 70/30 into a training and testing set. Model training was performed with 10-fold cross-validation to find optimum parameters minimizing area under the receiver operator characteristic curve (ROC-AUC). An ensemble prediction model was developed by optimally weighting established tree-based, neural network, clustering, and logistic regression-based machine learning classifiers (Supplemental Table 2), building on the open-source library AutoGluon to streamline model development and facilitate reproducibility. Missing data was handled with default settings in the AutoGluon architecture, using mean imputation for numerical values and one-hot encoding with a placeholder “Unknown” column for categorical values. Data analysis was performed in Python (v3.7) using AutoGluon, matplotlib, pandas, and numpy. Commercially available cloud computing resources were utilized for model construction and prediction.

An initial “static” prediction model was constructed using preoperative and operative data to predict occurrence of AKI in the first 72 hours after cardiac surgery. A “dynamic” model was then constructed which included preoperative, operative, and postoperative data for produced serial predictions over the postoperative recovery period. A 24-hour sliding window was used for postoperative data, with the prediction task at each time point set as development of AKI over the ensuing 48 hours (Figure 1, Graphical Abstract). For all models, only data points recorded prior to clinical detection of AKI were used for prediction. For patients who developed the outcome of interest, any stage or Stage 2 or greater AKI, the rolling prediction of future AKI risk was not calculated for time points occurring after AKI identification. Separate models were built to predict any AKI and Stage 2 or greater AKI. Prediction performance was assessed on the test dataset using ROC-AUC and balanced accuracy. Balanced accuracy is a performance metric for binary classification which accounts for imbalanced class frequency and is calculated as the arithmetic mean of sensitivity and specificity.

Figure 1:

Study Summary and Prediction Model Architecture with Sliding Window Format

Feature Importance

The relative predictive power of variables in the algorithm was estimated using permutation feature importance scores (FIS). Permutation FIS of a variable is the decrease in predictive performance with that variable “shuffled” to contain random, noninformative data prior to measuring model performance on the test dataset, with all other variables unchanged.²² Permutation FIS for each variable was calculated as the average degradation in performance over 30 iterations of the shuffling procedure.

RESULTS

Patient Characteristics and Acute Kidney Injury

After excluding patients with baseline ESRD (n=324) or missing baseline creatinine values (n=3,555), the study cohort included 4,267 cardiac surgery patients. Cardiac procedures performed were CABG in 58% (2,479/4,267), valve procedures in 36% (1,515/4,267), aortic procedures in in 4% (156/4,267), and other procedures in 3% (117/4,267). (Table 2)

Table 2.

Preoperative, intraoperative, and postoperative characteristics of patient cohort (n=4,267)

Preoperative Characteristics	Median [IQR] or % of cohort (n)
Age (years)	69 [61–76]
Female Sex	30.0% (1,277/4,267)
Comorbidities  Prior Myocardial Infarction  Heart Failure  Peripheral Vascular Disease  Cerebrovascular disease  Chronic pulmonary disease  Diabetes  Chronic kidney disease	38.3% (1,635/4,267) 35.2% (1,504/4,267) 14.4% (615/4,267) 11.1% (474/4,267) 22.3% (980/4,267) 39.3% (1,675/4,267) 17.2% (735/4,267)
Baseline Creatinine (g/dL)	1.0 [0.8–1.2]
Admission Category  Elective  Urgent  Emergent	58.1% (721/4,267) 25.0% (1066/4,267) 16.9% (2480/4,267)
Operative Characteristics	Median [IQR] or % of cohort (n)
Procedure  Isolated CABG  CABG + Valve(s)  Isolated Aortic Valve  Isolated Mitral Valve  Aortic Valve + Mitral Valve  Aorta Procedure  Other Procedure	58.1% (2,479/4,267) 14.1% (602/4,267) 10.5% (450/4,267) 8.8% (376/4,267) 2.0% (87/4,267) 3.7% (156/4,267) 2.7% (117/4,267)
Postoperative Stay	Median [IQR] or % of cohort (n)
Acute kidney injury (AKI)  Stage 1 or greater AKI  Stage 2 or greater AKI  Stage 3 AKI  New Hemodialysis	50% (2,129/4,267) 8% (324/4,267) 4% (183/4,267) 3% (119/4,267)
Intensive care length of stay (hours)	34 [27–71]
Hospital length of Stay (days)	5.0 [4.0–7.0]
In-hospital mortality (overall) *	1.6% (68/4,112)
In-hospital mortality (by AKI stage)*  No AKI  Any AKI  Stage 2 or greater AKI	0.5% (10/2,065) 2.6% (54/2,047) 16.2% (50/307)
Discharge to another health facility (overall)*	42.5% (1,720/4,048)
Discharge to another health facility (by AKI stage)*  No AKI  Any AKI  Stage 2 or greater AKI	31.7% (651/2,055) 53.6% (1,069/1,993) 76.3% (196/257)

^*Discharge disposition (including mortality) known for 96.4% (4,112/4,267) of patient cohort

The median baseline creatinine was 1.0 g/dL [IQR 0.8–1.2]. Serum creatinine was measured an average of once every 13.5 hours while in the ICU in the patient cohort. Postoperative Stage 1 AKI was diagnosed in 50% (2,129/4,267), Stage 2 AKI in 8% (324/4,267), and Stage 3 AKI in 4% (183/4,267). New-onset postoperative dialysis was required in 3% (119/4,267). Patients developed AKI of any severity a median of 36 hours [IQR 18–43] after surgery. Patients developed Stage 2 or greater AKI a median of 46 hours [IQR 11–83] after surgery. (Supplemental Figure 1)

Discharge disposition including mortality was known for 96.4% (4,112/4,267) of patients (Table 2). Overall, in-hospital mortality for the patient cohort was 1.6% (64/4,112). In-hospital mortality was 2.6% (54/2,047) for patients with any AKI and 16.2% (50/307) for patients with Stage 2 or greater AKI. Rates of discharge to another health facility, rather than home, increased with presence and severity of AKI (Table 2).

Time Series Data Parameters

Within the patient cohort, there were ~7 million data points available from the electronic medical record for prediction modeling, encompassing the pre-, intra-, and postoperative phases of care (Table 3). The postoperative phase contained physiologic parameters, medication administration, laboratory values, and fluid intake and output (Figure 2). Missingness of time series monitoring data is presented in Supplemental Table 3.

Table 3:

Predictor Variables

Static Parameters  (see Table 2)

Vital Signs  Routine (blood pressure, heart rate, pulse oximetry, respiratory rate, temperature)  Invasive (pulmonary artery catheter, cardiac output  Cardiac rhythm  Supplemental oxygen  Circulatory support (e.g., intra-aortic balloon pump)

Fluid Intake  Enteral fluids  Intravenous fluids (colloid/crystalloid)  Blood product transfusion Fluid Output  Urine output  Blood loss  Drains/tubes output (e.g., chest tubes)

Continuous intravenous medications dosage  Vasoactive/inotropic medications  Diuretics  Sedation/analgesia

Intermittent medications dosage  Enteral  Intravenous  Inhaled

Mechanical Ventilation  Intubation/Extubation  Ventilator settings and pressures

Laboratory Values  Routine (chemistry, blood count)  Respiratory (arterial blood gas)  Specialized (e.g., serum drug levels)

Figure 2:

Composition of postoperative data

Static Model Performance

An initial static model utilizing preoperative and operative variables to predict occurrence of any AKI in the first 72 hours after cardiac surgery achieved an ROC-AUC of 0.69 and balanced accuracy of 64%. The static model predicting Stage 2 or greater AKI within 72 hours after cardiac surgery with an ROC-AUC of 0.68 and balanced accuracy of 52%.

To incorporate postoperative care data in the above static model, an expanded ML model was developed containing pre- and intraoperative variables plus the first 24 hours of postoperative time series data for prediction of any AKI within 72 hours after cardiac surgery. The addition of initial postoperative data greatly improved predictive performance, yielding an ROC-AUC of 0.87 and balanced accuracy of 76%. Model performance for prediction of Stage 2/3 AKI was an ROC-AUC of 0.96 and balanced accuracy of 80%.

Dynamic Postoperative Prediction Model

To produce a model that provides updated predictions of AKI risk during postoperative recovery, ICU data was organized as multiple sliding windows as model inputs for prediction of future AKI over the next 48 hours (Figure 1). The computational time for generation of each model was approximately 4 hours total using commercial cloud computing resources. Computational time for model predictions was approximately 2 minutes per patient. For rolling prediction of any AKI over the next 48 hours, overall ROC-AUC was 0.82, and balanced accuracy 75%. For rolling prediction of stage 2 or greater AKI over the next 48 hours, overall ROC-AUC was 0.95, and balanced accuracy 86%. Predictive performance varied over time after surgery (Figure 3). Prediction capability was low in the first 24 hours after surgery but improved at 24 hours and remained high. Hourly variability in predictive performance subjectively increased after 100 hours of postoperative stay, reflecting the lower number of patients available for prediction and less frequent data recording due to discharge from intensive care unit. Sensitivity analysis of model performance in different clinical scenarios revealed no qualitatively significant differences in performance by patient age, gender, case urgency, procedure type, or specific comorbidities such as chronic kidney disease, congestive heart failure, or diabetes mellitus (Supplemental Table 4).

Figure 3:

Time-varying Performance of AKI Prediction Model After Cardiac Surgery

The ML model predicted the occurrence of any stage AKI earlier than clinical detection in 89.7% of cases, at the same time as clinical detection (“on-time” detection) in 9.4% of cases, and later than clinical detection in 0.9% of cases. The ML algorithm predicted AKI a median of 13 hours (IQR 3–26 hours) prior to clinical detection.

The most impactful variables for AKI prediction identified with feature importance score (FIS) analysis included static baseline parameters (e.g., patient age, gender, and baseline kidney disease), hemodynamic parameters or vasoactive medications (e.g., cardiac output, central venous pressure, and calcium channel blockers), and medications known to indicate or influence renal function (e.g., furosemide, metolazone) (Table 4).

Table 4:

Feature Importance Scores

Variable	Static or Dynamic Parameter*	Mean Permutation Feature Importance Score	Standard Deviation	p-value
Creatinine (Maximum)	Dynamic	0.1723	0.00213	<0.0001
Baseline Creatinine	Static	0.0159	0.00046	<0.0001
Cardiac Output (Minimum)	Dynamic	0.0033	0.00023	<0.0001
Creatinine (Minimum)	Dynamic	0.0031	0.00030	<0.0001
Nafcillin (Maximum dosage)	Dynamic	0.0029	0.00018	<0.0001
Cardiac Output (Maximum)	Dynamic	0.0025	0.00020	<0.0001
Age	Static	0.0019	0.00027	<0.0001
Calcium gluconate (Maximum dosage)	Dynamic	0.0018	0.00025	<0.0001
Central venous pressure (Maximum)	Dynamic	0.0016	0.00012	<0.0001
Chest tube output (Maximum)	Dynamic	0.0014	0.00010	<0.0001
Propofol (Maximum dosage)	Dynamic	0.0013	0.00008	<0.0001
Furosemide (Maximum dosage)	Dynamic	0.0012	0.00016	<0.0001
Gender	Static	0.0012	0.00015	<0.0001
Potassium chloride (Maximum dosage)	Dynamic	0.0010	0.00013	<0.0001
Nicardipine (Maximum dosage)	Dynamic	0.0010	0.00007	<0.0001
Chronic kidney disease	Static	0.0009	0.00023	0.0006
Anion gap (Minimum)	Dynamic	0.0009	0.00009	<0.0001
Chronic liver disease	Static	0.0009	0.00007	<0.0001
Famotidine (Maximum dosage)	Dynamic	0.0008	0.00008	0.0000
Metolazone (Maximum dosage)	Dynamic	0.0008	0.00017	0.0000

^*Dynamic parameters are input as the maximum (or minimum) value over the prior 24 hours, on a rolling basis. Static parameters are known prior to intensive care unit admission and remain constant throughout postoperative period

DISCUSSION

In this study, electronic medical record data was utilized to develop and test an algorithm for dynamic prediction of AKI risk after cardiac surgery with accurate predictions and earlier detection than traditional clinical diagnosis based on serum creatinine and/or urine output criteria. Incorporation of postoperative variables significantly improved predictive performance compared to models based on pre- and intra-operative variables alone. Incorporation of this approach in the postoperative setting may facilitate earlier intervention to prevent or mitigate renal injury after cardiac surgery (Video 1).

The development of AKI remains the most common complication after cardiac surgery and is associated with significantly worse short- and long-term clinical outcomes.^{1–4, 23, 24} A strong association with postoperative mortality and morbidity persists even for “mild” AKI.^{2, 23} The etiology of AKI after cardiac surgery is multi-faceted and includes non-pulsatile renal blood flow during cardiopulmonary bypass, altered perioperative hemodynamics, proinflammatory cytokines, embolization, hemolysis, acute blood loss anemia, and nephrotoxic medications.^{25, 26} Few options exist for treatment or reversal of established AKI, with treatment being largely supportive. Currently, AKI diagnosis relies on increased serum creatinine and/or decreased urine output signifying decreased glomerular filtration. However, these are late and insensitive findings and significant renal injury can occur prior to detection using these manifestations. Facilitating earlier detection of elevated risk or occurrence of AKI would facilitate timely treatment to potentially prevent or mitigate renal injury to improve postoperative outcomes.

Most risk calculators for AKI after cardiac surgery utilize baseline patient factors and details of planned operative procedure(s) to calculate a preoperative AKI risk score.^10–12 While preoperative risk assessment is essential for patient selection and operative planning, this static assessment provides minimal information to guide postoperative management in the dynamic post-cardiac surgery critical care environment. Relevant clinical guidelines recommend a reassessment of AKI risk in the early postoperative period incorporating intra- and postoperative variables.²⁶ The current study demonstrates that routinely collected EMR data can facilitate dynamic assessment of AKI risk after cardiac surgery, with accurate prediction of future AKI prior to clinical detection.

Multiple urinary biomarkers have been a focus of recent investigations to detect developing or “sub-clinical” renal injury as a strategy to facilitate early detection of AKI after cardiac surgery, with the most studied being neutrophil gelatinase–associated lipocalin (NGAL) and tissue inhibitor of metalloproteinases 2 plus insulin-like growth factor binding protein 7 (TIMP2*IGFBP7).^{8, 9, 27–29} Urine biomarkers are typically measured at set times in the initial 24 hours after cardiopulmonary bypass and used in isolation or in combination with traditional clinical statistical models to provide an updated AKI risk assessment. Despite significant research efforts and logical mechanistic relationships to AKI pathophysiology, urine biomarkers have demonstrated only modest discrimination (ROC-AUC 0.66–0.79) in adult patients undergoing cardiac surgery.⁹ The current study demonstrates that the necessary data for identification of AKI risk in the postoperative period may be found in routinely collected EMR data, and dynamic risk assessment based on EMR data can provide superior predictive performance without requiring costly analysis of biological samples. However, prediction models based on serum creatinine and/or urine output are, by definition, unable to capture subclinical renal damage. Additionally, urine biomarkers may capture a distinct subset of patients at risk for kidney injury compared to traditional clinical risk factors included in prediction models produced with EMR data.³⁰ Monitoring with real-time analysis of clinical data supplemented by biomarker measurements could potentially represent a useful strategy for maximizing predictive performance and clinical utility.

Importantly, the reported prediction model uses EMR data without requiring significant manual input or transformation, which could facilitate direct integration into clinical systems. Time consuming collection and calculation of current clinical support tools and predictive models has been identified as a significant barrier to widespread clinical use, with most established models not compatible with full EMR integration.³¹ With access to a real-time stream of clinical data, the algorithm employed in this study could facilitate augmented decision making to improve clinical management of patients at risk for AKI after cardiac surgery.

Future Directions

Development of the presented prediction model focused on maximizing predictive performance/accuracy. The employed ensemble ML method was chosen because this approach has been demonstrated to produce accurate predictions in a variety of use cases.^{32, 33} However, a drawback of the ensemble ML method approach is limited interpretability of the algorithm, meaning the key factors driving predictions are not easily described. Such a “black box” algorithm may be less likely to be trusted or adopted by clinicians. Additional model refinement will be necessary to enable interpretable predictions without sacrificing predictive performance. Similarly, the time window of postoperative parameters input into the prediction model was set as the preceding 24 hours based on currently accepted models of AKI pathophysiology and theorized timing of renal insult prior to detectable injury.³⁴ While the resultant model accurately predicted future AKI, it remains unknown if an alternative time window (e.g., 48 hours vs 24 hours) or modelling strategy (e.g., expanding window vs. sliding window) would further improve predictive performance. Integration of intraoperative variables (e.g., cardiopulmonary bypass time, nadir hematocrit, peak lactate) may further refine subsequent iterations of the model. Integration of renal biomarkers may further improve the model’s ability to predict Stage 1 AKI.

To be employed for real-time monitoring at the point of care, a prediction model will require close integration with existing EMR systems to minimize data latency. The specific configuration may depend on individual EMR systems, but is achievable using emerging application programming interfaces and data standards for medical data such as Fast Healthcare Interoperability Resources (FHIR).³⁵ Prospective evaluation and external validation will be required to confirm feasibility and validate model predictions prior to clinical deployment.

Limitations

The current study is limited by its retrospective nature and the use of a single-center dataset. The clinical utility of the proposed algorithm was unable to be directly tested due to the retrospective study design. Publicly available ICU datasets such as MIMIC-IV undergo preprocessing for anonymization and standardization prior to publication. External validity will need to be confirmed in future studies using directly accessed EMR data. A significant number of patients were excluded from analysis due to missing key variables, especially baseline creatinine values, which may introduce selection bias. Utilizing directly accessed EMR data with cross-referencing to existing surgery outcomes databases may address this limitation.

The MIMIC-IV database does not contain intraoperative variables which have previously been shown to be key factors for the incidence of severity of postoperative AKI, including cardiopulmonary bypass time.³ Directly accessing institutional EMR data may provide intraoperative data to further refine subsequent iterations of the model, including anesthesia and perfusionist data. Performance of the presented model constructed without intraoperative data may underestimate the potential clinical utility of the proposed strategy.

We utilized ensemble ML analysis to maximize predictive accuracy. However, this strategy is categorized as a “black box” model, meaning the relationships between variables and outcomes are not directly interpretable by humans or end users. While we utilize feature importance scores to assess the relative contribution of predictor variables to predictive accuracy, this method gives limited insight into interactions between predictors. Alternative modeling strategies may be able to provide more detailed and clinically relevant insight, such as the relative risk associated with key clinical variables or interventions.

CONCLUSIONS

In conclusion, this study has demonstrated that routine EMR data can be utilized to accurately predict the occurrence of renal injury prior to detection by traditional clinical parameters. Successful application of this approach in the clinical setting may facilitate earlier recognition of AKI and implementation of best practices to minimize ongoing renal damage and improve overall outcomes after cardiac surgery.

Supplementary Material

6

Video #1: The lead author provides an overview of the work and perspective on the potential clinical utility of the prediction model

CENTRAL MESSAGE

Machine learning models using electronic medical record data can dynamically predict AKI risk after cardiac surgery. Postoperative risk assessment could facilitate interventions to limit renal injury.

PERSPECTIVE STATEMENT

Acute kidney injury occurs frequently after cardiac surgery and is associated with increased short and long-term mortality. Diagnosis relies on decreased urine output or increased serum creatinine, findings which develop 48–72 hours after injury onset. Early detection may allow therapeutic interventions, such as volume resuscitation and hemodynamic optimization to reduce or prevent renal injury.

GLOSSARY OF ABBREVIATIONS

AKI

Acute kidney injury

EMR

Electronic medical record

ICU

Intensive care unit

IQR

Interquartile range

KDIGO

Kidney Disease: Improving Global Outcomes

MIMIC

Medical information mart for intensive care

ML

Machine learning

ROC-AUC

Area under the receiver operating characteristics curve