Artificial Intelligence in Acute Kidney Injury Prediction

Abstract

The use of artificial intelligence (AI) in nephrology and its associated clinical research is growing. Recent years have seen increased interest in utilizing AI to predict the development of hospital-based acute kidney injury (AKI). Several AI techniques have been employed to improve the ability to detect AKI across a variety of hospitalized settings. This review discusses the evolutions of AKI risk prediction discussing the static risk assessment models of yesteryear as well as the more recent trend toward AI and advanced learning techniques. We discuss the relative improvement in AKI detection as well as the relative dearth of data around the clinical implementation and patient outcomes using these models. The use of AI for AKI detection and clinical care is in its infancy, and this review describes how we arrived at our current position and hints at the promise of the future.

Artificial intelligence (AI) is defined as the theory and development of computer systems to perform tasks that normally require human intelligence. The term was first described by John McCarthy in 1956 as the science and engineering of making intelligent machines.¹ AI in medicine has been divided primarily into 2 subtypes, virtual and physical.¹ Physical AI includes tangible objects that can assist the patient or provider including robotassisted surgeries and intelligent bioprostheses, whereas virtual AI includes applications such as electronic health record systems to neural network-based guidance in treatment decisions.²

There are several important but developing roles for AI in nephrology. Current examples of AI in nephrology include using machine learning algorithms to improve detection of urinary tract infections, using image analysis to evaluate renal pathology in the setting of kidney transplantation (there are several other AI opportunities in the realm of transplantation), and risk prediction for chronic kidney disease progression.^3–5 In 2015, the Acute Disease Quality Initiative Consensus Conference recognized acute kidney injury (AKI) as an ideal syndrome to apply AI (including machine learning and big data) since it is common and has a standardized readily identifiable definition and is a major burden on the health care system.⁶ In theory, AI could help not only reduce the increased morbidity and mortality associated with hospital-based AKI but also lower the cost of AKI care (more resources, intensified monitoring, investigations, longer hospital stays).⁶

AKI is broadly defined as an abrupt (over hours to days) decrease in kidney function. There have been 3 consensus definitions with the most recent being the Kidney Disease Improving Global Outcomes (KDIGO) which defines AKI via changes in either serum creatinine (SCr) and/or urine output (UOP) (Table 1).⁷ While some have recently advocated for the inclusion of novel functional and damage biomarkers into the definition of AKI, unfortunately this has yet to gain wide-scale acceptance/implementation.⁸

Table 1.

Kidney Disease Improving Global Outcome (KDIGO) Definition of AKI⁷

Stage	Serum Creatinine	Urine Output
1	1.5–1.9 Times baseline OR ≧0.3-mg/dL increase	<0.5 mL/kg/h for 6–12 h
2	2.0–2.9 Times baseline	<0.5 mL/kg/h for ≧ 12 h
3	3.0 Times baseline OR increase in serum creatinine to ≧4.0 OR initiation of renal replacement therapy OR in patients <18 years, decrease in eGFR to <35 mL/min per 1.73 m2	<0.3 mL/kg/h for ≧ 24 h OR anuria ≧12 h

Abbreviations: AKI, acute kidney injury; eGFR, estimated glomerular filtration rate.

The KDIGO definition utilizes SCr and UOP; however, SCr increases are often delayed since it serves as a functional marker of glomerular filtration rate (GFR) and may take up to 24–36 hours for creatinine to rise even after a catastrophic renal insult.⁹ These delayed increases in the gold standard for diagnosis can potentially further delay care.

Given that AKI is a clinical syndrome that can occur across the complete spectrum of hospital encounters, it can present with nonspecific symptoms that vary on the clinical scenario. Despite its varied presentations, since it is simply defined via changes in SCr and UOP, it serves as an ideal target for a machine learning or augmented intelligence risk-prediction algorithm. In addition to the rising SCr and drop in UOP, AKI has often accompanied other changes (rising phosphorus and potassium levels and increasing metabolic acidosis and prior drops in blood pressure) that create identifiable patterns in high-throughput patient data.¹⁰ This approach has demonstrated efficacy for several specific clinical contexts such as hospital-acquired AKI, postoperative AKI, cancer patients with AKI, traumatic injuries, and critical illness.^11–21 In fact, looking at these patterns and trends in patient data, we may identify sub-phenotypes of AKI with larger AKI domains. For example, as discussed below, within patients with sepsis-associated AKI there may be several subsets of AKI due to the underlying insult (ischemia, inflammation, nephrotoxin exposure), and within these sub-phenotypes, there may be distinct responses to therapy.^22,23 Table 2 defines a few AI terms that the average clinician may not be familiar with.

Table 2.

Brief Definitions and Explanations of Artificial Intelligence Terminology

	Logistic Regression	LASSO	XGB	Neural Network	Deep Learning	Random Forest
Definition	Statistical analysis to predict a binary outcome based on prior observations of a data set. Model predicts a dependent data variable by analyzing the relationship between 1 or more independent variables	Shrinkage and variable selection method for linear regression models. Goal to obtain the subset of predictors that minimized prediction error for a quantitative response variable	Machine learning technique for regression and classification problems, which produces a prediction model in the form of an ensemble of weak prediction models, typically decision trees. It builds the model in a stage-wise fashion like other boosting methods do, and it generalizes them by allowing optimization of an arbitrary differentiable loss function.	Series of algorithms that endeavors to recognize underlying relationships in a set of data through a process that mimics the way the human brain operates	Neural network that consists of multiple layers of interconnected nodes, each building upon a previous layer to refine and optimize the prediction or categorization	Supervised machine learning algorithm, as it builds decision trees on different samples and takes their majority vote for classification and average in case of regression

Abbreviations: LASSO, least absolute shrinkage and selection operator; XGB, eXtreme Gradient Boosting.

Prior to these AI investigations, AKI risk prediction was done through static prediction based on baseline (or preoperative) data. However, newer AI models incorporate baseline (pre-AKI) as well as evolving, dynamic, real-time data collected during a hospital admission which show improved predicative abilities.²⁴ An example of a recent static (logistic regression-derived) risk score is provided by Bell and colleagues which incorporated 4 variables (age, baseline estimated GFR, diabetes, and heart failure) for the prediction of KDIGO-defined AKI.²⁵ These factors were used to construct a 4-variable AKI risk score in 273,450 Scottish patients with good predictive performance (C-statistic of 0.80, 95% confidence interval [CI] 0.80–0.81).²⁵ Furthermore, their score was still accurate when externally validated in 2 additional populations including 218,091 patients from Kent, United Kingdom (C-statistic 0.71 [95% CI 0.70–0.72]) and 1,173,607 patients from Alberta, Canada (C-statistic 0.76 [95% CI 0.75–0.76]).²⁵

Static AKI risk models have been developed in the perioperative settings, especially in cardiac and cardiothoracic procedures. One of the earlier studies in 1998 by Mangano and colleagues sought to determine the incidence and characteristics of postoperative renal dysfunction and failure.²⁶ In this prospective, observational, multicenter study in 24 university hospitals with 2222 patients having cardiac revascularization with or without valvular surgery measurements, the main outcome measure was renal dysfunction (defined as postoperative SCr ≧ 177 mmol/L [2 mg/dL] with a preoperative-to-postoperative increase ≧62 micro-mol/L [0.7 mg/dL]).²⁶ In their study, approximately 7.7% (n = 171) of patients had postoperative renal dysfunction, 1.4% (n = 30) of which required dialysis.²⁶ Multivariable analysis identified 5 independent preoperative predictors of renal dysfunction including age greater than 70 years, congestive heart failure, previous myocardial revascularization, type 1 diabetes mellitus, and elevated preoperative serum glucose levels.²⁶ In order to address the utility of these preoperative predictors of renal dysfunction, Palomba and colleagues created a prognostic score system following cardiac surgery with variables including age >65 years, preoperative creatinine above 1.2 mg/dL, preoperative glucose above 140 mg/dL, heart failure, combined surgeries (coronary artery bypass grafting + valve), cardiopulmonary bypass time above 2 hours, low cardiac output, and low central venous pressure.²⁷ Their score (Acute Kidney Injury following Cardiac Surgery) presented good calibration and discrimination, and in their validation cohort of 215 patients recruited by the same observer and institution, they reported their prognostic score with an area under the receiver operating characteristic curve (AUC ROC) of 0.85 (95% CI 0.79 – 0.90).²⁷

Moving beyond static preoperative models, Mathis and colleagues sought to determine whether the association between intraoperative hypotension and KDIGO AKI varies by preoperative risk by reviewing major noncardiac surgical procedures across 8 hospitals (2008–2015), and among the 138,021 cases reviewed, 9% (12,431) had developed postoperative AKI.²⁸ This study included several major risk factors in the risk score including anemia, estimated GFR, surgery type, American Society of Anesthesiologists Physical Status, and expected anesthesia duration. These factors were used for risk stratification by 2 separate multivariable logistic regression models which were then divided into 4 equal-sized preoperative risk quartiles: low, medium, high, and highest risk. Low baseline risk demonstrated no association between intraoperative hypotension and AKI, whereas medium risk had association with severe intraoperative hypotension (defined as mean arterial pressure less than 50 mm Hg) and AKI with adjusted odds ratio of 1.77 (95% CI 1.2 to 2.61; p = 0.004).²⁸ Furthermore, patients who were in the highest risk category with mild hypotension ranges (mean arterial pressure 55 to 59 mm Hg) were also associated with a 25–54% increased risk of AKI and an amplified associated increase of 79% to 150% in AKI for more than 10 minutes of hypotension of less than 50 mm Hg.²⁸ The authors concluded that these patients demonstrated varying associations with distinct levels of hypotension when stratified by preoperative risk factors and specific levels of absolute hypotension, but not relative hypotension, an important independent risk factor for AKI.²⁸

In the past there was intense investigation into AKI risk following cardiac surgery with several models being developed and the most widely used one being the Cleveland Clinic risk score which predicted postoperative dialysis requiring AKI.^29,30 While few scores focused on the lower stages of AKI, many scores contained commonalities including but not limited to age, gender, race, preoperative renal insufficiency, history of diabetes mellitus, surgery type, and intraoperative and postoperative parameters.^27,29–35 Recently, in a retrospective observation multicenter cohort, Demirjian and colleagues sought to improve the prior Cleveland Clinic model via the derivation and validation of a model for moderate to severe AKI (KDIGO stage 2 and 3 AKI) after cardiac surgery.³⁶ In addition to preoperative risk factors, they included perioperative change in laboratory values including SCr, blood urea nitrogen, sodium/potassium/bicarbonate, and albumin from the first postoperative metabolic panel.³⁶ The model provided excellent discrimination for stage 2/3 AKI in the validation cohort (4734 patients from 3 hospitals) with an AUC of 0.86 (95% CI 0.84–0.91) within 72 hours and 0.84 (95% CI 0.82–0.87) within 14 days. The model’s performance for AKI requiring dialysis was excellent with an AUC of 0.88 (95% CI 0.84–0.92) within 72 hours and 0.87 (95% CI 0.84–0.91) within 14 days. This performance was similar compared to their derivation cohort which consisted of 58,526 patients from the Cleveland Clinic Hospital.³⁶ However, further research is required to determine whether the use of this new prediction tool improves clinical outcomes.

Others have sought to use AI while including perioperative features in their models to predict postoperative AKI. Adhikari and colleagues developed a machine learning model using a retrospective single-center cohort of 2911 postoperative adults at the University of Florida between 2000 and 2010.¹⁵ They specifically sought to identify AKI during the first 3 and 7 postoperative days after surgery. They included preoperative data as well as integration of intraoperative statistical features (including intraoperative pulse, systolic and meanarterial pressures) and created an AKI risk score through a machine learning stacking approach inside a random forest classifier. Model performance was evaluated using AUC, accuracy, and net reclassification improvement in which the proposed model had an AUC of 0.86 (accuracy of 0.78) for the 7-day AKI outcome while the model using only preoperative data had an AUC of 0.84 (accuracy of 0.76).¹⁵ The net reclassification improvement for each outcome was AKI at 3 days (8%) and 7 days (7%).³⁷ By integration of the intraoperative features, the algorithm was able to reclassify 40% of the false-negative patients from the preoperative model, thereby improving postoperative AKI prediction with high sensitivity and specificity.¹⁵

Similarly, Lei and colleagues examined whether adding preoperative and intraoperative data is associated with improved prediction of noncardiac postoperative AKI across several statistical model methods.³⁷ They generated models using logistic regression with elastic net selection, gradient boosting machine, and random forest approaches.³⁷ They studied 42,615 patients who underwent noncardiac surgery, and progressive addition of clinical data improved model performance for the prediction of KDIGO-AKI across all modeling approaches, with gradient boosting machine providing the highest discrimination by AUC which increased from 0.71 (95% CI 0.69–0.73) using prehospitalization variables to 0.80 (95% CI 0.79–0.82) using preoperative and prehospitalization variables (p < 0.001) and further increased to 0.82 (95% CI 0.80–0.83) when adding intraoperative variables (p < 0.001 for comparison vs model using preoperative variables). However, statistically significant improvements in discrimination did not appear to be clinically significant; for example, the AKI rate among patients classified as high risk only improved from 29.1% to 30%.³⁷ The study, like others, suggests that using AI techniques with EHR (electronic health record) data may accurately stratify patients at risk of perioperative AKI but highlight that this additional accuracy may only lead to modest improvements and unclear clinical benefits.³⁷

Cronin and colleagues sought to move beyond surgical patients and look at hospital-based AKI.¹¹ They studied a national retrospective cohort of 1,620,898 patient hospitalizations from 116 Veterans Affairs hospitals from 2003 to 2012 and developed several models using EHR data and several techniques (logistic regression, least absolute shrinkage and selection operator [LASSO] regression, and random forests).¹¹ The AUC for different model outcomes ranged from 0.746 to 0.758 in stage 1+ (defined as KDIGO classification stages 1, 2, or 3), 0.714 to 0.720 in stage 2+ (defined as KDIGO stage 2 or 3), and 0.823 to 0.825 in new dialysis requirements. Logistic regression had the best AUC in stage 11 and dialysis (AUC 0.76 [95% CI 0.76–0.76]), whereas random forests had the best AUC in stage 2+ (AUC 0.72 [95% CI 0.72–0.72]) but the least favorable calibration plots.¹¹ When the study compared observed to expected risks across models, the random forests overpredicted risk as predicted probability increased. This study concluded that several model-generation methods provide good discrimination and performance characteristics but that the random forest models did not calibrate as well as LASSO or logistic regression models.¹¹

Flechet and colleagues reanalyzed data collected as part of the Early vs Late Parenteral Nutrition in Critically Ill Adults (EPANIC) trial database and developed and validated AKI prediction models in adult intensive care unit (ICU) patients.³⁸ Subsequently they made the models available via an online prognostic calculator.³⁸ The model development cohort (n = 2123) and validation cohort (n= 2367) had data from before and after ICU admission (up to day 1).³⁸ The primary outcome was the comparison of predictive performance between models and the AKI biomarker neutrophil gelatinase-associated lipocalin (NGAL) for the development of any AKI and AKI stages 2 or 3 during the first week of ICU stay. NGAL is a well-studied AKI biomarker which has been shown to detect early kidney damage after cardiopulmonary bypass in both pediatric and adult populations.^39–41 The AKI-123 model (their random forest machine-learning algorithm to detect all 3 stages of AKI) outperformed NGAL, with AUCs of 0.82 (95% CI 0.82–0.82) for AKI-123 and 0.84 (95% CI 0.83–0.84) for AKI-23 after 24 hours compared to NGAL with AUCs of 0.74 (95% CI 0.74–0.74) for AKI-123 and 0.79 (95% CI 0.79–0.79) for AKI-23.³⁸ In a separate study by Flechet and colleagues, the AKI-predictor model was tested against physician prediction in an ICU setting and demonstrated improved discriminative performance compared to physicians for prediction of AKI stages 2 or 3, and higher benefit overall, with bedside physicians frequently overestimating the risk of AKI.²¹

More advanced models can use techniques such as deep learning with neural networks to accurately detect creatinine-based AKI 1–3 days in advance of changes in SCr.^42,43 This lead time can create the ideal situation giving clinicians time to deliver focused care which promotes kidney health and prevents further kidney injury. Deep learning, a subset of machine learning, is a neural network that consists of multiple layers of interconnected nodes, each building upon a previous layer to refine and optimize the prediction or categorization. Tomasev and colleagues demonstrated a deep learning approach using a recurrent neural network that operates sequentially over individual electronic health records processing the data one step at a time and building an internal memory that keeps track of relevant information seen up to that point, and at each time point, the model outputs a probability of AKI occurring at any stage of severity within the next 48 hours.⁴³ The final curated data set contained 315 base features (demographics, vital signs, laboratory tests, medications, chronic diseases etc.) to study the continuous prediction of AKI risk within clinically actionable windows of up to 48 hours. They utilized longitudinal data from 703,782 adults in the US Veterans Administration Health Care System.⁴³ Their model predicted 55.8% of all inpatient AKI within a window of up to 48 hours, with a ratio of 2 false predictions for every true positive, corresponding to an AUC of 0.921. In addition, the model provided correct early predictions in 84.3% of episodes where administration of inhospital or outpatient dialysis was required within 30 days of onset of AKI of any stage and in 90.2% of cases where regular outpatient administration of dialysis was scheduled within 90 days of the onset of AKI, corresponding to an AUC ROC of 0.835 and AUC ROC 0.838, respectively.⁴³ Despite the limitations of a primarily male population (93.62%), 2:1 false alert rate, and a complex deep learning model with a wide range of features which may be difficult to implement in real time, this study demonstrates the inspiring potential of machine learning in AKI clinical care.⁴³

Koyner and colleagues similarly sought to develop an AKI prediction model using EHR data in hospitalized patients in an observational cohort study in a tertiary urban academic medical center from November 2008 to January 2016.⁴² They used a gradient boosting machine learning algorithm (including demographics, vital signs, laboratory values diagnostics, and therapeutic interventions) to predict KDIGO stage 2 AKI. They divided their single-center cohort of 121,158 patients into a 60% derivation and 40% validation cohort.⁴² In this study, 14.4% developed any form of creatinine-based KDIGO AKI, with 3.5% developing SCr-based stage 2 AKI. In the validation cohort, the AUC was 0.90 (95% CI 0.90–0.90) for predicting stage 2 AKI within 24 hours and 0.87 (95% CI 0.87–0.87) within 48 hours. In addition, the AUC for receipt of renal replacement therapy (821 patients) was 0.96 (95% CI 0.96–0.96) in the next 48 hours. This AKI prediction machine learning risk algorithm performed well for both ward and ICU patients and surgical and nonsurgical patients.⁴² Furthermore, while the risk score provided excellent discrimination, unfortunately, it had over 130 features and would be cumbersome to implement in real time for clinical use. As such, Churpek and colleagues validated a simplified model in a multicenter diagnostic study of 495,971 admissions from 6 hospitals in 3 health systems across Chicago, Illinois.⁴⁴ The abridged (47 features) machine learning algorithm provided similarly excellent discrimination in all cohorts, with alerts identifying patients over a day and a half before stage 2 AKI.⁴⁴ Depending on the cohort, the AUC for predicting at least stage 2 AKI in the next 48 hours ranged from 0.85 (95% CI 0.840.85) to 0.86 (95% CI 0.86–0.86).⁴⁴ However, as with the other aforementioned models, the clinical utility of implementing this model to improve AKI remains unknown. Future studies are needed to determine if a real-time AKI risk score can improve patient outcomes. Importantly, many of these risk scores are for all patients across the entire hospital, and while data support the generalizability of these scores, it remains unclear if a similar hospital-wide approach in terms of AKI interventions can also yield improved outcomes.

Many of the machine learning models discussed above have focused on large groups of patients (across the entire hospital or ICU); however, the coronavirus disease pandemic (severe acute respiratory syndrome coronavirus 2) brought forth a subset of hospitalized patients with increased risk of AKI. Some groups capitalized on this unique pandemic to operationalize AI and advanced learning techniques to identify high-risk patients. Ponce and colleagues aimed to develop a prognostic score by using a machine learning approach to fit models in a training set using 10-fold cross-validation and validated the accuracy using AUC ROC, with the coefficients of the best model (elastic net) to build a predictive score for inhospital mortality in COVID-19 patients with AKI.⁴⁵ They utilized demographic data, comorbidities/conditions at admission, laboratory exams within 24 hours, and characteristics and causes of AKI (defined by KDIGO criteria) in a cross-sectional multicenter prospective cohort study in the Latin America AKI COVID-19 Registry.⁴⁵ In their study of 870 COVID-19 patients, they found their prognostic score for predicting in-hospital mortality in hospitalized COVID-19 patients with AKI to have an AUC ROC of 0.82 (95% CI 0.76–0.89) in the validation cohort which they believe may help health care workers with intensive monitoring and resource allocation.⁴⁵ Vaid and colleagues developed and validated several models including logistic regression, LASSO, random forest, and eXtreme Gradient Boosting (with and without imputation), for predicting treatment with dialysis or death at various time horizons (days 1, 3, 5, 7) after hospital admission in hospitalized COVID-19 patients with AKI.⁴⁶ In their multicenter cohort of 6093 patients (2442 training and 3651 external validation), the eXtreme Gradient Boosting model without imputation for prediction of a composite outcome of either death or dialysis in patients positive for COVID-19 (AUC ROC range of 0.85–0.87 and area under precision-recall curve range of 0.27–0.54) had the best performance as compared with standard and other machine learning models.⁴⁶ As with prior AKI models, the impact of implementation of these models remains unknown. Table 3 summarizes many of the aforementioned studies.

Table 3.

Summary of AI in AKI Risk Assessment Papers

Study	Setting	Definition	# Of Patients	AI Methods Employed	Results
Bell et al25	Hospitalized International Scotland, United Kingdom, Canada	KDIGO serum creatinine-based criteria	273,450	LR	Risk score including 4 variables (older age, lower baseline eGFR, DM, heart failure) had a good predictive performance C-Statistic 0.80 (95% CI 0.80–0.81) in the development cohort
Palomba et al27	Postoperative coronary artery bypass graft surgery, valve surgery, or both	Serum creatinine > 2 mg/dL or >50% baseline	215	Univariate and multivariate analyses	Risk score following cardiac surgery including age >65, preoperative creatinine above 1.2 mg/dL, preoperative glucose above 140 mg/dL, heart failure, combined surgeries (CABG + valve), cardiopulmonary bypass time above 2 h, low cardiac output, and low central venous pressure, prognostic score with an AUC ROC 0.85 (95% CI 0.79–0.90)
Mathis et al28	Perioperative risk for major noncardiac surgical procedures across 8 hospitals	KDIGO serum creatinine-based criteria	138,021	LR	Risk stratification by 2 separate multivariable LR models divided into 4 equal-sized preoperative risk quartiles: low, medium, high, and highest risk, where medium risk had association with severe intraoperative hypotension with medium risk AOR 1.77 (95% CI 1.2–2.61)
Thakar et al30	Postoperative open-heart surgery	Renal failure requiring dialysis	33,217	LR	Scoring model in a randomly selected test set then validated with score ranges between 0 and 17 points and the acute renal failure frequency at each score level in the validation set had an AUC ROC: 0.82 (95% CI 0.80–0.85)
Demirjian et al36	Postoperative cardiac surgery	KDIGO serum creatinine-based criteria and AKI requiring dialysis	4734	LR	In the validation cohort, the models for moderate to severe AKI after the surgical procedure showed AUCs of 0.860 (95% CI 0.838–0.882) within 72 h and 0.842 (95% CI 0.820–0.865) within 14 d, and the models for AKI requiring dialysis and 14 d had an AUC of 0.879 (95% CI 0.840–0.918) within 72 h and 0.873 (95% CI 0.836–0.910) within 14 d after the surgical procedure
Adhikari et al15	Postoperative retrospective single-center university cohort	KDIGO serum creatinine-based criteria	2911	ML, RF	Predictive performance of a proposed model that enriched a preoperative model by integrating intraoperative statistical features that had an AUC 0.86 (95% CI 0.84–0.89)
Lei et al37	Postoperative noncardiac surgery	KDIGO serum creatinine-based criteria	42,615	LR w/ENS, GBM, RF	Progressive addition of clinical data improved model performance for prediction of KDIGO-AKI across all modelling approaches with GBM providing the highest discrimination by AUC when using prehospitalization, preoperative, and intraoperative variables: AUC 0.82 (95% CI 0.80–0.83)
Cronin et al11	National Retrospective Cohort of Veteran Affairs Hospitals	KDIGO serum creatinine-based criteria Stage 1 + (defined as KDIGO classification stages 1, 2, or 3); stage 2+ (defined as KDIGO stage 2 or 3)	1,620,898	LR, LAS, LASSO, RF	AUC for different model outcomes ranged from 0.746 to 0.758 in stage 1+, 0.714 to 0.720 in stage 2+, and 0.823 to 0.825 in new dialysis requirement. Logistic regression had the best AUC in stage 1+ and dialysis AUC 0.76 (95% CI 0.76–0.76) whereas random forests had the best AUC in stage 2+ with AUC 0.72 (95% CI 0.72–0.72) but the least favorable calibration plots
Flechet et al38	Adult ICU patients collected as part of EPANIC trial database	KDIGO serum creatinine-based criteria and NGAL	2367	ML, RF	AKI-123 model (their random forest machine-learning algorithm to detect all 3 stages of AKI) outperformed NGAL, with AUCs of 0.82 (95% CI 0.82–0.82) for AKI-123 and 0.84 (95% CI 0.83–0.84) for AKI-23 after 24 h compared to NGAL with AUCs of 0.74 (95% CI 0.74–0.74) for AKI-123 and 0.79 (95% CI 0.79–0.79) for AKI-23
Tomasev et al43	US Veterans Administration Health Care System	KDIGO serum creatinine-based criteria	703,782	RNN	Model (with final curated data set of 315 base features) provided early predictions correctly in 84.3% of episodes where administration of in-hospital or outpatient dialysis was required within 30 d of onset of AKI of any stage and 90.2% of cases where regular outpatient administration of dialysis was scheduled within 90 d of the onset of AKI, corresponding to an AUC ROC of 0.835 and 0.838, respectively
Koyner et al42	Single-center urban academic university hospital	KDIGO serum creatinine-based criteria	121,158	ML, GBM	In validation cohort, the AUC was 0.90 (95% CI 0.90–0.90) for predicting stage 2 AKI within 24 h and 0.87 (95% CI 0.87–0.87) within 48 h. AUC for receipt of renal replacement therapy was 0.96 (95% CI 0.96–0.96) in the next 48 h
Churpek et al44	6 Hospitals in 3 health systems	KDIGO serum creatinine-based criteria	495,971	ML	Depending on the cohort, the AUCs for predicting at least stage 2 AKI in the next 48 h were 0.85 (95% CI 0.84–0.85) to 0.86 (95% CI 0.86–0.86)
Ponce et al45	Multicenter prospective Latin America registry for hospitalized COVID-19	KDIGO serum creatinine-based criteria	870	ML	Prognostic score for predicting in-hospital mortality in hospitalized COVID-19 patients with AKI to have an AUC ROC of 0.82 (95% CI 0.76–0.89)
Vaid et al46	Multicenter hospitalized COVID-19	KDIGO serum creatinine-based criteria	6093	LR, LASSO, RF, XGBoost	XGBoost model without imputation for prediction of composite outcome of either death or dialysis in patients positive for COVID-19 (AUC ROC range of 0.85–0.87 and AUPRC range of 0.27–0.54) had the best performance as compared with standard and other machine learning models

Summary of studies with the setting, AKI definition, number of patients, methods, and results.

Abbreviations: AI, artificial intelligence; AKI, acute kidney injury; AOR, adjusted odds ratio; AUC, area under the curve; AUPRC, area under precision-recall curve; CABG, coronary artery bypass grafting; CI, confidence interval; DM, diabetes mellitus; eGFR, estimated glomerular filtration rate; ENS, elastic net selection; EPANIC, Early vs Late Parenteral Nutrition in Critically Ill Adults trial; GBM, gradient boosting machine; ICU, intensive care unit; KDIGO, Kidney Disease Improving Global Outcome; LAS, least absolute shrinkage; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; ML, machine learning; NGAL, neutrophil gelatinase-associated lipocalin; RF, random forest; RNN, recurrent neural network; ROC, receiver operating characteristic; XGBoost, eXtreme Gradient Boosting.

While the future of AKI care appears to hold many potential options to improve AKI risk stratification, there is a separate benefit to several of these AI investigations, the identification of AKI sub-phenotypes. While AKI is classically defined solely based on changes in creatinine and UOP, buried within the data are other clues to better understand injury patterns, responses to therapeutic interventions, and forecasting outcomes. As discussed above, in the main subheadings of AKI (eg, cardiac surgery, sepsis, surgical nephrotoxin), there are clusters that likely represent patient subsets with distinct pathophysiologic mechanisms. In those with sepsis-associated AKI, some may exhibit an inflammatory injury pattern, some others may be more ischemic from hypotension, and some may have nephrotoxin injury from antibiotics or the infection itself in the setting of septic shock.⁴⁷ Our current AKI classification does not allow for the differentiation of these sub-phenotypes.

This process has already begun to take shape, with several papers using advanced learning to identify AKI sub-phenotypes. Bhatraju and colleagues used data from the Vasopressin in Septic Shock Trial and latent class analysis to identify 2 distinct sub-phenotypes with different clinical outcomes (7-day renal nonrecovery and 28-day mortality), each with a unique response to vasopressin therapy.²³ The sub-phenotypes which were discovered (n = 794) and validated (n = 425) in unique cohorts discriminated clinical outcome better than KDIGO staging, and this performance was further enhanced when the phenotype classification was combined with 3 biochemical biomarkers (endothelial [angiopoietin 1 and 2 (Ang-1 and Ang-2)] and inflammatory [interleukin-8]). Building on this, Bhatraju and colleagues then looked at single-nucleotide polymorphisms and genetic variation within the Ang-1 and Ang-2 and tumor necrosis factor receptor 1A genes.²³ They demonstrated that a single-nucleotide polymorphism within the Ang-2 gene was associated with decreased plasma Ang-2 concentrations and potentially plays a causal role with regard to subphenotype.⁴⁸ It remains to be seen if phenotype-specific interventions targeting Ang-2 production or other aspects of these sepsis-associated AKI phenotypes can improve or alter care.

Separately Nadkarni and colleagues used data from the Medical Information Mart for Intensive Care III database and identified sub-phenotypes of the sepsis-associated AKI in 4001 patients with sepsis-associated AKI within the first 48 hours of ICU admission.²² Using vital signs and laboratory measurements from the first 48 hours along with comorbidities (188 variables/features), they employed deep learning techniques to identify 3 unique clinical AKI sub-phenotypes. Within these 3 subphenotypes, they demonstrated difference in preadmission comorbidities as well as differences in lab values and patient outcomes (including need for dialysis and 28-day mortality).²² While these data are limited in that they come from a retrospective cohort, it demonstrates the promise of AI in AKI; the ability to identify high-risk patients who may need changes in their medication dosing, different clinical resources (nephrology consult, dialysis), or those who may be the ideal candidate for future clinical trials. Prospective investigation of these sub-phenotypes is needed.

Decision support and electronic alert systems (e-alerts) represent an opportunity to pair with AI risk scores to improve early recognition and early AKI care. e-Alerts may improve patient outcomes and alter physician decisions, but to date, there are mixed results. Initially e-alerts in AKI were focused on medication management, e-mail alerts for patients or computer alerts for physicians about nephrotoxic or renal-excreted medications.^49,50 Colpaert and colleagues were among the first to describe using the Risk Injury Failure Loss and End-stage criteria to alert ICU physicians about AKI by triggering an automatic message sent by cordless telephone.⁵¹ While this trial describes the successful implementation of the alert system and acceptance of the alert by the end user, it did not seek to evaluate if such alerts could lead to earlier interventions or improved outcomes.⁵¹ Several years later, Selby and colleagues reported a hospital-wide e-alert system using the AKI-Network diagnostic criteria that had a false-positive rate of 1.7% and a false-negative rate of 0.2%.⁵² Furthermore, Hodgson and colleagues investigated if an AKI e-alert which was triggered by an electronic clinical prediction rule in patients at high risk of developing AKI could reduce hospital-acquired AKI and improve patient outcomes.⁵³ This controlled beforeand-after study of 30,295 acute medical admissions in 2 South of England hospitals (2014–2016) demonstrated a reduction in incidence of hospital acquired AKI and in-hospital mortality via a difference-in-differences analysis.⁵³ However, e-alerts had no effect on outcomes for patients admitted with established AKI from the community.⁵³ Separately, Wilson and colleagues sought to determine whether e-alerts for patients with newly diagnosed AKI would improve the primary (composite) outcome of mortality, receipt of dialysis, and progression of AKI. This study enrolled 6030 adult patients from 6 hospitals (Yale New Haven Health System) including both small community hospitals and large tertiary care centers.⁵⁴ These alerts did not reduce the risk of the primary outcome with heterogeneity of effects across clinical centers.⁵⁴ Including a signal for harm in patients who received an alert in nonteaching hospitals, perhaps highlighting the importance of having a structured intervention linked to an AKI alert.⁵⁴ Collectively, these e-alert studies reveal conflicting data on the efficacy of e-alerts in the setting of established AKI; however, combining a pre-emptive AKI risk score with e-alerts and a directed care bundle may help prevent or recognize AKI earlier and lead to improved clinical outcomes.

While AI holds much promise for the AKI and nephrology in general, the work described above is not without its challenges. There are hurdles that exist in obtaining and aggregating the large data sets needed to perform AI work. As with all clinical research data, care must be taken to ensure that all data are deidentified and that patient’s private health information cannot become public. However, data aggregation and sharing between different institutions has its own unique challenges. Once data use agreements and legal issues are settled, data cleaning and harmonization can be started. Ensuring that the data aligned are the first step, a variable that 1 institution calls a pulse may be labeled as heart rate in another system, so care must be taken to aggregate date correctly. Similarly, care must be taken to make sure that data are accurate and physiologic. Nonphysiologic values (whether vital signs or laboratory values) need to be identified and removed. While there is no accepted standard for how to handle these and other missing data, some have suggested imputation using the mean or median while others have chosen to keep them as missing.^42,55 Lastly it is important to ensure that algorithms are built as bias-free tools; as a nephrologist, we are well aware of the damage that a biased tool (eg, race-based equations for GFR) can cause.^56,57 All AKI algorithms should be constructed fairly and not disadvantage any patient group based on race, gender, social-economic status, or preexisting medical condition. While a complete discussion around building bias-free AI tools and the hurdles of conducting AI research is beyond the scope of this review, it is important to recognize some of the major issues confronting this line of investigation.

The use of AI to predict AKI is an ongoing area of intense investigation. The limited existing literature suggests risk scores can easily be developed and validated using EHR data and appear to be very generalizable and highly accurate. Contemporary AI applications can predict AKI before changes in SCr, a century-old functional biomarker of glomerular filtration. Pairing these cutting-edge risk-assessment scores with e-alerts, care-bundles, and kidney focused care (eg, avoidance of nephrotoxins) can possibly reduce AKI severity as well as its associated morbidity and mortality. Modern use of AI requires access to data sources, ability to clean and interrogate that data with knowledge of advanced biostatistics, language processing, deep learning, and understanding how to develop and implement a clinical risk score. Moving forward, it is imperative for us to work to create an AI-competent workforce as currently there is a paucity of training in the areas of AI and clinical decision support in nephrology and all other training programs (save for informatics fellowships). Furthermore, the data sets and data interpretation should attempt to minimize bias, and gaining the skills to recognize and avoid biases requires further technical skills. AI-developed AKI risk scores are ready for clinical implementation. There are several options already in the literature, and we anticipate further growth over the next decade. While these AI tools often suffer due to provider mistrust and uncertainty, as such these risk scores need to prove that they improve patient outcomes when clinically implemented.

CLINICAL SUMMARY.

• Static risk models for detecting acute kidney injury are slowly being replaced by dynamic models which use advanced learning and real-time data.

• In several clinical investigations, artificial intelligence has improved our ability to predict the future development of acute kidney injury in hospitalized patients.

• Implementation of these advanced learning models has yet to demonstrate a clinically meaningful benefit.