Prediction of Acute Kidney Injury With a Machine Learning Algorithm Using Electronic Health Record Data.

BACKGROUND: A major problem in treating acute kidney injury (AKI) is that clinical criteria for recognition are markers of established kidney damage or impaired function; treatment before such damage manifests is desirable. Clinicians could intervene during what may be a crucial stage for preventing permanent kidney injury if patients with incipient AKI and those at high risk of developing AKI could be identified.
OBJECTIVE: In this study, we evaluate a machine learning algorithm for early detection and prediction of AKI.
DESIGN: We used a machine learning technique, boosted ensembles of decision trees, to train an AKI prediction tool on retrospective data taken from more than 300 000 inpatient encounters.
SETTING: Data were collected from inpatient wards at Stanford Medical Center and intensive care unit patients at Beth Israel Deaconess Medical Center. PATIENTS: Patients older than the age of 18 whose hospital stays lasted between 5 and 1000 hours and who had at least one documented measurement of heart rate, respiratory rate, temperature, serum creatinine (SCr), and Glasgow Coma Scale (GCS). MEASUREMENTS: We tested the algorithm's ability to detect AKI at onset and to predict AKI 12, 24, 48, and 72 hours before onset.
METHODS: We tested AKI detection and prediction using the National Health Service (NHS) England AKI Algorithm as a gold standard. We additionally tested the algorithm's ability to detect AKI as defined by the Kidney Disease: Improving Global Outcomes (KDIGO) guidelines. We compared the algorithm's 3-fold cross-validation performance to the Sequential Organ Failure Assessment (SOFA) score for AKI identification in terms of area under the receiver operating characteristic (AUROC).
RESULTS: The algorithm demonstrated high AUROC for detecting and predicting NHS-defined AKI at all tested time points. The algorithm achieves AUROC of 0.872 (95% confidence interval [CI], 0.867-0.878) for AKI detection at time of onset. For prediction 12 hours before onset, the algorithm achieves an AUROC of 0.800 (95% CI, 0.792-0.809). For 24-hour predictions, the algorithm achieves AUROC of 0.795 (95% CI, 0.785-0.804). For 48-hour and 72-hour predictions, the algorithm achieves AUROC values of 0.761 (95% CI, 0.753-0.768) and 0.728 (95% CI, 0.719-0.737), respectively. LIMITATIONS: Because of the retrospective nature of this study, we cannot draw any conclusions about the impact the algorithm's predictions will have on patient outcomes in a clinical setting.
CONCLUSIONS: The results of these experiments suggest that a machine learning-based AKI prediction tool may offer important prognostic capabilities for determining which patients are likely to suffer AKI, potentially allowing clinicians to intervene before kidney damage manifests.

What was known before

Early detection of acute kidney injury (AKI) is necessary for clinicians to intervene in early stages of disease progression and prevent kidney damage. However, AKI can be difficult to detect before kidney damage and impaired function are present.

What this adds

The machine learning algorithm described in this study is capable of predicting an AKI up to 72 hours before onset, allowing for early clinical intervention.

Introduction

Acute kidney injury (AKI) is common, affecting 5% to 7% of all hospitalizations and causing $10 billion of additional health care–related expenditures per year through per-hospitalization excess costs of $7933.[1-3] Acute kidney injury is associated with increased mortality, end-stage renal disease, and chronic kidney disease, which can require ongoing dialysis and kidney replacement.[4-6] There exists some controversy as to how to best treat patients experiencing AKI. Standard approaches include reducing or eliminating nephrotoxic and antibiotic medications, relieving possible obstruction, and correcting electrolyte and fluid imbalances.[7,8] However, the effectiveness of these interventions may be limited by an inability to consistently identify patients with active or incipient AKI.[9] A system which identifies incipient AKI or predicts clinical manifestations of AKI with a substantial lead time may enable clinicians to better assess existing and novel interventions, and to ultimately provide more effective therapy which mitigates or avoids AKI and long-term kidney damage.

It has been recognized that early identification of AKI is desirable in hospital settings and that even small increases in serum creatinine (SCr) levels are associated with long-term damage and increased mortality.[2,10] Furthermore, accurate prediction of AKI onset before patients meet clinical criteria for recognition is advantageous, as such current clinical criteria represent markers of established kidney damage or impaired function.[11,12] Electronic health records present an opportunity to utilize machine learning techniques for predicting AKI and sending automated alerts for individual patients at risk of developing AKI. Several studies have assessed clinical decision support (CDS) tools for early detection of AKI, but many of these tools suffer from a variety of design and performance problems. These issues include lack of predictive ability, lack of an e-alert implementation, heavy tradeoffs between sensitivity and specificity, and restrictions to limited patient populations such as intensive care unit (ICU), postcardiac surgical, or elderly patients.[13-16]

In this paper, we describe an approach based on a machine learning algorithm (MLA)—a procedure which, in this case, identifies the statistical patterns in electronic health record data corresponding to AKI-related outcomes and the result of which is a software-based prediction tool intended to provide significant, accurate advance warning of AKI. Machine learning methods may provide advantages for AKI detection, as they can be trained to predict AKI far in advance of onset, can maintain concurrently high sensitivity and specificity, and can be customized to specific populations for increased accuracy. The machine learning method assessed in this study was that of gradient boosted trees, a method that iteratively combines the results of multiple decision trees into an overall risk prediction score.

The machine learning method assessed in this study was that of gradient boosted trees, a method that iteratively combines the results of multiple decision trees into an overall risk prediction score. The objective of this study was to assess the performance of the MLA for detecting AKI onset and predicting an impending AKI 12, 24, 48, and 72 hours before onset. In addition, we compared the performance of this prediction tool to the Sequential Organ Failure Assessment (SOFA) score.[17] The SOFA score is a commonly used disease severity scoring system which was developed to assess organ function in hospitalized patients. The SOFA score is tabulated from subscores for each of respiratory, coagulation, liver, cardiovascular, renal, and neurological systems. In past work, the SOFA score has been shown to independently predict AKI risk and outcomes, and thus serves as an important comparator for our approach.[12,18]

Materials and Methods

Data Sources

Data used in this study were drawn from the 651-bed Beth Israel Deaconess Medical Center (BIDMC; Boston, Massachusetts) and from the 613-bed Stanford University Medical Center (Stanford, California). BIDMC data were collected from the Medical Information Mart for Intensive Care III (MIMIC-III) v1.3 database.[19] This database was compiled by the MIT Laboratory for computational physiology and contains 61 532 inpatient ICU encounters collected between 2001 and 2012. The Stanford University data set contains 286 797 inpatient encounters from all hospital wards between December 2008 and May 2017.

For both data sets, we included only those patients whose hospital stays lasted between 5 and 1000 hours. Of those patients, we included only those who had at least one measurement of each required measurement (see “Imputation and Feature Creation” section), and who had age data available and were older than 18 years of age. Patients with chronic kidney disease were not excluded from the final patient population. The inclusion flowchart is presented in Figure 1.

Figure 1.

Inclusion criteria for patients in the BIDMC and Stanford data sets.

Note. Patients who met all inclusion criteria were included in this study. BIDMC = Beth Israel Deaconess Medical Center.

aRequired measurements include heart rate, respiratory rate, temperature, Glasgow Coma Scale, and serum creatinine.

Data collection for both data sets was passive and had no impact on patient safety. Both data sets were deidentified in compliance with the Health Insurance Portability and Accountability Act (HIPAA) Privacy Rule. Studies performed on deidentified data constitute nonhuman subject research, thus no institutional or ethical approvals were required for this study.

Statistical Analysis

Data Processing

All data from both data sets were processed by custom database queries. The retrieved data were converted into flat .csv files, which were in turn loaded into a custom data processing code written in the programming language Python (Python Software Foundation, https://www.python.org/). This code associated each measurement or observation with a timestamp and a measurement type key. Demographics and other patient characteristics (eg, age) were stored with a similar keyed retrieval mechanism.

Imputation and Feature Creation

Beginning at the time of the first recorded patient measurement, all data were discretized into 1-hour intervals. If multiple observations of the same patient measurement were taken within a given hour, those measurements were averaged to produce a single value for that hour. This ensured that the rate at which measurements were fed into the algorithm was standardized across patients. If no measurement of a clinical variable was available for a given hour, a carry-forward imputation method was employed to fill the missing measurement with the most recently available previous measurement, a causal procedure. Details of these data processing methods have been described in a previous publication on the use of gradient boosted trees for sepsis detection and prediction.[20]

For all experiments, we generated MLA predictions using patient data on heart rate, respiratory rate, temperature, SCr, and Glasgow Coma Scale (GCS). These measurements were selected because they are frequently and easily collected at the bedsides, even before clinical suspicion of AKI is present. After imputation and averaging, for each prediction time, we took our causal feature vector to include the previous 5 hourly values of each of heart rate, respiratory rate, temperature, SCr, and GCS as well as the patient’s age.

Gold Standard

We implemented the National Health Service (NHS) England AKI Algorithm as our gold standard.[21] This system is based on Kidney Disease: Improving Global Outcomes (KDIGO) guidelines,[22] but relies exclusively on changes in SCr levels to determine the presence and staging of AKI. The NHS algorithm is an appropriate gold standard for this work because it was designed explicitly for early AKI detection and generation of e-alerts for affected patients, and because it does not rely on urine output, which has been shown to be a poorer indicator of AKI than SCr and is subject to poor documentation, particularly in the emergency department.[23,24]

We determined the presence of AKI for adult inpatients only. Using either the lowest value from the past 0 to 7 days or the median value from the past 8 to 365 days as a baseline reference value, the ratio of current SCr levels to the reference value was calculated as in the NHS Algorithm (Supplemental Table 1). We computed these ratios using SCr measurements from the past 0 to 7 days whenever these data were available, using measurements from the past 8 to 365 days in all other cases. We determined the MLA’s ability to predict stage 2 or stage 3 AKI at 0, 12, 24, 48, and 72 hours before onset.

We additionally assessed the algorithm’s ability to detect stage 2 or stage 3 AKI as determined by the KDIGO criteria (Supplemental Table 2). For these experiments, we included all patients who had at least 1 recorded observation of the 5 required measurements, plus at least 1 observation of urine output (Supplemental Tables 2 and 3). The KDIGO criteria require a premorbid SCr level to be used as a baseline; for patients for whom no premorbid SCr measurements were available, a baseline SCr was estimated using the Modification of Diet in Renal Disease (MDRD) equation.[25] We tested the MLA’s ability to detect KDIGO stage 2 or stage 3 AKI at 0, 12, and 24 hours before onset.

Machine Learning and Experimental Methods

All predictors trained in this work are boosted ensembles of decision trees produced using the XGBoost package for Python.[26] The boosting process improves predictions by successively adding new decision trees to a growing ensemble of trees, where new trees are trained to perform better on those patients who are misclassified by the current ensemble.

We used 3-fold cross-validation to assess the performance of the algorithm under the NHS gold standard separately on the BIDMC and Stanford data sets, in which we divided each data set into thirds, trained a predictor on two of the thirds, and tested the trained predictor on the remaining third. This process was repeated so that training occurred on all possible combinations of thirds; thus, all included patients were cycled through during the training process. The data sets were randomly divided, with randomization based on patient identification number. We measured area under the receiver operating characteristic (AUROC), accuracy, diagnostic odds ratio (DOR), and positive and negative likelihood ratios (LR+ and LR−) obtained by the MLA via this method. Our reported metrics are the average metrics of 30 independently trained models, each trained using 3-fold cross-validation. Randomization was performed before training of each model.

We compared these results with those same measures obtained by the SOFA organ dysfunction score[17] on both data sets. The SOFA score was calculated as in study of Pandharipande et al,[27] with SpO2/FiO2 ratios used in place of PaO2/FiO2 ratios due to data availability. Statistical comparisons were performed using pairwise, single-tailed t tests with significance set at P < .01.

To assess the algorithm under the KDIGO criteria, we used 10-fold cross-validation to separately train the algorithm on the BIDMC and Stanford data sets. We then assessed algorithm performance on test sets from the BIDMC data set. Reported performance metrics are the average of each of the 10 models generated by the 10-fold cross-validation process.

Results

Participants

The final patient population used to train and test the algorithm included 48 582 patients from the BIDMC data set and 19 737 patients from the Stanford data set. Patient demographics differed in several important ways between the two data sets (Table 1). The BIDMC data set contains only patients admitted to the ICU, while the Stanford data set contains all inpatients; BIDMC patients therefore represent a more critically ill population. In addition, the data sets display differences in age and gender. The Stanford data set skewed younger than the BIDMC data set, with around 15% of Stanford patients in the 18- to 29-year-old group and only around 4.5% of BIDMC patients in this group. Around 41% of BIDMC patients were older than the age of 70, while only around 14% of Stanford patients fell into this age group. The BIDMC data set also skewed more heavily male than the Stanford data set, with more than 56% of BIDMC patients male. Around 49% of Stanford patients were male. Prevalence of AKI was higher in the BIDMC than in the Stanford data set.

Table 1.

Patient Demographic Information for Complete BIDMC and Stanford Cohorts.

Characteristic	BIDMC (%)	Stanford (%)
Gender
Female	43.66	51.19
Male	56.44	48.81
Age (years)
18-29	4.51	15.23
30-39	5.26	11.22
40-49	10.64	11.22
50-59	17.50	13.20
60-69	20.98	12.69
70+	40.91	14.07
Severe AKI based on NHS England algorithm[a]
Yes	2.7%	0.5%
No	97.3%	99.5%
In-hospital death
Yes	9.2%	2.78%
No	90.8%	97.22%

Note. BIDMC = Beth Israel Deaconess Medical Center; AKI = acute kidney injury; NHS = National Health Service.

Prevalence of stage 2 or stage 3 AKI before filtering patients according to inclusion criteria.

Main Results

For detecting severe AKI under the NHS gold standard, the MLA demonstrated higher AUROC, accuracy, and DOR than the SOFA score at all prediction windows and for each data set. When tested on data collected from BIDMC, the MLA demonstrated an AUROC of 0.841 at time of onset while the SOFA score achieved an AUROC of 0.762 at time of onset. The MLA AUROC improved upon that of SOFA for all prediction windows (P < .01 for all windows). In addition, MLA accuracy and DOR remained superior for all prediction windows (12, 24, 48, and 72 hours prior to onset) (Tables 2 and 3). The algorithm had higher or comparable positive likelihood ratios (LR+) and comparable negative likelihood ratios (LR−) at onset and for all prediction windows. Full performance metrics for the MLA and SOFA when tested on BIDMC data are presented in Table 2.

Table 2.

Comparison of Performance Metrics for the MLA and for the SOFA Score Measured on Patient Data From Beth Israel Deaconess Medical Center.

Prediction time	Onset		12 hours		24 hours		48 hours		72 hours
Predictor	MLA	SOFA	MLA	SOFA	MLA	SOFA	MLA	SOFA	MLA	SOFA
AUROC (95% CI)	0.841 (0.837-0.844)	0.762	0.749 (0.744-0.755)	0.734	0.758 (0.754-0.762)	0.716	0.707 (0.701-0.713)	0.675	0.674 (0.669-0.679)	0.653
Sensitivity	0.81	0.55	0.77	0.54	0.83	0.78	0.83	0.84	0.82	0.82
Specificity	0.75	0.79	0.62	0.78	0.56	0.57	0.48	0.41	0.45	0.39
Accuracy	0.81	0.57	0.76	0.55	0.82	0.76	0.82	0.81	0.80	0.79
DOR	13.1	4.8	5.5	4.2	6.2	4.7	4.5	3.6	3.7	3.0
LR+	3.3	2.7	2.0	2.5	1.9	1.8	1.6	1.4	1.5	1.3
LR−	0.25	0.56	0.37	0.59	0.30	0.39	0.35	0.39	0.40	0.46

Note. Predictions were made at 0, 12, 24, 48, and 72 hours before stage 2 or stage 3 AKI onset. Operating points for the MLA were chosen to keep sensitivities close to 0.80. 95% CIs were calculated only for the MLA. MLA = machine learning algorithm; SOFA = sequential organ failure assessment; AUROC = area under the receiver operating characteristic; CI = confidence interval; DOR = diagnostic odds ratio; LR = likelihood ratios; AKI = acute kidney injury.

Table 3.

Comparison of Performance Metrics for the MLA and for the SOFA Score Measured on Patient Data From Stanford Medical Center.

Prediction time	Onset		12 hours		24 hours		48 hours		72 hours
Predictor	MLA	SOFA	MLA	SOFA	MLA	SOFA	MLA	SOFA	MLA	SOFA
AUROC (95% CI)	0.872 (0.867-0.878)	0.815	0.800 (0.792-0.809)	0.781	0.795 (0.785-0.804)	0.764	0.761 (0.753-0.768)	0.732	0.728 (0.719-0.737)	0.720
Sensitivity	0.77	0.73	0.75	0.73	0.79	0.55	0.85	0.53	0.78	0.51
Specificity	0.82	0.78	0.73	0.74	0.64	0.83	0.51	0.79	0.53	0.81
Accuracy	0.78	0.73	0.75	0.73	0.79	0.56	0.84	0.54	0.79	0.53
DOR	15.5	9.7	8.0	7.3	6.9	5.9	5.8	4.3	4.4	4.3
LR+	4.3	3.4	2.7	2.7	2.2	3.2	1.7	2.6	1.7	2.7
LR−	0.28	0.35	0.34	0.37	0.32	0.55	0.30	0.60	0.38	0.61

Note. Predictions were made at 0, 12, 24, 48, and 72 hours before stage 2 or stage 3 AKI onset. Operating points were chosen to keep sensitivities close to 0.80. 95% CIs were calculated only for the MLA. MLA = machine learning algorithm; SOFA = sequential organ failure assessment; AUROC = area under the receiver operating characteristic; CI = confidence interval; DOR = diagnostic odds ratio; LR = likelihood ratio; AKI = acute kidney injury.

The algorithm also demonstrated superior performance when trained and tested for NHS gold standard AKI on patient data from Stanford Medical Center. At time of onset, the MLA demonstrated an AUROC of 0.872, while the SOFA score demonstrated an AUROC of 0.815. As on BIDMC data, the MLA AUROC exceeded that of the SOFA score for all prediction windows (P < .01 for all windows). MLA accuracy was higher than that of the SOFA score for all prediction windows (Table 3). The MLA also demonstrated improved DOR compared with the SOFA score. Full performance measures for the algorithm and SOFA when tested on Stanford data are presented in Table 3.

We note that the performance metrics in Tables 2 and 3 were measured for prediction sensitivities held near 0.80, to facilitate comparison of metrics across prediction times. The MLA performance across all such operating points (ie, choices of sensitivity) is summarized in Figures 2 and 3. Figure 2 provides a receiver operating characteristic (ROC) curve comparison across prediction times for the MLA trained and tested on BIDMC data, and algorithm performance on Stanford data is displayed in Figure 3. On both the BIDMC and Stanford data sets, MLA performance declined gradually as the prediction window was lengthened from 0 hours to 72 hours before AKI onset.

Figure 2.

Comparison of the receiver operating characteristic and area under the receiver operating characteristic for machine learning algorithm 0-, 12-, 24-, 48-, and 72-hour advance prediction of stage 2 or stage 3 acute kidney injury development for BIDMC patient data.

Note. BIDMC = Beth Israel Deaconess Medical Center.

Figure 3.

Comparison of the receiver operating characteristic and area under the receiver operating characteristic for machine learning algorithm 0-, 12-, 24-, 48-, and 72-hour advance prediction of stage 2 or stage 3 acute kidney injury development for Stanford Medical Center patient data.

The algorithm also demonstrated high AUROC, sensitivity, and specificity for detecting stage 2 or stage 3 AKI under the KDIGO criteria (Table 4). When trained on patient data from Stanford and tested on data from BIDMC, the algorithm demonstrated AUROC above 0.75 for AKI detection up to 24 hours in advance of onset.

Table 4.

Algorithm Performance for Detection of Stage 2 or Stage 3 Acute Kidney Injury Under the Kidney Disease: Improving Global Outcomes Criteria, Measured on Patient Data From BIDMC.

	Trained on BIDMC			Trained on Stanford
	Onset	12 hours	24 hours	Onset	12 hours	24 hours
AUROC (95% CI)	0.924 (0. 872-0.975)	0.914 (0.814-0.999)	0.882 (0.669-0.999)	0.844 (0.716-0.972)	0.826 (0.716-0.935)	0.760 (0.591, 0.929)
Sensitivity	0.987	0.999	0.900	0.981	0.971	0.933
Specificity	0.912	0.907	0.879	0.715	0.719	0.602

Note. BIDMC = Beth Israel Deaconess Medical Center; AUROC = area under the receiver operating characteristic; CI = confidence interval.

Discussion

The machine learning approach described here results in a prediction tool which demonstrates strong predictive performance, in terms of AUROC, up to 72 hours in advance of stage 2 or stage 3 AKI onset, under both the NHS and KDIGO criteria for AKI. Furthermore, for a given degree of sensitivity, the MLA outperforms the commonly used SOFA score in terms of specificity, accuracy, and other metrics. This performance was achieved using only 5 commonly collected patient measurements as inputs. By requiring the presence of only these 5 measurements for AKI predictions, this algorithm is designed to be able to predict AKI risk on a large portion of the hospital population in future clinical work. Based on these results, we believe this MLA could provide clinicians the opportunity to improve patient outcomes through earlier AKI detection and subsequent intervention, which may include volume resuscitation or avoidance of nephrotoxic medications to minimize further kidney injury.[28]

We emphasize that the MLA performed similarly well on the BIDMC and Stanford data sets, an observation which has important clinical implications. The BIDMC data included only patients admitted to the ICU, while the Stanford data set contained information about inpatient stays from all hospital wards. These two data sets thus represent hospital settings with different demographics, frequency of patient measurement collection, levels of care provision, and disease severity in patients. The predictive ability of the algorithm across these data sets suggests that the algorithm can identify patients at risk of AKI onset in a variety of hospital settings. Because AKI is a common complication of hospital stays of a diverse nature,[1] this ability is of central importance in an AKI prediction tool.

In previous studies, we assessed a MLA for sepsis detection in both retrospective[29-31] and prospective settings.[32,33] Our prospective work includes a randomized controlled trial (RCT) conducted in 2 adult ICUs at the University of California, San Francisco. In this study, there was a statistically significant decrease in the primary endpoint of mean hospital length of stay (13.0 days vs 10.3 days, P = .0421) and in the secondary endpoint of in-hospital mortality rate (21.3% vs 8.96%, P = .0176).[32] This study demonstrates the feasibility of bedside use of a machine learning–based prediction system, as well as the potential for such a system to improve patient outcomes through earlier and more accurate detection of patient conditions. Future clinical implementation of the system described in this work is intended not to disrupt clinical workflow or require additional work from care providers. As with past implementations of our sepsis detection system, more frequent collection of patient measurements is not required; the same forward-filling imputation method described in this paper can be used to make accurate predictions in clinical settings.

Machine learning methods have previously been applied to AKI detection. One such study used a random forest model to successfully predict AKI with AUROC up to 0.84.[14] However, this model was developed and validated using only ICU data, thus may have limited applicability to other settings. Another study demonstrated the successful application of a discrete-time survival model to predict AKI development on the ward.[13] Additional work has compared logistic regression, support vector machine, naive Bayes, and decision tree models for AKI detection on a population of patients 60 years or older.[15] The work presented here advances the field by evaluating a machine learning method trained and tested on a mixed-ward population that includes adults of all ages.

A machine learning approach provides advantages over currently used systems. Unlike the SOFA score, which is a generalized disease severity score, this method is specific to AKI. The measurable benefits (Tables 2 and 3) of focusing on AKI prediction could allow clinicians to more rapidly determine the cause of patient deterioration and, thus, administer appropriate treatments in a more timely manner. In addition, the ROC curves of Figures 2 and 3 provide a continuous range of sensitivity-specificity pairs at which the MLA can operate. If fewer alerts, greater specificity, and 72-hour notice were preferable over more alerts, greater sensitivity, and nearer-onset notice, the MLA could function accordingly. This flexibility is not available for a rules-based score like SOFA. The MLA also may provide advantages over manual AKI detection methods, which may not be implemented unless a physician already suspects AKI, and are subject to human error. Future work will investigate ensemble learning methods for further improved accuracy in AKI detection.

Limitations

Because this work presents a retrospective study, we cannot draw strong conclusions about this algorithm’s performance in a live clinical setting. We cannot determine from the nature of this study what impact the algorithm might have on clinicians and the care which they provide. This study assesses algorithm performance only on US patients older than the age of 18, and results may not be generalizable to additional populations. In prospective settings, if the algorithm is implemented on patient populations which differ substantially from those used in this study, the predictive performance of the algorithm may differ. Indeed, our cross-validation analysis only allows us to conclude that the performance we report would generalize well to patient populations similar to the BIDMC and Stanford data sets. Because this study does not examine variables such as cause of admission or patient comorbidities, we cannot determine from this study whether there is a subpopulation of patients for whom this algorithm may be most useful.

Because there have been several proposed consensus definitions for AKI, our predictive algorithm may have different results when compared against various gold standard definitions, or in prospective clinical settings which utilize a different gold standard in their diagnostic procedures. Due to missing observations of urine output in the data sets used in these experiments, the training and testing sets for the KDIGO criteria[21] were limited in size. Performance against this criteria in a different retrospective or clinical setting may therefore differ from the results presented here.

Conclusion

The machine learning approach described in this study accurately predicts stage 2 or stage 3 AKI up to 72 hours in advance of onset on when trained and tested on two distinct data sets. This algorithm may improve detection of AKI in clinical settings, allowing for earlier intervention and improved patient outcomes.

Click here for additional data file.

Supplementary Material, cjkhd_supplemental_tables for Prediction of Acute Kidney Injury With a Machine Learning Algorithm Using Electronic Health Record Data by Hamid Mohamadlou, Anna Lynn-Palevsky, Christopher Barton, Uli Chettipally, Lisa Shieh, Jacob Calvert, Nicholas R. Saber, and Ritankar Das in Canadian Journal of Kidney Health and Disease