Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Aug 7.
Published in final edited form as: Am J Kidney Dis. 2026 Jun 19;88(3):414–423.e1. doi: 10.1053/j.ajkd.2026.04.010

Electronic Phenotype for Detection, Staging, and Subtyping of Acute Kidney Injury

Ning Shang 1, Katherine Xu 1, Jacob S Stevens 1, Jonathan Barasch 1, Krzysztof Kiryluk 1
PMCID: PMC13445448  NIHMSID: NIHMS2200733  PMID: 42320578

Abstract

Rationale & Objective:

Distinguishing between transient and sustained subtypes of acute kidney injury (AKI) among hospitalized patients is valuable for clinical management and risk stratification. The objective of this study was to develop and validate a pragmatic electronic phenotype (e-phenotype) for the diagnosis, staging, and subtyping of AKI using electronic health record (EHR) data.

Study Design:

Development of a computable rule-based algorithm to diagnose, stage, and subtype AKI using longitudinal changes of serum creatinine values recorded in an electronic health record. Assessment of the test characteristics of these algorithm-based diagnoses was implemented using a set of patients admitted to the Emergency Department (ED) with or without clinically diagnosed AKI. Validation of AKI diagnoses was implemented in two ways: using a set of patients hospitalized for COVID-19 and using a dataset of patients hospitalized for any cause following an ED visit. These analyses examined the associations of AKI stage and subtype with mortality.

Setting & Participants:

Assessment of the test characteristics of the algorithm-based diagnoses: 90 ED visits for AKI and 376 visits without clinical evidence of AKI at Columbia University. Validation in the setting of COVID-19: EHR data from 117,514 instances of COVID-19 infection diagnosed at Columbia University Medical Center throughout the pandemic. Validation in the setting of general hospital admissions: 405,467 general hospital admissions at Beth Israel Medical Center, Boston, MA.

Tests Compared:

The algorithm-based diagnosis, stage, and subtype of AKI were compared with the clinically adjudicated AKI diagnosis, stage, and subtype.

Outcomes:

AKI detected, staged, and subtyped by an electronic algorithm, and 30-day mortality.

Results:

The AKI e-phenotype had a positive predictive value of 95.4%, sensitivity of 69.0%, specificity of 99.2%, and overall accuracy of 93.4%. In COVID-19 patients, pre-existing CKD was an independent predictor of AKI. COVID-19-related AKI was associated with mortality in a stage-dependent manner. The subtype of sustained AKI was associated with higher mortality compared to transient AKI within and across all pandemic waves. These results were reproducible in the dataset of patients hospitalized for any condition.

Limitations:

Reliance on serum creatinine patterns alone and the inability to incorporate urine output or molecular markers to diagnose and subtype AKI.

Conclusions:

The AKI e-phenotype is an accurate, scalable, and generalizable to diverse EHR datasets with reproducible associations with mortality.

Keywords: Acute Kidney Injury (AKI), Chronic Kidney Disease (CKD), Electronic Phenotyping, COVID-19

Plain-Langage Summary

Acute kidney injury (AKI) is a common and serious complication in hospitalized patients, but identifying and tracking it accurately in electronic health records is challenging. We developed an algorithm that uses patterns in kidney function test data to detect, classify, and distinguish AKI subtypes. After testing and validating the method, we applied it to two large hospital databases: one including patients hospitalized with COVID-19 and another including hospital admissions for any cause. Our approach reliably identified AKI and its association with poor outcomes. Patients with more severe or sustained AKI were more likely to die. This tool may help researchers study kidney injury more consistently across different healthcare settings.

Introduction

Acute kidney injury (AKI) is characterized by a rapid decline in kidney function and is clinically defined by an acute increase in serum Cr. The causes of AKI are highly heterogeneous, ranging from reversible hemodynamic causes, such as reduced blood flow to the kidneys from sepsis or volume depletion to severe tissue injury such as in acute tubular necrosis. The clinical distinction between hemodynamic etiologies and early intra-renal injury can be difficult. The major challenge is due to the diagnostic use of serum Cr (a functional biomarker produced by muscle that builds up in serum with reduced kidney filtration) rather than an injury biomarker directly reflective of kidney damage1–3. In approximately 75% of AKI cases, serum Cr improves within 48 hours, but the remaining 25% of cases experience sustained elevation of serum Cr despite treatment2. Sustained AKI correlates with urinary biomarkers of tissue injury and with serious complications, including electrolyte imbalances, fluid overload, and even death4,5. Thus, accurate distinction between transient and sustained subtypes is critical for proper management and risk stratification of hospitalized patients. This premise provides a strong motivation for the design of electronic algorithms that not only detect and stage AKI but also distinguish transient from sustained AKI subtypes.

We have previously designed an electronic phenotype for the diagnosis and staging of chronic kidney disease (CKD) that utilized the KDIGO (Kidney Disease: Improving Global Outcomes) staging on a classification grid of albuminuria (A-stage) by glomerular filtration rate (G-stage)6,7. Similarly, the KDIGO classification of AKI based on temporal changes in serum Cr levels or urine output is widely used by clinicians8,9. While serum creatinine is available in electronic health records, urine output is not routinely recorded for most patients outside the intensive care units, thus urine output criteria are typically not considered in electronic AKI detection algorithms10,11. Importantly, although KDIGO classification provides a standardized framework for the AKI diagnosis, the reliance on serum Cr has major limitations due to its poor sensitivity and delayed rise after injury12,13. Moreover, the KDIGO criteria do not differentiate sustained AKI (more likely due to tissue damage) from transient AKI (more likely due to reversible causes).

The objective of this study was to develop and validate a pragmatic electronic phenotype (e-phenotype) that leverages longitudinal SCr trajectories to detect, stage, and subtype AKI. To demonstrate the utility of our proposed e-phenotype, we applied it to two large independent datasets. First, we analyzed patients admitted with coronavirus disease 2019 (COVID-19) during the entire pandemic period and examined the risk factors and outcomes of COVID-19-related AKI. Second, we performed parallel analyses in the MIMIC-IV dataset of general hospital admissions. Both applications illustrate the reproducibility and scalability of our phenotyping approach. We demonstrate that sustained AKI is associated with worse clinical outcomes in both datasets. In combination with the previously published CKD e-phenotype, this work provides a broadly applicable set of computable tools for studies of acute and chronic kidney diseases based on electronic health records (EHR).

Methods

Our study protocol was approved by the Columbia University Internal Review Board, protocol number AAAS9994. Our reporting of fulfills the requirements of the TRIPOD checklist.

Electronic Phenotyping of AKI

To automate AKI diagnosis and staging based on structured data contained in EHR, we used daily serum Cr increments as defined by the KDIGO guidelines8,9, but with several pragmatic modifications to make the algorithm robust when applied to real-life EHR data. First, we do not use urine output criteria because urine output is not reliably available in EHR for the vast majority of hospitalized patients. Second, we used AKI diagnosis and staging criteria based only on relative changes in kidney function disregarding the absolute cut-offs. This choice aims to reduce false positives for mild AKI that arise from the use of absolute increment in serum creatinine of 0.3 mg/dL without accounting for nonrenal factors (e.g. BMI, medications, diet, or cardiovascular hemodynamics), a known limitation of the KDIGO definition14–18. Third, we modified the operational definition of baseline serum creatinine used in the algorithm. The KDIGO defines the baseline as the median serum creatinine level in the 7 to 365 days prior to presentation (hospital admission), but alternative methods have been proposed to compensate for data missingness19. We used the sequential algorithmic approach, where we first applied the KDIGO baseline definition as the preferred strategy. However, if no creatinine data were available 7 to 365 days prior to presentation, we used a minimum serum creatinine value for the 7 days prior to admission. If these data were not available, we used a minimum serum creatinine level from the time of admission to the time of an AKI event. This data-driven approach is more accurate than the alternative methods that propose imputing baseline serum Cr based on age, sex, and race20.

AKI severity was staged using the established AKIN system (stages 1–3)10,11. The algorithm also defines AKI subtypes based on duration: transient AKI (tAKI, lasting less than 48 hours) and sustained AKI (sAKI, lasting more than 48 hours). These subtypes were based on the definitions proposed by the consensus report of the Acute Disease Quality Initiative (ADQI) 16 Workgroup21 as subsequently operationalized by Stevens et al.19. After defining the first AKI event, the algorithm also assesses the AKI recurrence and counts the number of AKI occurrences during a given hospitalization. Recurrence of AKI was defined as another increase in serum creatinine 1.5-times over baseline more than two days after resolution of the previous AKI event. The severity and subtype are determined for each AKI occurrence. Figure 1 summarizes the overall workflow of our phenotyping algorithm.

Figure 1. Acute Kidney Injury (AKI) Electronic Phenotype Algorithm Summary.

Figure 1.

Open in a new tab

This rule-based algorithm first pre-filters patient records based on pre-existing ESRD and available SCr data. Next, “Baseline SCr” is defined using ordered priority rules based on available retrospective EHR kidney function data. Once the first instance of SCr >50% baseline is detected, an “AKI Block” is defined as a number of consecutive days with SCr >50% baseline; the block is terminated on the day that SCr returns below 50% baseline. Recurrent AKI is defined by a subsequent elevation of SCr >50% baseline more than 2 days after the previous block. For each block, AKI stage and subtype are defined. AKI stage is defined based on the maximum SCr per block. AKI subtype is defined based on the rules depicted in the table, with “YES” (red) indicating a day with SCr exceeding 50% baseline, “NA” (blue) indicating a day with missing kidney function data, and “No” indicating a day with SCr below 50% baseline. We use average SCr for the days with multiple SCr measurements. The algorithm’s outputs are depicted in orange. The staging and subtyping for only block 1 (first AKI event after admission) was used in the analysis of COVID-19 data.

Validation of AKI Electronic Phenotype

To validate our algorithm, we used the Columbia University emergency department (ED) admission-related AKI dataset previously adjudicated by Stevens et al.19. This gold-standard dataset contained 90 AKI and 376 non-AKI ED visits verified by manual chart reviews by two expert nephrologists blinded to our algorithm’s output. We applied our algorithm to EHR data for these patients and calculated sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall accuracy against manual adjudication. To additionally assess how our modified definition of baseline renal function affects the algorithm’s performance, we limited validations to a subset of 61 cases and 235 controls with baseline serum Cr measured 7–365 days prior to presentation.

Application I: COVID-19 Dataset

For the first application, we utilized EHR data for 117,514 COVID-19 occurrences in 89,400 patients diagnosed at Columbia University Irving Medical Center (CUIMC) during the pandemic period (Table S1). COVID-19 electronic phenotype was developed based on the World Health Organization (WHO) guidelines, N3C COVID-19 Phenotype Documentation (Version 2.2)22 and our local data characteristics. The cases were defined as having at least one positive SARS-CoV-2 RT-PCR or antibody test or, if no tests were available, at least three COVID-19 diagnosis or problem list codes. COVID-19 reinfection was identified if COVID-19 occurred more than 90 days after the onset of the previous infection23. In this study, we considered both COVID-19 primary infections and reinfections as distinct COVID-19 occurrences. All clinical encounters that occurred within eight weeks of COVID-19 occurrence were extracted, and consecutive daily encounters were merged to define a hospital admission. ICU-level care was defined by either ICU admission, or the need for vasopressors, high-flow nasal cannula (HFNC) oxygen, bilevel positive airway pressure (BiPAP), mechanical ventilation, extracorporeal membrane oxygenation (ECMO), or tracheal intubation. Severe COVID-19 was defined as ICU-level care or COVID-19-related death. COVID-19-related death was defined as death during COVID-19 hospitalization or within 30 days after the COVID-19 occurrence. For CKD diagnosis and staging, we used our previously validated electronic algorithm applied to pre-COVID-19 diagnosis data6,7. Other comorbid conditions were electronically phenotyped based on diagnosis and procedure codes (ICD-9, ICD-10, and CPT-4) with detailed algorithms accessible through the Phenotype Knowledgebase (PheKB: https://phekb.org/phenotype/acute-kidney-injury-aki) and GitHub (https://github.com/sunnyshang/pheAKI). We calculated the Charlson and Elixhauser comorbidity index, as proposed previously24–26.

The World Health Organization (WHO) declared COVID-19 a pandemic on March 11, 202027, and the US Federal COVID-19 Public Health Emergency ended on March 11, 202328,29, delineating a 3-year pandemic period for our analyses. To identify the COVID-19 pandemic waves in the US, we utilized the COVID-19 Data Repository by the Center for Systems Science and Engineering (CSSE) based at Johns Hopkins University30. We implemented the method proposed by Zhang et al.31 of using the effective reproduction number (R) to define each pandemic wave. By integrating our local trends in COVID-19 diagnoses at CUIMC, we mapped nationwide pandemic waves to our local data (Figure S1). Notably, New York City (NYC) was the initial epicenter of the US outbreak32–34, followed by a rapid spread of the virus to the rest of the country, producing two distinct waves early in the pandemic28,35. Accordingly, by plotting CUIMC case counts in relation to the national waves, these two waves together represented the first NYC wave; thus, only five local waves were analyzed. The SARS-CoV-2 variant emergence timeline36 was mapped to these five waves as follows: 1) the Initial (Alpha) Wave, 2) the Alpha/Beta Wave, 3) the Delta Wave, 4) the Omicron BA.1/2 Wave, and 5) the Omicron BA.4/5 Wave.

Logistic regression was used for testing the outcome of COVID-19 related AKI. Cox proportional-hazards models were used to assess associations with COVID-19-related mortality. The fully adjusted models included both A-stage and G-stage, demographics (age, sex, race, ethnicity), and pre-existing comorbidities (based on univariate screening with inclusion P<0.05). Statistical testing was performed individually for each wave of the pandemic. Subsequently, the association statistics were tested for heterogeneity and meta-analyzed across all waves under a fixed effects model.

Application II: MIMIC-IV Dataset

To demonstrate generalizability of our findings, we next analyzed the MIMIC-IV dataset (V3.1) and MIMIC-IV ED and Note (V2.2), a publicly available, de-identified EHR repository from the Beth Israel Deaconess Medical Center. The MIMIC-IV encompassed 405,467 hospital admissions through the emergency room (Figure S2). We implemented both, the AKI and CKD e-phenotypes in this dataset. To ensure data completeness, only patients with at least 3 serum Cr values, diagnosis ICD codes, and existing discharge summary were included in the analysis. First, we assessed the performance of the AKI e-phenotype against the documented diagnosis of AKI. We defined a documented diagnosis of AKI by the presence of any ICD code consistent with AKI and AKI being listed in the discharge summary as one of the hospital diagnoses. Controls were defined as having a discharge summary without a mention of AKI and no ICD codes consistent with the AKI diagnosis. Sensitivity, specificity, PPV, NPV and overall accuracy were calculated for AKI e-phenotype against the documented diagnosis. The performance was also assessed in a subset of patients admitted to the ICU, and in those with and without baseline serum Cr available 7–365 days prior to presentation (Table S2). We then tested for associations of AKI diagnosis, stage, and subtype with 30-day hospital mortality. Similar to COVID-19 dataset, we also tested for associations of pre-existing CKD A- and G-stages with the risk of AKI. The statistical models were similar to the COVID-19 dataset. Baseline A- and G-stages were determined using our CKD e-phenotype and were tested for association with AKI using logistic regression. Electronically-defined AKI stages and subtypes were tested for association with 30-day all-cause mortality using Cox proportional hazards models. All models were adjusted for demographics (age, sex, race/ethnicity), Elixhauser Comorbidity Index, and pre-existing comorbidities found significant (P<0.05) in corresponding univariate analyses.

All statistical analyses were conducted using R version 4.0.4 and Python 3.8.

Results

Electronic phenotype for AKI and its performance

Our proposed AKI e-phenotype is summarized in Figure 1. This algorithm achieved excellent overall performance in identifying AKI in our gold standard validation cohort, with a positive predictive value (PPV) of 95.4%, specificity of 99.2%, sensitivity of 69.0%, and accuracy of 93.4% (Table 1). Our subgroup analyses further demonstrate that the algorithm maintains very high specificity even when applied to patients without baseline serum Cr 7–365 days prior to presentation (specificity 99.6% vs. 98.6% for sub-cohorts with and without baseline data). For the classification of AKI stage and subtype, the algorithm correctly classified 90.3% (28/31) of AKI stages and 83.9% (26/31) of AKI subtypes among the true positives defined by manual chart review. These results provide a robust validation of our algorithm in the setting of a new hospital admission.

Table 1.

The performance of the new AKI e-phenotype in the clinically adjudicated validation dataset of 466 emergency department (ED) admissions.

Metrics AKI-ED Validation Cohort
(N=466)		 AKI-ED Validation Cohort with baseline SCr 7–365 days prior to presentation
(N=296)		 AKI-ED Validation Cohort without baseline SCr 7–365 days prior to presentation
(N=170)
AKI prevalence 19.3% 20.6% 17.1%
Positive Predictive Value (PPV) 95.4% 98.1% 83.3%
Sensitivity 68.9% 85.3% 34.5%
Specificity 99.2% 99.6% 98.6%
Negative Predictive Value (NPV) 93.0% 96.3% 88.0%
Accuracy 93.4% 96.6% 87.7%
Open in a new tab

Sensitivity analyses were also performed for a subgroup of cases with and without baseline SCr measurements 7–365 days prior to admission.

Application to the COVID-19 dataset

In total, we identified 89,400 COVID-19 patients and 117,514 COVID-19 occurrences at CUIMC during the pandemic period. Table S1 describes the demographics, comorbidities, and outcomes of the CUIMC study cohort. Overall, 10.7% of COVID-19 occurrences were hospitalized, 7.9% were classified as severe COVID-19 (defined by ICU-level care or COVID-19-related death), and 5.2% had COVID-19-related AKI defined by our e-phenotype. AKI occurred in 26.4% of hospitalized COVID-19 occurrences, and in 31.6% of severe COVID-19 occurrences. Notably, COVID-19 occurrences from the Initial Wave had the highest hospitalization rate (16.6%) and experienced high rates of shock (4.9%), transient AKI (5.3%) and sustained AKI (6.4%). Consistent with published findings, patients who developed COVID-19-related AKI were significantly older and more likely to have pre-existing diabetes, heart failure, and CKD compared to patients without AKI across most waves. Self-reported Black/African American race was also associated with higher risk of COVID-19-related AKI across all waves. Males were more frequently affected by COVID-19-related AKI, but only in the Initial Wave. Other co-morbidities associated with increased risk of AKI predominantly in the earlier waves of the pandemic included pre-existing coronary artery disease, hypertension, obesity, and cancer.

Association of pre-existing CKD with COVID-19–related AKI and mortality

We combined both the AKI and CKD e-phenotypes to explore the effect of pre-existing CKD (A and G stage) on the risk of COVID-19-related AKI and COVID-19 mortality (Figure S3). In the pooled analysis of all pandemic waves, we observed mutually independent effects of A- and G-stage on all outcomes. In a subgroup analysis by each wave, we observed a consistent pattern of higher risk for COVID-19-related AKI with increasing A and G stages across all waves. The multivariate analyses confirmed that pre-existing A- and G-stages were independently associated with higher risk of COVID-19-related AKI across the pandemic (adjusted ORA-stage 1.68, 95%CI: 1.52–1.84 and adjusted OR G-stage 1.61, 95%CI 1.45–1.78, respectively), as well as within each wave (Figure 2). In the analysis of mortality, the A-stage remained a significant independent predictor of death across the entire pandemic (adjusted HRA-stage 1.62, 95%CI: 1.39–1.90), while the G-stage became non-significant after accounting for A-stage and other risk factors (adjusted HRG-stage 1.11, 95%CI: 0.95–1.31).

Figure 2. Effects of pre-existing CKD stages on COVID-19-related AKI (top), and mortality (bottom) in fully adjusted models: (a) A-stage; (b) G-stage.

Figure 2.

Open in a new tab

All models were adjusted for demographics (age, age squared, gender, ethnicity), mutually adjusted for A- and G- stages, and additionally adjusted for the pre-existing comorbidities (diabetes, obesity, hypertension, cancer, and cardiovascular diseases) selected based on univariate analyses; # indicates estimates that were not statistically significant.

Association of COVID-19–related AKI stages and subtypes with 30-day mortality

Consistent with prior reports for non-COVID-19 hospitalizations, patients with COVID-19-related AKI had significantly worse clinical outcomes and higher mortality compared to patients without AKI. To demonstrate the utility of our algorithmic AKI staging and subtyping, we examined the association of computed AKI severity and subtype with 30-day mortality (Figure 3). We confirmed that the degree of risk correlated with AKI severity, with stage 1 being associated with a 6-fold increased risk (OR 6.23, 95%CI: 5.12–7.58) and stage 3 with over 15-fold increased risk of COVID-19-related death (OR 15.3, 95%CI: 12.0–19.4) across all waves combined. The subtype of sustained AKI was also associated with significantly higher mortality (HR 11.6, 95%CI: 9.58–14.1) compared to the subtype of transient AKI (HR 5.21, 95%CI 4.15–6.54). These associations were statistically significant for each pandemic wave, except for the Delta Wave that had the lowest number of COVID-19 cases, AKI events, and deaths. Furthermore, our sensitivity analyses demonstrated that these findings were robust to a more stringent definition of baseline renal function (Figure S4).

Figure 3. Association between AKI staging (a) and subtype (b) and COVID-19-related mortality.

Figure 3.

Open in a new tab

The vertical dashed line corresponds to Hazards Ratio (HR) of 1.0 (i.e., no effect). All estimates were adjusted for demographics (age, age squared, gender, ethnicity), A- and G-stage, and pre-existing diseases significant by univariate analyses (diabetes, obesity, hypertension, cancer, and cardiovascular diseases). The number of outcomes was too low for meaningful AKI subgroup analyses during the Delta Wave thus this wave was omitted. All estimates are statistically significant.

Application to the MIMIC-IV dataset

We first evaluated the performance of the AKI e-phenotype against the “documented diagnosis of AKI” among ED admissions in the MIMIC-IV dataset. Despite imperfect documentation of AKI used as a comparator, our algorithm achieved sensitivity of 43.2%, specificity of 91.2% and overall accuracy of 86.9%. Performance was further improved in the subgroup of patients with a baseline SCr measured 7–365 days before presentation, achieving sensitivity 51.8%, specificity 93.0%, and accuracy 88.4%. Similar performance was observed across ICU and non-ICU subgroups. Overall, the AKI e-phenotype achieved high specificity (>90%) across all subgroups, supporting its robustness for identifying AKI phenotypes within heterogeneous ED populations (Table S2).

Next, we applied both the AKI and CKD e-phenotypes to the analysis of 405,467 MIMIC-IV hospitalized patients admitted through the emergency department. Similar to COVID-19 dataset, both A- and G-stage exhibited mutually independent associations with increased risk of AKI. Additionally, A-stage showed a significant association with 30-day mortality, consistent with our findings in COVID-19 (Figure 4B and 4D). More severe AKI stage and the subtype of sustained AKI were each independently associated with worse clinical outcomes (Figure 4A and 4C). Stage 1 AKI was associated with a 2-fold increased risk of death (HR 2.28, 95%CI: 2.11–2.46) while stage 3 conferred a 5-fold increased risk (HR 5.18, 95%CI: 4.53–5.92). The sustained AKI subtype was also strongly associated with mortality (HR 3.42, 95%CI: 3.14–3.73) and had a larger effect size compared to transient AKI (HR 1.73, 95%CI 1.57–1.90). Thus, we closely recapitulate the patterns of association observed in the COVID-19 dataset, but with lower effect sizes for MIMIC-IV general hospital admissions, suggesting that COVID-19 infection considerably amplifies the mortality risk associated with AKI.

Figure 4: MIMIC-IV analysis of 405,467 hospital admissions: (a) associations of AKI stages with 30-day mortality after hospital admission, (b) A-stage and G-stage with 30-day mortality after hospital admission, (c) AKI subtypes (transient and sustained) with 30-day mortality after hospital admission, and (d) A-stage and G-stage with the risk of AKI during hospital admission.

Figure 4:

Open in a new tab

The vertical dashed line represents no effect (HR or OR=1). Cox proportional-hazards models were used to test associations with mortality, while logistic regression models were used to test associations with AKI. All models were adjusted for demographics, Elixhauser Comorbidity Index, and pre-existing comorbidities that tested as significant in corresponding univariate analyses.

Taken together, our findings indicate that the AKI e-phenotyping algorithm is robust and generalizable to non-COVID-19 populations, effectively capturing clinically meaningful associations between AKI diagnosis, severity, and subtype, and adverse clinical outcomes.

Discussion

In this study, we developed and validated a new strategy for computationally detecting, staging, and subtyping AKI using structured EHR data. The definition of AKI encompasses highly heterogeneous diagnoses with variable etiology, course and outcomes37–39. By subtyping AKI based on duration, we aimed to distill this complexity into clinically meaningful categories. A major innovation of our approach is the distinction between transient and sustained AKI subtypes, defined by improvement or further loss of renal function within the initial 48 hours. This temporal dimension provides a more granular view of kidney dysfunction than is currently specified in the KDIGO-recommended framework.

Another challenge addressed by our algorithm is frequent absence of prior serum creatinine values, which usually necessitates imputation or surrogate measures to baseline kidney function40,41. Prior studies have shown that varying definitions of baseline serum Cr can compromise diagnostic performance38,42–45. Our sensitivity analysis confirmed that our approach to baseline determination is effective, maintaining a high level of specificity and thereby enhancing the robustness compared to prior work.

We subsequently applied our AKI e-phenotype to a large-scale, real-world dataset of patients admitted with COVID-19. This application was chosen because kidney involvement in COVID-19 is common46–49, pre-existing CKD is a major risk factor for COVID-19-related AKI50,51, and AKI in the setting of COVID-19 confers increased in-hospital mortality52–56. A unique strength of our study is that we retrospectively analyzed entire EHR during all pandemic waves, demonstrating both the flexibility and scalability of our algorithms. Our wave-specific analysis revealed dynamic shifts in comorbidities, AKI risk factors, and outcomes across the course of the pandemic57. Importantly, our algorithm allowed us to dissect the independent and joint contribution of pre-existing CKD (stratified by albuminuria and eGFR stage) and COVID-19-associated AKI (staged and subtyped) to mortality at an unprecedented scale and resolution. Our results highlight several consistent findings. First, pre-existing CKD remained one of the most reliable predictors of adverse COVID-19 outcomes across all waves, including COVID-19-associated AKI. Second, more severe or sustained AKI were associated with higher mortality compared to mild or transient AKI, even after adjusting for demographics, comorbidities, and illness severity. In an independent large-scale application to the MIMIC-IV dataset, we demonstrated that these findings are broadly generalizable to diverse groups of hospitalized patients. Our results suggest that a significant refinement of the KDIGO classification may be warranted to incorporate AKI subtypes as a clinically meaningful dimension of risk stratification. Our study has several limitations. First, pragmatic choices were necessary, including omitting the urine output criteria and extrapolating baseline renal function in the absence of data for a smaller number of cases. These choices reflect real-world EHR data constraints and may reduce sensitivity. Second, our AKI definition relies on the serum creatinine alone, which cannot fully distinguish reversible functional changes from structural kidney injury. Our subtyping by duration partially addresses this issue, but future incorporation of molecular biomarkers may further enhance this discrimination58,59.

In summary, we developed, validated, and implemented a new electronic phenotype for AKI that incorporates both staging and subtyping, and demonstrated its scalability and integration with a CKD e-phenotype. Applied at scale, these tools provide novel insights into the interplay between acute and chronic kidney disease in COVID-19, highlighting the prognostic importance of AKI duration and CKD staging. By making these algorithms openly available, we provide a framework for future large-scale EHR-based studies of kidney diseases.

Supplementary Material

1

Supplementary File (PDF)

Figure S1. Timeline of the COVID-19 pandemic: pandemic waves identified in the US, corresponding COVID-19 occurrence counts at CUIMC, and (C) COVID-19 variant emergence.

Figure S2. Workflow for implementing the electronic AKI (e-AKI) phenotype in the MIMIC-IV dataset

Figure S3. Frequency of COVID-19 outcomes for patients algorithmically placed on the A-by-G grid based on pre-existing CKD.

Figure S4. Sensitivity analysis for association between AKI stages and subtypes and COVID-19-related mortality for a subgroup of patients with available baseline SCr between 7 to 365 days prior to presentation.

Table S1. COVID-19 study cohort baseline characteristics, COVID-19 severity, and outcomes by the pandemic wave.

Table S2. Performance metrics for the AKI e-phenotype against the “documented diagnosis” of AKI in the MIMIC-IV dataset.

Acknowledgments:

We would like to acknowledge the Columbia University Biobank (CUB), which operated as a Columbia COVID-19 Biobank during the pandemic and provided useful feedback on the methods for identifying COVID-19 cases and their hospitalization periods.

Support:

NS is funded by NIH grant U01-HG013201; KX is funded by NIH grant K01-DK135917; JB is funded by NIH grant TL1-DK136048, R01-DK124667, and U54-DK104309; KK is funded by NIH grants R01 DK144793, R01-DK105124, R01-DK136765, U01-HG008680, U01-HG013201, U01-DK100876, and UL1-TR001873. The funders had no role in study design, data collection, analysis, reporting, or the decision to submit the manuscript for publication.

Footnotes

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Financial Disclosure: The authors declare that they have no relevant financial interests.

Peer Review: Received March 26, 2025. Evaluated by 2 external peer reviewers, with direct editorial input from a Statistics/Methods Editor, an Associate Editor, and the Editor-in-Chief. Accepted in revised form April 8, 2026.

Data Sharing:

The Acute Kidney Injury (AKI) electronic phenotype algorithm code, SQL server implementation, and relevant documentation are available on GitHub (https://github.com/sunnyshang/pheAKI) and PheKB (https://phekb.org/phenotype/acute-kidney-injury-aki).

References

  • 1.Barasch J, Zager R, Bonventre JV. Acute Kidney Injury: A Problem of Definition. Lancet. 2017/02/25/ 2017;389(10071):779–781. doi: 10.1016/S0140-6736(17)30543-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Desanti De Oliveira B, Xu K, Shen TH, et al. Molecular nephrology: types of acute tubular injury. Nat Rev Nephrol. Oct 2019;15(10):599–612. doi: 10.1038/s41581-019-0184-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Kiryluk K, Bomback AS, Cheng YL, et al. Precision Medicine for Acute Kidney Injury (AKI): Redefining AKI by Agnostic Kidney Tissue Interrogation and Genetics. Semin Nephrol. Jan 2018;38(1):40–51. doi: 10.1016/j.semnephrol.2017.09.006 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Xu K, Shang N, Levitman A, et al. Elevated Neutrophil Gelatinase-Associated Lipocalin Is Associated With the Severity of Kidney Injury and Poor Prognosis of Patients With COVID-19. Kidney Int Rep. 2021/12/01/ 2021;6(12):2979–2992. doi: 10.1016/j.ekir.2021.09.005 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Stevens JS, Xu K, Corker A, et al. Rule Out Acute Kidney Injury in the Emergency Department With a Urinary Dipstick. Kidney Int Rep. Nov 2020;5(11):1982–1992. doi: 10.1016/j.ekir.2020.09.006 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Shang N, Khan A, Polubriaginof F, et al. Medical records-based chronic kidney disease phenotype for clinical care and “big data” observational and genetic studies. npj Digit Med. 2021/04/13/ 2021;4(1):1–13. doi: 10.1038/s41746-021-00428-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Shang N, Mohan S, Gharav AG, et al. PheKB: Chronic Kidney Disease. Updated 2016. https://phekb.org/phenotype/chronic-kidney-disease
  • 8.Kidney Disease: Improving Global Outcomes Acute Kidney Injury Work G. KDIGO clinical practice guideline for acute kidney injury. Kidney Inter Suppl. 2012 2012;2:1–138. [Google Scholar]
  • 9.Wainstein M, MacDonald S, Fryer D, et al. Use of an extended KDIGO definition to diagnose acute kidney injury in patients with COVID-19: A multinational study using the ISARIC–WHO clinical characterisation protocol. PLOS Medicine. 2022/04/20/ 2022;19(4):e1003969. doi: 10.1371/journal.pmed.1003969 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Bellomo R, Ronco C, Kellum JA, Mehta RL, Palevsky P, the Aw. Acute renal failure – definition, outcome measures, animal models, fluid therapy and information technology needs: the Second International Consensus Conference of the Acute Dialysis Quality Initiative (ADQI) Group. Critical Care. 2004/05/24/ 2004;8(4):R204. doi: 10.1186/cc2872 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Mehta RL, Kellum JA, Shah SV, et al. Acute Kidney Injury Network: report of an initiative to improve outcomes in acute kidney injury. Critical Care. 2007/03/01/ 2007;11(2):R31. doi: 10.1186/cc5713 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Xu K, Rosenstiel P, Paragas N, et al. Unique Transcriptional Programs Identify Subtypes of AKI. JASN. 2017/06// 2017;28(6):1729. doi: 10.1681/ASN.2016090974 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Nickolas TL, Schmidt-Ott KM, Canetta P, et al. Diagnostic and prognostic stratification in the emergency department using urinary biomarkers of nephron damage: a multicenter prospective cohort study. J Am Coll Cardiol. 2012/01/17/ 2012;59(3):246–255. doi: 10.1016/j.jacc.2011.10.854 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Edelstein CL. Biomarkers of Acute Kidney Injury. Adv Chronic Kidney Dis. 2008/07// 2008;15(3):222–234. doi: 10.1053/j.ackd.2008.04.003 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Bouchard J. Estimating Baseline Serum Creatinine for Assessing Acute Kidney Injury: Not a One Size Fits All Approach. Kidney Int Rep. 2021/02/06/ 2021;6(3):562–564. doi: 10.1016/j.ekir.2021.01.030 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Nix DE, Erstad BL. Creatinine Assessment in Non–Steady-State Conditions: A Critical Review. Ann Pharmacother. 2021/12/01/ 2021;55(12):1536–1544. doi: 10.1177/1060028021999644 [DOI] [PubMed] [Google Scholar]
  • 17.Mishra RC, Sodhi K, Prakash KC, et al. ISCCM Guidelines on Acute Kidney Injury and Renal Replacement Therapy. Indian J Crit Care Med. 2022/10// 2022;26(Suppl 2):S13–S42. doi: 10.5005/jp-journals-10071-24109 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Devarajan P. Acute kidney injury in children: Clinical features, etiology, evaluation, and diagnosis. UpToDate. Updated 2024/02/20/. https://www.uptodate.com/contents/acute-kidney-injury-in-children-clinical-features-etiology-evaluation-and-diagnosis
  • 19.Stevens JS. Rule Out Acute Kidney Injury in the Emergency Department With a Urinary Dipstick. Clinical Research. 2020 2020;5(11):1982–1992. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Kellum JA, Lameire N, Aspelin P, et al. Kidney disease: improving global outcomes (KDIGO) acute kidney injury work group. KDIGO clinical practice guideline for acute kidney injury. Kidney international supplements. 2012 2012;2(1):1–138. doi: 10.1038/kisup.2012.1 [DOI] [Google Scholar]
  • 21.Chawla LS, Bellomo R, Bihorac A, et al. Acute kidney disease and renal recovery: consensus report of the Acute Disease Quality Initiative (ADQI) 16 Workgroup. Nat Rev Nephrol. Apr 2017;13(4):241–257. doi: 10.1038/nrneph.2017.2 [DOI] [PubMed] [Google Scholar]
  • 22.Haendel MA, Chute CG, Bennett TD, et al. The National COVID Cohort Collaborative (N3C): Rationale, design, infrastructure, and deployment. Journal of the American Medical Informatics Association. 2021/03/01/ 2021;28(3):427–443. doi: 10.1093/jamia/ocaa196 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Yahav D, Yelin D, Eckerle I, et al. Definitions for coronavirus disease 2019 reinfection, relapse and PCR re-positivity. Clinical Microbiology and Infection. 2021/03/01/ 2021;27(3):315–318. doi: 10.1016/j.cmi.2020.11.028 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Quan H, Sundararajan V, Halfon P, et al. Coding Algorithms for Defining Comorbidities in ICD-9-CM and ICD-10 Administrative Data. Medical Care. 2005/11// 2005;43(11):1130–1139. doi: 10.1097/01.mlr.0000182534.19832.83 [DOI] [PubMed] [Google Scholar]
  • 25.Gasparini A. Comorbidity: An R package for computing comorbidity scores. JOSS. 2018/03/29/ 2018;3(23):648. doi: 10.21105/joss.00648 [DOI] [Google Scholar]
  • 26.Charlson ME, Carrozzino D, Guidi J, Patierno C. Charlson Comorbidity Index: A Critical Review of Clinimetric Properties. PPS. 2022 2022;91(1):8–35. doi: 10.1159/000521288 [DOI] [PubMed] [Google Scholar]
  • 27.World Health O. WHO Director-General’s opening remarks at the media briefing on COVID-19 - 11 March 2020. WHO. Updated 2020/03/11/. https://www.who.int/director-general/speeches/detail/who-director-general-s-opening-remarks-at-the-media-briefing-on-covid-19---11-march-2020 [Google Scholar]
  • 28.Wikipedia. Timeline of the COVID-19 pandemic in New York City. https://en.wikipedia.org/wiki/Timeline_of_the_COVID-19_pandemic_in_New_York_City2023.
  • 29.NYC Health. As the Federal COVID-19 Emergency Ends, New York City Ensures COVID-19 Care and Services Remain Available - NYC Health. Updated 2023/03/11/. https://www.nyc.gov/site/doh/about/press/pr2023/nyc-ensures-covid-19-care-and-services-remain-available.page
  • 30.Dong E, Du H, Gardner L. An interactive web-based dashboard to track COVID-19 in real time. The Lancet Infectious Diseases. 2020/05/01/ 2020;20(5):533–534. doi: 10.1016/S1473-3099(20)30120-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Zhang SX, Arroyo Marioli F, Gao R, Wang S. A Second Wave? What Do People Mean by COVID Waves? – A Working Definition of Epidemic Waves. Risk Manag Healthc Policy. 2021/09/13/ 2021;14:3775–3782. doi: 10.2147/RMHP.S326051 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Thompson CN, Baumgartner J, Pichardo C, et al. COVID-19 Outbreak — New York City, February 29–June 1, 2020. MMWR Morb Mortal Wkly Rep. 2020/11/20/ 2020;69(46):1725–1729. doi: 10.15585/mmwr.mm6946a2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Holshue ML, DeBolt C, Lindquist S, et al. First Case of 2019 Novel Coronavirus in the United States. New England Journal of Medicine. 2020/03/05/ 2020;382(10):929–936. doi: 10.1056/NEJMoa2001191 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Gross SA, Robbins DH, Greenwald DA, Schnoll-Sussman FH, Pochapin MB. Preparation in the Big Apple: New York City, a New Epicenter of the COVID-19 Pandemic. Official journal of the American College of Gastroenterology | ACG. 2020/06// 2020;115(6):801. doi: 10.14309/ajg.0000000000000636 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Razavian N, Major VJ, Sudarshan M, et al. A validated, real-time prediction model for favorable outcomes in hospitalized COVID-19 patients. npj Digit Med. 2020/12// 2020;3(1):130. doi: 10.1038/s41746-020-00343-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Centers for Disease Control and Prevention (CDC). CDC Museum COVID-19 Timeline. Updated 2023/03/15/. https://www.cdc.gov/museum/timeline/covid19.html
  • 37.Vaara ST, Bhatraju PK, Stanski NL, et al. Subphenotypes in acute kidney injury: a narrative review. Critical Care. 2022/08/19/ 2022;26(1):251. doi: 10.1186/s13054-022-04121-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Endre ZH, Mehta RL. Identification of acute kidney injury subphenotypes. Current Opinion in Critical Care. 2020/12// 2020;26(6):519–524. doi: 10.1097/MCC.0000000000000772 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Birkelo BC, Koyner JL, Ostermann M, Bhatraju PK. The Road to Precision Medicine for Acute Kidney Injury. Critical Care Medicine. 2024/07// 2024;52(7):1127–1137. doi: 10.1097/CCM.0000000000006328 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Ahmed A, Vairavan S, Akhoundi A, et al. Development and validation of electronic surveillance tool for acute kidney injury: A retrospective analysis. Journal of Critical Care. 2015/10/01/ 2015;30(5):988–993. doi: 10.1016/j.jcrc.2015.05.007 [DOI] [PubMed] [Google Scholar]
  • 41.Kellum JA, Prowle JR. Paradigms of acute kidney injury in the intensive care setting. Nature Reviews Nephrology. 2018/04// 2018;14(4):217–230. doi: 10.1038/nrneph.2017.184 [DOI] [PubMed] [Google Scholar]
  • 42.Kashani K, Shao M, Li G, et al. No increase in the incidence of acute kidney injury in a population-based annual temporal trends epidemiology study. Kidney International. 2017/09/01/ 2017;92(3):721–728. doi: 10.1016/j.kint.2017.03.020 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43.Smith TD, Soriano VO, Neyra JA, Chen J. Identifying KDIGO Trajectory Phenotypes Associated with Increased Inpatient Mortality. Proc (IEEE Int Conf Healthc Inform). 2019/06// 2019;2019:10.1109/ichi.2019.8904739. doi: 10.1109/ichi.2019.8904739 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Chan HY, Liu M. Interpretable Sub-phenotype Identification in Acute Kidney Injury. AMIA Annu Symp Proc. 2023/04/29/ 2023;2022:339–348. [PMC free article] [PubMed] [Google Scholar]
  • 45.Shah AP, Snead W, Daga A, et al. Methods for phenotyping adult patients with acute kidney injury: a systematic review. J Nephrol. 2024/09/16/ 2024;doi: 10.1007/s40620-024-02080-w [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46.Nadim MK, Forni LG, Mehta RL, et al. COVID-19-associated acute kidney injury: consensus report of the 25th Acute Disease Quality Initiative (ADQI) Workgroup. Nature Reviews Nephrology. 2020/12// 2020;16(12):747–764. doi: 10.1038/s41581-020-00356-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47.Hassanein M, Radhakrishnan Y, Sedor J, et al. COVID-19 and the kidney. Cleve Clin J Med. 2020/10/01/ 2020;87(10):619–631. doi: 10.3949/ccjm.87a.20072 [DOI] [PubMed] [Google Scholar]
  • 48.Alfano G, Ferrari A, Fontana F, et al. Incidence, risk factors and outcome of acute kidney injury (AKI) in patients with COVID-19. Clin Exp Nephrol. 2021/11/01/ 2021;25(11):1203–1214. doi: 10.1007/s10157-021-02092-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49.Almeida DCd, Franco MdCP, Santos DRPd, et al. Acute kidney injury: Incidence, risk factors, and outcomes in severe COVID-19 patients. PLOS ONE. 2021/05/25/ 2021;16(5):e0251048. doi: 10.1371/journal.pone.0251048 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50.Kang SH, Kim SW, Kim AY, Cho KH, Park JW, Do JY. Association between Chronic Kidney Disease or Acute Kidney Injury and Clinical Outcomes in COVID-19 Patients. Journal of Korean Medical Science. 2020/11/25/ 2020;35(50)doi: 10.3346/jkms.2020.35.e434 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51.Era-Edta Council, Eracoda Working Group. Chronic kidney disease is a key risk factor for severe COVID-19: a call to action by the ERA-EDTA. Nephrology Dialysis Transplantation. 2021/01/01/ 2021;36(1):87–94. doi: 10.1093/ndt/gfaa314 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 52.Fabrizi F, Alfieri CM, Cerutti R, Lunghi G, Messa P. COVID-19 and Acute Kidney Injury: A Systematic Review and Meta-Analysis. Pathogens. 2020/12// 2020;9(12):1052. doi: 10.3390/pathogens9121052 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 53.Li Q, Zhang T, Li F, et al. Acute Kidney Injury Can Predict In-Hospital Mortality in Elderly Patients with COVID-19 in the ICU: A Single-Center Study. Clin Interv Aging. 2020/11/09/ 2020;15:2095–2107. doi: 10.2147/CIA.S273720 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 54.Yan Q, Zuo P, Cheng L, et al. Acute Kidney Injury Is Associated With In-hospital Mortality in Older Patients With COVID-19. The Journals of Gerontology: Series A. 2020/07/16/ 2020;(glaa181)doi: 10.1093/gerona/glaa181 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 55.Cheng Y, Luo R, Wang X, et al. The Incidence, Risk Factors, and Prognosis of Acute Kidney Injury in Adult Patients with Coronavirus Disease 2019. CJASN. 2020/10// 2020;15(10):1394. doi: 10.2215/CJN.04650420 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 56.Dai Y, Liu Z, Du X, et al. Acute Kidney Injury in Hospitalized Patients Infected with COVID-19 from Wuhan, China: A Retrospective Study. Research Article. BioMed Research International. 2021/01/12/ 2021; [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 57.Bowe B, Cai M, Xie Y, Gibson AK, Maddukuri G, Al-Aly Z. Acute Kidney Injury in a National Cohort of Hospitalized US Veterans with COVID-19. CJASN. 2021/01/07/ 2021;16(1):14–25. doi: 10.2215/CJN.09610620 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 58.Haase M, Bellomo R, Devarajan P, Schlattmann P, Haase-Fielitz A. Accuracy of Neutrophil Gelatinase-Associated Lipocalin (NGAL) in Diagnosis and Prognosis in Acute Kidney Injury: A Systematic Review and Meta-analysis. American Journal of Kidney Diseases. 2009/12/01/ 2009;54(6):1012–1024. doi: 10.1053/j.ajkd.2009.07.020 [DOI] [PubMed] [Google Scholar]
  • 59.Koyner JL, Zarbock A, Basu RK, Ronco C. The impact of biomarkers of acute kidney injury on individual patient care. Nephrology Dialysis Transplantation. 2020/08/01/ 2020;35(8):1295–1305. doi: 10.1093/ndt/gfz188 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

1

Supplementary File (PDF)

Figure S1. Timeline of the COVID-19 pandemic: pandemic waves identified in the US, corresponding COVID-19 occurrence counts at CUIMC, and (C) COVID-19 variant emergence.

Figure S2. Workflow for implementing the electronic AKI (e-AKI) phenotype in the MIMIC-IV dataset

Figure S3. Frequency of COVID-19 outcomes for patients algorithmically placed on the A-by-G grid based on pre-existing CKD.

Figure S4. Sensitivity analysis for association between AKI stages and subtypes and COVID-19-related mortality for a subgroup of patients with available baseline SCr between 7 to 365 days prior to presentation.

Table S1. COVID-19 study cohort baseline characteristics, COVID-19 severity, and outcomes by the pandemic wave.

Table S2. Performance metrics for the AKI e-phenotype against the “documented diagnosis” of AKI in the MIMIC-IV dataset.

NIHMS2200733-supplement-1.pdf (2MB, pdf)

Data Availability Statement

The Acute Kidney Injury (AKI) electronic phenotype algorithm code, SQL server implementation, and relevant documentation are available on GitHub (https://github.com/sunnyshang/pheAKI) and PheKB (https://phekb.org/phenotype/acute-kidney-injury-aki).

RESOURCES