Utilization of Deep Learning for Subphenotype Identification in Sepsis-Associated Acute Kidney Injury : Clinical Journal of the American Society of Nephrology

Journal Logo

Original Articles: Acute Kidney Injury and ICU Nephrology

Utilization of Deep Learning for Subphenotype Identification in Sepsis-Associated Acute Kidney Injury

Chaudhary, Kumardeep1; Vaid, Akhil2; Duffy, Áine1; Paranjpe, Ishan1; Jaladanki, Suraj1; Paranjpe, Manish3,4; Johnson, Kipp1; Gokhale, Avantee5; Pattharanitima, Pattharawin5; Chauhan, Kinsuk5; O’Hagan, Ross1; Van Vleck, Tielman1; Coca, Steven G.5; Cooper, Richard6; Glicksberg, Benjamin2; Bottinger, Erwin P.2; Chan, Lili5; Nadkarni, Girish N.2,5

Author Information
CJASN 15(11):p 1557-1565, November 2020. | DOI: 10.2215/CJN.09330819
  • Free
  • Infographic
  • SDC



AKI is common in hospitalized patients and is associated with higher morbidity and mortality (1–4). AKI is frequently associated with sepsis in critically ill patients admitted to intensive care units (ICUs). It was initially thought that sepsis-associated AKI is due to systemic hypotension leading to a decrease in kidney perfusion causing ischemia and acute tubular necrosis (ATN). However, there is growing evidence of different mechanisms of sepsis-associated AKI with potentially different clinical characteristics and outcomes (5).

Thus, AKI is likely not a single clinical entity but an overarching clinical syndrome composed of several different subtypes. Previous work has demonstrated that latent class analysis (a statistical method to group patients into classes on the basis of the results of a set of categorical indicator variables) identifies subphenotypes in AKI with differing outcomes and response to therapy (6). However, this analysis utilized carefully curated and prospectively collected data and biomarkers instead of electronic health record (EHR) data generated as a part of routine clinical care. Deep learning techniques—a subset of the artificial intelligence universe where a computer program uses multiple layers to progressively extract higher-level features from the raw input—can leverage these data to identify patterns within complex diseases, revealing subphenotypes (7). To our knowledge, there have not been any studies that leverage routine EHR data using deep learning to identify subphenotypes in sepsis-associated AKI.

Our primary aim was to determine if we could identify subphenotypes of sepsis-associated AKI utilizing measurements done as part of patients’ routine care outside of traditional features, such as age, sex, and race. We sought to incorporate hundreds of data features collected routinely in the EHR to identify subphenotypes of AKI in patients admitted to the ICU with sepsis and explore differences in patient outcomes among them.

Materials and Methods

Study Population

We utilized the Medical Information Mart for Intensive Care III (MIMIC III) database to identify patients with sepsis-associated AKI. MIMIC III is a freely accessible critical care database of patients from a large, single-center tertiary care hospital from 2001 to 2012 (8). This includes patient demographics, vital signs, laboratory results, billing codes, and notes. We included patients if they had AKI within 48 hours of ICU admission as per Kidney Disease Improving Global Outcomes (KDIGO) guidelines (9). We used both creatinine and urine output values to identify patients with AKI. We used the lowest creatinine 7 days prior to ICU admission as baseline and compared it with highest creatinine that occurred within 48 hours of ICU admission. Urine output 48 hours after ICU admission was binned into 6-hour time periods, and if any 6-hour time interval was below the limit as per KDIGO guidelines, patients were considered to have AKI. We then defined sepsis using International Classification of Diagnosis Ninth Revision (ICD9) codes as previously validated (10). We defined comorbidities using the Elixhauser Comorbidity Software, which identifies comorbidities by grouping International Classification of Diseases Clinical Modification codes (11). We excluded patients <18 or >89 years old, admitted for ≤24 hours, with ESKD, or missing vital signs. We also excluded patients who had any dialysis at any point before 48 hours after AKI diagnosis, as well as those who died in that window (Figure 1). We did so to ensure the results were not skewed by inclusion of terminal patients. As patients could be admitted several times during the period, we considered only the data from the first admission with AKI per patient.

Figure 1.:
Flow sheet for data processing and subphenotype identification. ICD9, International Classification of Diseases Ninth Revision; KDIGO, Kidney Disease Improving Global Outcomes; MIMIC III, Medical Information Mart for Intensive Care III.

Data Processing

We utilized laboratory values and vital sign measurements to identify the clusters and included all laboratory values and vital sign measurements from admission to 48 hours after the diagnosis of AKI. A feature space is all of the features that are put into a deep learning architecture to enable it to discern patterns. For both laboratory values and vital sign measurements, we calculated derived features, such as median, variability, and number of times measured, leading to a feature space of 52 for vital sign measurements and 2464 for laboratory values. Only features that were present in ≥70% of patients were included, reducing the feature space down to 188 features. Additionally, we considered comorbidities, results for blood and urine culture, mechanical ventilation, and use of vasopressor drugs, leading to a final feature space of 225. Missing values were imputed using K-nearest neighbor imputation.

Following this, we performed MinMax scaling on the resultant data to bring all values in the feature space to a comparable scale (12). Several machine learning algorithms are sensitive to distances in higher-dimensional space. Scaling is important both to ensure that the algorithm does not misinterpret the importance of one feature relative to another on the basis of unit of measurement alone, as well as to speed up calculations for training neural networks. An autoencoder is a nonlinear dimensional decomposition deep learning architecture that takes the many features and combines them into fewer generated features in lower dimensions. Clustering takes into account distances in higher-dimensional spaces. An increase in the dimensionality of feature space also increases its sparsity. This is known as the “curse of dimensionality.” Consequently, clustering algorithms become both slower and inaccurate. Utilizing an autoencoder allows us to circumvent this and detect patterns inherent to the data that could not be detected in higher dimensions. We used a five-layer-deep autoencoder with 32, eight, two, eight, and 32 neurons per hidden layer, respectively. We considered the output from the layer prior to the middle layer of the autoencoder for further analysis (Figure 1).


With the final dimensionally reduced feature matrix of combined laboratory values, vital sign measurements, and comorbidities for all of the samples, we performed factor analysis to further decompose the data frame to aid the performance of the clustering algorithms. We then performed unsupervised K-means clustering on the resultant data testing with cluster size ranging from K=2 to K=5. To validate the stability of the found subphenotypes, we calculated the silhouette score (a measure of how similar a sample is to its own cluster), the Davies–Bouldin Score (a measure of the average similarity of each cluster with its most similar cluster), and the Calinski–Harabasz score (the ratio between the within-cluster dispersion and the between-cluster dispersion). After obtaining the subphenotype labels, we used the t-distributed stochastic neighbor embedding technique to reduce data to three dimensions for better visualization. We used the t-distributed stochastic neighbor embedding from scikit-learn to decompose the eight-dimensional autoencoder output (latent features from bottleneck layer) into three dimensions (13). Finally, clusters were visualized in three-dimensional space using the matplotlib package in Python (14).

Manual Chart Review

In order to ensure that these clusters were not completely driven by known pathophysiology of kidney disease (a prerenal phenotype versus ATN), we randomly selected 30 patient charts from each subphenotype (approximately 2% of total population) for extensive physician clinical review. Two independent physicians, blinded to subphenotype assignment, classified patients into ATN, prerenal, other etiology documented, or other after review of all available physician progress notes and discharge summaries.

Analysis of Persistence among Subphenotypes

To assess whether the differences in subphenotypes were driven by AKI duration, we examined differences in persistence (15). We defined persistence if the last available creatinine was greater than the lowest creatinine 7 days prior to ICU admission. This analysis was only performed on patients who met the creatinine KDIGO criteria.

Statistical Analyses

After cluster identification, we conducted analysis to explore differences among clusters. We used the Kruskal–Wallis test for continuous variables and the Fisher exact or chi-squared test for categorical variables. We included need for dialysis and 28-day post-AKI mortality as outcomes of interest. We used logistic regression to determine the association between cluster and mortality adjusting for sex; age; race/ethnicity; CKD; liver disease; hypertension; congestive heart failure; AKI stage; and AKI diagnosis by creatinine, urine output, or both. We used the chi-squared test to assess differences in AKI stage and persistence. As this study was done on publicly available, deidentified data, it was considered institutional review board exempt. Analysis was done using SAS 9.4 and R 3.4.3 software.


Clinical Features of Patients with Sepsis-Associated AKI

Of 46,520 patients, 4858 (10%) had sepsis-associated AKI. After all exclusions, we included 4001 patients (Figure 1). Patients had a mean age of 66 years; 57% were men, and 73% were White. Most patients’ admission type was through the emergency department. Patients had a high prevalence of hypertension (49%), congestive heart failure (37%), and diabetes mellitus (31%); 139 (4%) did not have sufficient information to calculate urine output rates and were therefore identified by creatinine criteria only.

Unsupervised Clustering to Identify Subphenotypes

From these combined features with transformed values, we implemented K-means clustering ranging from K=2 to K=5. Clustering with K=3 was found to have a silhouette score of 0.61, a low Davies–Bouldin score of 0.53, and a high Calinski–Harabasz score of 6939. (Silhouette scores are measured from −1 to one—values closer to one being better. Davies–Bouldin scores are measured from zero upward—values closer to zero being better. Calinski–Harabasz scores are measured from zero upward—higher values being better for a dataset.) Subphenotype 1 had 1443 (36%) patients, subphenotype 2 had 1898 (47%) patients, and subphenotype 3 had 660 (16%) patients (Figure 2).

Figure 2.:
t-Distributed stochastic neighbor embedding (t-SNE) visual representation of subphenotype 1 in blue, subphenotype 2 in orange, and subphenotype 3 in green demonstrates separation among subphenotypes. The t-SNE plot is a method of visualizing high-dimensional data in a low-dimensional space (in this case, three dimensions). Each dot represents a patient and displays clusters within the scaled down and dimensionally reduced space of the autoencoder embeddings. As such, this space does not have any real-world units.

Clinical and Biologic Characteristics of Each Phenotype

The baseline characteristics of the three clusters are presented in Table 1. Patients in subphenotype 3 were the youngest (63 years old; interquartile range [IQR], 52–73 years old versus 66 years old; IQR, 54–77 years old [subphenotype 2] versus 70 years old; IQR, 58–79 years old [subphenotype 1]; P<0.001). Although subphenotype 3 had lower proportions of patients with hypertension, congestive heart failure, and diabetes, it had a significantly higher proportion of patients with liver disease (27% versus 18% [subphenotype 2] versus 5% [subphenotype 1]; P<0.001). Subphenotype 2 had the largest proportion of patients with CKD (21% versus 15% [subphenotype 1] versus 15% [subphenotype 3]; P<0.001). Simplified Acute Physiology Score II (SAPS II) scores were highest in subphenotype 3 (54 versus 47 [subphenotype 2] versus 38 [subphenotype 1]; P<0.001). There were small but significant differences in BP across the three subphenotypes; however, there were large differences in the proportions of patients requiring vasopressor support (76% [subphenotype 3] versus 62% [subphenotype 2] versus 39% [subphenotype 1]; P<0.001). Unspecified septicemia was the primary discharge diagnosis in all three subphenotypes (Supplemental Table 1).

Table 1. - Characteristics of 4001 patients with sepsis-associated AKI from the Medical Information Mart for Intensive Care III database
Patient Characteristics Subphenotype 1, n=1443 Subphenotype 2, n=1898 Subphenotype 3, n=660 P Value
Median age, yr (IQR) 70 (58 to 79) 66 (54 to 77) 63 (52 to 73) <0.001
Men, n (%) 830 (58) 1033 (54) 424 (64) <0.001
Race, n (%) 0.03
 White 1088 (75) 1384 (73) 472 (72)
 Black 124 (9) 166 (9) 47 (7)
 Hispanic 36 (3) 58 (3) 26 (4)
 Asian 17 (1) 47 (3) 12 (2)
 Other or unknown 178 (12) 243 (13) 103 (16)
Comorbidities, n (%)
 Hypertension 798 (55) 906 (48) 261 (40) <0.001
 Congestive heart failure 596 (41) 678 (36) 194 (29) <0.001
 Diabetes mellitus 489 (34) 595 (31) 175 (27) 0.003
 Liver disease 67 (5) 349 (18) 180 (27) <0.001
 Peripheral vascular disease 152 (11) 176 (9) 102 (15) <0.001
 HIV infection 8 (1) 49 (93) 9 (1) <0.001
 CKD 221 (15) 394 (21) 97 (15) <0.001
 SAPS II score 38 (30–47) 47 (37–57) 54 (44–66) <0.001
First service in intensive care, n (%) <0.001
 Medical 867 (60) 1246 (66) 302 (46)
 Surgical 187 (13) 228 (12) 152 (23)
 Cardiac 183 (13) 190 (10) 68 (10)
 Other 206 (14) 234 (12) 138 (21)
Admission type, n (%) 0.03
 Elective 64 (4) 87 (5) 46 (7)
 Urgent/emergency 1379 (96) 1811 (95) 614 (93)
Laboratory values, median (IQR)
 White blood cell count, 103/µl 9.0 (8.1–13.6) 6.6 (4.0–9.2) 13.0 (7.3–14.2) <0.001
 Hemoglobin, g/dl 10.5 (9.4–11.9) 9.7 (8.8–10.9) 10.1 (9.0–11.3) <0.001
 Hematocrit, % 31.4 (28.2–35.3) 28.7 (25.9–32.0) 29.7 (26.8–33.2) <0.001
 Platelets, 103/µl 232 (168–324) 140 (63–236) 126 (69–249) <0.001
 Sodium, mEq/L 137 (134–140) 136 (133–139) 136 (133–139) <0.001
 Potassium, mEq/L 4.0 (3.7–4.4) 4.0 (3.7–4.5) 4.2 (3.7–4.7) <0.001
 Chloride, mEq/L 104 (100–108) 106 (101–110) 103 (99–108) <0.001
 Bicarbonate, mEq/L 25 (23–28) 22 (18–25) 22 (18–26) <0.001
 Anion gap 13.0 (11.0–16.0) 14.0 (12.0–17.0) 16.0 (13.0–20.0) <0.001
 BUN, mg/dl 22 (14–35) 29 (17–48) 32 (19–54) <0.001
 Phosphate 3.1 (2.5–3.9) 3.4 (2.7–4.3) 4.0 (3.1–5.3) <0.001
 Eosinophils, % 0.3 (0.0–1.0) 0.1 (0.0–1.0) 0.1 (0.0–1.0) <0.001
 Basophils, % 0.1 (0.0–0.3) 0.0 (0.0–0.2) 0.0 (0.0–0.2) <0.001
 Neutrophils, % 85.0 (78.3–90.2) 79.0 (61.0–87.0) 80.6 (68.05–88.0) <0.001
 Aspartate aminotransferase, IU/L 38 (23–77) 44 (24–95) 103 (43–397) <0.001
 Alanine aminotransferase, IU/L 31 (17–67) 33 (18–69) 71 (31–262) <0.001
 Alkaline phosphatase, IU/L 97 (68–161) 110 (72–184) 109 (73–177) <0.001
 Bilirubin, mg/dl 0.7 (0.4–1.5) 1.1 (0.5–3.3) 2.7 (0.9–7.05) <0.001
 Creatinine, mg/dl 1.0 (0.7–1.5) 1.2 (0.8–2.1) 1.6 (1.0–2.8) <0.001
 Albumin, g/dl 2.9 (2.5–3.3) 2.7 (2.3–3.2) 2.7 (2.2–3.1) <0.001
 Prothrombin time, s 14.6 (13.3–17.1) 15.4 (13.8–18.8) 16.7 (14.5–20.8) <0.001
 Partial thromboplastin time, s 32.1 (27.5–42.42) 35.2 (29.4–46.0) 41.1 (32.8–57.6) <0.001
 pCO2, mmHg 43.0 (38.0–51.0) 37.0 (32.0–43.0) 38.0 (33.0–45.0) <0.001
 Lactate, mmol/L 1.6 (1.1–2.3) 2.1 (1.4–3.3) 3.6 (2.1–6.4) <0.001
 Lactate dehydrogenase, U/L 262 (199–385) 289 (199–467) 384 (246–892) <0.001
Blood culture positive, n (%) 436 (30) 702 (37) 199 (30) <0.001
Urine culture positive, n (%) 314 (22) 504 (27) 143 (22) 0.002
Vital signs, median (IQR)
 Temperature, °C 37.1 (36.5–37.7) 36.9 (36.3–37.6) 37 (36.4–37.6) <0.001
 Systolic BP, mm Hg 114 (101–131) 109 (97–124) 109 (96–124) <0.001
 Diastolic BP, mm Hg 58 (46–68) 56 (48–67) 56 (46–66) <0.001
 Heart rate 87 (75–103) 91 (78–105) 80 (74–83) <0.001
 Respiratory rate 20 (16–24) 20 (16–25) 21 (16–26) <0.001
 Oxygen saturation, % 97 (94–99) 97 (95–99) 97 (95–99) <0.001
Median admission length of stay, d (IQR) 11 (7–20) 14 (8–24) 19 (9–32) <0.001
On vasopressors, n (%) 562 (39) 1177 (62) 499 (76) <0.001
Required mechanical ventilation 768 (53) 1155 (61) 564 (86) <0.001
KDIGO diagnostic criteria, n (%) <0.001
 Urine output 1015 (70) 749 (39) 87 (13)
 Serum creatinine 102 (7) 246 (13) 76 (12)
 Both 326 (23) 903 (48) 497 (75)
KDIGO AKI stage <0.001
 1 356 (25) 425 (22) 102 (15)
 2 795 (55) 812 (43) 173 (26)
 3 292 (20) 661 (35) 385 (58)
AKI persistence <0.001
 Persistence 71 (17) 257 (22) 190 (33)
 No persistence 357 (83) 892 (78) 383 (67)
IQR, interquartile range; SAPS II, simplified acute physiology score; pCO2, partial pressure of carbon dioxide; KDIGO, Kidney Disease Improving Global Outcomes.

There were significant differences in several laboratory features (Table 1). We calculated the top 18 features that had the largest differences between subphenotypes 1 and 3 (Figure 3). Subphenotype 3 had significantly higher median bilirubin levels (2.7 versus 1.1 versus 0.7 mg/dl; P<0.001), higher median aspartate aminotransferase (103 versus 44 versus 38 U/L; P<0.001), and higher median alanine aminotransferase (71 versus 33 versus 31 U/L; P<0.001) compared with subphenotypes 2 and 1, respectively. Patients in subphenotype 3 also had higher median lactate, lactate dehydrogenase, and white blood cell count than patients in subphenotypes 1 and 2. Patients in subphenotype 3 also had worse kidney function parameters, including higher creatinine (1.6 versus 1.2 versus 1.0 mg/dl; P<0.001), higher BUN (32 versus 29 versus 22 mg/dl; P<0.001), and lower bicarbonate (22 versus 22 versus 25 mEq/L; P<0.001) than patients in subphenotypes 2 and 1, respectively. A full list of all laboratory features that were considered for inclusion and their associated P values and missingness is included in Supplemental Table 2.

Figure 3.:
Top 18 features found to have the largest differences between subphenotype 1 and subphenotype 3 include labs representative of kidney and liver function. Violin plots for the median value of 18 laboratory features. The 18 laboratory features with the largest log-transformed differences between subphenotypes 1 and 3 are displayed. Subphenotype 1 is represented in blue, subphenotype 2 is represented in orange, and subphenotype 3 is represented in red. pCO2, partial pressure of carbone dioxide; PT, prothrombin time; PTT, partial thromboplastin time; RDW, red cell distribution width.

AKI Characteristics among Subphenotypes

The predominant cause of AKI in all three subphenotypes was ATN: 14 of 30 (47%) patients in subphenotype 1, 19 of 30 (63%) patients in subphenotype 2, and 23 of 30 (77%) patients in subphenotype 3. There were significant differences in AKI etiology among the three subphenotypes (P=0.001) (Table 2). There was a statistically significant difference in KDIGO AKI stages, with a higher proportion of stage 3 AKI in subphenotype 3 (58% versus 35% [subphenotype 2] versus 20% [subphenotype 1]; P<0.001). More patients in subphenotype 3 had AKI by KDIGO creatinine or creatinine and urine output criteria: 87% in subphenotype 3, 61% in subphenotype 2, and 30% in subphenotype 1 (P<0.001). Characteristics of patients identified by KDIGO urine output criteria alone, creatinine criteria alone, or urine output and creatinine criteria are presented in Supplemental Table 3. Small but significant differences were found for age, sex, and race. Patients who had AKI by both creatinine and urine output criteria had higher proportions of hypertension, congestive heart failure, liver disease, and CKD (Supplemental Table 3). Subphenotype 3 had a higher proportion of patients with AKI persistence (33% versus 22% [subphenotype 2] versus 17% [subphenotype 1]; P=0.03).

Table 2. - Causes of AKI as assessed by review of 30 charts from each subphenotype
Cause of AKI Subphenotype 1 Subphenotype 2 Subphenotype 3
ATN, n (%) 14 (47) 19 (63) 23 (77)
Prerenal, n (%) 13 (43) 6 (20) 3 (10)
Other, n (%) 3 (10) 5 (17) 4 (13)
P=0.001. ATN, acute tubular necrosis.

Association between Subphenotype and Outcomes

Patients in subphenotype 3 had a significantly higher mortality (49% versus 35% versus 23%; P<0.001) and more patients received dialysis (26% versus 7% versus 4%; P<0.001) compared with subphenotypes 2 and 1, respectively (Figure 4). In unadjusted analysis, compared with subphenotype 1, subphenotype 2 had nearly double the odds of in-hospital mortality (odds ratio, 1.8; 95% confidence interval [95% CI], 1.5 to 2.1), and subphenotype 3 had three times the odds (odds ratio, 3.2; 95% CI, 2.6 to 3.9). After adjustment, subphenotypes 2 (adjusted odds ratio [aOR], 1.4; 95% CI, 1.2 to 1.6) and 3 (aOR, 1.93; 95% CI, 1.5 to 2.4) had higher odds or mortality than subphenotype 1. Subphenotypes 2 and 3 also had higher odds for dialysis than subphenotype 1 in unadjusted analysis (aOR, 2; 95% CI, 1.5 to 2.7 and aOR, 8.8; 95% CI, 6.4 to 12.1, respectively); however, after adjustment, only the association between subphenotypes 3 and 1 remained significant (aOR, 3.6; 95% CI, 2.5 to 5.4).

Figure 4.:
Large differences on dialysis and mortality across AKI subphenotypes. Gray bars represent the proportion of patients who required dialysis, and black bars represent the proportion of patients who died within each subphenotype.


Using deep learning techniques to analyze routinely measured laboratory measurements and vital signs, we identified three distinct subphenotypes of patients within the larger syndrome of sepsis-associated AKI. We find that clusters were significantly different in regard to comorbidities, laboratory measurements, and vital signs. We also found that these subphenotypes differed significantly in terms of mortality, which persisted even after adjusting for demographics, comorbidities, and stage and definition of AKI.

There has been speculation that sepsis-associated AKI is not a single clinical entity but likely a complex syndrome composed of several different subtypes (16). Because of widespread use of EHRs, granular data are collected on every patient in the ICU. This provides us with the opportunity to investigate this hypothesis in a data-driven manner. Subphenotyping in other diseases, such as diabetes, has been conducted using EHR data with great success, and patient subgroups are found to have differing outcomes and genetic pathways (17). Additionally, previous work using clinical trial data from an acute respiratory distress syndrome trial has shown different subtypes within AKI with differing outcomes (18). Although others have applied machine learning to identify postoperative sepsis and AKI, to the best of our knowledge, this is the first instance of utilizing routinely collected EHR data from the ICU setting for subphenotyping of sepsis-associated AKI using deep learning.

We identified three subphenotypes of patients with low, moderate, and high risk of dialysis and 28-day mortality. We accomplished this by agnostically identifying subphenotypes in a data-driven manner using physiologic and biochemical measurements and comorbidities, without inclusion of features, such as age and sex, that are commonly included in most ICU prediction models (19,20). The inclusion of inflammatory biomarkers may provide additional insights; however, these were not available as this was a clinical database and not a research database (21). We also chose to use deep learning for dimensionality reduction on these complex data to generate latent features for subphenotyping.

Through this inclusive data-driven approach, we identified several laboratory results that were substantially different across the three subphenotypes. Markers of liver injury were common among the features with largest differences; these included bilirubin, aspartate aminotransferase, and alanine aminotransferase. This is consistent with the higher prevalence of liver disease in subphenotype 3. The higher risk of in-hospital sepsis-related mortality of patients who are cirrhotic compared with the general population may explain some of the differences seen in outcomes among our subphenotypes (22).

CKD was common in all three subphenotypes. Although subphenotype 2 had the highest proportion of patients with CKD, subphenotype 3 had worse kidney function parameters, including higher creatinine, higher BUN, and worse acidosis. Additionally, subphenotype 3 had more patients who had AKI persistence and more patients with stage 3 AKI. Several studies have found that duration of AKI and AKI stage are associated with higher mortality (15,23–25).

The highest-risk group, subphenotype 3, had a higher severity of sepsis indicated by higher sepsis markers and acute-phase reactants, such as lactate, white blood cell count, and lactate dehydrogenase. These features have been documented to be associated with higher risk of mortality (26–28). These laboratory tests were likely markers of more severe sepsis or the development of organ failure. Additionally, subphenotype 3 had a much higher proportion of patients who required vasopressor therapy compared with the other subphenotypes. The high mortality seen in subphenotype 3 is likely driven by differences in comorbidities, severity and duration of AKI, and severity of sepsis.

We used both KDIGO urine output and serum creatinine criteria to identify our AKI cohort. Although this increases the number of patients we may have, this may also increase the false-positive rate because urine output is not always documented accurately. We recognize that there were differences in patients identified by creatinine and urine output criteria. However, even after adjustment for method used for AKI identification, there was still a significantly higher mortality among patients in subphenotype 3 compared with those in subphenotype 1.

The widespread use of EHRs allows for the implementation of machine-learned models via clinical decision support systems. Such systems have been used for AKI prediction and alerting, resulting in improvement in length of stay and mortality (29,30). Accurate risk assessment early in the ICU course will allow providers to allocate scarce resources, such as continuous RRT, and aid in discussions with patients and their families regarding goals of care. However, the model presented here would require additional testing and external validation before clinical implementation can be considered.

We also acknowledge several limitations. We used the MIMIC III database, which has undergone extensive cleaning; therefore, it may not be reflective of other EHRs, which decreases the generalizability of our findings. We used ICD9 codes to define our sepsis population; therefore, we are unable to determine the timing of sepsis diagnosis. Because MIMIC III is an exclusive inpatient database, we are unable to determine patients’ true baseline serum creatinine and kidney function. This will lead us to miss patients who have community-acquired AKI. However, this was mitigated by limiting AKI diagnosis to within 48 hours of ICU admissions and, thus, identifying patients with sepsis-associated AKI at the expense of a smaller sample size. AKI diagnosis was done using both creatinine and urine output KDIGO criteria, and urine output recordings may be inaccurate. However, we used only urine output measurements within 48 hours of ICU admission, which is presumably better than the 40% documentation prior to ICU admission (31). Although the MIMIC III database is a large ICU database, it is a single-center database, and external validation in ICU databases from different centers is needed.

In summary, we used a data-driven deep learning approach to identify three subphenotypes of sepsis-associated AKI with significantly different outcomes even after adjusting for AKI stage, acute severity, and demographics from publicly available data. To the best of our knowledge, this is the first study to demonstrate clinical subphenotypes within sepsis-associated AKI syndrome in the ICU using routinely collected EHR data. When combined with other biomarker and omics data, this approach could further expedite investigations into discovery of novel biomarkers and dysregulated pathways for sepsis-associated AKI.


L. Chan reports receiving grant funding for work outside of this project. S.G. Coca reports receiving personal fees, equity, and stock options from RenalytixAI; consulting fees from Bayer, Boehringer-Ingelheim, CHF Solutions, Goldfinch Bio, inRegen, Relypsa, and Takeda; and personal fees and stock options from pulseData, outside the submitted work. S. Jaladanki reports receiving grants from Icahn School of Medicine at Mount Sinai during the conduct of the study. K. Johnson reports receiving personal fees from Tempus Labs and Thorne Research, outside the submitted work. G.N. Nadkarni reports receiving financial compensation as a consultant and advisory board member for RenalytixAI and has equity in RenalytixAI as a cofounder; personal fees from AstraZeneca for serving on the advisory board for Sodium Glucose Co-Transporter 2 inhibitors; personal fees from Reata for serving on the virtual advisory board; consulting fees from BioVie and GLG Consulting; and grant support from Goldfinch Bio during the conduct of the study. T. Van Vleck was one of the founding employees of Clinithink, the company producing the Natural Language Processing software used to identify phenotypes in this study. T. Van Vleck continues to consult for Clinithink and remains a minor shareholder. Although the software was used in the study, Clinithink was not involved in the study and does not benefit from it. All remaining authors have nothing to disclose.



Published online ahead of print. Publication date available at www.cjasn.org.

See related editorial, “Not All Sepsis-Associated Acute Kidney Injury Is the Same: There May Be an App for That,” on pages .


We thank the Massachusetts Institute of Technology Laboratory for Computational Physiology for making the MIMIC III database freely available and the larger scientific community for generating methods and code repositories to utilize this resource.

Supplemental Material

This article contains the following supplemental material online at http://cjasn.asnjournals.org/lookup/suppl/doi:10.2215/CJN.09330819/-/DCSupplemental.

Supplemental Table 1. Top 20 primary discharge diagnoses for the three subphenotypes.

Supplemental Table 2. List of all laboratory features, P values, and percentage missing for the three subphenotypes.

Supplemental Table 3. Patient characteristics of patients identified by urine output only, serum creatinine only, or both urine output and creatinine.


1. Chertow GM, Levy EM, Hammermeister KE, Grover F, Daley J: Independent association between acute renal failure and mortality following cardiac surgery. Am J Med 104: 343–348, 1998 9576407
2. de Mendonça A, Vincent JL, Suter PM, Moreno R, Dearden NM, Antonelli M, Takala J, Sprung C, Cantraine F: Acute renal failure in the ICU: Risk factors and outcome evaluated by the SOFA score. Intensive Care Med 26: 915–921, 2000 10990106
3. Chertow GM, Burdick E, Honour M, Bonventre JV, Bates DW: Acute kidney injury, mortality, length of stay, and costs in hospitalized patients. J Am Soc Nephrol 16: 3365–3370, 2005 16177006
4. Mansuri U, Patel A, Shah H, Chauhan K, Poojary P, Saha A, Dave M, Hazra A, Mishra T, Annapureddy N, Nadkarni G, Chan L: Trends and outcomes of sepsis hospitalizations complicated by acute kidney injury requiring hemodialysis. J Crit Care 38: 353–355, 2017 27986342
5. Prowle JR, Ishikawa K, May CN, Bellomo R: Renal blood flow during acute renal failure in man. Blood Purif 28: 216–225, 2009 19648741
6. Bhatraju PK, Zelnick LR, Herting J, Katz R, Mikacenic C, Kosamo S, Morrell ED, Robinson-Cohen C, Calfee CS, Christie JD, Liu KD, Matthay MA, Hahn WO, Dmyterko V, Slivinski NSJ, Russell JA, Walley KR, Christiani DC, Liles WC, Himmelfarb J, Wurfel MM: Identification of acute kidney injury subphenotypes with differing molecular signatures and responses to vasopressin therapy. Am J Respir Crit Care Med 199: 863–872, 2019 30334632
7. Chan L, Vaid A, Nadkarni GN: Applications of machine learning methods in kidney disease: Hope or hype? Curr Opin Nephrol Hypertens 29: 319–326, 2020 32235273
8. Johnson AEW, Pollard TJ, Shen L, Lehman LWH, Feng M, Ghassemi M, Moody B, Szolovits P, Celi LA, Mark RG: MIMIC-III, a freely accessible critical care database. Sci Data 3: 160035, 2016 27219127
9. Kidney Disease: Improving Global Outcomes (KDIGO): KDIGO Clinical Practice Guideline for Acute Kidney Injury. Available at: https://kdigo.org/wp-content/uploads/2016/10/KDIGO-2012-AKI-Guideline-English.pdf. Accessed April 24, 2019
10. Martin GS, Mannino DM, Eaton S, Moss M: The epidemiology of sepsis in the United States from 1979 through 2000. N Engl J Med 348: 1546–1554, 2003 12700374
11. Agency for Health Care and Research Quality: Elixhauser Comorbidity Software, Version 3.7.
12. Kuhn M: Building predictive models in R using the caret package. J Stat Softw 28: 1–26, 2008
13. Krijthe JH: Rtsne: T-distributed stochastic neighbor embedding using a Barnes-Hut implementation, 2018. Available at: https://github.com/jkrijthe/Rtsne. Accessed March 19, 2019
14. Ligges U, Mächler M: Scatterplot3d—An R package for visualizing multivariate data. J Stat Softw 8: 1–20, 2003
15. Mehta S, Chauhan K, Patel A, Patel S, Pinotti R, Nadkarni GN, Parikh CR, Coca SG: The prognostic importance of duration of AKI: A systematic review and meta-analysis. BMC Nephrol 19: 91, 2018 29673338
16. Singbartl K, Kellum JA: AKI in the ICU: Definition, epidemiology, risk stratification, and outcomes. Kidney Int 81: 819–825, 2012 21975865
17. Li L, Cheng WY, Glicksberg BS, Gottesman O, Tamler R, Chen R, Bottinger EP, Dudley JT: Identification of type 2 diabetes subgroups through topological analysis of patient similarity. Sci Transl Med 7: 311ra174, 2015 26511511
18. Calfee CS, Delucchi K, Parsons PE, Thompson BT, Ware LB, Matthay MA; NHLBI ARDS Network: Subphenotypes in acute respiratory distress syndrome: Latent class analysis of data from two randomised controlled trials. Lancet Respir Med 2: 611–620, 2014 24853585
19. Le Gall J-R, Lemeshow S, Saulnier F: A new Simplified Acute Physiology Score (SAPS II) based on a European/North American multicenter study. JAMA 270: 2957–2963, 1993 8254858
20. Knaus WA, Draper EA, Wagner DP, Zimmerman JE: APACHE II: A severity of disease classification system. Crit Care Med 13: 818–829, 1985 3928249
21. Wiersema R, Jukarainen S, Vaara ST, Poukkanen M, Lakkisto P, Wong H, Linder A, van der Horst ICC, Pettilä V: Two subphenotypes of septic acute kidney injury are associated with different 90-day mortality and renal recovery. Crit Care 24: 150, 2020 32295614
22. Foreman MG, Mannino DM, Moss M: Cirrhosis as a risk factor for sepsis and death: Analysis of the national hospital discharge survey. Chest 124: 1016–1020, 2003 12970032
23. Han SS, Kim S, Ahn SY, Lee J, Kim DK, Chin HJ, Chae DW, Na KY: Duration of acute kidney injury and mortality in critically ill patients: A retrospective observational study. BMC Nephrol 14: 133, 2013 23802916
24. Hoste EAJ, Clermont G, Kersten A, Venkataraman R, Angus DC, De Bacquer D, Kellum JA: RIFLE criteria for acute kidney injury are associated with hospital mortality in critically ill patients: A cohort analysis. Crit Care 10: R73, 2006 16696865
25. Levi TM, de Souza SP, de Magalhães JG, de Carvalho MS, Cunha AL, Dantas JG, Cruz MG, Guimarães YL, Cruz CM: Comparison of the RIFLE, AKIN and KDIGO criteria to predict mortality in critically ill patients. Rev Bras Ter Intensiva 25: 290–296, 2013 24553510
26. Mikkelsen ME, Miltiades AN, Gaieski DF, Goyal M, Fuchs BD, Shah CV, Bellamy SL, Christie JD: Serum lactate is associated with mortality in severe sepsis independent of organ failure and shock. Crit Care Med 37: 1670–1677, 2009 19325467
27. Lu J, Wei Z, Jiang H, Cheng L, Chen Q, Chen M, Yan J, Sun Z: Lactate dehydrogenase is associated with 28-day mortality in patients with sepsis: A retrospective observational study. J Surg Res 228: 314–321, 2018 29907227
28. Chang R, Wong GY: Prognostic significance of marked leukocytosis in hospitalized patients. J Gen Intern Med 6: 199–203, 1991 2066822
29. Al-Jaghbeer M, Dealmeida D, Bilderback A, Ambrosino R, Kellum JA: Clinical decision support for in-hospital AKI. J Am Soc Nephrol 29: 654–660, 2018
30. Wilson FP, Greenberg JH: Acute kidney injury in real time: Prediction, alerts, and clinical decision support. Nephron 140: 116–119, 2018 30071528
31. Jonsson T, Jonsdottir H, Möller AD, Baldursdottir L: Nursing documentation prior to emergency admissions to the intensive care unit. Nurs Crit Care 16: 164–169, 2011 21651656

acute renal failure; dialysis; mortality; AKI; deep learning; subtypes; acute kidney injury

Copyright © 2020 by the American Society of Nephrology