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

Abstract

Visual Abstract

Background and objectives 

Sepsis-associated AKI is a heterogeneous clinical entity. We aimed to agnostically identify sepsis-associated AKI subphenotypes using deep learning on routinely collected data in electronic health records.

Design, setting, participants, & measurements 

We used the Medical Information Mart for Intensive Care III database, which consists of electronic health record data from intensive care units in a tertiary care hospital in the United States. We included patients ≥18 years with sepsis who developed AKI within 48 hours of intensive care unit admission. We then used deep learning to utilize all available vital signs, laboratory measurements, and comorbidities to identify subphenotypes. Outcomes were mortality 28 days after AKI and dialysis requirement.

Results 

We identified 4001 patients with sepsis-associated AKI. We utilized 2546 combined features for K-means clustering, identifying three subphenotypes. Subphenotype 1 had 1443 patients, and subphenotype 2 had 1898 patients, whereas subphenotype 3 had 660 patients. Subphenotype 1 had the lowest proportion of liver disease and lowest Simplified Acute Physiology Score II scores compared with subphenotypes 2 and 3. The proportions of patients with CKD were similar between subphenotypes 1 and 3 (15%) but highest in subphenotype 2 (21%). Subphenotype 1 had lower median bilirubin levels, aspartate aminotransferase, and alanine aminotransferase compared with subphenotypes 2 and 3. Patients in subphenotype 1 also had lower median lactate, lactate dehydrogenase, and white blood cell count than patients in subphenotypes 2 and 3. Subphenotype 1 also had lower creatinine and BUN than subphenotypes 2 and 3. Dialysis requirement was lowest in subphenotype 1 (4% versus 7% [subphenotype 2] versus 26% [subphenotype 3]). The mortality 28 days after AKI was lowest in subphenotype 1 (23% versus 35% [subphenotype 2] versus 49% [subphenotype 3]). After adjustment, the adjusted odds ratio for mortality for subphenotype 3, with subphenotype 1 as a reference, was 1.9 (95% confidence interval, 1.5 to 2.4).

Conclusions 

Utilizing routinely collected laboratory variables, vital signs, and comorbidities, we were able to identify three distinct subphenotypes of sepsis-associated AKI with differing outcomes.

Introduction

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).

Clustering

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.

Results

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.

Discussion

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.

Disclosures

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.

Funding

None.