ARTICLE

https://doi.org/10.1038/s43856-023-00278-w OPEN

Deep learning-based electrocardiographic

screening for chronic kidney disease
Lauri Holmstrom1,2,3,4,10, Matthew Christensen1,4,10, Neal Yuan5, J. Weston Hughes6, John Theurer1,4,
Melvin Jujjavarapu7, Pedram Fatehi8, Alan Kwan 1, Roopinder K. Sandhu1, Joseph Ebinger1, Susan Cheng1,
James Zou 6,9, Sumeet S. Chugh1,2,4,11 & David Ouyang 1,4,11 ✉

Abstract Plain language summary

Background Undiagnosed chronic kidney disease (CKD) is a common and usually asymp- Chronic kidney disease (CKD) is a
common condition involving loss of
1234567890():,;

tomatic disorder that causes a high burden of morbidity and early mortality worldwide. We
kidney function over time and results
developed a deep learning model for CKD screening from routinely acquired ECGs.
in a substantial number of deaths.
Methods We collected data from a primary cohort with 111,370 patients which had 247,655
However, CKD often has no symp-
ECGs between 2005 and 2019. Using this data, we developed, trained, validated, and tested a toms during its early stages. To
deep learning model to predict whether an ECG was taken within one year of the patient detect CKD earlier, we developed a
receiving a CKD diagnosis. The model was additionally validated using an external cohort computational approach for CKD
from another healthcare system which had 312,145 patients with 896,620 ECGs between screening using routinely acquired
electrocardiograms (ECGs), a cheap,
2005 and 2018.
rapid, non-invasive, and commonly
Results Using 12-lead ECG waveforms, our deep learning algorithm achieves discrimination
obtained test of the heart’s electrical
for CKD of any stage with an AUC of 0.767 (95% CI 0.760–0.773) in a held-out test set and activity. Our model achieved good
an AUC of 0.709 (0.708–0.710) in the external cohort. Our 12-lead ECG-based model accuracy in identifying any stage of
performance is consistent across the severity of CKD, with an AUC of 0.753 (0.735–0.770) CKD, with especially high accuracy in
for mild CKD, AUC of 0.759 (0.750–0.767) for moderate-severe CKD, and an AUC of 0.783 younger patients and more severe
stages of CKD. Given the high global
(0.773–0.793) for ESRD. In patients under 60 years old, our model achieves high perfor-
burden of undiagnosed CKD, novel
mance in detecting any stage CKD with both 12-lead (AUC 0.843 [0.836–0.852]) and 1-lead
and accessible CKD screening stra-
ECG waveform (0.824 [0.815–0.832]). tegies have the potential to help
Conclusions Our deep learning algorithm is able to detect CKD using ECG waveforms, with prevent disease progression and
stronger performance in younger patients and more severe CKD stages. This ECG algorithm reduce premature deaths related
has the potential to augment screening for CKD. to CKD.

1 Department of Cardiology, Smidt Heart Institute, Cedars-Sinai Medical Center, Los Angeles, CA, USA. 2 Center for Cardiac Arrest Prevention, Department of

Cardiology, Smidt Heart Institute, Cedars-Sinai Medical Center, Los Angeles, CA, USA. 3 Research Unit of Internal Medicine, Medical Research Center Oulu,
University of Oulu and Oulu University Hospital, Oulu, Finland. 4 Division of Artificial Intelligence in Medicine, Department of Medicine, Cedars-Sinai Medical
Center, Los Angeles, CA, USA. 5 Department of Medicine, Division of Cardiology, San Francisco VA, UCSF, San Francisco, CA, USA. 6 Department of
Computer Science, Stanford University, Palo Alto, CA, USA. 7 Enterprise Information Service, Cedars-Sinai Medical Center, Los Angeles, CA, USA. 8 Division
of Nephrology, Department of Medicine, Stanford University, Palo Alto, CA, USA. 9 Department of Biomedical Data Science, Stanford University, Palo Alto,
CA, USA. 10These authors contributed equally: Lauri Holmstrom, Matthew Christensen. 11These authors jointly supervised this work: Sumeet S. Chugh, David
Ouyang. ✉email: [email protected]

COMMUNICATIONS MEDICINE | (2023)3:73 | https://doi.org/10.1038/s43856-023-00278-w | www.nature.com/commsmed 1

ARTICLE COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00278-w

A
lmost 700 million individuals globally have chronic kid- model to evaluate the entire spectrum of kidney disease. To
ney disease (CKD), an important but often unrecognized further evaluate our model, we validated its performance using
cause of morbidity and early mortality1. The initial pre- corresponding data from a separate healthcare system.
sentation of CKD is usually asymptomatic and without overt
clinical manifestations especially in the early stages of the disease. Methods
Recently, the Global Burden of Diseases, Injuries and Risk Factors Data sources and study population. We retrospectively identified
Study (GBD) estimated that CKD accounts for 4.6% of total 64,308 ECGs among 7816 patients between 2005 and 2019 which
mortality worldwide, with a 41.5% increase between 1990 and were linked to a diagnosis of CKD within a 1-year window at
20171. Delayed diagnosis and limited patient recognition of the Cedars-Sinai Medical Center. We also identified 183,290 ECGs
condition contribute significantly to the burden of morbidity2,3. among 103,554 patients between 2008 and 2019 with no CKD
Early detection can potentially change the disease trajectory. The diagnoses at any point, which were used as matched negative
most common causes of CKD, such as hypertension and diabetes, controls. Study cohorts included both ambulatory and in-hospital
can be reversible or treatable, and early diagnosis is crucial for patients. If a patient had multiple ECGs taken within a year of a
avoiding renal replacement therapy4,5. There are few methods to CKD diagnosis, each ECG-CKD pair was considered an indepen-
cheaply or non-invasively screen for CKD, with conventional risk dent case during model training, but only one was used in the test
calculators lacking specificity and requiring both serum and urine datasets. The study population from Cedars-Sinai Medical Center
laboratory testing6. was randomly split 8:1:1 into training, validation, and test cohorts
Electrocardiograms (ECGs) are inexpensive, non-invasive, by patient such that the multiple ECGs from the same patient were
widely available, and rapid diagnostic tests frequently obtained limited to one cohort. In addition, we identified 896,620 ECGs
during routine visits, prior to exercise, during preoperative eva- among 312,145 patients at Stanford Healthcare from 8/2005 to 6/
luation, and for patients at increased risk of cardiovascular dis- 2018, which were used for external validation (Fig. 1).
ease. Deep learning algorithms (DLA) have recently been applied ECGs from Cedars-Sinai Medical Center were obtained from
to medical imaging and clinical data to achieve high precision, MUSE Cardiology Information System (GE Healthcare), and the
and to identify additional information beyond the interpretation model used the original ECG waveforms stored for training the
of human experts7,8. Deep learning analysis of ECG waveforms model in CKD prediction. In external validation at Stanford
has had potentially promising performance in prognosticating University, ECGs were stored using the Phillips TraceMaster
outcomes9, identifying subclinical disease10,11, and identifying system, and were independently used as input examples for
systemic phenotypes not traditionally associated with ECGs12,13. external validation. The ECG waveform data were acquired at a
Given the prior success in identifying occult arrhythmias14,15, sampling rate of 500 Hz and extracted as 10 second, 12 × 5000
ventricular dysfunction10, anemia13, and age12, DLA applied to matrices of amplitude values. ECGs with missing leads were
screening ECGs could potentially identify patients who would excluded from the study cohort. Associated clinical data for each
benefit from further evaluation for kidney disease. patient was obtained from the electronic health record. The data on
The high prevalence of concomitant cardiovascular disease and medical diagnoses was extracted from the electronic health records
the well-established changes that accompany electrolyte using International Classification of Diseases (ICD) 9/10th edition
abnormalities suggest that the ECG is also altered in the setting of codes, which are listed in Supplementary Table 1. Demographic
CKD and that discrete electrocardiographic signatures could be and clinical characteristics (e.g., age, gender, BMI, cardiovascular
identifiable with deep learning techniques. Patients with CKD disease) were also extracted from the electronic health records. The
have a disproportionate accumulation of cardiovascular risk institutional review boards of Cedars-Sinai Medical Center and
factors, such as diabetes and hypertension, as well as subclinical Stanford Healthcare approved the study protocol (Cedars Protocol
cardiovascular changes such as left ventricular hypertrophy, 1506 and Stanford Protocol 43721). Informed consent was waived
myocardial fibrosis, and diastolic dysfunction16. It is not fully for analysis of de-identified retrospective data.
clear at which stage CKD patients start to develop manifest car-
diovascular changes. However, recent studies have reported that
in addition to coronary artery disease and left ventricle hyper- AI model design and training. We designed a convolutional
trophy, patients with early-stage CKD may already have an neural network, for ECG interpretation with potential for clinical
increase in diffuse myocardial fibrosis on cardiac MRI as well17. It data integration to predict the primary outcomes of chronic kidney
is hence likely that already early-stage CKD associates with non- disease and end-stage renal disease (Fig. 2). The model was trained
specific ECG signals. In addition to myocardial remodeling, CKD to predict outcomes with the input of one 12-lead ECG obtained
associates with a variety of electrolyte abnormalities that also within 1 year of diagnosis. Please see Supplementary information
cause widespread ECG abnormalities (e.g., decreased T-wave for additional details on model training. If the same patient had
amplitudes in hypokalemia, large-amplitude T-waves, and pro- multiple ECGs, each was considered an independent case. Models
longed QRS duration in hyperkalemia, and QTc prolongation in were trained using the PyTorch deep learning framework. The
hypocalcemia)18. Prior work has shown such patterns are model was initialized with random weights and trained using a
detectable on ECG waveforms, contributing to the AI-ECG binary cross-entropy loss function for up to 100 epochs with an
detection of hyperkalemia, which might augment a model’s ability ADAM optimizer and an initial learning rate of 1e-4. Early stop-
to detect CKD19,20. However, given the relative infrequency of ping was performed based on the validation dataset’s area under
overt abnormalities, likely not the primary feature analyzed in the receiver operating curve. Local Interpretable Model-agnostic
detecting CKD. Given such observations, it may be possible that Explanations (LIME)15,21 was used with 1000 samples per study to
asymptomatic CKD presents with subtle ECG alterations that are identify relevant features in the ECG waveform by iteratively ran-
not visible to the human eye. domly perturbing 0.5% of the waveform and identifying which
To overcome current limitations in screening for occult CKD, changes most impacted model performance.
we designed, trained, and validated a deep learning model to
predict CKD, including end-stage renal disease (ESRD), by ana- Statistical analysis. All analyses were performed on the held-out
lysis of waveform signals from a single 12-lead and 1-lead ECG. test dataset, which was never seen during model training. The
Incorporating both structured information from medical diag- performance of the model in predicting the primary outcomes
noses as well as laboratory data, we assessed the ability of our was mathematically assessed by the area under the curve (AUC)

2 COMMUNICATIONS MEDICINE | (2023)3:73 | https://doi.org/10.1038/s43856-023-00278-w | www.nature.com/commsmed

COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00278-w ARTICLE

Primary cohort External validation cohort

247,655 ECG recordings among 111,370 paents 896,620 ECG recordings among 312,145 paents

Positive patients Negative patients

•CKD diagnosis within 1 year of ECG •No CKD diagnosis at any me

• 64,308 ECGs among 7,816 paents •183,290 ECGs among 103,554 paents

Positive patients Negative patients

•CKD diagnosis within 1 year of ECG •No CKD diagnosis at any me

80:10:10 Train:Validaon:Test split • 108,306 ECGs among 18,002 paents •788,314 ECGs among 294,143 paents

Training dataset Validation dataset Testing dataset External validation dataset

•197,434 ECGs from 89,064 •24,490 ECGs from 11,126 •25,681 ECGs from 11,132
paents paents paents
•7,306 posive paents and •897 posive paents and •933 posive paents and
81,758 negave paents 10,229 negave paents 10,199 negave paents

Fig. 1 Study subject selection. Our primary cohort consists of 111,370 patients and 247,655 ECGs between 2005 and 2019 from Cedars-Sinai Medical
Center. The primary cohort was randomly split 8:1:1 into training, validation, and test cohorts. We also used 896,620 ECGs among 312,145 patients at
Stanford Healthcare from 8/2005 to 6/2018 as external validation cohort. CKD Chronic kidney disease.

of the receiver operating characteristic (ROC) curve. After model achieving an AUC of 0.753 (0.735–0.770) in discriminating mild
derivation and training, primary and secondary analyses were CKD, AUC of 0.759 (0.750–0.767) in discriminating moderate-
performed on trained models using the held-out test cohort. severe CKD, and AUC of 0.783 (0.773–0.793) in discriminating
Secondary sensitivity analyses were limited to procedures per- ESRD. In all cases, negative examples were defined as ECGs
formed in patients with diabetes, hypertension, male, and age without CKD diagnoses.
greater or lower than 60 years old. We computed two-sided 95% Given the increased prevalence of wearable technologies,
confidence intervals using 1000 bootstrapped samples for each particularly devices that include single lead ECG information, we
calculation. Statistical analysis was performed in R and Python. trained an additional deep learning model with information from
only single lead ECG information to simulate the DLA’s
Reporting summary. Further information on research design is performance with single-lead wearable information. With 1-lead
available in the Nature Portfolio Reporting Summary linked to ECG waveform data, DLA achieved an AUC of 0.744 (0.737–0.751)
this article. in detecting any stage CKD, with sensitivity and specificity of 0.723
(0.723–0.723) and 0.643 (0.643–0.643), respectively.
Since early detection of CKD is crucial to prevent disease
Results progression and complications in older age, we tested the
Primary cohort characteristics. In total, we identified 17,860 performance of our model in younger patients (<60 years of
patients with a CKD diagnosis at Cedars-Sinai Medical Center age). 12-lead and 1-lead ECG-based DLAs were able to detect any
(7.8% of the total patient sample), among which 7816 had an stage CKD with AUCs of 0.843 (0.836–0.852) and 0.824
ECG taken within a 1-year window of CKD diagnosis. Our pri- (0.815–0.832) among patients under 60 years of age, respectively.
mary cohort consisted of a total of 247,655 ECGs, of which We also tested the performance of our model separately among
221,974 were randomized to the training set (for both training diabetic, hypertensive, older patients, who are generally con-
and validation) and 25,681 to the testing set. Of the primary sidered as high-risk subgroups. 12-lead based model detected
cohort ECGs, 74.3% had no serum creatinine or eGFR estimation CKD with an AUC of 0.747 (0.707–0.783) among diabetic
within 30 days and 50.7% of ECGs had no serum creatinine or patients, an AUC of 0.711 (0.696–0.725) among patients with
eGFR estimation at any point in the EHR, however this does not hypertension, and an AUC of 0.706 (0.697–0.716) among patients
capture outside hospital or paper clinic records of laboratory greater than 60 years old. When the model was trained with 12-
testing that might have been used in the diagnosis of CKD. The lead ECG waveform, age, sex, diabetes, and hypertension, the
mean age of the primary cohort was 61.3 ± 19.7 years and 48% model achieved similar discrimination of any stage CKD in the
were female. Demographic and clinical characteristics are pre- held-out test set with an AUC of 0.79 (0.781–0.798). Detailed
sented in Table 1. Demographics and clinical characteristics results for 1-lead and 12-lead ECG-based DLA performance in
according to age group are presented in Supplementary Table 2. the held-out test set are presented in Tables 2 and 3, while AUC
curves are illustrated in Supplementary Fig. 1.
Model performance in the primary cohort. Our 12-lead ECG- The model performed similarly in detecting CKD in subset
based model achieved discrimination of any stage CKD with an populations of patients with albuminuria, patients with correspond-
AUC of 0.767 (95% CI 0.76–0.773). The model performance was ing laboratory testing and documented eGFR, and in both
consistent across the range of CKD stage, with our model ambulatory and in-hospital patients (Supplementary Table 3).

COMMUNICATIONS MEDICINE | (2023)3:73 | https://doi.org/10.1038/s43856-023-00278-w | www.nature.com/commsmed 3

ARTICLE COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00278-w

Training
Input Training dataset Training ECG Inclusion/Exclusion Model Architecture Diagram
12-lead ECG Inclusion/exclusion of ECGs for x2 x3
corresponding CKD diagnosis

Atrous Convoluon

Mobile Boleneck
Mobile Boleneck

Mobile Boleneck
Mobile Boleneck
Positive

Mobile Boleneck
Mobile Boleneck
ECG ECG CKD dg ECG ECG

Max Pool

Max Pool
patients

±1 year window

No CKD diagnosis at any point

1-lead
1 lead ECG (lead I) of electronic health record

Negative ECG ECG ECG

Mobile Boleneck

Mobile Boleneck

Mobile Boleneck
Mobile Boleneck

Fully Connected
Average Pool
patients

Conv1D
CKD diagnosis

Testing
Internal New Never-Seen dataset Subset Analysis Explanation/Interpretability
Validation
ECG ECG CKD dg ECG
Positive
patients
±1 year window
Negative ECG ECG ECG > 60 years < 60 years
patients
LIME test

External Validation
Female Male CKD+ CKD-
ECG ECG CKD dg ECG
Positive
patients
±1 year window
Negative ECG ECG ECG
patients
Diabetic Hypertensive
CKD-related ECG abnormalities

Fig. 2 Schematic illustration of deep learning model training, testing, and validation. We designed a convolutional neural network for ECG interpretation
with potential for clinical data integration. The model was trained to predict CKD with the input of one 12-lead ECG within 1 year of CKD diagnosis. CKD
Chronic kidney disease.

In patients with both a CKD diagnosis and eGFR estimated to be less External validation cohort characteristics. The external valida-
than 60 mL/min, the AUC was 0.754 (0.737–0.771), and this tion cohort consisted of a total of 896,620 ECGs among 312,145
performance was similar in patients with hyperkalemia with an AUC patients. The prevalence of mild CKD was 1.2% while 3.6% had
of 0.741 (0.698–0.787) and without hyperkalemia with an AUC of moderate-severe CKD, and 0.9% had ESRD. The mean age of the
0.758 (0.747–0.768). The model also performed well in patients with external validation cohort was 56.7 ± 18.7 years and 50.4% were
known albuminuria, with an AUC of 0.734 (0.723–0.745) and had female. The proportion of Caucasians was 47.5%, while 3.6%
similar performance regardless of the positive to negative ratio in the were black, 12.3% were Asians, and 36.6% had other or unknown
training set (Supplementary Table 4). race. Demographic and clinical characteristics are presented in
Table 1.
Electrocardiographic features in CKD. To understand the key
features of relevance for our deep learning model to be able to Model performance in the external validation dataset. In the
detect CKD, we performed two sets of experiments to evaluate the external validation dataset, our 12-lead and 1-lead models’ per-
ECG parameters that are important for identifying CKD. We formances were comparable to the primary cohort. 12-lead ECG-
found statistically significant differences in all available ECG based model achieved an AUC of 0.709 (0.708-0.710) in dis-
variables (heart rate, PR interval, P wave duration, QRS duration, criminating any stage CKD. 1-lead ECG-based model detected
QTc interval, P-wave axis, R-wave axis, T-wave axis) between any stage CKD with an AUC of 0.701 (0.700–0.702).
CKD stages (Supplementary Table 5). Consistent with the primary cohort in which our model achieved
Secondly, we used LIME to identify which ECG segments were higher CKD detection accuracy among younger patients, 12-lead
particularly used in the identification of CKD. Supplementary and 1-lead ECG-based models achieved AUCs of 0.784
Fig. 2 shows examples of LIME-highlighted ECG segments in 12- (0.782–0.786) and 0.777 (0.775–0.779) in detecting any stage CKD
lead and 1-lead ECG waveforms taken from correctly recognized among subjects under 60 years of age, respectively. Detailed results
CKD and healthy control patients in the held-out test set. In for 1-lead and 12-lead ECG-based DLA performance in the external
both examples, the LIME-highlighted ECG features focused validation cohort are presented in Supplementary Tables 6 and 7.
mostly on QRS complexes and PR intervals. In addition, QRS
complexes and PR intervals in limb leads were most frequently Discussion
highlighted, potentially denoting CKD-associated electrophysio- In the present study, we investigated the performance of a
logic alterations. deep learning model to detect CKD using ECG waveforms.

4 COMMUNICATIONS MEDICINE | (2023)3:73 | https://doi.org/10.1038/s43856-023-00278-w | www.nature.com/commsmed

COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00278-w ARTICLE

Table 1 Demographic and clinical characteristics in the internal and external dataset.

Characteristic Internal training and validation datasets External validation dataset

Total Training Test
Number of patients 111,370 100,233 11,137 312,145
Number of ECGs 247,655 221,974 25,681 896,620
Demographics
Age, years 61.3 ± 19.7 61.3 ± 19.6 61.5 ± 19.9 56.7 ± 18.7
Female, n (%) 53,476 (48.0%) 48,133 (48.0%) 5,343 (48.0%) 139,235 (50.4%)
BMI, kg/m2 26.7 ± 16.6 26.7 ± 18.2 26.6 ± 7.2 27.6 ± 6.6
Caucasian, n (%) 67,253 (60.4%) 60,550 (60.4%) 6,703 (60.2%) 148,367 (47.5%)
Black, n (%) 15,323 (13.8%) 13,751 (13.7%) 1,572 (14.1%) 11,191 (3.6%)
Asian, n (%) 6,135 (5.5%) 5,558 (5.5%) 577 (5.2%) 38,229 (12.3%)
Other/unknown race, n (%) 22,659 (20.3%) 20,374 (20.3%) 2,285 (20.5%) 114,358 (36.6%)
Clinical characteristics, n (%)
Hypertension 25,446 (26.4%) 22,623 (26.4%) 2,823 (26.3%) 110,311 (35.3%)
Diabetes Mellitus 8,728 (7.8%) 6,970 (7.8%) 914 (8.2%) 14,152 (4.5%)
Cardiovascular disease 15,719 (14.1%) 12,561 (14.1%) 1,577 (14.2%) 34,582 (11.1%)
Heart Failure 9,604 (8.6%) 7,634 (8.6%) 995 (8.9%) 20,167 (6.5%)
Proteinuria 1,007 (0.9%) 795 (0.9%) 110 (1.0%) 2,086 (0.7%)
Anemia 11,933 (10.7%) 9,537 (10.7%) 1,194 (10.7%) 18,653 (6.0%)
Chronic kidney disease, n (%)
Mild (Stage 1–2) 864 (0.8%) 782 (0.8%) 82 (0.7%) 3,727 (1.2%)
Moderate-severe (Stage 3–5) 3918 (3.5%) 3515 (3.5%) 403 (3.6%) 11,335 (3.6%)
ESRD 3034 (2.7%) 2722 (2.7%) 312 (2.8%) 2,940 (0.9%)
eGFR*, ml/min/1.73 m2
<15 1,744 (7.0%) 1,531 (7.0%) 213 (7.7%)
15-29 1,789 (7.2%) 1,563 (7.1%) 226 (2.0%)
30-60 5,477 (22.1%) 4,831 (22.0%) 646 (23.3%)
>60 15,575 (62.9%) 13,904 (63.3%) 1,671 (60.3%)
Potassium (mmol/L)
>5.5 2,356 (2.7%) 2,075 (2.7%) 281 (2.9%)
<=5.5 85,573 (97.3%) 76,106 (97.3%) 9,467 (97.1%)

Continuous variables are presented as mean±standard deviation. *within a month of ECG (data available from 24,585 patients). BMI Body mass index, eGFR estimated glomerular filtration rate, ESRD End-
stage renal disease.

Table 2 Performance of the 12-lead ECG-based deep learning algorithm in the internal dataset.

Model AUC (95% CI) Sensitivity (95% CI) Specificity (95% CI) PPV (95% CI) NPV (95% CI)
12-lead ECG models
Any stage CKD 0.767 (0.76–0.773) 0.699 (0.699–0.699) 0.698 (0.698–0.698) 0.443 (0.443–0.443) 0.871 (0.871–0.871)
Mild CKD 0.753 (0.735–0.77) 0.75 (0.75–0.75) 0.644 (0.644–0.644) 0.064 (0.064–0.064) 0.987 (0.987–0.987)
Moderate-severe CKD 0.759 (0.75–0.767) 0.785 (0.785–0.785) 0.598 (0.598–0.598) 0.271 (0.271–0.271) 0.936 (0.936–0.936)
ESRD 0.783 (0.773–0.793) 0.704 (0.704–0.704) 0.726 (0.726–0.726) 0.237 (0.237–0.237) 0.953 (0.953–0.953)
High Risk Subgroup analyses
for any stage CKD
Diabetic patients 0.747 (0.707–0.783) 0.699 (0.699–0.699) 0.682 (0.682–0.682) 0.906 (0.906–0.906) 0.342 (0.342–0.342)
Hypertensive patients 0.711 (0.696–0.725) 0.648 (0.629–0.668) 0.662 (0.644–0.68) 0.638 (0.618–0.658) 0.672 (0.653–0.691)
Age > 60 years 0.706 (0.697–0.716) 0.604 (0.604–0.604) 0.701 (0.701–0.701) 0.397 (0.397–0.397) 0.844 (0.844–0.844)
Male 0.764 (0.755–0.772) 0.727 (0.727–0.727) 0.666 (0.666–0.666) 0.485 (0.485–0.485) 0.849 (0.849–0.849)
Female 0.756 (0.745–0.768) 0.699 (0.699–0.699) 0.680 (0.680–0.680) 0.351 (0.351–0.351) 0.902 (0.902–0.902)
Screening cohort (age <60)
Any stage CKD 0.843 (0.836–0.852) 0.761 (0.761–0.761) 0.787 (0.787–0.787) 0.57 (0.57–0.57) 0.899 (0.899–0.899)
Mild CKD 0.795 (0.766–0.823) 0.702 (0.702–0.702) 0.748 (0.748–0.748) 0.075 (0.075–0.075) 0.989 (0.989–0.989)
Moderate-severe CKD 0.854 (0.842–0.865) 0.792 (0.792–0.792) 0.787 (0.787–0.787) 0.373 (0.373–0.373) 0.959 (0.959–0.959)
ESRD 0.842 (0.831–0.853) 0.746 (0.746–0.746) 0.791 (0.791–0.791) 0.394 (0.394–0.394) 0.945 (0.945–0.945)

AUC area under the receiver operating characteristics curve, CI Confidence interval, CKD Chronic kidney disease, ESRD End-stage renal disease, PPV positive predictive value, NPV negative
predictive value.

Our 12-lead ECG-based model had good accuracy in identifying under 60 years of age. While 12-lead ECGs are widely available
any stage CKD and higher accuracy in detecting CKD in patients in the healthcare unit settings, rapid adoption of wearable tech-
under 60 years of age. Accuracy also improved along with the nology has also introduced opportunities for large-scale data
worsening CKD stage. These results were validated in a separate collection outside of formal healthcare settings. Our 1-lead ECG-
health care system, that also showed good discrimination accu- based DLA showed good discrimination accuracy for CKD in
racy for the presence of any stage CKD in the whole study young patients, suggesting artificial intelligence may possess
population and higher discrimination accuracy among patients significant potential in widescale screening in this patient

COMMUNICATIONS MEDICINE | (2023)3:73 | https://doi.org/10.1038/s43856-023-00278-w | www.nature.com/commsmed 5

ARTICLE COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00278-w

Table 3 Performance of the 1-lead ECG-based deep learning algorithm in the internal dataset.

Model AUC (95% CI) Sensitivity (95% CI) Specificity (95% CI) PPV (95% CI) NPV (95% CI)
1-lead ECG models
Any stage CKD 0.744 (0.737–0.751) 0.723 (0.723–0.723) 0.643 (0.643–0.643) 0.41 (0.41–0.41) 0.871 (0.871–0.871)
Mild CKD 0.746 (0.728–0.764) 0.735 (0.735–0.735) 0.66 (0.66–0.66) 0.066 (0.066–0.066) 0.987 (0.987–0.987)
Moderate-severe CKD 0.735 (0.726–0.744) 0.732 (0.732–0.732) 0.618 (0.618–0.618) 0.267 (0.267–0.267) 0.924 (0.924–0.924)
ESRD 0.757 (0.748–0.767) 0.738 (0.738–0.738) 0.647 (0.647–0.647) 0.202 (0.202–0.202) 0.953 (0.953–0.953)
High Risk Subgroup analyses
for any stage CKD
Diabetic patients 0.663 (0.625–0.707) 0.819 (0.819–0.819) 0.457 (0.457–0.457) 0.868 (0.868–0.868) 0.367 (0.367–0.367)
Hypertensive patients 0.684 (0.668–0.699) 0.658 (0.639–0.677) 0.614 (0.594–0.633) 0.61 (0.589–0.629) 0.662 (0.641–0.68)
Age > 60 years 0.681 (0.671–0.691) 0.679 (0.679–0.679) 0.588 (0.588–0.588) 0.35 (0.35–0.35) 0.849 (0.849–0.849)
Male 0.742 (0.733–0.75) 0.747 (0.747–0.747) 0.612 (0.612–0.612) 0.455 (0.455–0.455) 0.848 (0.848–0.848)
Female 0.735 (0.723–0.747) 0.672 (0.672–0.672) 0.681 (0.681–0.681) 0.342 (0.342–0.342) 0.894 (0.894–0.894)
Screening cohort (age <60)
Any stage CKD 0.824 (0.815–0.832) 0.812 (0.812–0.812) 0.705 (0.705–0.705) 0.506 (0.506–0.506) 0.91 (0.91–0.91)
Mild CKD 0.819 (0.791–0.844) 0.771 (0.771–0.771) 0.758 (0.758–0.758) 0.085 (0.085–0.085) 0.991 (0.991–0.991)
Moderate-severe CKD 0.828 (0.816–0.84) 0.814 (0.814–0.814) 0.713 (0.713–0.713) 0.313 (0.313–0.313) 0.96 (0.96–0.96)
ESRD 0.82 (0.808–0.831) 0.82 (0.82–0.82) 0.693 (0.693–0.693) 0.328 (0.328–0.328) 0.955 (0.955–0.955)

AUC area under the receiver operating characteristics curve, CI Confidence interval, CKD Chronic kidney disease, ESRD End-stage renal disease. PPV positive predictive value. NPV negative
predictive value.

population. One-lead ECGs could also increase screening rates in can be detected by CKD within large cohorts across multiple
high-risk patients (Supplementary Figs. 3 and 4). However, the international sites.
integration of artificial intelligence in electronic devices requires a The strengths of our study include the large cohort of patients
more detailed evaluation of accuracy in a real-life setting. undergoing ECG recording across a decade and the use of state-
Low awareness of CKD and limitations in current screening of-the-art deep learning architectures. We also used two separate
measures highlight the urgency of novel screening strategies to approaches to understand the key features of relevance for our
increase detection rates of early-stage CKD. Being non-invasive deep learning model. While previous studies have reported that
and often obtained in the clinic, ECGs are often the first line of patients with CKD have high rates of P wave abnormalities,
clinical evaluation. In our healthcare system, 74% of ECGs prolonged PR interval, QTc prolongation, QT dispersion, and left
obtained did not have laboratory testing of kidney function ventricle hypertrophy28–31, in the present study CKD was asso-
within 30 days. Previous studies have demonstrated that the cost- ciated with skewed P-, R-, and T-wave axes in addition to pro-
effectiveness of CKD screening is highly dependent on patient longed QRS, PR, and QTc intervals. However, a few limitations
risk factor profile and CKD probability, and there has been debate warrant consideration. Our study is retrospective, and study
on whether CKD screening should be targeted only to high-risk populations are derived from two large academic medical centers
patients, or also extend to patients without risk factors for situated in dense urban metropolitan areas using ICD-9 codes. By
CKD22–25. Although screening high-risk patients is guideline- prioritizing priority codes, we sought to avoid incidences of acute
recommended, testing rates remain low as only about 20% of rather than chronic kidney injury, however we cannot exclude the
high-risk patients receive guideline-recommended assessment in possibility that some of the study subjects without CKD diagnosis
the U.S.26. Consequently, most of the high-risk patients are likely in electronic health records have an undiagnosed disease, as
to be unaware of underlying CKD2,3. Moreover, a substantial especially mild-stage CKD can often be undiagnosed, particularly
proportion of all CKD patients are not high-risk patients and using an ICD9 code-based adjudication. In the subset with both
hence not recommended to be screened regularly, which further ICD-9 code adjudication of CKD as well as laboratory testing, the
highlights the need for novel screening methods. ICD-9 codes were consistent with the calculated eGFR, however
Our model performed better at detecting CKD in younger only a minority of patients were able to be linked to data
patients, whereas detection accuracy was lower in older and high- regarding microalbuminuria.
risk patients. Reasons for this observation are not fully clear but Validation in prospective general population cohorts in out-
may be due to the fact that younger patients in general have fewer patient settings is required to confirm an ECG-based DLA’s
comorbidities, meaning that any detected ECG abnormalities may ability to recognize patients with CKD. Although the prevalence
be especially meaningful and specific. Although older age is a of CKD was low in our training cohort with ECGs, and this
well-known risk marker for CKD, the prevalence of CKD in prevalence not directly comparable to epidemiological cohorts
younger patients is also notably high in the U.S. (8–10% in <65 with CKD (as ECGs are more commonly obtained in patients
years)3. Remarkably, however, awareness of underlying CKD is without CKD), we show that our disease definitions is consistent
also very low in younger patients, as only about 8% are aware of with laboratory testing and documented eGFR (Supplementary
the disease3. Given the availability of effective low-risk CKD Table 8) and that our deep learning approach is relatively
treatments and the reversibility of CKD, there are substantial insensitive in model accuracy to disease prevalence in the training
potential benefits for detecting and treating CKD, especially in the set (Supplementary Table 4). The prevalence of hypertension
young. A recent paper by Kwon et al.27 also used data from ECG diagnosis may be underestimated in the internal cohort, however
waveforms in addition to age and sex to develop a DLA to detect our model performed similarly well in internal and external test
changes in eGFR, which can include both patients with acute cohorts with different prevalences of hypertension.
kidney injury (e.g., dehydration, pharmacotherapy, urinary tract By 2030, the UN’s Sustainable Development Goals are to
obstruction) as well as chronic kidney disease. Their model reduce premature mortality related to non-communicable dis-
achieved a slightly higher performance with an AUC of 0.86–0.91, eases by a third. Given the high prevalence of asymptomatic
however reaffirms the overall conclusion that renal abnormalities CKD, serious consequences of untreated disease, presence of

6 COMMUNICATIONS MEDICINE | (2023)3:73 | https://doi.org/10.1038/s43856-023-00278-w | www.nature.com/commsmed

COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00278-w ARTICLE

effective low-risk treatment, and detectable preclinical state with 11. Kwon, J. M. et al. Comparing the performance of artificial intelligence and
inexpensive and simple diagnostic tests, CKD represents a good conventional diagnosis criteria for detecting left ventricular hypertrophy using
target for large-scale population screening and harbors the electrocardiography. Europace 22, 412–419 (2020).
12. Attia, Z. I. et al. Age and sex estimation using artificial intelligence
potential for reducing premature mortality related to non- from standard 12-Lead ECGs. Circ. Arrhythm Electrophysiol. 12, e007284
communicable diseases. In addition to the high mortality and (2019).
morbidity due to CKD, treatment costs for CKD are also high and 13. Kwon, J. M. et al. A deep learning algorithm to detect anaemia with
have increased during the last decades32. Especially, the increas- ECGs: a retrospective, multicentre study. Lancet Digit Health 2, e358–e367
ing number of patients requiring renal replacement creates (2020).
14. Attia, Z. I. et al. An artificial intelligence-enabled ECG algorithm for the
challenges for health care systems worldwide, and the shortage of identification of patients with atrial fibrillation during sinus rhythm: a
sufficient replacement services may cause at least 2 million pre- retrospective analysis of outcome prediction. Lancet. 394, 861–867 (2019).
mature deaths annually33. Therefore, widely available, inexpen- 15. Hughes, J. W. et al. Performance of a convolutional neural network and
sive, and effective CKD prevention and management strategies explainability technique for 12-lead electrocardiogram interpretation. JAMA
are warranted to enable equal opportunities in reducing CKD- Cardiol. 6, 1285–1295 (2021).
16. Jankowski, J., Floege, J., Fliser, D., Bohm, M. & Marx, N. Cardiovascular
related disability-adjusted life years. disease in chronic kidney disease: pathophysiological insights and therapeutic
options. Circulation. 143, 1157–1172 (2021).
Conclusions 17. Hayer, M. K. et al. Defining myocardial abnormalities across the stages of
Our ECG-based deep learning model was able to detect CKD with chronic kidney disease: a cardiac magnetic resonance imaging study. JACC
Cardiovasc. Imaging 13, 2357–2367 (2020).
good discrimination accuracy in multiple study populations and 18. Diercks, D. B., Shumaik, G. M., Harrigan, R. A., Brady, W. J. & Chan, T. C.
with particularly high accuracy in patients under 60 years of age. Electrocardiographic manifestations: electrolyte abnormalities. J. Emerg. Med.
These results suggest that deep learning-based ECG analysis may 27, 153–160 (2004).
provide additional value in detecting various CKD stages, espe- 19. Galloway, C. D. et al. Development and validation of a deep-learning model to
cially in younger patients. The clinical significance of this study screen for hyperkalemia from the electrocardiogram. JAMA Cardiol. 4,
428–436 (2019).
lies in the potential enhancement of screening methods for the 20. Urtnasan, E. et al. Noninvasive screening tool for hyperkalemia using a single-
early detection of CKD, which is crucial to enable early treatment lead electrocardiogram and deep learning: development and usability study.
and prevent disease progression. JMIR Med. Inform. 10, e34724 (2022).
21. Ribeiro MT, Singh S & Guestrin C. “Why Should I Trust You?”: Explaining
the Predictions of Any Classifier. Paper presented at: Proceedings of the 22nd
ACM SIGKDD International Conference on Knowledge Discovery and Data
Data availability Mining; 2016; San Francisco, California, USA.
All analytical methods applied for the deep learning algorithm are included in this 22. Berns, J. S. Routine screening for CKD should be done in asymptomatic
published article, supplementary files. The patient data is not publicly available due to adults… selectively. Clin. J. Am. Soc. Nephrol. 9, 1988–1992 (2014).
potentially identifiable nature of the associated data. De-identified data is available from 23. Komenda, P. et al. Cost-effectiveness of primary screening for CKD: a
the corresponding author on reasonable request. systematic review. Am. J. Kidney Dis. 63, 789–797 (2014).
24. Moyer, V. A., Force USPST. Screening for chronic kidney disease: U.S.
Preventive Services Task Force recommendation statement. Ann. Intern. Med
157, 567–570 (2012).
Code availability 25. Qaseem, A., Hopkins, R. H. Jr, Sweet, D. E., Starkey, M. & Shekelle, P. and
Code is available at https://github.com/ecg-net/CKDscreening34.
Clinical Guidelines Committee of the American College of P. Screening,
monitoring, and treatment of stage 1 to 3 chronic kidney disease: a clinical
Received: 22 August 2022; Accepted: 10 March 2023; practice guideline from the American College of Physicians. Ann. Intern. Med.
159, 835–47 (2013).
26. Alfego, D. et al. Chronic kidney disease testing among at-risk adults in the
U.S. remains low: real-world evidence from a national laboratory database.
Diabetes Care 44, 2025–2032 (2021).
27. Kwon, J. M. et al. Artificial intelligence assessment for early detection and
References prediction of renal impairment using electrocardiography. Int. Urol. Nephrol.
1. Collaboration GBDCKD. Global, regional, and national burden of chronic 54, 2733–2744 (2022).
kidney disease, 1990−2017: a systematic analysis for the Global Burden of 28. Huang, J. C. et al. P wave dispersion and maximum P wave duration are
Disease Study 2017. Lancet. 395, 709–733 (2020). associated with renal outcomes in chronic kidney disease. PLoS ONE 9,
2. Chu, C. D. et al. Centers for disease C and prevention chronic kidney disease e101962 (2014).
surveillance T. CKD awareness among US adults by future risk of kidney 29. Kestenbaum, B. et al. Kidney function, electrocardiographic findings, and
failure. Am. J. Kidney Dis. 76, 174–183 (2020). cardiovascular events among older adults. Clin. J. Am. Soc. Nephrol. 2,
3. Dharmarajan, S. H. et al. Centers for disease C and prevention CKDSS. state- 501–508 (2007).
level awareness of chronic kidney disease in the U.S. Am. J. Prev. Med. 53, 30. Sherif, K. A., Abo-Salem, E., Panikkath, R., Nusrat, M. & Tuncel, M. Cardiac
300–307 (2017). repolarization abnormalities among patients with various stages of chronic
4. Chronic Kidney Disease Prognosis, C. et al. Association of estimated kidney disease. Clin. Cardiol. 37, 417–421 (2014).
glomerular filtration rate and albuminuria with all-cause and cardiovascular 31. Stewart, G. A. et al. Electrocardiographic abnormalities and uremic
mortality in general population cohorts: a collaborative meta-analysis. Lancet. cardiomyopathy. Kidney Int. 67, 217–226 (2005).
375, 2073–2081 (2010). 32. Himmelfarb, J. & Ikizler, T. A. Hemodialysis. New Engl. J. Med. 363,
5. Gerstein, H. C. et al. and Investigators HS. Albuminuria and risk of 1833–1845 (2010).
cardiovascular events, death, and heart failure in diabetic and nondiabetic 33. Liyanage, T. et al. Worldwide access to treatment for end-stage kidney disease:
individuals. JAMA. 286, 421–426 (2001). a systematic review. Lancet. 385, 1975–1982 (2015).
6. Bang, H. et al. SCreening for Occult REnal Disease (SCORED): a simple 34. theurerjohn & Christensen, K. ecg-net/CKDscreening: COMMSMED
prediction model for chronic kidney disease. Arch Intern Med 167, 374–381 Publication Release. https://doi.org/10.5281/zenodo.7713558.
(2007).
7. McKinney, S. M. et al. International evaluation of an AI system for breast
cancer screening. Nature. 577, 89–94 (2020).
8. Ouyang, D. et al. Video-based AI for beat-to-beat assessment of cardiac Acknowledgements
function. Nature. 580, 252–256 (2020). This work was supported in part by the National Institutes of Health NIH K99
9. Raghunath, S. et al. Prediction of mortality from 12-lead electrocardiogram HL157421-01. L.H. was supported by Sigrid Juselius Foundation, The Finnish Cultural
voltage data using a deep neural network. Nat. Med. 26, 886–891 (2020). Foundation, Instrumentarium Science Foundation, Orion Research Foundation, and
10. Attia, Z. I. et al. Screening for cardiac contractile dysfunction using an artificial Paavo Nurmi Foundation. The funding sources had no involvement in the preparation of
intelligence-enabled electrocardiogram. Nat. Med. 25, 70–74 (2019). this work or the decision to submit for publication.

COMMUNICATIONS MEDICINE | (2023)3:73 | https://doi.org/10.1038/s43856-023-00278-w | www.nature.com/commsmed 7

ARTICLE COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00278-w

Author contributions Reprints and permission information is available at http://www.nature.com/reprints

Study design and conception: L.H., M.C., S.S.C., and D.O. Acquisition, analysis or
interpretation of data: L.H., M.C., N.Y., J.W.H., J.T., M.J., P.F., A.K., R.K.S., J.E., S.C., J.Z., Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
S.S.C., and D.O. Drafting of the manuscript: L.H., M.C., and D.O. Statistical analysis: published maps and institutional affiliations.
M.C., J.W.H., and D.O. Critical revision of the manuscript for important intellectual
content: N.Y., J.W.H., J.T., M.J., P.F., A.K., R.K.S., J.E., S.C., J.Z., and S.S.C. Adminis-
trative and material support: S.S.C. and D.O. Obtained funding: D.O. Supervision: S.S.C. Open Access This article is licensed under a Creative Commons
and D.O. Full access to the data: M.C. and D.O. All authors reviewed and approved the Attribution 4.0 International License, which permits use, sharing,
final version of the manuscript. adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative
Competing interests Commons license, and indicate if changes were made. The images or other third party
The authors declare no competing interests. material in this article are included in the article’s Creative Commons license, unless
indicated otherwise in a credit line to the material. If material is not included in the
article’s Creative Commons license and your intended use is not permitted by statutory
Additional information regulation or exceeds the permitted use, you will need to obtain permission directly from
Supplementary information The online version contains supplementary material the copyright holder. To view a copy of this license, visit http://creativecommons.org/
available at https://doi.org/10.1038/s43856-023-00278-w. licenses/by/4.0/.

Correspondence and requests for materials should be addressed to David Ouyang.

© The Author(s) 2023
Peer review information Communications Medicine thanks the anonymous reviewers
for their contribution to the peer review of this work. Peer reviewer reports are available.

8 COMMUNICATIONS MEDICINE | (2023)3:73 | https://doi.org/10.1038/s43856-023-00278-w | www.nature.com/commsmed

 © The Author(s) 2023. This work is published under
http://creativecommons.org/licenses/by/4.0/(the “License”). Notwithstanding
the ProQuest Terms and Conditions, you may use this content in accordance
with the terms of the License.