ARTICLE

https://doi.org/10.1038/s43856-023-00241-9 OPEN

Cardiothoracic ratio values and trajectories are

associated with risk of requiring dialysis and
mortality in chronic kidney disease
Che-Yi Chou1,2,3, Charles C. N. Wang4, Hsiu-Yin Chiang5 ✉, Chien-Fong Huang5, Ya-Luan Hsiao6,
Chuan-Hu Sun5, Chun-Sheng Hu5, Min-Yen Wu 5, Sheng-Hsuan Chen5, Chun-Min Chang7, Yu-Ting Lin5,
Jie-Sian Wang3, Yu-Cuyan Hong3, I-Wen Ting3,8, Hung-Chieh Yeh3,8 & Chin-Chi Kuo3,4,5,8 ✉

Abstract
1234567890():,;

Plain language summary

Background The prognostic role of the cardiothoracic ratio (CTR) in chronic kidney disease An enlarged heart occurs during
(CKD) remains undetermined. various medical conditions and can
result in early death. However, it is
Methods We conducted a retrospective cohort study of 3117 patients with CKD aged
unclear whether this is also the case
18–89 years who participated in an Advanced CKD Care Program in Taiwan between 2003
in patients with chronic kidney dis-
and 2017 with a median follow up of 1.3(0.7–2.5) and 3.3(1.8–5.3) (IQR) years for outcome of ease (CKD). Although the size of the
end-stage renal disease (ESRD) and overall death, respectively. We developed a machine heart can be measured on chest X-
learning (ML)–based algorithm to calculate the baseline and serial CTRs, which were then rays, this process is time consuming.
used to classify patients into trajectory groups based on latent class mixed modelling. We used artificial intelligence to
quantify the heart size of 3117 CKD
Association and discrimination were evaluated using multivariable Cox proportional hazards
patients based on their chest X-rays
regression analyses and C-statistics, respectively.
within hours. We found that CKD
Results The median (interquartile range) age of 3117 patients is 69.5 (59.2–77.4) years. We patients with an enlarged heart were
create 3 CTR trajectory groups (low [30.1%], medium [48.1%], and high [21.8%]) for the more likely to develop end-stage
2474 patients with at least 2 CTR measurements. The adjusted hazard ratios for ESRD, kidney disease or die. This could
cardiovascular mortality, and all-cause mortality in patients with baseline CTRs ≥0.57 improve monitoring of CKD patients
with an enlarged heart and improve
(vs CTRs <0.47) are 1.35 (95% confidence interval, 1.06–1.72), 2.89 (1.78–4.71), and 1.50
their care.
(1.22–1.83), respectively. Similarly, greater effect sizes, particularly for cardiovascular mor-
tality, are observed for high (vs low) CTR trajectories. Compared with a reference model, one
with CTR as a continuous variable yields significantly higher C-statistics of 0.719 (vs 0.698,
P = 0.04) for cardiovascular mortality and 0.697 (vs 0.693, P < 0.001) for all-cause
mortality.
Conclusions Our findings support the real-world prognostic value of the CTR, as calculated
by a ML annotation tool, in CKD. Our research presents a methodological foundation for
using machine learning to improve cardioprotection among patients with CKD.

1 Division of Nephrology, Department of Internal Medicine, Asia University Hospital, Wufeng, Taichung, Taiwan. 2 Department of Post-baccalaureate Veterinary

Medicine, Asia University, Wufeng, Taichung, Taiwan. 3 Division of Nephrology, Department of Internal Medicine, China Medical University Hospital and College
of Medicine, China Medical University, Taichung, Taiwan. 4 Department of Bioinformatics and Medical Engineering, Asia University, Taichung, Taiwan. 5 Big Data
Center, China Medical University Hospital and College of Medicine, China Medical University, Taichung, Taiwan. 6 Department of Health Administration, Johns
Hopkins Bloomberg School of Public Health, Johns Hopkins University, Baltimore, MD, USA. 7 Department of Electrical and Computer Engineering, University of
Wisconsin-Madison, Madison, WI, USA. 8 AKI-CARE (Clinical Advancement, Research and Education) Center, Department of Internal Medicine, China Medical
University Hospital and College of Medicine, China Medical University, Taichung, Taiwan. ✉email: [email protected]; [email protected]

COMMUNICATIONS MEDICINE | (2023)3:19 | https://doi.org/10.1038/s43856-023-00241-9 | www.nature.com/commsmed 1

ARTICLE COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00241-9

T
he cardiothoracic ratio (CTR) was first defined in 1919 and CKD stages 3b to 5. China Medical University Hospital (CMUH),
has since been accepted as a means of quantifying heart a tertiary medical center in central Taiwan, joined this care
size and volume on posterior-anterior chest X-rays (PA- project in 2003 and prospectively enrolled consecutive patients
CXRs)1,2. A CTR of <0.5 is considered normal, whereas a CTR >0.5 with CKD who agreed to participate in this Advanced CKD Care
indicates cardiomegaly. The CTR is a key prognostic indicator in Program. The details of the program were described in a previous
many conditions, such as coronary artery disease3, end-stage renal study23. Biochemical markers of renal injury, including serum
disease (ESRD)4,5, and conditions related to aging6. A higher CTR creatinine (S-Cre), blood urea nitrogen, and spot uPCR, were
value is generally associated with poor prognosis, and 2 longitudinal measured at least once every 3 months. In the present study, we
studies of healthy populations have reported significant associations further integrated the data from the Advanced CKD Care Pro-
between high CTR values and cardiovascular (CV) mortality7,8. gram with data from the CMUH Clinical Research Data Repo-
Several studies have challenged the current CTR threshold; their sitory, including data on laboratory tests, medications, special
results have revealed CTRs below 0.5 to be associated with increased procedures, and admission records24–26. The index date was the
CV mortality in the general population7,9. Zaman et al. observed date of patient enrollment in the Advanced CKD Care Program.
that a preangiographic CTR above 0.42 to be associated with All enrolled patients were followed up until the initiation of long-
increased all-cause mortality3. term renal replacement therapy (hemodialysis, peritoneal dialysis,
Studies have demonstrated that the CTR is a significant or transplantation), death, loss to follow-up, or December 31st,
indicator of mortality and deteriorating functional status for 2019, whichever occurred first. Dates of death were verified at the
patients with incident dialysis10,11 and ESRD requiring regular National Death Registry of the Ministry of Health and Welfare of
hemodialysis4,12–14 or peritoneal dialysis5. In addition, the CTR Taiwan25,26.
has been used to monitor fluid overload, malnutrition, and heart We included the data of patients aged 18–89 years participating
disease in patients receiving dialysis4,10,12. However, the inde- in the Advanced CKD Care Program from January 1st, 2003, to
pendent value of the CTR for predicting disease progression and December 31st, 2017, who had no history of dialysis. For
mortality among patients with nondialysis chronic kidney disease population of baseline CTR analysis, we included patients who
(CKD-ND) is unclear. A recent study reported a CTR > 0.5 and had at least one CTR measurement during the baseline period
the presence of aortic arch calcification are associated with rapid (−1 year to +6 months of the index date) and one CTR during
renal progression as well as CV and all-cause mortality in CKD- the follow-up period (+6 months of the index date to December
ND15. However, the researchers could not assess the risk of 31st, 2017), with the two CTR measurements were at least
progression to dialysis or determine an optimal risk cutoff for the 6 months apart. Baseline CTR was the CTR in the baseline period
CTR because the few patients enrolled in the study and the short and closest to the index date. In total, 3117 patients with 6234
follow-up. To date, much of the evidence on the role of the CTR CTR records were included for analysis (Fig. 1). All CTR records
in ESRD originates from studies conducted in Taiwan because were derived from PA-CXRs, which were taken with the patients
semiannual CTR measurement is required by the Taiwan Society in upright position. The average number of days between the
of Nephrology for evaluation of dialysis adequacy. Nevertheless, dates of baseline CTR and serum creatinine (S-Cre) was 83.0
the CTR is not considered in the routine care of the Advanced (IQR 46.0–131.0). The frequency of repeated CTR measurements
CKD Care Program in Taiwan or similar programs in other for the longitudinal arm was shown in the Supplementary
countries; this may partially explain the limited availability of Table 1. For population of CTR trajectory analysis, we further
related data in CKD-ND. Furthermore, despite ongoing efforts to excluded patients who did not have at least two CTR
investigate the prognostic role of the CTR in various disease measurements within the 2 years after the index date and a total
states, the validity and generalizability of these studies are of 2474 participants with 12391 CTR records were included in the
uncertain because sample sizes are often small; CTR measure- trajectory analysis (Fig. 1). This study was approved by the Big
ment requires labor-intensive annotation by physicians. Although Data Center of CMUH and the Research Ethics Committee and
several machine learning (ML) algorithms have been developed to Institutional Review Board (REC/IRB) of CMUH, and informed
simplify and standardize the annotation process, adoption of consent was waived due to retrospective nature of the study
these algorithms has been slow because the algorithms require (CMUH105-REC3-068 & CMUH108-REC2-022). All methods
further validation before they can be fully integrated into were performed in accordance with the relevant guidelines and
the clinical workflow16–20. To address this research gap and regulations of the REC/IRB.
evaluate the risks of requiring dialysis and mortality associated
with the various baseline CTRs and CTR trajectories in patients
with CKD-ND, we developed an artificially intelligent CTR Automatic determination of CTR using deep learning. We used
(iCTR) assessment system to annotate PA-CXRs. We further deep learning to perform automatic cardiothoracic cavity seg-
evaluated whether the incorporation of the CTR into the work- mentation in PA-CXR images by using a U-Net model27. The
flow could improve the predictive performance of the Kidney U-Net architecture is illustrated in Supplementary Fig. 1 27. The
Failure Risk Equation (KFRE) and a conventional mortality risk model consists of a contracting path (left side) and an expanding
model for CKD21,22. Our findings identify the positive associa- path (right side). The contracting path follows the typical archi-
tions between CTR and accelerated progression to ESRD or tecture of a convolutional network, consisting of repeated appli-
mortality in CKD and demonstrate the real-world prognostic cation of two 3 × 3 convolutions (unpadded convolutions), each
value of CTR in predicting mortality. of which is followed by a rectified linear unit (ReLU)28 function
and a 2 × 2 max-pooling operation with a stride of 2 for down-
sampling. In each downsampling step, the number of feature
Methods channels is doubled. Each step in the expanding path consists of
Study patients and design. Taiwan’s National Health Insurance upsampling of the feature map followed by a 2 × 2 convolution
launched the Project of Integrated Care of CKD in 2002. The (“up-convolution”) that halves the number of feature channels,
project initially targeted patients with an estimated glomerular concatenation with a correspondingly cropped feature map from
filtration rate (eGFR) of <60 mL/min/1.73 m2 or with proteinuria the contracting path, and two 3 × 3 convolutions, each of which is
(urine protein-to-creatinine ratio [uPCR] > 1 [g/g]). In 2007, the followed by a ReLU function. The cropping is necessary because
project adopted a multidisciplinary approach and focused on of the loss of border pixels with each convolution. In the final

2 COMMUNICATIONS MEDICINE | (2023)3:19 | https://doi.org/10.1038/s43856-023-00241-9 | www.nature.com/commsmed

COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00241-9 ARTICLE

Fig. 1 Flow diagram for included patients. Patients and chest X-ray images are identified through deep data cleaning process to form the final study
population.

layer, a 1 × 1 convolution is used to map each 32-component CMUH central laboratory by using a Beckman UniCel DxC 800
feature vector to the desired number of classes. In total, the immunoassay system (Beckman Coulter, Brea, CA, USA).
network has 23 convolutional layers. To evaluate the performance eGFRs were calculated using the Chronic Kidney Disease Epi-
of the model, we evaluated CTR agreement as an absolute dif- demiology Collaboration creatinine equation29. The S-Cre level
ference (AD) of <0.03 between the ML-estimated CTR and one at enrollment was used to determine the baseline eGFR and
calculated through physician annotation. From the testing data- corresponding CKD stage according to the following cutoff
base, 510 of 537 PA-CXRs had a CTR AD < 0.03, for an accuracy values: eGFRs of >90, 60 to 89.9, 45 to 59.9, 30 to 44.9, 15 to
of 94.97%. The iCTR Assessment System that incorporates the 29.9, and <15 mL/min/1.73 m2 were respectively considered
CTR estimation algorithm has received premarket approval stages 1, 2, 3a, 3b, 4, and 5. Proteinuria was defined as a uPCR
for software as a medical device (SaMD) from the US Food of >0.5 [g/g cre]. For patients with only a urine albumin-to-
and Drug Administration (FDA; 510(k) number, K212624) and creatinine ratio (uACR) available, we converted the uACR into
Taiwan FDA (TFDA; medical device license number, 007443). uPCR by using the following equation derived from a Japanese
The ML-estimated CTR of actual cases were illustrated in Sup- study: lnðACRÞ ¼ 1:32  lnðPCRÞ  2:6430.
plementary Fig. 2.
Statistical analyses. Continuous variables were compared using
Determination of kidney function. S-Cre levels were measured the Wilcoxon rank sum test and are expressed as medians with
using the Jaffe rate method (kinetic alkaline picrate) at the IQRs. Categorical variables were compared using the chi-square

COMMUNICATIONS MEDICINE | (2023)3:19 | https://doi.org/10.1038/s43856-023-00241-9 | www.nature.com/commsmed 3

ARTICLE COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00241-9

test and are expressed as frequencies with percentages. The (Supplementary Data 2). By contrast, the proportions of patients
baseline CTR was categorized into quartiles. To characterize the who were active smokers, were alcohol drinkers, and had total
trajectories of all CTR measurements among patients in the study education of ≥16 years decreased with CTR quartile. Compared
population for the trajectory analysis, Latent Class Mixed Mod- with patients with a CTR of <0.47, patients with a CTR of ≥0.57
elling (LCMM) was used to fit a semiparametric mixture model to were more likely to use antiplatelet agents and had higher base-
the longitudinal data using the maximum likelihood method31. line systolic blood pressure (median: 142.0 [123.6–158.0] vs 132.5
The LCMM package (version 1.9.3) in R was applied32. This [119.0–148.0] mmHg) and lower baseline diastolic blood pressure
approach is appropriate when the number of subgroups or other (median: 73.8 [65.7–81.2] vs 76.3 [70.0–86.0] mmHg). Significant
information, such as the trajectory shape of each subgroup, is decreasing trends in baseline eGFR, hemoglobin, and urinary
unknown. We empirically compared 2-, 3- and 4-group solutions creatinine levels were concomitant with increasing trends in
and optimized the number of subgroups by using Bayesian proteinuria and the baseline serum levels of phosphorus, blood
information criterion values (with a number close to 0 indicating urea nitrogen, and uric acid (Supplementary Data 1). In the
a good fit); the trajectory shapes were determined by the order of sensitivity analysis of matching CTR values with available echo-
the polynomial (e.g., linear, quadratic, or cubic). The CTR tra- cardiographic parameters such as ejection fraction, left ven-
jectories were determined before the risk analysis for dialysis and tricular mass index (LVMI), and left atrial volume index (LAVI),
mortality. the increasing trends of LVMI, and LAVI were concordant with
For time-to-event analysis, the number of person-years free from the increasing baseline CTR groups and CTR trajectory group
dialysis and mortality after enrollment in the Advanced CKD (Supplementary Table 2). Detailed information about the CMUH
Program was computed. We evaluated the prospective associations echocardiographic cohort has been published previously24,34.
of both the baseline CTR and longitudinal CTR trajectories with
dialysis initiation and mortality by using multiple Cox proportional
Clinical characteristics by CTR trajectory. Among 2474 patients
hazards models. The models were adjusted based on a priori
with at least 2 CTR measurements, the median (IQR) number of
knowledge to control potential confounding variables. To char-
CTR measurements was 4.0 (2.0–6.0). We identified 3 CTR tra-
acterize the dialysis risk associated with the exposures of interest,
jectories through LCMM and subdivided patients into low-
we performed a competing risk analysis using cause-specific Cox
(30.1%), medium- (48.1%), and high-(21.8%) trajectory groups
proportional hazards modeling, with deaths considered censoring
(Supplementary Fig. 3). The clinicodemographic and laboratory
events to minimize the potential bias introduced by a competing
trends from the low to high trajectory were similar to those
death risk. Model fit was evaluated using the Akaike information
observed across the baseline CTR quartiles (Supplementary
criterion. Because data were missing for some explanatory variables
Data 3). The coefficient of variation in serial CTR measurements
(eg, 15.4% missing for pooled uPCR; Supplementary Data 1), we
was the greatest among patients with medium CTR trajectories.
further performed multiple imputation by using a fully conditional
When the change in the variation of the CTR was defined using
method in SAS, namely, an iterative Markov chain Monte Carlo
the absolute difference (AD), a significant increasing trend in AD
procedure, for the missing variables. We set the number of
was observed with CTR trajectory from low to high. However,
imputations to 20 and the number of iterations to 100. To explore
other variation indicators, including slope and percentage change,
the potential effect modifiers in the association between both the
were not significantly different among the CTR trajectory groups
baseline and longitudinal CTRs and the main outcomes, we
(Supplementary Data 3).
stratified patients based on age (older or younger than 65 years),
sex, hypertension, diabetes, CV disease (CVD) status, CKD stage
(3 vs 4 or 5), and uPCR (> vs ≤0.5 [g/g cre]) at baseline. To evaluate Associations of baseline CTR and longitudinal CTR trajectories
the predictive value of the CTR, we compared the Uno’s C-statistics with ESRD and mortality. In the fully adjusted model, each 10%
from the reference KFRE, which considers age, sex, eGFR, and increase in baseline CTR was significantly associated with a higher
uPCR (log-transformed), with that from the same model with the hazard ratio (HR) for progression to ESRD (adjusted HR [aHR]
addition of the CTR as a categorical or continuous variable33. 1.15, 95% CI, 1.03–1.28), CV mortality (1.83, 95% CI, 1.49–2.24),
The reference mortality predictive model (RMPM) considered and all-cause mortality (1.27, 95% CI, 1.16–1.40), respectively
age, sex, eGFR, diabetes, hypertension, and anemia (hemoglobin < (Supplementary Data 4). The aHRs for progression to ESRD, CV
12 g/dL)21. We also plotted the observed versus the predicted risk mortality, and all-cause mortality in patients with a baseline CTR of
to reveal the differences in the calibration of the three prediction ≥0.57 were 1.35 (95% CI, 1.06–1.72), 2.89 (95% CI, 1.78–4.71), and
models. All statistical analyses were performed using SAS version 1.50 (95% CI, 1.22–1.83), respectively, relative to patients with a
9.4 (SAS Institute, Cary, NC, USA) and R version 3.5.1 (R CTR of <0.47. Furthermore, compared with the low-trajectory
Foundation for Statistical Computing, Vienna, Austria). Statistical group, the aHRs for progression to ESRD, CV mortality, and all-
significance was defined as a 2-tailed P-value of <0.05. cause mortality were 1.70 (95% CI, 1.28–2.25), 3.84 (95% CI,
2.28–6.45), and 1.78 (95% CI, 1.44–2.22), respectively, in the high-
trajectory group (Supplementary Data 4). The dose-response
Reporting summary. Further information on research design is
relationship between ESRD and baseline CTR was curvilinear up
available in the Nature Portfolio Reporting Summary linked to
to a CTR of 0.6, at which it plateaued (Fig. 2a). By contrast, a linear
this article.
nonthreshold dose-response relationship of baseline CTR with CV
and all-cause mortality was observed beyond the ratios of 0.47 and
Results 0.50, respectively (Fig. 2b and c). In subgroup analysis for baseline
Clinical characteristics by baseline CTR quartile. The median CTR, patients without hypertension and advanced CKD stage 4–5
(IQR) age at enrollment for all 3117 patients was 69.5 were more vulnerable to increased baseline CTR for the outcome of
(59.2–77.4) years; the median duration of follow-up was progression to ESRD (Supplementary Data 5). By contrast, the
1.3 (0.7–2.5) and 3.3 (1.8–5.3) (IQR) years for ESRD events and associations between baseline CTR and CV- or all-cause mortality
death, respectively, as events usually occurred before death. were not modified by age, sex, hypertension, diabetes, CVD, CKD
Increasing trends in age, body mass index, and the proportions of stage, and proteinuria (Supplementary Data 5). In subgroup ana-
female sex and of DM, hypertension, CVD, or stage 4 or 5 CKD lysis for CTR trajectory, female and hypertension-free patients were
at baseline were observed across the increasing quartiles of CTR more susceptible to high CTR trajectory regarding the dialysis

4 COMMUNICATIONS MEDICINE | (2023)3:19 | https://doi.org/10.1038/s43856-023-00241-9 | www.nature.com/commsmed

COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00241-9 ARTICLE

Fig. 2 Dose-response relationship between baseline CTR and risk of developing outcomes of interest based on the study population of 3117 patients.
Adjusted HRs for (a) progression to ESRD, (b) CV mortality, and (c) all-cause mortality based on baseline CTR. Solid black lines represent adjusted HRs
based on restricted cubic splines for baseline CTR with knots at the 10th, 50th, and 90th percentiles. Dashed black lines represent 95% CI. The reference is
set at the 10th percentile of baseline CTR.

outcome (Supplementary Data 6). Patients without hypertension optimal cutoff based on large study samples with or without kidney
were also more vulnerable to CV mortality (Supplementary disease should be an urgent research priority.
Data 6). Left ventricular dilation is one of the most common causes of a
CTR > 0.5 and is an independent risk factor for CVD and heart
Performance of baseline CTR in predicting dialysis and mor- failure–related mortality38,39. A pooled analysis of 466 patients
tality. Adding the baseline CTR to the conventional 4-variable demonstrated moderate sensitivity and specificity of 83.3% and
KFRE for predicting ESRD did not significantly improve the 45.4%, respectively, for predicting left ventricular dilation with
predictive performance of the KFRE (C-statistics of 0.842 to 0.843 the CTR35. However, among 272 patients that were hospitalized,
for baseline CTR categories, P = 0.60 with similar calibration) the area under the receiver operating characteristic curve for
(Table 1 and Supplementary Fig. 4). However, when a reference moderate and severe left ventricular or right ventricular dys-
model comprising age, sex, eGFR, diabetes, hypertension, and function based on CTRs was only 0.70, with the best cut-off of
anemia status was used for predicting mortality outcomes, CTR 0.55 having the highest Youden index40. Similarly, among
incorporating baseline CTR as a continuous variable significantly patients receiving hemodialysis, significant associations have been
increased the C-statistics from 0.698 to 0.719 (P = 0.04) for CV consistently observed between a high CTR and all-cause
mortality (Table 1) and from 0.693 to 0.697 (P < 0.001) for all- mortality13,41,42; nevertheless, another study discovered that the
cause mortality (Table 1). Integrating baseline CTR also improved CTR did not improve the predictive value for mortality41. The
the calibration performance for the outcome of CV mortality moderate diagnostic accuracy and prognostic value of the CTR
(Supplementary Fig. 4). have limited its clinical utility in conventional practice.
In our study, ML streamlined CTR calculation such that we
Discussion could evaluate >70,000 PA-CXRs within 24 h. As far as we are
Our findings support the independent association of the baseline aware, this is the first report of linear dose-response associations
CTR and the first 2-year longitudinal CTR trajectory in patients of the CTR with CV and all-cause mortality among patients with
with CKD-ND with the risk of progression to ESRD and the CKD-ND. We discovered that this strong association significantly
mortality outcomes. The critical cutoff (0.44 for dialysis, 0.49 for improved the predictive performance of the conventional models
CV mortality, and 0.45 for all-cause mortality), defined by the for mortality risk assessment in CKD21. However, the CTR only
cross-point of the lines of HR = 1 and the lower bound of 95% CI marginally improved the predictive power of the KFRE among
in the dose-response plots (Fig. 2), was lower than the widely patients with CKD. Indeed, the effect size of HR does not lend
accepted value of 0.54,10,12,15,35. We observed a gradient increase itself to drawing direct conclusions regarding the contribution of
in the risk of requiring dialysis, CV mortality, and all-cause the CTR to prognosis prediction43. Varga et al. has systematically
mortality for CKD patients in the second to the fourth quartile of reviewed this issue recently44. The reason underlying the dis-
baseline CTR compared to those with a baseline CTR < 0.47. In crepancy of the significantly positive association but poor pre-
addition, the CV mortality log-linearly increased with baseline dictive capability of the CTR with all outcomes of interest may be
CTR among patients with advanced CKD-ND. The consistent explained by the follows: (1) the predictive variables of KFRE or
risk patterns for longitudinal CTR trajectories strengthen the RMPM have predicted the outcomes of ESRD and mortality well
causal inference. Despite baseline CTR significantly enhances the and the predictive role of CTR is therefore not standing out; (2)
prognostic prediction for mortality outcomes, the predictive there were large overlapping distributions of CTR when strati-
performance for progression to ESRD is not improved compared fying the population based on the outcomes of interest (Supple-
with the conventional KFRE. mentary Fig. 5); (3) CTR is a summarized and relatively low-
Because a lack of automation, the statistical power of related resolution digital cardiac marker measuring a simple ratio
studies has been insufficient to evaluate the prognostic value of the between cardiac and thoracic horizontal diameter on the PA-
CTR. Use of the CTR has therefore heavily depended on individual CXR, rather than providing detailed information about anato-
expertise, and the cutoff of 0.5 has become a convention, leaving mical abnormalities such as left ventricular mass index or left
the CTR’s potential for serving as a rapid, low-cost, and repro- atrial volume index and functional impairments such as systolic
ducible digital marker beyond heart disease unexplored36. or diastolic dysfunction. Therefore, it may be insufficient to
Although a CTR value of 0.42–0.50 is considered within “normal provide additional prognostic benefit on top of the original KFRE
range”, our findings showed the critical threshold may be lower or RMPM models. Nevertheless, our findings encourage future
than the conventional cutoff of 0.5, which is consistent with some studies to identify ML-driven cardiac digital markers that are
previous research3,37. Verifying our study results and finding an more disease prognosis specific to the development of ESRD and

COMMUNICATIONS MEDICINE | (2023)3:19 | https://doi.org/10.1038/s43856-023-00241-9 | www.nature.com/commsmed 5

ARTICLE COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00241-9

Table 1 C-statistics from KFRE and Reference Mortality Predictive Model (RMPM) with continuous or categorical CTR variable for predicting ESRD, CV mortality, and all-cause mortality, particularly using common clinical imaging modalities
to infer important anatomical and functional features.

0.983 (0.980, 0.987)

0.695 (0.674, 0.716)

Abbreviations: AIC Akaike’s information criterion, BIC Bayesian information criterion, CI confidence interval, CTR cardiothoracic ratio, eGFR estimated glomerular filtration rate, HR hazard ratio, KFRE Kidney Failure Risk Estimation, uPCR urine protein-to-creatinine ratio.
1.040 (1.034, 1.046)
1.566 (1.289, 1.903)

1.276 (1.128, 1.444)

1.171 (0.962, 1.425)
1.236 (1.013, 1.507)

1.416 (1.197, 1.674)

1.431 (1.263, 1.621)

1.552 (1.311, 1.838)

Our results suggest that CTR can be used to personalize CKD

Categorical CTR
HR (95% CI)
care; this suggestion is also supported by the observations of a

16423.000
16468.420
RMPM + large epidemiological study in which ~50% of patients with CKD-
ND experienced CV mortality45. Although cardiorenal syndrome

0.22
Ref
–
may explain this phenomenon46, the exact pathogenesis remains

–
unclear. A recent animal study revealed that the monocytic

0.984 (0.980, 0.988)

expression of G protein-coupled receptor 68 (GPR68) triggered

0.697 (0.676, 0.717)

1.039 (1.033, 1.045)

1.386 (1.172, 1.640)

1.546 (1.306, 1.831)
1.277 (1.129, 1.444)
1.461 (1.289, 1.657)
1.297 (1.184, 1.421)

by retinol and its binding protein (RBP4) exacerbates CKD-

Continuous CTR
HR (95% CI)

induced inflammation and fibrosis of the heart47. In human,

16448.310
16412.990
RMPM +

single-nucleotide polymorphisms of RBP4 have been associated

0.0005
with obesity, diabetes, and, particularly, cardiovascular disease48.
–
–
–
–

–
Chronic myocardial inflammation has been reported to involve
b

0.693 (0.673, 0.714)

1.043 (1.037, 1.049)

1.586 (1.340, 1.877)

All-cause Mortality

1.485 (1.257, 1.754)

in the pathophysiological process of cardiomegaly and

1.291 (1.142, 1.460)
1.319 (1.170, 1.488)

remodeling49. In addition, low calcitriol levels among patients

(0.979, 0.986)
HR (95% CI)

16472.200 with CKD may be associated with decreases in the number and
16441.920
RMPM

functions of peripheral endothelial progenitor cells (EPCs), which

0.983

Ref
are linked to CVD50. It is also reported the dysfunction of EPCs
–

–
–
–
–

may associate with adverse cardiac remodeling and impaired

myocardial perfusion, leading to heart failure51. To fully under-
0.977 (0.968, 0.986)

0.720 (0.679, 0.761)

1.414 (0.853, 2.344)

1.074 (0.739, 1.559)

1.263 (0.962, 1.659)
1.722 (1.046, 2.835)
3.237 (2.016, 5.198)

1.697 (1.130, 2.548)

1.030 (1.017, 1.043)

1.441 (1.094, 1.898)

stand cardiorenal interactions throughout the entire course of

Categorical CTR

CKD, regular and systematic evaluation of heart function among

HR (95% CI)

patients with CKD is essential. Research should focus on opti-

RMPM +

3267.781
3236.761

mizing heart functional evaluation and clarifying the detrimental

0.10
Ref

role of cardiac dysfunction in different CKD stages52. As the

–

healthcare industry has entered the era of digital transformation

and innovation, integrating ML powered annotation into the
0.977 (0.968, 0.986)

0.719 (0.678, 0.761)

1.059 (0.728, 1.539)
1.977 (1.625, 2.406)

1.590 (1.057, 2.392)

1.279 (0.975, 1.679)
1.028 (1.015, 1.041)

1.531 (1.158, 2.023)

existing healthcare paradigms is essential to achieve universal

Continuous CTR
HR (95% CI)

automated disease screening services and harness the potential to

3224.883
3249.010
RMPM +

significantly improve the healthcare efficiency and equity. In the

0.04

current practice, a CXR image does not provide an automated

–
–
–
–

measurement of CTR, not to mention the prediction of future risk

0.698 (0.656, 0.741)
1.295 (0.987, 1.700)
1.926 (1.285, 2.887)

probability ranging from CKD progression, CVD, to all-cause

1.153 (0.796, 1.670)
1.167 (0.893, 1.526)
1.038 (1.025, 1.051)
a,b

mortality and the design of actionable targets. This is exactly the

(0.966, 0.984)
HR (95% CI)
CV Mortality

reason why CTR is currently underused in the clinical practice. If

3267.648
3288.328

the clinical effectiveness through the CTR-driven cardiorenal care

RMPM

0.975

Ref

bundle can be confirmed, CTR will likely become a routine digital

–

–
–
–
–

marker in CKD care.

KFRE + Categorical CTR

The findings of vulnerabilities of increasing CTR among

0.843 (0.823, 0.864)
0.968 (0.962, 0.973)

0.891 (0.883, 0.900)

1.024 (0.812, 1.291)

1.686 (1.567, 1.814)

1.369 (1.092, 1.716)
1.385 (1.114, 1.723)

1.355 (1.171, 1.568)

patients without hypertension deserved attention. This pheno-

HR (95% CI)

type, cardiomegaly without a hypertension history, represents a

10578.860
10612.120

unique clinical challenge as the opportunity to optimize heart size

is limited. It is also likely that hypertension-free patients with
0.60
Ref
–

–
–
–

significant cardiomegaly and CKD may exhibit hemodynamic

instability and compromise the kidney perfusion that accelerates
0.892 (0.884, 0.900)
0.968 (0.963, 0.974)

0.843 (0.822, 0.863)

the progression to ESRD. Given scarce evidence, more research

1.687 (1.569, 1.814)
1.162 (1.054, 1.282)

1.356 (1.171, 1.570)

Continuous CTR

efforts are warranted to confirm our findings and establish the

HR (95% CI)

pathogenic basis underlying this phenomenon.

10583.000
10606.750
KFRE +

The strengths of this study include its high-quality data col-

0.80

lection methods, which comprise longitudinal measurement of

–
–
–
–

–
–
–

the CTR, a large sample size, and rigorous and systematic sta-
0.842 (0.822, 0.863)

Using cause specific model for competing risk of death.

1.707 (1.587, 1.837)

All models use multiple imputation for missing values.

1.287 (1.117, 1.483)

tistical and ML analyses. The study also has several limitations. A

(0.883, 0.900)
(0.965, 0.976)

lack of external validation limits the generalizability of the find-

HR (95% CI)

10609.590
10590.580

ings from this single-center study. Furthermore, we encountered

a,b

Using Uno’s method to calculate c-statistics.

0.970

0.892
ESRD

KFRE

potential confounding, with the decision to undergo CXRs being

Ref
–

–
–
–
–

–
–
–

associated with the patient characteristics that influenced out-

comes. However, in subgroup analysis, a greater risk of devel-
oping dialysis was observed among patients without CVD or
p-value comparing C-statistics
Pooled uPCR (per 1 log unit

hypertension; with respect to mortality outcomes, the prognostic

eGFR (mL/min/1.73m2)

value of the CTR was not affected by CVD status. This study is
further limited by the possibility of residual confounding (eg,
CTR (Continuous)
CTR (Categorical)

Male (ref female)

mortality.

0.47 ≤ ~ <0.52
0.52 ≤ ~ <0.57

physical activity, environmental exposure, and heritability

Hypertension

C-statistics c

information)53 and overadjustment for variables that may lie

increase)
Diabetes

Anemia
≥ 0.57
<0.47

along the causal pathway (eg, CVD). Last but not least, the dis-
Age

AIC
BIC

tribution of CTR may be differential between men and women

b
a

6 COMMUNICATIONS MEDICINE | (2023)3:19 | https://doi.org/10.1038/s43856-023-00241-9 | www.nature.com/commsmed

COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00241-9 ARTICLE

and therefore using sex-specific cutoffs of CTR may help improve 13. Ogata, H. et al. The cardiothoracic ratio and all-cause and cardiovascular
the accuracy of risk assessment. However, due to the reduced disease mortality in patients undergoing maintenance hemodialysis: results of
sample size in the sex-stratified analysis, it would be difficult to the MBD-5D study. Clin. Exp Nephrol. 21, 797–806 (2017).
14. Yen, T. H., Lin, J. L., Lin-Tan, D. T. & Hsu, K. H. Cardiothoracic ratio,
provide robust evaluation of appropriate sex-specific cutoff values inflammation, malnutrition, and mortality in diabetes patients on
of CTR for clinical practice in the present study. maintenance hemodialysis. Am. J. Med. Sci. 337, 421–428 (2009).
15. Chen, S. C. et al. Increased aortic arch calcification and cardiomegaly is
Conclusions associated with rapid renal progression and increased cardiovascular mortality
in chronic kidney disease. Sci. Rep. 9, 5354 (2019).
This study demonstrated the prognostic value of the CTR in 16. Candemir, S., Rajaraman, S., Thoma, G. & Antani, S. in 2018 IEEE Life
patients with CKD by using an ML-powered annotation tool. Sciences Conference (LSC) 109–113 (IEEE, 2018).
Validation of the predictive performance and clinical effectiveness 17. Chamveha, I., Promwiset, T., Tongdee, T., Saiviroonporn, P. &
of the CTR in multi-center clinical trials for progression to ESRD Chaisangmongkon, W. Automated cardiothoracic ratio calculation and
and mortality among patients with CKD-ND warrants adoption cardiomegaly detection using deep learning approach. arXiv https://doi.org/
10.48550/arXiv.2002.07468 (2020).
of this ML application in clinical CKD care. Nevertheless, our 18. Gupte, T. N. M., Gawali, M., Kulkarni, V., Kharat, A., Pant, A. Deep learning
research presents a methodological foundation for using ML to models for calculation of cardiothoracic ratio from chest radiographs for
improve cardioprotection among patients with CKD. assisted diagnosis of cardiomegaly. arXiv https://doi.org/10.48550/arXiv.2101.
07606 (2021).
19. Li, Z. et al. Automatic cardiothoracic ratio calculation with deep learning.
IEEE Access 7, 37749–37756 (2019).
Data availability 20. Que, Q. et al. CardioXNet: Automated detection for cardiomegaly based on
The data that support the findings of this study are available on request from the deep learning. Annu Int. Conf IEEE Eng. Med. Biol. Soc. 2018, 612–615 (2018).
corresponding author, CCK. The data are not publicly available due to them containing 21. Johnson, E. S., Thorp, M. L., Yang, X., Charansonney, O. L. & Smith, D. H.
information that could compromise research participant privacy. Addition data of dose- Predicting renal replacement therapy and mortality in CKD. Am. J. Kidney
response plots are provided at https://figshare.com/articles/dataset/Dose_response_xlsx/ Dis. 50, 559–565 (2007).
21679154. 22. Tangri, N. et al. Multinational assessment of accuracy of equations for
predicting risk of kidney failure: a meta-analysis. JAMA 315, 164–174 (2016).
23. Ting, I. W. et al. Joint longitudinal low Calcium high phosphorus trajectory
associates with accelerated progression, acute coronary syndrome and
Code availability mortality in chronic kidney disease. Sci. Rep. 10, 9682 (2020).
Code for prediction model development is provided at https://zenodo.org/record/
24. Liang, H. Y., Lo, Y. C., Chiang, H. Y., Chen, M. F. & Kuo, C. C. Validation and
7403096#.Y47s0HZByUk54. Further detail of all the code developed during the study is
comparison of the 2003 and 2016 diastolic functional assessments for
available from [email protected] on reasonable request.
cardiovascular mortality in a large single-center cohort. J. Am. Soc.
Echocardiogr. 33, 469–480 (2020).
Received: 4 December 2021; Accepted: 10 January 2023; 25. Chiang, H. Y. et al. Association between preoperative blood glucose level and
hospital length of stay for patients undergoing appendectomy or laparoscopic
cholecystectomy. Diabetes Care 44, 107–115 (2021).
26. Chiang, H. Y. et al. Electronic medical record-based deep data cleaning and
phenotyping improve the diagnostic validity and mortality assessment of
infective endocarditis: Medical big data initiative of CMUH. BioMedicine
References (Taiwan) 11, 59–67 (2021).
1. Hammermeister, K. E., Chikos, P. M., Fisher, L. & Dodge, H. T. Relationship 27. Ronneberger, O., Fischer, P. & Brox, T. in International Conference on Medical
of cardiothoracic ratio and plain film heart volume to late survival. Circulation image Computing and Computer-Assisted Intervention 234–241 (Springer,
59, 89–95 (1979). 2015).
2. Danzer, C. S. The cardiothoracic ratio: an index of cardiac enlargement. Am J. 28. Agarap, A.F. Deep learning using rectified linear units (relu). arXiv https://doi.
Med. Sci. 157, 513–521 (1919). org/10.48550/arXiv.1803.08375 (2018).
3. Zaman, M. J. et al. Cardiothoracic ratio within the “normal” range 29. Levey, A. S. et al. A new equation to estimate glomerular filtration rate. Ann.
independently predicts mortality in patients undergoing coronary Intern. Med. 150, 604–612 (2009).
angiography. Heart 93, 491–494 (2007). 30. Yamamoto, K. et al. The total urine protein-to-creatinine ratio can predict the
4. Yotsueda, R. et al. Cardiothoracic ratio and all-cause mortality and presence of microalbuminuria. PLoS One 9, e91067 (2014).
cardiovascular disease events in Hemodialysis patients: the Q-Cohort study. 31. Proust-Lima, C., Philipps, V. & Liquet, B. Estimation of extended mixed
Am. J. Kidney Dis. 70, 84–92 (2017). models using latent classes and latent processes: the R package lcmm. J. Statist.
5. Chen, K. H. et al. Cardiothoracic ratio association with mortality in patients Softw. 78, 1–56 (2017).
on maintenance peritoneal dialysis. Ther. Apher. Dial. 15, 81–88 (2011). 32. Proust-Lima, C., Philipps, V., Diakite, A. & Liquet, B. Extended mixed models
6. Frishman, W. H. et al. Cardiomegaly on chest x-ray: prognostic implications using latent classes and latent processes. J. Stat. Softw. https://doi.org/10.
from a ten-year cohort study of elderly subjects: a report from the Bronx 48550/arXiv.1503.00890 (2021).
longitudinal aging study. Am. Heart J. 124, 1026–1030 (1992). 33. Guo, C., So, Y. C. & Jang, W. Evaluating Predictive Accuracy of Survival
7. Hemingway, H., Shipley, M., Christie, D. & Marmot, M. Is cardiothoracic ratio Models with PROC PHREG. https://www.semanticscholar.org/paper/
in healthy middle aged men an independent predictor of coronary heart Evaluating-Predictive-Accuracy-of-Survival-Models-Guo-So/
disease mortality? Whitehall study 25 year follow up. BMJ 316, 1353–1354 0f637c13f7eac0dbbeb1a691da46197593fa131b (2017).
(1998). 34. Liang, H. Y. et al. Associations between myocardial diastolic dysfunction and
8. Rautaharju, P. M. et al. Electrocardiographic estimate of left ventricular mass cardiovascular mortality in chronic kidney disease: a large single-center cohort
versus radiographic cardiac size and the risk of cardiovascular disease study. J. Am.Soc. Echocardiogr. 35, 395–407 (2022).
mortality in the epidemiologic follow-up study of the First National Health 35. Loomba, R. S., Shah, P. H., Nijhawan, K., Aggarwal, S. & Arora, R.
and Nutrition Examination Survey. Am. J. Cardiol. 62, 59–66 (1988). Cardiothoracic ratio for prediction of left ventricular dilation: a systematic
9. Hemingway, H., Shipley, M., Christie, D. & Marmot, M. Cardiothoracic ratio review and pooled analysis. Future Cardiol. 11, 171–175 (2015).
and relative heart volume as predictors of coronary heart disease mortality. the 36. Jotterand, M., Faouzi, M., Dedouit, F. & Michaud, K. New formula for
Whitehall study 25 year follow-up. Eur. Heart J. 19, 859–869 (1998). cardiothoracic ratio for the diagnosis of cardiomegaly on post-mortem CT.
10. Quinn, R. R., Laupacis, A., Hux, J. E., Oliver, M. J. & Austin, P. C. Predicting Int. J. Legal Med. 134, 663–667 (2020).
the risk of 1-year mortality in incident dialysis patients: accounting for case- 37. Truszkiewicz, K., Poreba, R. & Gac, P. Radiological cardiothoracic ratio in
mix severity in studies using administrative data. Med Care 49, 257–266 evidence-based medicine. J. Clin. Med. 10, 2016 (2021).
(2011). 38. Kajimoto, K., Minami, Y., Otsubo, S. & Sato, N., Investigators of the Acute
11. Ito, K. et al. A higher cardiothoracic ratio is associated with 2-year mortality Decompensated Heart Failure Syndromes Registry Sex differences in left
after hemodialysis initiation. Nephron Extra 5, 100–110 (2015). ventricular cavity dilation and outcomes in acute heart failure patients with
12. Chen, K. H. et al. Cardiothoracic ratio, malnutrition, inflammation, and two- left ventricular systolic dysfunction. Can. J. Cardiol. 34, 477–484 (2018).
year mortality in non-diabetic patients on maintenance hemodialysis. Kidney 39. Davies, M. K., Gibbs, C. R. & Lip, G. Y. ABC of heart failure. BMJ 320,
Blood Press Res. 31, 143–151 (2008). 297–300 (2000).

COMMUNICATIONS MEDICINE | (2023)3:19 | https://doi.org/10.1038/s43856-023-00241-9 | www.nature.com/commsmed 7

ARTICLE COMMUNICATIONS MEDICINE | https://doi.org/10.1038/s43856-023-00241-9

40. Chana, H. S., Martin, C. A., Cakebread, H. E., Adjei, F. D. & Gajendragadkar, Author contributions
P. R. Diagnostic accuracy of cardiothoracic ratio on admission chest C.-Y.C., C.C.N.W., J.-S.W., Y.-C.H., I.W.T., H.-C.Y., and C.-C.K. established study
radiography to detect left or right ventricular systolic dysfunction: a direction and designed the study. C.-F.H. and C.-H.S. collected and managed data;
retrospective study. J. R Soc. Med. 108, 317–324 (2015). C.-F.H. and Y.-T.L. conducted statistical analysis. H.-Y.C. and Y.-T.L. monitored data
41. Bohn, E. et al. Predicting risk of mortality in dialysis patients: a retrospective analytic quality. C.-H.S., C.-S.H., M.-Y.W., and S.-H.C. developed the iCTR algorithm.
cohort study evaluating the prognostic value of a simple chest X-ray. BMC C.-Y.C. and H.-C.Y. participated in validation trial of iCTR algorithm. C.C.N.W. Y.-L.H.,
Nephrol 14, 263 (2013). and C.-M.C. provided technical advice. C.-Y.C., C.C.N.W., H.-Y.C., H.-C.Y., C.-C.K.
42. Okute, Y. et al. Cardiothoracic ratio as a predictor of cardiovascular events in a provided interpretation of data. C.-M.C. and C.-C.K. drafted the manuscript. C.-Y.C. and
cohort of hemodialysis patients. J. Atheroscler. Thromb. 24, 412–421 (2017). C.-C.K. coordinated and supervised study. H.-Y.C., C.-F.H, Y.-L.H., and C.-C.K. revised
43. Bansal, A. & Heagerty, P. J. A comparison of landmark methods and time- the manuscript. All authors reviewed and approved final version of the manuscript.
dependent ROC methods to evaluate the time-varying performance of
prognostic markers for survival outcomes. Diagn. Progn. Res. 3, 14 (2019).
44. Varga, T. V., Niss, K., Estampador, A. C., Collin, C. B. & Moseley, P. L.
Competing interests
The authors declare no competing interests.
Association is not prediction: a landscape of confused reporting in diabetes—a
systematic review. Diabetes Res. Clin Pract. 170, 108497 (2020).
45. Tanaka, K. et al. Cardiovascular events and death in Japanese patients with Ethical approval
chronic kidney disease. Kidney Int. 91, 227–234 (2017). The study was approved by the Research Ethical Committee/Institutional Review Board
46. Zannad, F. & Rossignol, P. Cardiorenal syndrome revisited. Circulation 138, of China Medical University Hospital (CMUH105-REC3-068 & CMUH108-REC2-022).
929–944 (2018).
47. Yoshida, Y. et al. Alteration of circadian machinery in monocytes underlies
chronic kidney disease-associated cardiac inflammation and fibrosis. Nat. Additional information
Commun. 12, 2783 (2021). Supplementary information The online version contains supplementary material
48. Steinhoff, J. S., Lass, A. & Schupp, M. Biological functions of RBP4 and its available at https://doi.org/10.1038/s43856-023-00241-9.
relevance for human diseases. Front. Physiol. 12, 659977 (2021).
49. Frieler, R. A. & Mortensen, R. M. Immune cell and other noncardiomyocyte Correspondence and requests for materials should be addressed to Hsiu-Yin Chiang or
regulation of cardiac hypertrophy and remodeling. Circulation 131, Chin-Chi Kuo.
1019–1030 (2015).
50. Shantsila, E., Watson, T. & Lip, G. Y. Endothelial progenitor cells in Peer review information Communications Medicine thanks Francesco Bellocchio,
cardiovascular disorders. J. Am. Coll Cardiol 49, 741–752 (2007). Ryusuke Yotsueda and Takuhiro Moromizato for their contribution to the peer review of
51. Broughton, K. M. et al. Mechanisms of cardiac repair and regeneration. Circ. this work.
Res. 122, 1151–1163 (2018).
52. Rangaswami, J. & McCullough, P. A. Heart failure in end-stage kidney disease: Reprints and permission information is available at http://www.nature.com/reprints
pathophysiology, diagnosis, and therapeutic strategies. Semin Nephrol 38,
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
600–617 (2018).
published maps and institutional affiliations.
53. Jokkel, Z. et al. Heritability of cardiothoracic ratio and aortic arch calcification
in twins. Medicina (Kaunas) 57, 421 (2021).
54. C.Y., C. et al. Cardiothoracic ratio values and trajectories are associated with
risk of requiring dialysis and mortality in chronic kidney disease. Am. J. Open Access This article is licensed under a Creative Commons
Kidney Dis. 70, 84–92 (2022). Attribution 4.0 International License, which permits use, sharing,
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
Commons license, and indicate if changes were made. The images or other third party
Acknowledgements material in this article are included in the article’s Creative Commons license, unless
We appreciate the data exploration, statistical analysis, manuscript preparation, and the indicated otherwise in a credit line to the material. If material is not included in the
support of the iHi Clinical Research Platform from the Big Data Center of CMUH. We article’s Creative Commons license and your intended use is not permitted by statutory
would like to thank the Health and Welfare Data Science Center (HWDC), Ministry of regulation or exceeds the permitted use, you will need to obtain permission directly from
Health Welfare, and Health Data Science Center, China Medical University Hospital for the copyright holder. To view a copy of this license, visit http://creativecommons.org/
providing administrative, technical, and funding support. This study was supported by the licenses/by/4.0/.
National Science and Technology Council of Taiwan (grant number: 108-2314-B-039-038-
MY3, 111-2321-B-468-001, and 111-2320-B-039-052-MY3) and the intramural grant from
China Medical University Hospital (grant number: DMR-111-037, DMR-112-188). © The Author(s) 2023

8 COMMUNICATIONS MEDICINE | (2023)3:19 | https://doi.org/10.1038/s43856-023-00241-9 | www.nature.com/commsmed