Coronary artery calcium measurement on attenuation correction computed tomography using artificial intelligence: correlation with coronary flow capacity and prognosis

Abstract

Purpose

This study aimed to test whether the coronary artery calcium (CAC) burden on attenuation correction computed tomography (CTac), measured using artificial intelligence (AI-CACac), correlates with coronary flow capacity (CFC) and prognosis.

Materials and methods

We retrospectively enrolled patients who underwent [¹³N]ammonia positron emission tomography (PET) between September 2021 and May 2023. CTac data were obtained from all the patients. Patients with (history of) acute coronary syndrome, previous coronary stent insertion or bypass surgery, or left ventricular ejection fraction < 40% were excluded. The total Agatston score measured using a dedicated AI-CAC quantification software on CTac was defined as AI-CACac. The correlations between AI-CACac and PET-measured myocardial blood flow (MBF) and CFC and significant ischaemia (summed difference score ≥ 7) were analysed. Their prognostic values for major cardiovascular events (MACE), including death, nonfatal myocardial infarction, hospitalisation due to angina pectoris or heart failure, and late (> 90 days) revascularisation, were also evaluated.

Results

In total, 289 patients were included in this study. Significant negative correlations were found between AI-CACac and stress MBF (ρ = −0.363, p < 0.001) and MFR (ρ = −0.305, p < 0.001). AI-CACac > 10 was associated with a significantly higher prevalence of impaired CFC (31% vs. 7%, p < 0.001) and significant ischaemia (20% vs. 7%), which remained significant after adjusting for other risk factors. MACE occurred in 49 (17%) patients (median follow-up, 284 days), and those who experienced MACE had significantly higher AI-CACac (median, 166 vs. 56; p = 0.039). However, multivariable analysis revealed an independent prognostic association among impaired CFC, diabetes, smoking, but not for AI-CACac.

Conclusion

AI-measured CACac correlates well with PET-measured MBF and CFC, but its prognostic significance requires further validation.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00259-024-06948-8.

Introduction

Coronary artery calcium (CAC) measurement is a widely accepted strategy for stratifying cardiovascular risk. The association between higher CAC burden and incremental risks of major adverse cardiovascular events (MACE) has been repeatedly demonstrated in various populations. CAC quantification can be used as a gatekeeper test for the initial evaluation of coronary artery disease and as a guide for determining preventive medical therapy in asymptomatic individuals [1–4].

Based on its well-validated prognostic value, CAC quantification using non-cardiac computed tomography (CT) scans is gaining clinical relevance. Attenuation correction CT (CTac) is a good candidate for CAC quantification. In addition to its original role of providing a spatial map of photon attenuation, this non-gated low-dose CT scan can provide useful information regarding atherosclerotic plaque burden. CAC measured on CTac (CACac) improves cardiovascular risk stratification in combination with myocardial perfusion information [5–9]. It may be particularly helpful for positron emission tomography (PET), which has superior image quality in comparison with single-photon emission tomography (SPECT) and can quantify myocardial blood flow (MBF) in absolute values. Notably, CTac is invariably obtained during PET scan.

However, CAC analysis is time-consuming and prone to inter- or intra-user variabilities for CTac compared with conventional CAC CT [10]. Thus, recent studies have focused on utilizing artificial intelligence (AI)-based CAC measurements. It has been proven to be convenient in the analysis process and is in excellent agreement with human measurements [11]. By adopting AI to measure CAC with CTac from PET, CAC burden and absolute MBF can be simultaneously quantified in a single imaging study without increasing scan time, radiation exposure, and additional workload. Although several recent studies have shown the potential prognostic value of AI-measured CACac (AI-CACac) and its correlation with myocardial perfusion [5], the latter is limited to relative perfusion defect analysis. A combined analysis of AI-CACac and coronary flow capacity (CFC), a comprehensive metric of the haemodynamic severity of coronary artery disease that can be measured with PET, would further validate AI-based CAC measurement and its application to CTac. Thus, this study aimed to test whether AI-CACac also correlates with PET CFC and whether AI-CACac has an independent prognostic value against CFC.

Materials and methods

Patients

Consecutive [¹³N]ammonia PET scans (n = 439) obtained between September 2021 and May 2023 were retrospectively screened. The inclusion criteria were (1) age at least 18 years old and (2) availability of rest and stress PET scans and MBF quantification. The exclusion criteria were as follows: (1) (history of) acute coronary syndrome, (2) previous coronary stent insertion or bypass surgery, and (3) left ventricular ejection fraction < 40%. When repeated PET scans were obtained from a patient, the first scan was used for the analysis. Patient selection and data analyses as the following descriptions were approved by the Chonnam National University Hospital Institutional Review Board (CNUH-2023-105), which waived the requirement for obtaining signed consent.

[¹³N]ammonia positron emission tomography (PET) and coronary flow capacity

All scans were performed using a 1-day rest–stress protocol. With the patient in the supine position, 370 MBq of [¹³N]ammonia was injected as a bolus into either of the upper extremity veins. Dynamic image acquisition was simultaneously initiated and lasted for 10 min in a list-mode. Immediately after the resting image acquisition, pharmacological stress was induced with either adenosine or dobutamine. The same dose (370 MBq) of [¹³N]ammonia was injected as a bolus at peak stress, and dynamic image acquisition was simultaneously initiated. The total elapsed time ranged from approximately 45 min from the patient positioning to the end of the PET scan.

The dynamic PET images were retrospectively divided into 16 frames. MBF was measured using one-compartment model [12] equipped on a cardiac PET quantification software package (Carimas™, Turku PET Centre, Finland). Myocardial flow reserve (MFR) was calculated by dividing the stress MBF with rest MBF. The CFC was classified as excellent, minimal reduction, mild reduction, moderate reduction, and severe reduction as suggested by Gould et al. [13]. We adopted the same cutoffs of stress MBF and MFR for [¹³N]ammonia PET, following the report by Miura et al. [14]. However, the CFC stratification was performed at the level of global left ventricle, not regional coronary territories, to effectively assess the correlation between global atherosclerotic burden and myocardial perfusion without interference due to individual variations in coronary anatomy.

We classified myocardial perfusion abnormalities by using quantitative and visual (semiquantitative) definitions, impaired CFC and significant ischaemia, respectively. Impaired CFC was defined as having mild or worse reduction in global CFC [14]; significant ischaemia was defined as having summed difference score (SDS) ≥ 7, which corresponds to %SDS ≥ 10% [15, 16]. PET images were analysed by a nuclear medicine physician with 13 years of clinical experience in nuclear cardiology.

Coronary artery calcium (CAC) measurement

CTac was obtained prior to both rest and stress PET image acquisition, and the rest CTac images were used for AI-based CAC quantification in the present investigation. The protocol and specifications of CTac are summarised in Online Resource 1 and are compared with those of the conventional CAC CT protocol at our institution. Using a commercially available CAC measurement AI software package (AVIEW-CAC version 1.1.42.6 [Coreline Soft, Seoul, Korea]), the plaques with attenuation > 130 Hounsfield unit were automatically delineated and quantified to calculate the Agatston score [17, 18]. The total Agatston score obtained from CTac was defined as the AI-CACac. Image analysis was performed on an intrahospital server where AI-based CAC quantification was automatically performed. After the analysis, a formalised CAC report of quantitative CAC parameters, such as the Agatston score, volume score, arterial age, and risk grouping, was automatically generated and transferred to our institution’s Picture Archiving and Communications System where physicians could see the results.

In patients with available conventional CAC CT scans within 1 month of [¹³N]ammonia PET, the total Agatston score was measured by an experienced cardiovascular radiologist without the aid of an AI (CACconv). AI-CACac and CACconv were compared in patients for whom CACconv was available.

Prognostic follow-up

Clinical outcomes were assessed using retrospective medical chart reviews and phone interviews, as necessary. The primary endpoint of the current analyses was MACE, which were defined as a composite of all-cause death, non-fatal acute myocardial infarction, hospitalisation due to angina pectoris or heart failure, and late (> 90 days) revascularisation.

Statistical analyses

Linear correlations between AI-CACac and PET MBF parameters were analysed. The prevalence of impaired CFC and significant ischaemia according to AI-CACac was analysed, and the related factors were evaluated using logistic regression analysis. To dichotomize AI-CACac, the cutoffs of 0, 10, 100, 400, and 1000 were used, according to a previous CAC stratification [19], to select the most relevant value according to prognosis. For patients with available CACconv, the correlation, agreement, and bias between AI-CACac and CACconv were analysed. AI-CACac, PET parameters, and other clinical factors were compared between patients who experienced MACE and those who did not. A multivariable Cox proportional hazards model was used to reveal the independent risk factors for MACE. For further analysis of the prognostic value of AI-CACac and continuous variables from PET, the optimal cutoff value was obtained using the R function surv_cutpoint, from the package survminer [20].

Linear correlations were assessed using Pearson’s or Spearman’s correlation analysis. The frequencies of the categorical variables were compared using the chi-square test. The bias between the two different CAC metrics (mean difference and 95% confidence interval [CI]) was calculated using Bland–Altman plots. The agreement between the two metrics was assessed by calculating intraclass coefficient with a 95% CI. The selection of analytic methodology (parametric vs. non-parametric) and presentation of analytic results were based on the data distribution assessed using the Shapiro–Wilk normality test. Statistical significance was set at p < 0.05, and statistical calculations were performed using R version 4.3.1.

Results

Baseline characteristics

Of the 439 [¹³N]ammonia PET scans initially screened (repeated scans excluded), 150 were excluded because of (history of) acute coronary syndrome (n = 100), left ventricular ejection fraction < 40% (n = 29), previous coronary stent insertion or bypass surgery (n = 15), or age < 18 years (n = 4). Two more cases were excluded because of technical failure to detect CAC (n = 2, Online Resource 2). The final study population (n = 289) consisted of 169 males (58%), with a median age of 66 years. The patients’ baseline characteristics are listed in Table 1.

Table 1.

Baseline characteristics

Age (y) This should be moved to the 1st row of the table contents, not the label of the column	66 (59–73)
Male sex	169 (58%)
Body-mass index (kg/m2)	24.4 (22.3–26.1)
Overweight (≥ 25 kg/m2)	117 (40%)
Obesity (≥ 30 kg/m2)	14 (5%)
Smoking	62 (21%)
Stable angina	124 (43%)
Exertional dyspnea	55 (19%)
Asymptomatic	110 (38%)
Hypertension	188 (65%)
Diabetes	102 (35%)
Dyslipidemia	131 (45%)
Chronic kidney disease	39 (13%)
On dialysis	18 (6%)
Post-kidney transplantation	10 (3%)

Data are presented as means ± standard deviations, medians (interquartile ranges) or n (%)

Artificial intelligence (AI)-measured CAC on attenuation correction computed tomography (AI-CACac) vs. PET myocardial blood flow and ischaemia

AI-CACac was positive (> 0) in most of the study patients (n = 271, 94%). Among patients with available conventional CAC CT within 1 month (n = 36), AI-CACac showed an excellent positive correlation (ρ = 0.935, p < 0.001). The agreement between AI-CACac and CACconv was also high, with an intraclass coefficient of 0.87 (95% CI, 0.64–0.94). However, a trend of underestimation was noted in AI-CACac compared with CACconv as the reference standard, with a mean difference of − 262 (–1078–553). The absolute CAC difference increased as coronary calcification became heavier, but such a trend was not observed in terms of %CAC difference (Online Resource 3).

AI-CACac showed significant negative linear correlations with both stress MBF and MFR, whereas no correlation was found with rest MBF (Fig. 1). Patients with an AI-CACac > 10 were also associated with a higher prevalence of impaired CFC (n = 70) and significant ischaemia (n = 48) (Fig. 2).

Fig. 1.

Correlations between AI-CACac and stress MBF, rest MBF, and MFR. AI-CACac showed negative correlations with stress MBF (A) and MFR (C). However, no correlation was observed between the AI-CACac and resting MBF (B). AI-CACac, coronary artery calcium measured by artificial intelligence on attenuation correction computed tomography; MBF, myocardial blood flow; MFR, myocardial flow reserve

Fig. 2.

Prevalence of significant myocardial ischaemia according to AI-CACac. AI-CACac is associated with a significantly higher prevalence of impaired CFC (A) and significant ischaemia (B)

Patients with impaired CFC had a significantly higher prevalence of AI-CACac > 10 (91% vs. 65%, p < 0.001) (Fig. 3), male sex (77% vs. 53%, p < 0.001), and higher prevalence of diabetes (49% vs. 31%, p = 0.012) and chronic kidney disease (21% vs. 11%, p = 0.042) than those with preserved CFC. Multivariable logistic regression analysis showed that only AI-CACac > 10 (odds ratio [95% CI], 3.84 [1.53–9.59]) and male sex (2.20 [1.15–4.22]) were independent predictors of impaired CFC.

Fig. 3.

Correlation between AI-CACac and PET-defined CFC. AI-CACac > 10 (yellow dots) was more prevalent in patients with an impaired CFC (pink area). An impaired CFC corresponds to the area defined by the perpendicular lines of stress MBF < 1.82 mL/min/g and MFR < 2.38, as suggested by Gould et al. [13]

CFC, coronary flow capacity; PET, positron emission tomography

Patients with significant ischaemia had a higher prevalence of AI-CACac > 10 (88% vs. 69%, p < 0.013) and were more frequently male (73% vs. 56%, p < 0.001) than those without significant ischaemia. Multivariable logistic regression analysis showed that only AI-CACac > 10 (2.70 [1.07–6.83]) was an independent predictor of significant ischaemia.

AI-CACac for prognostic prediction

During a median follow-up of 284 days following [¹³N]ammonia PET, MACE occurred in 49 (17%) patients (death [n = 1], non-fatal myocardial infarction [n = 2], hospitalisation due to angina or heart failure [n = 43], late revascularisation [n = 3]). The patients who experienced MACE had a higher prevalence of smoking, diabetes, and chronic kidney disease. Higher AI-CACac, lower stress MBF, lower MFR, and impaired CFC were associated with the incidence of MACE (Table 2), and the Kaplan–Meier survival curves revealed significantly different prognoses according to AI-CACac and CFC (Fig. 4). However, in the Cox proportional hazards model, an AI-CACac > 10 was not an independent predictor of MACE (Table 3). Even when the cutoff was optimzed to 28, AI-CACac failed to show independent prognostic values when adjusted for clinical risk factors with or without PET parameters (Online Resource 4). AI-CACac remained insignificant when the multivariable analysis included stress MBF or MFR instead of CFC; stress MBF and MFR did not demonstrate independent prognostic values in these analyses, either.

Table 2.

Comparison of patients with and without MACE

	Total (n = 289)	MACE (+) (n = 49)	MACE (–) (n = 240)	P
Age (y)	66 (59–73)	67 (59–71)	66 (59–74)	0.819
Male sex	169 (58%)	32 (65%)	137 (57%)	0.365
Obesity	14 (5%)	1 (1%)	13 (5%)	0.524
Smoking	62 (21%)	20 (41%)	42 (18%)	0.001*
Hypertension	188 (65%)	36 (74%)	152 (63%)	0.233
Diabetes	102 (35%)	29 (59%)	73 (30%)	< 0.001*
Dyslipidaemia	131 (45%)	23 (47%)	108 (45%)	0.928
Chronic kidney disease	39 (13%)	12 (25%)	27 (11%)	0.025*
CTac
AI-CACac	56 (7–461)	166 (15–716)	56 (7–461)	0.039*
log(AI-CACac + 1)	4.2 (2.2–6.2)	5.1 (2.8–6.6)	4.0 (2.1–6.1)	0.039*
> 0	271 (94%)	48 (98%)	223 (93%)	0.314
> 10	207 (72%)	41 (84%)	166 (69%)	0.060
> 28**	165 (57%)	35 (71%)	130 (54%)	0.039*
> 100	136 (47%)	28 (57%)	108 (45%)	0.163
> 400	81 (28%)	17 (35%)	64 (27%)	0.334
> 1000	35 (12%)	10 (20%)	25 (10%)	0.087
Age-adjusted risk  > 75 percentile†	84 (29%)	19 (39%)	65 (27%)	0.142
[13N]ammonia PET
Stress MBF (mL/min/g)	2.2 (1.8–2.6)	2.0 (1.6–2.3)	2.2 (1.9–2.7)	0.005*
Rest MBF (mL/min/g)	0.9 (0.7–1.0)	0.9 (0.7–1.1)	0.9 (0.7–1.0)	0.682
MFR	2.4 (2.0–3.0)	2.1 (1.8–2.7)	2.5 (2.0–3.0)	0.006*
MFR < 2.2**	93 (32%)	25 (52%)	68 (30%)	0.006*
Minimal to severe CFC reduction	133 (46%)	31 (63%)	102 (43%)	0.012*
Mild to severe CFC reduction (= impaired CFC)	70 (24%)	20 (41%)	50 (21%)	0.005*
Moderate to severe CFC reduction	11 (4%)	1 (2%)	10 (4%)	0.765
Significant ischaemia (SDS ≥ 7)	48 (17%)	13 (27%)	35 (15%)	0.087
SSS ≥ 5**	106 (37%)	27 (56%)	79 (35%)	0.010*
Stress EF (%)	58 ± 9	57 ± 10	59 ± 9	0.156
Rest EF (%)	57 ± 7	57 ± 7	57 ± 6	0.860

Data are presented as means ± standard deviations, medians (interquartile ranges) or n (%)

^*p < 0.05

^**Optimal cutoff obtained by surv_cutpoint function from the survminer package, R statistics

^†Hoff et al. Am J Cardiol 2001;87:1335–39

Fig. 4.

Survival curves according to AI-CACac and CFC. Patients with AI-CACac > 10 (A) and impaired CFC (B) showed significantly worse MACE-free survival. MACE, major adverse cardiac event

Table 3.

Univariable and multivariable cox hazard proportional analysis of major prognostic factors

	Univariable analysis		Multivariable analysis
	HR (95% CI)	p	HR (95% CI)	p
AI-CACac > 10	2.00 (0.94–4.27)	0.072	1.11 (0.49–2.54)	0.802
Impaired CFC	2.53 (1.43–4.47)	0.001	1.85 (1.01–3.39)	0.047*
Diabetes	2.85 (1.61–5.04)	< 0.001	2.22 (1. 19–4.12)	0.012*
Chronic kidney disease	2.14 (1.12–4.11)	0.022	1.23 (0.61–2.48)	0.566
Smoking	1.85 (1.29–2.65)	< 0.001	1.54 (1.07–2.24)	0.021*

*p < 0.05 in multivariable analysis

Discussion

The present study evaluated the correlation between the CAC burden measured using AI and MBF, quantified using PET, and prognosis. The AI-measured total CAC burden, AI-CACac, showed negative correlations with stress MBF and MFR, which are determinants of CFC. Higher AI-CACac levels were associated with an increased prevalence of impaired CFC and significant ischaemia. Patients who experienced MACE showed a higher AI-CACac burden, whereas independent associations between AI-CACac and MACE were not observed after adjustment for other risk factors.

Previous studies have demonstrated a correlation between CACac and myocardial perfusion abnormalities. Patchett et al. [6] showed that true-positive SPECT findings were more frequently observed in patients with positive CACac results on visual estimation. This led to significantly different diagnostic performances of SPECT between patients with and without visible CACac and an excellent positive predictive value for SPECT with positive CACac but a suboptimal negative predictive value, and vice versa. A similar correlation was found between quantitative MBF and CACconv. Danad et al. [21] described a stepwise decline in stress MBF and MFR in incremental CAC strata. This correlation remained significant among patients without visual perfusion abnormalities in a study by Naya et al. [22]. This suggests that a diffuse atherosclerotic burden can diminish the vasodilatory capacity of the coronary system, even in the absence of obstructive stenosis. We further extended the relationship between CAC and myocardial perfusion to PET-measured CFC, a comprehensive metric that integrates MBF and MFR. In our analysis, higher AI-CACac was independently associated with impaired CFC. AI-CACac < 10 could exclude CFC impairment (negative predictive value 93%, Fig. 2), which may be a more relevant prognostic index than visual ischaemia, stress MBF, or MFR alone [13, 23, 24].

Recent studies have reported the prognostic significance of AI-CACac. Miller et al. [5] reported that AI-CACac measured using a novel convolutional long short-term algorithm reduced the time required for analysis to < 2 s, with a categorical agreement of kappa 0.80. It added prognostic value to the total perfusion deficit, an automated myocardial perfusion metric, with an impressive prognostic net reclassification ranging from 40 to 50%. Pieszko et al. [25] also reported that AI-CACac was comparable to CAC quantification by human reading in terms of prognostic negative predictive value and net reclassification. This prognostic prediction was independent of PET MBF or MFR. By integrating CACac using machine learning algorithms, including clinical risk factors, haemodynamics during stress testing, and SPECT features (perfusion and volume), Feher et al. [26] found that CACac was the most important factor contributing to prognostic prediction.

In contrast, our study did not find an independent prognostic value for AI-CACac, with or without threshold adjustment. The relatively short follow-up period may have contributed to this finding: a recent study [27] showed the prognostic value of PET MBF was particularly remarkable within 4 years, in contrast to coronary CT, which showed stronger prognostic prediction for later time points (5–8 years). However, it is most important that MBF and CAC represent different aspects of atherosclerosis and they do not necessarily share the same prognostic significances, despite their mutual correlations. Notably, some studies have indicated only modest added value of CAC to quantitative PET MBF indices. Patel et al. [28] recently reported a similar observation; the additive prognostic value of combining CAC with MFR (from ⁸²Rb PET) was only modest. Adding CAC (from a separate CT scan with gating) did not significantly improve the C-index of the prediction model that included the clinical risk factors and PET variables (p = 0.16). In contrast, by adding MFR to the model that included clinical risk factors and CAC, the C-index increased from 0.764 to 0.784 (p < 0.0001). An earlier study [22] also showed that adding CAC to a prediction model that included clinical risk factors and MFR did not improve the C-index (p = 0.97). Therefore, it is not conclusive whether the combination of PET and CACac can improve risk stratification. Current evidence is based on substantially different CTac protocols, slice thicknesses, different quantitation methods, and different parameters from SPECT or PET, and further validation studies should be conducted to overcome the heterogeneity of the studies.

Some technical issues remain regarding our results. The slice thickness of CTac (3.75 mm) was almost two-fold thicker than that of conventional CAC CT at our institution and beyond the recommendation (up to 3 mm) of the software used in our analyses (AVIEW-CAC) [29]. Calcified plaques located between thicker CT slices may have been missed, thereby underestimating the actual CAC burden. Moreover, AI-based delineation of CAC burden can be affected by increased noise from ungated images (Online Resource 5). Such underestimation of CAC burden was more prominent in patients with high CAC burden especially in terms of absolute CAC, although %CAC difference varied according to CACconv. However, the studies by Patchett et al. (7.5-mm slices) [6] and Miller et al. (5-mm slices) [5] showed that CACac still had prognostic significance despite even thicker CT slices. It is possible that the prognostic values of CACac in these studies resulted from the relative difference in CAC burden within specific study populations rather than the absolute CAC scores. Thus, it is not appropriate to directly compare absolute CAC values or prognostic thresholds among different studies using various protocols. In addition, the two cases with failed AI analysis (Online Resource 2) suggested the possibility of body profile (e.g. body mass index) affecting the technical availability of AI in CAC quantification.

This study had some limitations. In the present analysis, the incidence of MACE was relatively high (17%), although most of the events were soft events (hospitalisation due to angina or heart failure). In the current study, patients undergoing PET had a high prevalence of positive CACac and hypertension, and a substantial number of patients had chronic kidney disease, which may have contributed to the high MACE rate. The CFC thresholds were adopted from cardiac PET studies using ⁸²Rb or [¹⁵O]water, which have different first-pass extractions and MBF values. However, Miura et al. demonstrated the prognostic value of the CFC classification with [¹³N]ammonia PET using the same threshold values as those from ⁸²Rb PET [14].

Conclusions

AI-CACac correlates well with the PET-measured CFC and significant ischaemia. A higher AI-CACac is apparently associated with the occurrence of MACE, but its independent prognostic significance is not demonstrated in the current study. However, further clinical validation is required to determine the prognostic values.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: Sang-Geon Cho, Jong Eun Lee, Kyung Hoon Cho, Ju Han Kim, Ho-Chun Song. Data curation: Sang-Geon Cho, Jong Eun Lee, Kyung Hoon Cho. Formal analysis: Sang-Geon Cho, Ki Seong Park. Funding acquisition: Jahae Kim, Ju Han Kim, Ho-Chun Song. Investigation: Sang-Geon Cho, Jong Eun Lee, Kyung Hoon Cho, Jang Bae Moon. Methodology: Sang-Geon Cho, Jong Eun Lee, Kyung Hoon Cho, Jang Bae Moon. Project administration: Ju Han Kim. Resources: Jahae Kim. Software: Sang-Geon Cho, Jang Bae Moon, Kang Bin Kim. Supervision: Ju Han Kim, Ho-Chun Song. Validation: Ki Seong Park, Jahae Kim, Jang Bae Moon. Visualization: Sang-Geon Cho. Writing-original draft: Sang-Geon Cho, Jong Eun Lee. Writing-review & editing: Kyung Hoon Cho, Ki Seong Park, Jahae Kim, Jang Bae Moon, Kang Bin Kim, Ju Han Kim, Ho-Chun Song.

Funding

This work was supported by the Establishment of K-Health National Medical Care Service and Industrial Ecosystem funded by the Ministry of Science and ICT, Republic of Korea (Project No. ITAH0603230110010001000100100).

Data availability

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

Financial interests: The authors have no relevant financial or non-financial interests to disclose regarding the vendor of the software used in the current study.

Ethics approval and consent to participate

Patient selection and data analyses as the following descriptions were approved by the Chonnam National University Hospital Institutional Review Board (CNUH-2023-105), which waived the requirement for obtaining signed consent.