Computers in Biology and Medicine 151 (2022) 106297

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier.com/locate/compbiomed

Using a machine learning-based risk prediction model to analyze the

coronary artery calcification score and predict coronary heart disease and
risk assessment
Yue Huang a, 1, YingBo Ren a, 1, Hai Yang a, YiJie Ding b, Yan Liu a, YunChun Yang a,
AnQiong Mao a, Tan Yang d, YingZi Wang c, Feng Xiao c, QiZhou He e, **, Ying Zhang a, *
a
Department of Anesthesiology, Hospital (T.C.M) Affiliated to Southwest Medical University, Luzhou, 646000, Sichuan, China
b
Yangtze Delta Region Institute (Quzhou), University of Electronic Science and Technology of China, 324000, Quzhou, Zhejiang, China
c
Southwest Medical University, Luzhou, 646099, Sichuan, China
d
Department of Cardiac and Vascular Surgery, Hospital (T.C.M) Affiliated to Southwest Medical University, Luzhou, 646000, Sichuan, China
e
Department of Radiology,Hospital (T.C.M) Affiliated to Southwest Medical University, Luzhou, 646000, Sichuan, China

A R T I C L E I N F O A B S T R A C T

Keywords: Objectives: To calculate the coronary artery calcification score (CACS) obtained from coronary artery computed
Coronary artery calcification(CAC) tomography angiography (CCTA) examination and combine it with the influencing factors of coronary artery
Machine learning(ML) calcification (CAC), which is then analyzed by machine learning (ML) to predict the probability of coronary heart
Coronary artery calcification score (CACS)
disease(CHD).
Coronary atherosclerotic heart disease(CHD)
Coronary artery computed tomography
Methods: All patients who were admitted to the Affiliated Hospital of Traditional Chinese Medicine of Southwest
angiography (CCTA) Medical University from January 2019 to March 2022, suspected of CHD, and underwent CCTA inspection were
retrospectively selected. The degree of CAC was quantified based on the Agatston score. To compare the cor­
relation between the CACS and clinical-related factors, we collected 31 variables, including hypertension, dia­
betes, smoking, hyperlipidemia, among others. ML models containing the random forest (RF), radial basis
function neural network (RBFNN),support vector machine (SVM),K-Nearest Neighbor algorithm (KNN) and
kernel ridge regression (KRR) were used to assess the risk of CHD based on CACS and clinical-related factors.
Results: Among the five ML models, RF achieves the best performance about accuracy (ACC) (78.96%), sensitivity
(SN) (93.86%), specificity(Spe) (51.13%), and Matthew’s correlation coefficient (MCC) (0.5192).It also has the
best area under the receiver operator characteristic curve (ROC) (0.8375), which is far superior to the other four
ML models.
Conclusion: Computer ML model analysis confirmed the importance of CACS in predicting the occurrence of CHD,
especially the outstanding RF model, making it another advancement of the ML model in the field of medical
analysis.

1. Introduction safety and quality of life.

CAC is a significant marker of coronary atherosclerosis, and CACS is
Global statistics demonstrated that among the leading causes of also the most important predictor of CHD and all atherosclerotic car­
death, cardiovascular disease is dominant. Meanwhile, the incidence of diovascular diseases (ASCVD) [2,3]. The presence and amount of CAC
coronary heart disease (CHD) was 7.2% according to the global de­ detected by CCTA correlate with the severity of coronary atheroscle­
mographic disease statistics from 2015 to 2018, and the number of rosis, which is strongly associated with cardiovascular events. Therefore
deaths from CHD reached 360,900 in 2019 alone [1]. Risk assessment CAC is an important predictor of future cardiovascular events [4].
and early diagnosis of CHD are of great significance to improve the Therefore, CCTA has become a viable noninvasive alternative to

* Corresponding author.
** Corresponding author.
E-mail addresses: [email protected] (Q. He), [email protected] (Y. Zhang).
1
These authors contributed equally to this work.

https://doi.org/10.1016/j.compbiomed.2022.106297
Received 7 August 2022; Received in revised form 12 October 2022; Accepted 6 November 2022
Available online 15 November 2022
0010-4825/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
Y. Huang et al. Computers in Biology and Medicine 151 (2022) 106297

investigating coronary anatomy [5–7]. However, a wide range of factors random forests (RF) [17], artificial neural networks, k-nearest neighbor
affected CAC, some of which also can directly affect CHD occurrences, algorithms (KNN), Kernel Ridge Regression(KRR), support vector ma­
such as hypertension, diabetes, and dyslipidemia [8–10]. These may be chines (SVM) [18], and Bayesian learning. These methods are generally
one of the reasons for the prevalence of CAC and the high incidence of applied to the field of data mining. In this study, RF, KNN, SVM, KRR,
CHD. and radial basis function neural networks (RBFNN) [19] are selected as
In recent years, the rapid development of artificial intelligence(AI) prediction models.
has penetrated into various fields, especially in the field of medical The RF [17] is a classifier that contains multiple decision trees, and
applications [11,12]. machine learning (ML) prediction models have its output classes are determined by the mode of class outputs in the
shown the same or better performance as human beings in cardiology individual trees. First, sampling with replacement is taken from the
diagnosis, decision-making, risk prediction, and other medical tasks original dataset, and a sub-dataset is constructed. The data volume of the
[13]. The use of medical AI has shifted from research to daily clinical sub-dataset is the same as that of the original dataset. Elements in
treatment. AI is widely used in cardiac imaging as the number of CCTA different and similar sub-datasets can be repeated. Second, using the
exams increases [14]. By using imaging reports and clinical parameters, sub-dataset to build a sub-decision tree, these data were placed into each
ML algorithms can obtain better prediction results. This is beyond the sub-decision tree, and each sub-decision tree outputs a result. Finally, if
traditional risk scoring [15]. It has also emerged in the field of prediction there is new data and the classification result should be obtained
and risk assessment in medical applications; however, only a few studies through the RF, the output result can be obtained by voting on the
used ML with clinical factors and CACS to evaluate CHD. We hypothe­ judgment result of the sub-decision tree. The topological structure of RF
sized that ML could be a better predictor of CHD than the most advanced is shown in Fig. 1.
risk prediction methods using the CACS detected by CCTA combined KNN is a classic computer AI algorithm. The essential idea of this
with other clinical variables that influence CAC. technique is: if the k is the most similar specimen belonging to a certain
class in the feature space (mean and k-nearest neighbors) of the spec­
2. Methods imen to be classified, then the specimen also appertains the class.
SVM [18] is a class of generalized linear classifiers that perform bi­
Study population: We retrospectively selected all patients admitted nary classification on data in a supervised learning manner, and its de­
to the Affiliated Hospital of Traditional Chinese Medicine of Southwest cision boundary is the maximum margin hyperplane (Fig. 2.) that
Medical University from January 2019 to March 2022 suspected of CHD resolves the learning specimens. SVM utilizes the hinge loss function to
diagnosis and had undergone CCTA scanning. Simultaneously, the estimate the empirical venture and adds a regularization term to the
following criteria must be met: age >30 years and hospitalized. How­ solution system for structural risk optimization. It is a categorizer with
ever, those who meet the following characteristics were excluded: (1) on sparsity and robustness. SVM can perform nonlinear classification
out-of-hospital treatment; (2) who had previously undergone coronary through kernel methods and is one of the common kernel learning
angiographic stent implantation or coronary artery bypass grafting or methods.
other cardiac revascularization; (3) incomplete or biased Agatston score Radial basis function (RBF) [19] neural network should use RBF as
in CCTA; (4) with incomplete medical records; (5) who were hyper­ the hidden unit before forming the hidden layer space. The hidden layer
susceptible to iodine contrast media; (6) with difficulty undergoing the transforms the input vector and maps the low-dimensional space to the
scanning scheme; and (7) with a history of renal impairment and acute high-order space (Fig. 3); where x is the input, c is the high dimension
myocardial infarction in hemodynamic instability decompensated heart represent, w is the learnable parameters, and y is the final output.This
failure. figure shows the specific algorithm used for classification in this study.
All participants had been informed and the study was approved by thus, the nonlinear issue can be figured out. As the RBF neural network
the Ethics Review Committee of the Affiliated Hospital of Traditional has mighty nonlinear fitting competence, it has simple learning rules but
Chinese Medicine of Southwest Medical University (No.:BY2022009). can map any intricate nonlinear relationship.
Clinical demographics: Basic personal information, CCTA examina­ Since the data may be nonlinear, the effects of simple linear
tion reports and Laboratory test results were recorded during the first
examination post-admission.Patient record information from CISZYYSZ
5.5 2003 clinical database. Hypertension, diabetes, fatty liver, osteo­
porosis, and hyperlipidemia were defined by doctors based on the pa­
tient’s report or scan or previous medical treatment history. The CHD
diagnosis was made by doctors with >10 years of professional experi­
ence, who made a clear diagnosis by combining symptoms, signs, and
examination results. Hypertension was classified as follows: healthy,
<140/90 mmHg; level 1, 140–159/90–99 mmHg; level 2, 160–180/
100–110 mmHg; and level 3, 180/110 mmHg. The smoking history was
classified as never, quit, or current. Scoring criteria for the course of
hypertension or diabetes: none, 0; <1 year, 1 point; 1–4 years, 2 points;
5–9 years, 3 points; 10–14 years, 4 points; 15–19 years, 5 points; and 20
years, 6 points.
CCTA introduction: Calcification of coronary artery branches and
extra-coronary calcium were detected in sections using Siemens Dazzle
dual source CT-SOMATOM Definition Flash. The Agatston score was
used for CAC, which was calculated by the density score of the examined
calcification site multiplied by the area. First, CT values of the lesions
were assigned as follows: 130–199, 1 point; 200–299, 2 points; 300–399,
3 points; and >400, 4 points. Then, CT value multiplied by calcification
area. Finally, the scores of coronary branches were added together [16].
CT threshold of 130 HU was adopted in this study, and the equal-quality
correction factor was 0.736.
AI intelligent model: ML methods mainly include decision trees, Fig. 1. The topological structure of the random forest.

2
Y. Huang et al. Computers in Biology and Medicine 151 (2022) 106297

(NLR), C-reactive protein (CRP), Monocyte-lymphocyte ratio (MLR);

Imaging examinations [49] include: aortic calcification, aortic valve
calcification, and mitral valve calcification. First, we analyze the cor­
relation of these influencing factors with CAC and rank their correlation.
Then, ML models including RF, KNN, SVM, KRR and RBFNN assess the
risk of CHD, using the CACS and those clinical-related factors.

3. Results

We use the IBM SPSS Statistics 25 to analyze the relationship be­

tween CACS and variables. The mean ± standard deviation is used to
represent continuous variables, and frequency to represent categorical
variables.
3.1.1 A comparison of patients without and with CHD have higher
Fig. 2. The SVM schematic. CACS (P = 0.000) in Table 1.
3.1.2 The correlation analysis between CACS and clinical influencing
factor variables is listed in Table 1: Age, systolic blood pressure, hy­
pertension grade, aortic calcification, diabetes mellitus, osteoporosis,
ALP, CRP, aortic valve calcification, calcium-phosphorus product, NLR,
MLR and mitral valve calcification are positively correlated with CACS
(P < 0.05). Gender, random blood glucose, hyperlipidemia, diabetes
course, TG, and LP (a) are positively correlated with CACS, but the
differences are not statistically significant. The variables negatively
correlated with CACS are TC, LDL-C, and APO-A (P < 0.05), whereas

Table 1
Information on the dataset.
Feature Value r* p

CHD(yes/no) 979/532 0.2620 0.000

Age (years) 69.36 ± 10.85 0.2100 <0.0001
Gender (Males/females) 774/728 0.0438 0.092
Weight(kg) 61.14 ± 11.44 − 0.0287 0.267
BMI 23.97 ± 3.63 − 0.0257 0.320
Systolic blood pressure (mmHg) 135.74 ± 20.97 0.0738 0.042
Diastolic blood pressure (mmHg) 79.66 ± 13.05 − 0.0132 0.611
random blood glucose(mmol/l) 8.56 ± 3.93 0.0482 0.062
hypertension(0/1/2/3 level) 576/90/259/577 0.1497 0.000
Fig. 3. The topological structure of RBF network.x is the input, c is the high
diabetes mellitus(yes/no) 462/1040 0.0949 0.000
dimension represent, w is the learnable parameters, and y is the final output. Fatty Liver Disease(yes/no) 321/1181 − 0.0420 0.906
osteoporosis(yes/no) 246/1256 0.0669 0.005
regression may not be very good; therefore, data can be mapped to hypertension history(years) 5.47 ± 7.59 − 0.0177 0.523
DM history (years) 2.22 ± 5.32 0.0343 0.172
kernel space and can be linearly separable in this kernel space. We call
smoking(no/quit/continue) 959/181/362 − 0.0046 0.189
this model KRR, which is very similar to SVM, except that the optimi­ TG 1.97 ± 1.56 0.0227 0.379
zation objective is different. TC 4.60 ± 1.13 − 0.0569 0.028
For various types of data in different fields, various models have their HDL-C 1.30 ± 0.32 − 0.0333 0.198
own merits and demerits. To verify the applicability of ML in this field, LDL-C 2.76 ± 0.93 − 0.0686 0.008
APO-α1 1.41 ± 0.26 − 0.0592 0.022
the five ML models will be used to test their performance prediction on APO-β 0.93 ± 0.26 − 0.0450 0.081
the dataset in this study. LP(a) 214.01 ± 232.86 0.0401 0.121
A total of 2049 patients who underwent CCTA were collected, ALP 86.87 ± 38.06 0.0518 0.045
including 97 patients who underwent coronary revascularization, 449 hs-CRP 12.01 ± 28.95 0.0903 0.001
Ca*P 2.44 ± 0.55 0.0635 0.014
with incomplete information, and 1 with excessive calcification score
NLR 4.53 ± 5.09 0.0609 0.018
bias. Finally, after removing these samples, 1502 samples that met the MLR 0.42 ± 0.32 0.0812 0.002
criteria for data analysis were included. There are numerous factors Calcification of the aorta(yes/no) 1102/400 0.1162 0.000
influencing CAC, and we included four aspects of observation: basic Calcification of aortic valve(yes/no) 219/1283 0.0723 0.005
conditions, medical history information, laboratory tests (lipid meta­ calcification of mitral valve(yes/no) 64/1438 0.0746 0.001
Hyperlipemia(yes/no) 350/1152 0.0096 0.594
bolism [42], inflammation [43]), and imaging tests, respectively.Basic
information including: gender [ [27,28]], age [29], systolic blood pres­ CHD:Coronary atherosclerotic heart disease.
sure (SBP), weight [30], body mass indesx (BMI) [31], diastolic blood BMI:Body Mass Index; TG:triglyceride; TC:total cholesterol.
pressure (DBP). Medical history information included hypertension HDL-C:high density lipoprotein cholesterol.
LDL-C:low density lipoprotein cholesterin.
(grade and duration) [44], diabetes mellitus (duration) [45], fatty liver
APO-α1:Apolipoprotein-α1; APO-β:Apolipoprotein-β.
[46], osteoporosis [47], smoking [48], and hyperlipidemia [ [23,32]].
LP(a):lipoprotein-a; ALP:alkaline phosphatase.
Laboratory tests include: random blood glucose, triglyceride (TG),
hs-CRP:Hypersensitive C-reactive protein.
Apolipoprotein-α(APO-α), total cholesterol(TC), High density lipopro­ Ca*P:The calcium-phosphorus product.
tein cholesterol (HDL-C), Apolipoprotein-β(APO-β), Lipoprotein(a), Low NLR:Neutrophil-lymphocyte ratio.
density lipoprotein cholesterol (LDL-C), alkaline phosphatase (ALP), MLR:Monocyte-lymphocyte ratio.
calcium-phosphorus product(Ca*P), neutrophil-lymphocyte ratio denotes that each variable correlated with label (CACS) using Pearson correla­
tion coefficient (r).

3
Y. Huang et al. Computers in Biology and Medicine 151 (2022) 106297

weight, BMI, diastolic blood pressure, fatty liver, hypertension course, Table 2
smoking history, HDL-C, and APO-β are negatively correlated with Comparison between existing classifiers (10-CV).
CACS, and the differences are not statistically significant.We ranked the Methods ACC (%) SN (%) Spec (%) MCC
influence of these variables on CAC and concluded that age had the
RF 78.96 93.86 51.13 0. 5192
highest association, followed by hypertension. The specific ordering is SVM 69.31 88.34 33.78 0. 2685
shown in Fig. 4. The association of each clinical variable with CACS is KRR 69.91 89.37 33.58 0. 2812
detailed in the Supplementary materials (Figs. 6 and 7). RBFNN 68.78 87.01 34.72 0. 2561
KNN 66.71 89.88 23.48 0. 1772

3.1. Predictive performance of the models with 66.71% (ACC), 89.88% (SN), 23.48% (Spec), and 0. 1772 (MCC).
In Fig. 5, RF achieves the best area under the ROC curve value (AUC:
In this section, we test the predictive performance of these five 0.8375), which is far better than the other four ML models. The area
models under 10-fold cross-validation (10-CV). Matthew’s correlation under the ROC curve is represented by the AUC value. The classifier with
coefficient (MCC), accuracy (ACC), specificity (SP), and sensitivity (SN) a large AUC value has better performance. For SVM, KRR, RBFNN, and
are used to appraise the performance of classifiers. The calculation KNN, the AUCs are 0.7163, 0.7122, 0.7027, and 0.5668, respectively.
formulas are as follows: The performance of SVM, KRR, and RBFNN is very close. KNN also
TP + TN achieves the lowest performance on AUC.
ACC =
TP + FP + TN + FN
TP 4. Discussion
SN =
TP + FN
(1) 4.1. Relationship between CACS and CHD
TN
Spec =
TN + FP
TP × TN − FP × FN Agatston score, as a calculation method of CACS, was first introduced
MCC = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ by Agatston in 1991 [20]. It is a reliable method for the detection and
(TP + FN) × (TN + FP) × (TP + FP) × (TN + FN)
quantitative evaluation of CAC deposition and can evaluate the degree
TP and TN represent the true-positive and true-negative values, of vascular calcification in patients. Its effectiveness and accuracy have
whereas FN and FP represent false-negative and false-positive values. been widely proven, and it has been widely used in screening and
MCC is a measurement used to measure the classification performance of diagnosing CAD [21,22]. The 2018 AHA/ACC Guidelines for Blood
two categories in ML. This index considered TP, TN, FP, and FN. It is Cholesterol Management report included CACS in the risk assessment of
generally considered that this measurement is relatively balanced. It can ASCVD and presented corresponding treatment regiments based on
be applied even when the sample contents of the two categories are stratification [23]. In our study, the higher the CACS, the higher the
different. Besides, the ROC is also applied to assess models. incidence of CHD, thus indicating that CACS is closely related to the
Our data are tested on several models, including KNN, SVM, RBFNN, development of CHD. Therefore, the measurement of CACS can improve
and RF. The consequences are displayed in Table 2 and Fig. 3. The RF the accuracy of risk classification of CHD. Other previous studies
has achieved the best performance in ACC (78.96%), SN (93.86%), Spec demonstrated that CACS as a risk factor for CHD has also been fully
(51.13%), and MCC (0.5192). Since RF is an ensemble classification confirmed [24–26].
model, it has strong robustness. Moreover, noise variables and samples
are also not affected. SVM and KRR are the kernel methods; therefore,
4.2. Clinical factors affecting calcification
they have similar predictive performances. The ACC of SVM and KRR is
69.31% and 69.91%, respectively. For RBFNN, 68.78% (ACC), 87.01%
4.2.1. Correlation between baseline demographic characteristics and CACS
(SN), 34.72% (Spec), and 0. 2561 (MCC) were observed. As a feed-
First, the study on the relationship between gender and CACS found
forward neural network with the ability of universal approximation,
that males have a higher proportion, which is consistent with that of
RBFNN is also a frequently-used neural network. For noisy data, its
previous studies [27,28]. However, due to the high proportion of
performance is comparable to SVM and KRR. However, KNN, a simple
linear model, achieves the worst predictive performance on our data

Fig. 4. The correlation coefficients between variables and categories. Fig. 5. The ROC of different models.

4
Y. Huang et al. Computers in Biology and Medicine 151 (2022) 106297

females with scores of >3000 (80%) in our samples, no significant sta­ approximation. Furthermore, RBFNN is a neural network with universal
tistical difference was observed. In the future, we will collect more approximation capability. The KNN algorithm is sensitive to noisy
samples with scores of >3000 to evaluate the gender difference and samples. Compared with other classifiers, RF has the capability of
mechanism of ultra-high calcification scores. Second, coronary calcifi­ feature selection. Therefore, RF has better robustness in identifying
cation increases with age(p < 0.05). In an earlier study, the investigators noise variables. The influence of variables on the model is significant,
not only identified the effects of gender and age on CAC but also sum­ and the inappropriate introduction of noise variables can lead to biased
marized the percentile chat [29]. Furthermore, overweight and obesity predictions.
are also important influencing factors of CAC [30,31]. However, because
our data were obtained from patients with CHD risk and CCTA exami­ 5. Limitation
nation, this group was mainly the elderly (age 69.36 ± 10.85), and their
body weight tended to decrease with advanced age. This bias in sample At first, a significant selection bias is inevitable in a retrospective
inclusion ultimately resulted in a negative correlation between body single-center study. In addition, information and data collection may not
weight and CAC in our results, as did BMI. be comprehensive, and other factors related to coronary calcification,
such as extracardiac adipose tissue, cannot be collected, which will be
4.2.2. Correlation between lipid metabolism and CACS used as observation indicators for future prospective studies. Then, all
Hyperlipidemia will aggravate the CAC formation [23,32]. However, patients were assessed by professional doctors for indications of CCTA,
in this study, some patients had been previously diagnosed with CHD rather than random tests; therefore, positive tests were found in the
and received relevant drug treatment, especially the administration of majority, leading to screening bias. In the future, follow-up of patients
statins, a lipid-lowering drug; thus, the TC and LDL-C indexes in this and CCTA review post-treatment will be increased to identify the evo­
study did not increase significantly or return to the normal range. Be­ lution of coronary calcification and grasp the patient outcomes. In
sides, the use of statins will also aggravate the formation of calcification addition, the current ML does not have a good solution for sample
[41]. Based on the above two changes, TC and LDL-C did not increase missing data (variables). The predecessors often use the deletion or
with the increase of CACS, but even showed a decreasing trend in this mean filling method, which will bring noise data to the model. In future
study.However, HDL-C is responsible for transporting cholesterol and work, the impact of missing data on the model should be addressed.
preventing it from building up in plaques. Meanwhile, HDL-C is also an
anti-inflammatory molecule that prevents foam cell formation, reduces 6. Conclusion
calcification formation, and protects coronary arteries [33].
The analysis of computer ML models confirmed the importance of
4.2.3. Correlation between inflammation and CACS CACS in predicting the occurrence of CHD, especially the RF model,
When our bodies are exposed to chronic stress or inflammation, the which has made another progress of AI computer learning methods in
endothelial cells in our blood vessels are destroyed. Smooth muscle cells the field of medical analysis due to its excellent performance of high
release apoptotic bodies and necrotic fragments and become attachment accuracy, ability to handle large data sets, stochasticity, better noise
points of calcium and phosphorus complexes, which eventually aggra­ immunity, balancing error, and fast learning speed. Machine learning
vate the formation of vascular calcification [34]. In this study, increased has helped solve many analytical tasks in the fields of biology [36,37],
CRP, NLR, and MLR would aggravate the formation of calcification. medicine [38,39], and pharmacology [40]. This study provides new
Inflammatory factors stimulate osteoblastic differentiation of smooth ideas for the assessment method of CHD, prompting ML to have a higher
muscle cells, forming the stromal vesicles, and eventually resulting in prospect of application in medical fields such as predicting the risk of the
calcium-phosphorus complex deposition [33]. disease and diagnosing the treatment. In future studies, we will collect
more samples, optimize models, and build deep learning-based models
4.2.4. Correlation between other influencing factors and CACS to more accurately predict the development of CHD, as well as treatment
Calcification of coronary arteries can be caused by matrix calcifica­ and prevention recommendations.
tion due to calcium-phosphorus imbalance or calcium-phosphorus
complex deposition due to osteoblast-like cell differentiation. There­ Declaration of competing interest
fore, the calcium-phosphorus product is an independent influencing
factor of CAC [34], which has also been confirmed in this study [35].At The authors declare that the research was conducted in the absence
the same time, hydroxyapatite crystals accumulate in the intimal layer of any commercial or financial relationships that could be construed as a
of the artery and form calcified plaques.Inorganic pyrophosphate (PPi) potential conflict of interest.
inhibits its formation, but PPi is hydrolyzed by ALP to inorganic phos­
phate. Therefore, elevated ALP activity can lead to a disruption of the
Acknowledgments
balance between inorganic phosphate and pyrophosphate. This imbal­
ance can promote to an increase in ectopic calcification, leading to
This work was supported by the Luzhou Municipal People’s Gov­
arterial calcification and contributing to cardiovascular events [35].
ernment and Southwest Medical University Science and Technology
strategic cooperation project (No.2021LZXND-J14).
4.3. ML model comparison
Appendix A. Supplementary data
In this study, ML models of KNN, SVM [18], RBFNN [19], and RF
[17] are used to predict CHD and risk assessment. It can be seen from the
Supplementary data to this article can be found online at https://doi.
experiments that RF achieves the best ACC (78.96%), SN (93.86%), Spec org/10.1016/j.compbiomed.2022.106297.
(51.13%), and MCC (0.5192). The recognition rate for positive samples
reached 93.86%, which is of significant clinical value. Different from
References
other ML models, RF is an integrated model composed of multiple de­
cision trees, which can reduce the influence of noise samples and vari­ [1] C.W. Tsao, A.W. Aday, Z.I. Almarzooq, A. Alonso, A.Z. Beaton, M.S. Bittencourt, A.
ables on final decision results. Furthermore, the AUC value of RF also K. Boehme, A.E. Buxton, A.P. Carson, Y. Commodore-Mensah, M.S.V. Elkind, K.
reaches the maximum value, which is significantly better than other ML R. Evenson, C. Eze-Nliam, J.F. Ferguson, G. Generoso, J.E. Ho, R. Kalani, S.S. Khan,
B.M. Kissela, K.L. Knutson, D.A. Levine, T.T. Lewis, J. Liu, M.S. Loop, J. Ma, M.
models (in Fig. 5). SVM and KRR are relatively common kernel methods, E. Mussolino, S.D. Navaneethan, A.M. Perak, R. Poudel, M. Rezk-Hanna, G.A. Roth,
which have good results for feature mapping and nonlinear E.B. Schroeder, S.H. Shah, E.L. Thacker, L.B. VanWagner, S.S. Virani, J.H. Voecks,

5
Y. Huang et al. Computers in Biology and Medicine 151 (2022) 106297

N.Y. Wang, K. Yaffe, S.S. Martin, Heart disease and stroke statistics-2022 update: a [23] S.M. Grundy, N.J. Stone, A.L. Bailey, C. Beam, K.K. Birtcher, R.S. Blumenthal, L.
report from the American heart association, Feb 22, Circulation 145 (8) (2022) T. Braun, S. de Ferranti, J. Faiella-Tommasino, D.E. Forman, R. Goldberg, P.
e153–e639, https://doi.org/10.1161/CIR.0000000000001052. Epub 2022 Jan 26. A. Heidenreich, M.A. Hlatky, D.W. Jones, D. Lloyd-Jones, N. Lopez-Pajares, C.
[2] S.Y. Tay, P.Y. Chang, W.T. Lao, Y.C. Lin, Y.H. Chung, W.P. Chan, The proper use of E. Ndumele, C.E. Orringer, C.A. Peralta, J.J. Saseen, S.C. Smith Jr., L. Sperling, S.
coronary calcium score and coronary computed tomography angiography for S. Virani, J. Yeboah, AHA/ACC/AACVPR/AAPA/ABC/ACPM/ADA/AGS/APhA/
screening asymptomatic patients with cardiovascular risk factors, Sci. Rep. 7 (1) ASPC/NLA/PCNA guideline on the management of blood cholesterol: a report of
(2017 Dec 15), 17653, https://doi.org/10.1038/s41598-017-17655-w. the American college of cardiology/American heart association task force on
[3] S. Winther, S.E. Schmidt, T. Mayrhofer, H.E. Botker, U. Hoffmann, P.S. Douglas, clinical practice guidelines, Circulation 139 (25) (2018) e1082–e1143, https://doi.
W. Wijns, J. Bax, L. Nissen, V. Lynggaard, J.J. Christiansen, A. Saraste, M. Bottcher, org/10.1161/CIR.0000000000000625, 2019 Jun 18, Epub 2018 Nov 10. Erratum
J. Knuuti, Incorporating coronary calcification into pre-test assessment of the in: Circulation. 2019 Jun 18;139(25):e1182-e1186.
likelihood of coronary artery disease, J. Am. Coll. Cardiol. 76 (21) (2020) [24] Muscogiuri, Giuseppe et al. “Artificial intelligence in coronary computed
2421–2432, https://doi.org/10.1016/j.jacc.2020.09.585. Nov 24. tomography angiography: from anatomy to prognosis.” BioMed Res. Int. vol. 2020
[4] T.S. Polonsky, R.L. McClelland, N.W. Jorgensen, et al., Coronary artery calcium 6649410. 16 Dec. 2020, doi:10.1155/2020/6649410.
score and risk classification for coronary heart disease prediction, JAMA 303 (16) [25] M.S. Lo-Kioeng-Shioe, D. Rijlaarsdam-Hermsen, R.T. van Domburg,
(2010) 1610–1616, https://doi.org/10.1001/jama.2010.461. M. Hadamitzky, J.A.C. Lima, S.E. Hoeks, J.W. Deckers, Prognostic value of
[5] F. Baessato, M. Guglielmo, G. Muscogiuri, et al., The incremental role of coronary coronary artery calcium score in symptomatic individuals: a meta-analysis of
computed tomography in chronic coronary syndromes, J. Clin. Med. 9 (12) (2020) 34,000 subjects, Int. J. Cardiol. 299 (2020) 56–62, https://doi.org/10.1016/j.
3925, https://doi.org/10.3390/jcm9123925. . Published 2020 Dec 3. ijcard.2019.06.003. Jan 15, Epub 2019 Jun 4.
[6] M.J. Budoff, D. Dowe, J.G. Jollis, M. Gitter, J. Sutherland, E. Halamert, M. Scherer, [26] L. Baskaran, Y.P. Neo, J.K. Lee, Y.E. Yoon, Y. Jiang, S.J. Al’Aref, A.R. van
R. Bellinger, A. Martin, R. Benton, A. Delago, J.K. Min, Diagnostic performance of Rosendael, D. Han, F.Y. Lin, R. Nakanishi, P. Maurovich Horvat, S.Y. Tan, T.
64-multidetector row coronary computed tomographic angiography for evaluation C. Villines, M.S. Bittencourt, L.J. Shaw, Evaluating the coronary artery disease
of coronary artery stenosis in individuals without known coronary artery disease: consortium model and the coronary artery calcium score in predicting obstructive
results from the prospective multicenter ACCURACY (Assessment by Coronary coronary artery disease in a symptomatic mixed Asian cohort, J. Am. Heart Assoc.
Computed Tomographic Angiography of Individuals Undergoing Invasive Coronary 11 (8) (2022), e022697, https://doi.org/10.1161/JAHA.121.022697. Apr 19,
Angiography) trial, J. Am. Coll. Cardiol. 52 (21) (2008) 1724–1732, https://doi. Epub 2022 Apr 12.
org/10.1016/j.jacc.2008.07.031. Nov 18. [27] A. Bigeh, C. Shekar, M. Gulati, Sex differences in coronary artery calcium and long-
[7] J.M. Miller, C.E. Rochitte, M. Dewey, A. Arbab-Zadeh, H. Niinuma, I. Gottlieb, term CV mortality, Curr. Cardiol. Rep. 22 (4) (2020 Feb 12) 21, https://doi.org/
N. Paul, M.E. Clouse, E.P. Shapiro, J. Hoe, A.C. Lardo, D.E. Bush, A. de Roos, 10.1007/s11886-020-1267-9.
C. Cox, J. Brinker, J.A. Lima, Diagnostic performance of coronary angiography by [28] L.J. Shaw, J.K. Min, K. Nasir, J.X. Xie, D.S. Berman, M.D. Miedema, M.J. Blaha, Sex
64-row CT, N. Engl. J. Med. 359 (22) (2008) 2324–2336, https://doi.org/10.1056/ differences in calcified plaque and long-term cardiovascular mortality:
NEJMoa0806576. Nov 27. observations from the CAC Consortium, Eur. Heart J. 39 (41) (2018) 3727–3735,
[8] A. Voutilainen, C. Brester, M. Kolehmainen, T.P. Tuomainen, Epidemiological https://doi.org/10.1093/eurheartj/ehy534, 2018.
analysis of coronary heart disease and its main risk factors: are their associations [29] A. Bigeh, C. Shekar, M. Gulati, Sex differences in coronary artery calcium and long-
multiplicative, additive, or interactive? Ann. Med. 54 (1) (2022) 1500–1510, term CV mortality, Curr. Cardiol. Rep. 22 (2020) 21, https://doi.org/10.1007/
https://doi.org/10.1080/07853890.2022.2078875. Dec. s11886-020-1267-9.
[9] S. Babiker, Y. Eltayeb, N. Sayed-Ahmed, et al., Logit model in prospective coronary [30] A.A. Salman, M. Matter, N.I. Fayad, et al., Improvement of coronary calcium scores
heart disease (CHD) risk factors prediction in Saudi population, Saudi J. Biol. Sci. after bariatric surgery in people with severe obesity, Obes. Surg. 32 (2022)
28 (12) (2021) 7027–7036, https://doi.org/10.1016/j.sjbs.2021.07.089. 472–479, https://doi.org/10.1007/s11695-021-05801-3.
[10] L. Dou, Z. Mao, Q. Fu, G. Chen, S. Li, Health-related quality of life and its [31] J.W. Gao, S. You, Z.Y. Liu, Q.Y. Hao, J.F. Wang, D.A. Vuitton, S.L. Zhang, P.M. Liu,
influencing factors in patients with coronary heart disease in China, Patient Prefer. Different metabolic phenotypes of obesity and risk of coronary artery calcium
Adherence 16 (2022) 781–795, https://doi.org/10.2147/PPA.S347681. Published progression and incident cardiovascular disease events: the CARDIA study,
2022 Mar 25. Arterioscler. Thromb. Vasc. Biol. 42 (5) (2022) 677–688, https://doi.org/10.1161/
[11] A. Larentzakis, N. Lygeros, Artificial intelligence (AI) in medicine as a strategic ATVBAHA.122.317526. May, Epub 2022 Apr 7.
valuable tool, Pan Afr. Med. J. 38 (2021) 184, https://doi.org/10.11604/ [32] V.S. Nunes, I.M. Bensenor, P.A. Lotufo, M. Passarelli, E.R. Nakandakare, E.C.
pamj.2021.38.184.28197. . Published 2021 Feb 17. R. Quintao, The coronary artery calcium score is linked to plasma cholesterol
[12] M.P. Cote, J.H. Lubowitz, J.C. Brand, M.J. Rossi, Artificial intelligence, machine synthesis and absorption markers: Brazilian Longitudinal Study of Adult Health,
learning, and medicine: a little background goes a long way toward understanding, Biosci. Rep. 40 (7) (2020), BSR20201094, https://doi.org/10.1042/BSR20201094.
Arthroscopy 37 (6) (2021) 1699–1702, https://doi.org/10.1016/j. [33] T. Nakahara, M.R. Dweck, N. Narula, D. Pisapia, J. Narula, H.W. Strauss, Coronary
arthro.2021.04.022. Jun. artery calcification: from mechanism to molecular imaging, JACC Cardiovasc.
[13] H. Wang, Q. Zu, J. Chen, et al., Application of artificial intelligence in acute Imaging 10 (5) (2017) 582–593, https://doi.org/10.1016/j.jcmg.2017.03.005.
coronary syndrome: a brief literature review, Adv. Ther. 38 (2021) 5078–5086, May.
https://doi.org/10.1007/s12325-021-01908-2. [34] H. Mori, S. Torii, M. Kutyna, A. Sakamoto, A.V. Finn, R. Virmani, Coronary artery
[14] Giuseppe Muscogiuri, et al., Artificial intelligence in coronary computed calcification and its progression: what does it really mean? JACC Cardiovasc.
tomography angiography: from anatomy to prognosis, BioMed Res. Int. (2020), Imaging 11 (1) (2018) 127–142, https://doi.org/10.1016/j.jcmg.2017.10.012.
https://doi.org/10.1155/2020/6649410, 2020 6649410. 16 Dec. Jan.
[15] S.J. Al’Aref, G. Maliakal, G. Singh, A.R. van Rosendael, X. Ma, Z. Xu, O. Alawamlh, [35] L. Panh, J.B. Ruidavets, H. Rousseau, A. Petermann, V. Bongard, E. Berard,
B. Lee, M. Pandey, S. Achenbach, M.H. Al-Mallah, D. Andreini, J.J. Bax, D. D. Taraszkiewicz, O. Lairez, M. Galinier, D. Carrie, J. Ferrieres, Association
S. Berman, M.J. Budoff, F. Cademartiri, T.Q. Callister, H.J. Chang, K. Chinnaiyan, between serum alkaline phosphatase and coronary artery calcification in a sample
B. Chow, L.J. Shaw, Machine learning of clinical variables and coronary artery of primary cardiovascular prevention patients, Atherosclerosis 260 (2017) 81–86,
calcium scoring for the prediction of obstructive coronary artery disease on https://doi.org/10.1016/j.atherosclerosis.2017.03.030. May, Epub 2017 Mar 23.
coronary computed tomography angiography: analysis from the CONFIRM [36] Y. Wang, F.Q. Shi, L.Y. Cao, et al., Morphological segmentation analysis and
registry, Eur. Heart J. 41 (3) (2020) 359–367, https://doi.org/10.1093/eurheartj/ texture-based support vector machines classification on mice liver fibrosis
ehz565. microscopic images[J], Curr. Bioinf. 14 (4) (2019) 282–294, https://doi.org/
[16] M.J. Blaha, M.B. Mortensen, S. Kianoush, R. Tota-Maharaj, M. Cainzos-Achirica, 10.2174/1574893614666190304125221.
Coronary artery calcium scoring: is it time for a change in methodology? JACC [37] M.N.F. ajila, Gene subset selection for leukemia classification using microarray
Cardiovasc. Imaging 10 (8) (2017) 923–937, https://doi.org/10.1016/j. data[J], Curr. Bioinf. 14 (4) (2019) 353–358, https://doi.org/10.2174/
jcmg.2017.05.007. Aug. 1574893613666181031141717.
[17] L. Breiman, Random forests, Mach. Learn. 45 (2001) 5–32, https://doi.org/ [38] Y. Xiao, J. Wu, Z. Lin, et al., A deep learning-based multi-model ensemble method
10.1023/A:1010933404324. for cancer prediction[J], Comput. Methods Progr. Biomed. 153 (2017) 1–9,
[18] C. Cortes, V. Vapnik, Support-vector networks, Mach. Learn. 20 (1995) 273–297, https://doi.org/10.1016/j.cmpb.2017.09.005.
https://doi.org/10.1007/BF00994018. [39] Esteva A., Kuprel B., Novoa R.A., et al. Dermatologist-level classification of skin
[19] John Moody, J. Christian, Darken. Fast learning in networks of locally-tuned cancer with deep neural networks[J]. Nature, 542(7639): 115-118.doi:10.1038/
processing units, Neural Comput. 1 (2) (1989) 281–294, https://doi.org/10.1162/ nature21056.
neco.1989.1.2.281. [40] J.H. Wang, H. Wang, X.D. Wang, et al., Predicting drug-target interactions via FM-
[20] W.R. Janowitz, A.S. Agatston, M. Viamonte Jr., Comparison of serial quantitative DNN learning[J], Curr. Bioinf. 15 (1) (2020) 68–76, https://doi.org/10.2174/
evaluation of calcified coronary artery plaque by ultrafast computed tomography 1574893614666190227160538.
in persons with and without obstructive coronary artery disease, Am. J. Cardiol. 68 [41] R. Puri, S.J. Nicholls, M. Shao, Y. Kataoka, K. Uno, S.R. Kapadia, E.M. Tuzcu, S.
(1) (1991 Jul 1) 1–6, https://doi.org/10.1016/0002-9149(91)90700-u. E. Nissen, Impact of statins on serial coronary calcification during atheroma
[21] D. Shreya, D.I. Zamora, G.S. Patel, et al., Coronary artery calcium score - a reliable progression and regression, J. Am. Coll. Cardiol. 65 (13) (2015) 1273–1282,
indicator of coronary artery disease? Cureus 13 (12) (2021), e20149 https://doi. https://doi.org/10.1016/j.jacc.2015.01.036. Apr 7.
org/10.7759/cureus.20149. Published 2021 Dec 3. [42] J.S. Wang, H.Y. Chiang, Y.C. Wang, H.C. Yeh, I.W. Ting, C.C. Liang, M.C. Wang, C.
[22] S. Shea, A. Navas-Acien, D. Shimbo, et al., Spatially weighted coronary artery C. Lin, C.T. Hsiao, M.Y. Shen, C.C. Kuo, Dyslipidemia and coronary artery calcium:
calcium score and coronary heart disease events in the multi-ethnic study of from association to development of a risk-prediction nomogram, Nutr. Metabol.
atherosclerosis, Circ. Cardiovasc. Imag. 14 (1) (2021), e011981, https://doi.org/ Cardiovasc. Dis. 32 (8) (2022) 1944–1954, https://doi.org/10.1016/j.
10.1161/CIRCIMAGING.120.011981. numecd.2022.05.006. Aug, Epub 2022 May 19. PMID: 35752545.

6
Y. Huang et al. Computers in Biology and Medicine 151 (2022) 106297

[43] H.Y. Lee, S. Lim, S. Park, Role of inflammation in arterial calcification, Feb, Korean population cohort, Cardiovasc. Diabetol. 12 (2013) 162, https://doi.org/10.1186/
Circ. J. 51 (2) (2021) 114–125, https://doi.org/10.4070/kcj.2020.0517. Erratum 1475-2840-12-162. Nov 5.
in: Korean Circ J. 2021 Mar;51(3):286-287. PMID: 33525066. [47] T.L. Chuang, M. Koo, Y.F. Wang, Association of bone mineral density and coronary
[44] K.B. Won, D. Han, S.Y. Choi, E.J. Chun, S.H. Park, H.W. Han, J. Sung, H.O. Jung, H. artery calcification in patients with osteopenia and osteoporosis, Diagnostics 10 (9)
J. Chang, Association between blood pressure classification defined by the 2017 (2020) 699, https://doi.org/10.3390/diagnostics10090699. Sep. 16.
ACC/AHA guidelines and coronary artery calcification progression in an [48] A. Leigh, J.W. McEvoy, P. Garg, J.J. Carr, V. Sandfort, E.C. Oelsner, M. Budoff,
asymptomatic adult population, Aug 11, Eur. Heart J. Open 1 (1) (2021), https:// D. Herrington, J. Yeboah, Coronary artery calcium scores and atherosclerotic
doi.org/10.1093/ehjopen/oeab009. oeab009, PMID: 35919095; PMCID: cardiovascular disease risk stratification in smokers, JACC Cardiovasc. Imaging 12
PMC9242050. (5) (2019) 852–861, https://doi.org/10.1016/j.jcmg.2017.12.017. May.
[45] S. Schurgin, S. Rich, T. Mazzone, Increased prevalence of significant coronary [49] Williams MC, Massera D, Moss AJ, Bing R, Bularga A, Adamson PD, Hunter A,
artery calcification in patients with diabetes, Diabetes Care 24 (2) (2001) 335–338, Alam S, Shah ASV, Pawade T, Roditi G, van Beek EJR, Nicol ED, Newby DE, Dweck
https://doi.org/10.2337/diacare.24.2.335. Feb. MR. Prevalence and clinical implications of valvular calcification on coronary
[46] K.C. Sung, Y.H. Lim, S. Park, S.M. Kang, J.B. Park, B.J. Kim, J.H. Shin, Arterial computed tomography angiography. Eur. Heart J. Cardiovasc. Imaging. 2021 Feb
stiffness, fatty liver and the presence of coronary artery calcim in a large 22;22(3):262-270. doi: 10.1093/ehjci/jeaa263.