Discovering hidden patterns: Association rules for cardiovascular diseases in type 2 diabetes mellitus

Abstract

BACKGROUND

It is increasingly common to find patients affected by a combination of type 2 diabetes mellitus (T2DM) and coronary artery disease (CAD), and studies are able to correlate their relationships with available biological and clinical evidence. The aim of the current study was to apply association rule mining (ARM) to discover whether there are consistent patterns of clinical features relevant to these diseases. ARM leverages clinical and laboratory data to the meaningful patterns for diabetic CAD by harnessing the power help of data-driven algorithms to optimise the decision-making in patient care.

AIM

To reinforce the evidence of the T2DM-CAD interplay and demonstrate the ability of ARM to provide new insights into multivariate pattern discovery.

METHODS

This cross-sectional study was conducted at the Department of Biochemistry in a specialized tertiary care centre in Delhi, involving a total of 300 consented subjects categorized into three groups: CAD with diabetes, CAD without diabetes, and healthy controls, with 100 subjects in each group. The participants were enrolled from the Cardiology IPD & OPD for the sample collection. The study employed ARM technique to extract the meaningful patterns and relationships from the clinical data with its original value.

RESULTS

The clinical dataset comprised 35 attributes from enrolled subjects. The analysis produced rules with a maximum branching factor of 4 and a rule length of 5, necessitating a 1% probability increase for enhancement. Prominent patterns emerged, highlighting strong links between health indicators and diabetes likelihood, particularly elevated HbA1C and random blood sugar levels. The ARM technique identified individuals with a random blood sugar level > 175 and HbA1C > 6.6 are likely in the “CAD-with-diabetes” group, offering valuable insights into health indicators and influencing factors on disease outcomes.

CONCLUSION

The application of this method holds promise for healthcare practitioners to offer valuable insights for enhancing patient treatment targeting specific subtypes of CAD with diabetes. Implying artificial intelligence techniques with medical data, we have shown the potential for personalized healthcare and the development of user-friendly applications aimed at improving cardiovascular health outcomes for this high-risk population to optimise the decision-making in patient care.

Key Words: Coronary artery disease, Type 2 diabetes mellitus, Coronary angiography, Association rule mining, Artificial intelligence

Core Tip: The study is aimed to apply association rule mining (ARM) to discover the correlation between type 2 diabetes mellitus (T2DM) and coronary artery disease (CAD). ARM leverages clinical and laboratory data to the meaningful patterns for diabetic CAD by harnessing the power help of data-driven algorithms. It will assist to optimise the decision-making in patient care by providing new insights into multivariate pattern discovery. Thus, this research facilitates targeted medication for patients to regulate uncontrolled clinical parameters in case of CAD with T2DM.

INTRODUCTION

Heart disease has become a global threat and considered to be a major cause of mortality and death rate in modern society[1]. The importance of detailed medical assessment cannot be denied; however; in this era of artificial intelligence, if this skillful tool can be given the added advantage of data-driven algorithms and association rule mining (ARM) techniques, it can prove to be helpful to perform this complicated task to estimate the probability of disease based on data and facts provided to the system[2]. In ARM, frequent item sets are built from symptom transactions using a minimum threshold specified by the user, and a confidence level of 0.9 is established by researchers to rank rules. For positively correlated rules, a minimum support and lift of 0.001 and 1, respectively, can be set. Low-confidence and low-support criteria can capture rare or unusual things, while low-support and high-confidence thresholds yield a small number of customer analytics rules[3]. In medical diagnosis, where many symptom combinations are rare, ARM can help identify linked patterns and rules, making symptom analysis more focused[4].

In several previous studies, the authors used ARM techniques to analyze lung cancer data from the SEER program and used hotspots[5,6]. In healthcare, the relationship between diseases is represented by an association rule among symptoms (or diseases), and in medicine it is represented as P→Q, where P and Q are distinct sets of symptoms (or diseases). P is the rule’s antecedents; Q is its consequent. Also, known as “if→then”, “if” is antecedent, and “then” is consequent. Support, confidence, and lift are used to quantify the effectiveness of generated rules, which are respectively calculated as follows:

Support (P→Q) = Number of patients having P and Q /total number of patients

Confidence (P→Q) = Number of patients having P and Q/number of patients having P

Lift (P→Q) = [(Number of Patients having both P and Q)/(number of patients having P)] (fraction of patients having Q)

Support establishes a rule’s frequency (i.e., generality) for a particular data collection. Lift controls the frequency of symptom Q when illness P occurs. Lift determines P and Q’s similarity: Independent (= 1), positive (> 1), negative (< 1)[7]. The Hotspot algorithm, as described previously[8,9], follows a straightforward approach. It starts with the input data and proceeds in a depth-first manner using a greedy strategy. At each node, it branches on the attribute that results in the most significant improvement in the target value while respecting the constraints and then repeats the process recursively for each child node. Each node corresponds to a segment of an association, and the target value can be enhanced by adjusting the target fraction or average target value. To implement the algorithm, the Hotspot program from “Waika to Environment for Knowledge Analysis” (WEKA)[10] is employed.

This program[8,9] provides an overview of disease predictions based on computational intelligence. Available data suggest that neural networks were used for lumen wall evaluation, vessel contour determination, and plaque characterization, optimizing for phase with each vessel segment analysis[11].

Recent advances in data mining techniques have provided unprecedented opportunities to uncover hidden patterns and associations in complex datasets, leading to new insights into the pathogenesis and management of cardiovascular diseases (CVD). Despite the growing interest in this area, there remains a need for more comprehensive studies that explore the intricate relationships between various clinical and genetic factors influencing the development and progression of coronary artery disease (CAD) in type 2 diabetes mellitus (T2DM) patients. By applying ARM to a large dataset of T2DM patients, this study aimed to identify novel associations between clinical and genetic factors that may help in the early detection and management of CAD in this high-risk population. Understanding these hidden patterns could lead to the development of more personalized approaches for the prevention and treatment of CAD in T2DM patients, ultimately improving their overall prognosis and quality of life.

Thus, our work focused on the application of ARM techniques to evaluate the association between angiographic coronary artery stenosis patterns and its severity in diabetic and non-diabetic CAD patients based on their glycemic control. It can prove useful to assist doctors and patients in early detection and intervention of complicated multifactorial disease like CAD.

MATERIALS AND METHODS

This hospital-based cross-sectional study was performed at the Department of Biochemistry in a tertiary care referral cardiac centre in Delhi, India. We analysed the data of 200 consented, age- and sex-matched, angiographically proven patients diagnosed as having CAD with or without diabetes and 100 healthy control individuals.

Data and sample collection

Venous blood samples were collected from consented patients under proper aseptic techniques for the analysis of serum glycated haemoglobin (HbA1c), blood sugar levels, lipids, and other routine parameters. The level of HbA1c was taken as a serum biomarker for diabetes where HbA1c levels more than 6.5% were considered as diabetic and less than 6.5% as non-diabetic[12]. We enrolled patients who had undergone coronary angiography using Judkin’s technique[13] to study angiography patterns.

Proposed analytical model and statistical analysis

In the present investigation, the clinical dataset comprised 35 attributes from enrolled patients, with their real/original clinical values serving as input for the proposed model (Table 1). Data extraction and management are illustrated in Figure 1. Utilizing a Hotspot rule-based data mining method, the study successfully resolved the test, leading to the identification of novel and significant symptom patterns in patients with CAD and diabetes.

Figure 1 Patient data extraction and management: A flowchart illustrating the process of data extraction and management in the study. CAD: Coronary artery disease; DM: Diabetes mellitus.

Table 1 Dataset of coronary artery disease with and without diabetes.

S. No	Attribute name	Data type	Categorical value
1	Age	Numeric	Minimum = 23 years, maximum = 78 yr
2	Gender	Categorical	Female, male
3	Diet	Categorical	Non-veg, veg
4	Systolic_BP	Numeric	Minimum = 73, maximum = 200
5	Diastolic_BP	Numeric	Minimum = 24, maximum = 120
6	Smoking	Categorical	No, yes
7	Alcoholic	Categorical	No, yes
8	Tobacco	Categorical	No, yes
9	Hypertensive	Categorical	No, yes
10	Disease type	Categorical	Double-vessel, single-vessel, triple-vessel, no-disease
11	Gensini score	Numeric	Minimum = 0, maximum = 136
12	Total cholesterol	Numeric	Minimum = 45, maximum = 381
13	Triglycerides	Numeric	Minimum = 25.2, maximum = 924
14	High density lipoprotein	Numeric	Minimum = 11, maximum = 132
15	Low density lipoprotein	Numeric	Minimum = 15, maximum = 303
16	Very low density lipoprotein	Numeric	Minimum = 5.4, maximum = 184
17	Random blood sugar	Numeric	Minimum = 72, maximum = 683
18	HbA1C (glycated haemoglobin)	Numeric	Minimum = 3.1, maximum = 16
19	Sodium	Numeric	Minimum = 122, maximum = 156
20	Potassium	Numeric	Minimum = 1.8, maximum = 6
21	Haemoglobin	Numeric	Minimum = 1.8, maximum = 17.4
22	WBC	Numeric	Minimum = 1300, maximum = 66000
23	Neutrophils	Numeric	Minimum = 0, maximum = 34.6
24	Lymphocytes	Numeric	Minimum = 0, maximum = 11
25	Monocyte	Numeric	Minimum = 0.026, maximum = 5.26
26	Eosinophils	Numeric	Minimum = 0, maximum = 4.6
27	Platelet count	Numeric	Minimum = 0.67, maximum = 173
28	Left atrium	Numeric	Minimum = 1, maximum = 4.1
29	Aortic root	Numeric	Minimum = 1, maximum = 3.4
30	LVIDd	Numeric	Minimum = 3.9, maximum = 5.7
31	LVIDs	Numeric	Minimum = 1, maximum = 4.7
32	IVSdIVDd	Numeric	Minimum = 0.6, maximum = 1.5
33	PWTd_PWTs	Numeric	Minimum = 0.7, maximum = 1.5
34	Ejection fraction	Numeric	Minimum = 20, maximum = 60
35	Group	Categorical	CAD-with-diabetes, CAD-without-diabetes, healthy control

CAD: Coronary artery disease; LVIDd: Left ventricular internal dimension in diastole; LVIDs: Left ventricular internal dimension in systole; IVSd: Interventricular septal thickness in diastole; IVDd: Left ventricular posterior wall thickness in diastole; PWTs: Posterior wall thickness in systole.

As far as available literature is concerned, the current study is the main review to extract the most well-known manifestations of cardiac vessel blockade in cases of CAD with diabetes using straightforward yet incredible Hotspot-based mining. A Hotspot learns a bunch of rules that augment or limit an objective variable or worth of interest[6]. The Hotspot algorithm processes real/numeric clinical data, utilizing the actual values of clinical attributes as input to identify associations among attributes. It aims to discover various rules/patterns specifically for patient groups, including CAD with DM, CAD without DM, and healthy control (HC).

In Figure 2, when we used the clinical data of CAD with DM patients and HCs, the rule mining method produced two rules with support scores of 0.1 and 0.33, confidence of 1, and lift of 1.2 for the antecedent (X) Age = 23 and consequent (Y) “Group = CAD-with-DM” in rule 1. All patients had CAD with DM with a confidence level of 1. Similarly, lift 1.2 shows that Sex = F and Age = 23 are connected with “Group = CAD with DM”.

Figure 2 Schematic demonstrating calculation of support, confidence, and lift using clinical data for association rule mining. CAD: Coronary artery disease; DM: Diabetes mellitus; M: Male; F: Female; UM: Unmarried.

RESULTS

Figure 3 displays a dataset with 300 instances, where the "GROUP" attribute signifies "CAD-with-diabetes" at 33.33% support (100 instances). The figure helps doctors recommend medication for controlling clinical parameters, and pharmacists can develop customized medicines. Rules were generated in the analysis with a maximum branching factor of 4 and rule length of 5, requiring a 1% increase in probability for improvement. Notable rules were found, showing strong connections between health indicators and diabetes likelihood, including elevated HbA1C and random blood sugar levels. These rules provide insights into potential risk factors for CAD with diabetes. For instance, if random blood sugar exceeds 175 and HbA1C is above 6.6, the instance is likely classified as "CAD-with-diabetes." Overall, these rules demonstrate high confidence, lift, and conviction, showcasing their strong predictive capability.

Figure 3 Association rule mining results for patients in the coronary artery disease-with-diabetes group, showing a support of 33%. CAD: Coronary artery disease; HBA1C: Glycated hemoglobin.

Figure 4 shows a dataset with 300 instances, with the target attribute "GROUP" indicating "CAD-without-diabetes" with support value taken as 33.33% (100 instances) of cases. Rules provided outline various conditions, including blood sugar levels, blood pressure, and other biomarkers, each contributing to the prediction of "CAD-without-diabetes" with associated confidence, lift, and conviction measures. For example, one rule specifies that if lymphocytes are less than or equal to 2.83, random blood sugar is less than or equal to 167, and HbA1C levels are less than or equal to 6.2, the predicted group is "CAD-without-diabetes." The presented rules demonstrate diverse combinations involving lymphocytes (≤ 5.05), high-density lipoprotein (≤ 96), triglycerides (≤ 559), platelet count (≤ 5.35), diastolic blood pressure (≤ 100), systolic blood pressure (≤ 150), monocyte levels (> 0.189), and sodium (> 125). These rules unveil intricate relationships, providing insights into the factors influencing the presence of CAD without diabetes, along with associated confidence, lift, leverage, and conviction metrics.

Figure 4 Association rule mining results for patients in the coronary artery disease-without-diabetes group, showing a support of 33%. CAD: Coronary artery disease; HBA1C: Glycated hemoglobin.

Figure 5 shows a dataset that comprises 300 instances, with the target attribute "GROUP" indicating "HC" with support value taken as 33.33%(100 instances). It depicts significant rules, including associations such as [DISEASE_TYPE = no disease, HbA1C ≤ 5.3] and [High density lipoprotein > 58.5, smoking= no], which exhibit robust relationships with the target value HC. The tree incorporates factors like HbA1C levels, lipoprotein concentrations, and lifestyle indicators. Criteria involve conditions such as low HbA1C levels (≤ 4.8), normal blood sugar levels, and specific ranges for lipoproteins. The presented rules highlight connections between various health parameters and the likelihood of disease. These rules offer insights into potential health indicators and contribute to understanding factors influencing disease outcomes within the specified parameters.

Figure 5 Association rule mining results for subjects in the healthy control group, showing a support of 33%. HC: Healthy control; CAD: Coronary artery disease; HBA1C: Glycated hemoglobin.

The association rules, presented as segment descriptions in Table 2, pertain to clinical feature datasets with support values ranging from 0.33 to 0.99. A confidence level approaching 1 indicates a robust association between clinical attributes and the specified target class (Consequent). Additionally, a lift value exceeding 1 signifies a strong correlation with the clinical attribute[7].

Table 2 Association rules for all groups.

Rules	Antecedents ≥ consequents	Support	Confidence	Lift	Segment description
R1	[Random blood sugar > 187, HbA1C > 6.1] ≥ [Group = CAD-with-diabetes]	0.33	1	3	If random blood sugar > 187 and HbA1C > 6.1, then Group = CAD-with-diabetes
R2	[Random blood sugar > 175, High density lipoprotein > 11, HbA1C > 6.1] ≥ [Group = CAD-with-diabetes]	0.66	1	3	If random blood sugar > 175, high density lipoprotein > 11, and HbA1C > 6.1, then Group = CAD-with-diabetes
R3	[Eosinophils ≤ 0.792, sodium ≤ 145, HbA1C > 6.2] ≥ [Group = CAD-with-diabetes]	0.99	0.96	2.88	If eosinophils ≤ 0.792, sodium ≤ 145, and HbA1C > 6.2, then Group = CAD-with-diabetes
R4	[Random blood sugar ≤ 103, Systolic_BP ≤ 130, HbA1C ≤ 6.2] ≥ [Group = CAD-without-diabetes]	0.33	1	3	If random blood sugar ≤ 103, Systolic_BP ≤ 130, and HbA1C ≤ 6.2, then Group = CAD-without-diabetes
R5	[Lymphocytes ≤ 2.83, random blood sugar ≤ 167, HBA1C ≤ 6.4, Systolic_BP ≤ 150] ≥ [Group = CAD-without-diabetes]	0.66	1	3	If lymphocytes ≤ 2.83, random blood sugar ≤ 167, HbA1C ≤ 6.4, and Systolic_BP ≤ 150, then Group = CAD-without-diabetes
R6	[Lymphocytes ≤ 5.05, high density lipoprotein ≤ 96, HBA1C ≤ 6.6] ≥ [Group = cad-without-diabetes]	0.99	0.87	2.61	If lymphocytes ≤ 5.05, high density lipoprotein ≤ 96, and HBA1C ≤ 6.6, then Group = CAD-without-diabetes
R7	[DISEASE_TYPE = No-disease, sodium > 142.5] ≥ [Group = HC]	0.33	1	3	If DISEASE_TYPE = No-disease and sodium > 142.5, then [Group = HC]
R8	[HbA1C ≤ 4.8, low density lipoprotein > 52, DISEASE_TYPE = No-disease] ≥ [Group = HC]	0.66	1	3	If HbA1C ≤ 4.8, low-density lipoprotein > 52, and there is no disease DISEASE_TYPE = No−disease, then [Group = HC]
R9	[DISEASE_TYPE = No-disease, smoking = no, HbA1C ≤ 5.9, tobacco = no] ≥ [Group = HC]	0.99	1	3	If DISEASE_TYPE = No−disease, smoking = no, HbA1C ≤ 5.9, and tobacco = no, then [Group = HC]

Group = CAD-with-diabetes; Group = CAD-without-diabetes; and Group = Healthy Control (HC). HbA1C: Glycated haemoglobin.

DISCUSSION

Oxidative stress and endothelial dysfunction are frequently observed in diabetic patients and may directly contribute to the development of atherosclerosis and CVD[14]. Our study focused on investigating significant differences in diabetic-specific parameters (random blood sugar and serum HbA1c levels) between diabetic and non-diabetic CAD patients. Studies suggested that fasting blood glucose is a major contributor to HbA1c levels in T2DM patients. The underlying multifactorial mechanisms of the disease still pose a challenge[15]. The Hoorn study proposed that T2DM and endothelial dysfunction may interact in a bidirectional manner in the pathogenesis of cardiovascular events[16]. The endothelial dysfunction can be due to various mechanisms, including dyslipidemia, formation of advanced glycated end products, intra-endothelial accumulation of glucose, increased oxidative stress, and low-grade inflammatory responses[17].

Recent advancements in cardiovascular research have shown promising developments. Chen et al[18] utilized self-supervised learning to transfer knowledge of cardiovascular progression between cohorts, enhancing the detection of specific cardiovascular events and improving examination strategies and prognosis. Ma et al[19] introduced a highly accurate and lightweight deep learning model for phonocardiogram classification, crucial for screening valvular heart disease via mobile health applications, demonstrating exceptional accuracy and robustness.

In another study, six machine learning algorithms demonstrated comparable diagnostic accuracy, suggesting stress-only myocardial perfusion imaging supported by machine learning models as a promising alternative in clinical settings[20]. Lee et al[21] developed a deep learning model for detecting obstructive CAD on coronary computed tomography angiography, showing high diagnostic accuracy and improved discrimination. Veroneze et al[22] applied ARM to discover consistent patterns of clinical features and differentially expressed genes relevant to T2DM, dyslipidemia, and periodontitis, offering insights into these chronic inflammatory diseases. Reti-CVD, an artificial intelligence (AI) software as a medical device utilizing retinal images for personalized CVD risk assessment, demonstrated a significant association with increased CVD risk, leading to regulatory authorization[23]. Additionally, Singh et al[24] validated the ARM method, discovering hidden patterns and associations of early-onset myocardial infarction with hypertension and diabetes mellitus.

Our findings align with previous studies regarding the relationship between T2DM, endothelial dysfunction, and adverse cardiovascular outcomes[25]. Pathak et al[26] observed a significantly higher occurrence of triple vessel disease/multivessel disease in patients with a diabetes mellitus duration > 10 years compared to those with a duration < 10 years.

In summary, our study contributes to the growing body of evidence supporting the association between T2DM, endothelial dysfunction, and cardiovascular outcomes. Further research leveraging advanced technologies and methodologies, as demonstrated in recent studies, holds promise for improving cardiovascular risk assessment and management in diabetic patients.

A larger multicenter cohort clinical trial, spanning diverse geographic locations and ethnicities, would be beneficial. Nevertheless, our study lays the groundwork for future large-scale studies that can help healthcare providers establish plasma glucose testing goals based on specific HbA1c values, thereby improving patient care.

CONCLUSION

In conclusion, we demonstrated that ARM is a powerful data analysis technique to identify consistent patterns between the clinical profiles of patients affected by specific pathological panels. In addition, ARM is able to evidence relevant associations among important parameters of HbA1C and random blood sugar levels of the patients. These insights shed light on potential risk factors for cardiovascular disease with diabetes. A standout rule suggests that instances with random blood sugar > 175 and HBA1C > 6.6 are likely in the "CAD with diabetes" group. These findings can contribute to the valuable insights into health indicators and factors influencing disease outcomes. With the aid of this research, doctors can effortlessly prescribe medication to regulate uncontrolled clinical parameters, while pharmacists can innovate customized medicines tailored to address specific attributes with uncontrolled clinical parameters. This research facilitates targeted medication for patients. Thus, it is pertinent to widen the horizon of AI applications, while integrating with CAD research and development in terms of systems development, process modelling, and its optimization.