Cardiovascular Disease Preliminary Diagnosis Application Using SQL Queries: Filling Diagnostic Gaps in Resource-Constrained Environments

Abstract

Cardiovascular diseases (CVDs) are chronic diseases associated with a high risk of mortality and morbidity. Early detection of CVD is crucial to initiating timely interventions, such as appropriate counseling and medication, which can effectively manage the condition and improve patient outcomes. This study introduces an innovative ontology-based model for the diagnosis of CVD, aimed at improving decision support systems in healthcare. We developed a database model inspired by ontology principles, tailored for the efficient processing and analysis of CVD-related data. Our model’s effectiveness is demonstrated through its integration into a web application, showcasing significant improvements in diagnostic accuracy and utility in resource-limited settings. Our findings indicate a promising direction for the application of artificial intelligence (AI) in early CVD detection and management, offering a scalable solution to healthcare challenges in diverse environments.

1. Introduction

Cardiovascular diseases (CVDs) remain the leading cause of death worldwide [1,2]. Approximately 17.9 million people lost their lives due to cardiovascular disease in 2019, representing 32% of the total death. Among these deaths, 85% were attributed to heart attacks and strokes. Most deaths resulting from cardiovascular diseases, more than 75%, occur in low- and middle-income countries [3]. Of the 17 million premature deaths (accidentally occurring before 70 years of age) caused by non-communicable diseases (NCD) in 2019, 38% were due to cardiovascular disease.

Prevention of cardiovascular disease involves addressing modifiable risk factors, including smoking, unhealthy diet, obesity, sedentary lifestyle, and excessive alcohol consumption. The incidence of cardiovascular diseases can be significantly reduced by mitigating the aforementioned behavioral factors. Early detection of cardiovascular disease is essential for effective interventions, including timely implementation of appropriate advice and medications, which can effectively manage the condition and improve patient outcomes.

Prevention should be implemented at the general public level, promoting a healthy lifestyle [4] and at the individual level, that is, in people with moderate to high CVD risk or patients already diagnosed with CVD by addressing an unhealthy lifestyle (e.g., poor quality diet, lack of exercise, smoking tobacco) and optimizing risk factors. Prevention is effective—eliminating risk-related activities for health would prevent $\ge 80\%$ from CVD cases and even 40% from cancer cases [5,6].

A personalized early diagnosis system based on artificial intelligence (AI), ontology, and other medical information processing systems may be a great prevention measure. Ontologies and ontology database models have been applied to the diagnosis of CVD [7,8] and other diseases such as pneumonia [9]. Ontologies are formal representations of knowledge that can be used to define the structure and content of a database, allowing the representation of complex and hierarchical relationships between different concepts [10,11].

The ontology-database model combines ontology principles with those of database management systems, allowing for the integration of large and complex data sets within a structured and well-defined framework [12,13]. This type of model has been applied in a variety of fields, including biology, medicine, and healthcare, to facilitate the management and analysis of large and diverse data sets that exemplified the use of ontologies in the diagnosis of CVD using the Human Phenotype Ontology (HPO) to classify and classify phenotypes, including CVD-related phenotypes. HPO is a standardized vocabulary of phenotypic abnormalities that can be used to describe the clinical features of diseases, such as focus diseases. CVDs, which allows for a systematic and comprehensive description of patient phenotypes [14]. This can facilitate the identification of patients with similar clinical characteristics and help in the diagnosis of CVD. The advantages of ontology-based approaches compared to other methods include high performance in combination with machine and human interpretability, reflecting the hierarchical nature of the description of symptoms and the corresponding health conditions [8,15,16,17,18].

Ontologies allow for encapsulating domain-specific knowledge in human-interpretable and machine-interpretable formats. These formats could conform to the diagnostic guidelines established by healthcare professionals, using a natural language that describes the symptoms of diseases [7,8,9,15,18]. A recently popular alternative is large language models, which are trained on large sets of textual data and generate human-like texts, including diagnostic recommendations; however, they do not demonstrate accurate medical understanding and can introduce the risk of spreading misinformation [19,20].

This work evaluates the accuracy of the diagnosis of cardiovascular diseases based on the ontologies of symptoms and guidelines that are implemented in a web application.

2. Related Work

In this section, we present selected applications of ontologies for effective knowledge modeling and information retrieval in diagnosing diseases. In [21], the authors proposed a method to convert an ontology into a database using predefined transformation rules. We discuss the application of ontologies for representing knowledge within relational databases, as well as its potential for use in web application development.

In addition to known computational algorithms, there is also an alternative approach to predict CVD risk, which involves the use of risk prediction models. The World Health Organization/International Society of Hypertension (WHO/ISH) risk score charts are widely used models to predict CVD risk [22]. These graphs are accompanied by practice notes that provide guidance on interpreting and adjusting scores, as well as on managing risk factors that are not included in the risk score calculations.

A recent study by [23] evaluated the use of WHO/ISH risk prediction charts to assess cardiovascular risk in three low- and middle-income countries in Asia. The study found that these graphs have certain advantages. One limitation is that they may underestimate the risk of CVD if individuals already on treatment are not taken into account. The degree of underestimation varies depending on the extent of coverage and the effectiveness of drug treatment.

The Framingham Risk Score (FRS) is another widely used model to calculate CVD risk. A study by [24] investigated the performance of this model by conducting a validation study on 959 randomly selected individuals aged 30–74 years. The results of the experiment showed that the model tended to overestimate the risk in older women and middle-aged men. On the basis of these findings, the researchers suggest that the model should be reclassified to account for these slight errors.

The systematic coronary risk evaluation (SCORE) model, which is used to predict the risk of the first fatal atherosclerotic cardiovascular event in 10 years, takes into account factors such as age, gender, smoking, total cholesterol, and systolic blood pressure (SBP). In [25], Conroy et al. examined the effectiveness of this model in different regions of Europe with varying levels of CVD risk. The study proposed improvements to the model, but it should be noted that the results may not fully reflect an individual’s typical risk due to the nature of the study being conducted in a single instance.

Selvarajah et al. [26] conducted an evaluation of the FRS, SCORE, and WHO/International Society of Hypertension (WHO/ISH) models. The results showed that the WHO/ISH model was inadequate for the stratification of cardiovascular risk.

However, the Framingham and SCORE models successfully identified the risk levels of Asian men and women. In particular, the SCORE high model accurately predicted the risk for men, but not for women. On the other hand, the Framingham model improved the detection of women’s risk compared with the SCORE model. In general, it was found that the Framingham and SCORE models were more effective at identifying cardiovascular risk in Asian men and women than the WHO/ISH labels [27,28,29].

In [8], El Massari et al. compared the performance of random Forest, logistic regression, decision tree, Naïve Bayes, k-Nearest Neighbors, artificial neural network, and support vector machine with ontology. The results show that ontology model achieved a higher accuracy (0.755) than the other examined models (0.731 for decision tree, 0.721 for logistic regression, 0.715 for random forest, 0.648 for support vector machines, 0.645 for artificial neural networks, 0.59 for Naïve Bayes, and 0.571 for k-nearest neighbors).

3. Materials and Methods

In the development of a diagnostic tool for diseases, several considerations were taken into account. One of the main decisions was the choice between using quantitative or qualitative data. Quantitative data for CVD diagnosis are generally collected through objective measurements, such as electrocardiographic (ECG) signals, and are often used in machine learning algorithms [30,31,32,33,34,35]. However, raw sensor data are not readily available to patients and can only be accessed by medical personnel with specialized software provided by the medical device manufacturer. Therefore, we decided to use patient symptoms as the main source of information for the diagnostic tool.

3.1. Dataset Preparation

To integrate this information into an existing behavioral risk factor calculator, a non-machine-learning-based diagnostic method was sought to address the problem with access to raw sensor data and ease of use for patients. After evaluating different options, it was determined that a knowledge base structured following an ontology-inspired approach would be the most suitable. This involved creating a knowledge base/ontology database model by extracting pertinent data from various disease research studies presented in

Table 1. Research items that mention the use of ontology-based approach for major diseases.

Disease	References
Heart Attack	[36,37]
Early Stroke	[37]
Coronary Heart Disease (CHD)	[27]
Hypertension	[38]
Hypotension	[39]
Heart Failure	[40,41,42]
Hole In Heart	[43]
Aortic Disease	[44]
Arrhythmia	[45]
Cerebrovascular Disease (CD)	[46]
Panic Attack	[47]
COVID-19	[48]
Anemia	[49]
Cirrhosis	[50]

and organizing them into tables in the database schema (see Table 1).

Each disease was assigned its own table, which includes details on symptoms, duration of symptoms, and the disease itself. Subsequent investigations have revealed instances of misdiagnosis in which certain diseases resemble cardiovascular disease, as well as associations between specific diseases, such as the correlation between coronary heart disease and panic attacks [51].

3.2. Software

C# and Microsoft SQL-based applications consist of three main components: the database, the application server, and the client user interface. The database is created using Microsoft SQL Server 2022 and is designed to ensure data integrity and efficient data retrieval. The application server processes client requests and handles business logic and data processing. The client user interface communicates with the application server and provides a user-friendly way to interact with the application.

4. System Architecture

4.1. Database Architecture

The CVD database is a collection of structured data that are designed to support the management and analysis of CVD-related information. This database consists of two main types of tables: fixed tables and updatable tables.

Fixed tables, also known as the knowledge base, are a set of predetermined tables within the database that contain structured data that should not be modified by users. Instead, these tables serve as a reference source of information for the database or system, providing a reliable and consistent source of information on CVD-related topics. Examples of data that can be stored in fixed tables include information on different types of CVD, risk factors for CVD, and CVD treatments.

Updatable tables, on the other hand, are the tables within the database that can be modified by users. These tables may contain data entered by users, such as personal data, symptoms, or diagnosis generated by the system per session. The data in these tables can be updated, inserted, or deleted by users as needed, allowing flexibility in the management and analysis of CVD-related information. In the database, tables contain key fields by which we can update the most crucial information regarding symptoms of the disease, disease entities, and patient identification data.

4.2. Applying Inner Join Principle in Diagnosis

Figure 1 shows the use of an inner join in a database to find a match between two tables. The first table, Table A (which also represents all disease tables in the database), contains a list of symptoms for each disease that was collected to create a knowledge base, as illustrated in Figure 2.

The second table, Table B, contains the symptoms of a patient, which were input by the user and stored in the database. This database is dynamic and allows the use of SQL commands such as UPDATE and DELETE, meaning that it can change with each user and is reset when the application is closed.

However, during the diagnosis process, Table B is assumed to be static and cannot be changed. This is to ensure that the diagnosis remains accurate and consistent. To find a match between the two tables, the current user’s symptoms in Table B are compared with the symptoms in the disease table (Table A) using an inner join. The inner join function returns only the rows that match the two tables.

To further analyze the match between the two tables, the “count” function is used to count the number of symptoms that match. These values are stored in a table of “disease_count” using the UPDATE command. There is one query for each major CVD and to handle the counting process for each disease (see Figure 3).

Utilizing the database of CVD symptoms, we align corresponding symptom descriptions from patient data. The maximum correlation identifies the most likely disease entity. To ascertain the most probable disease, the ’max’ function is employed to identify the highest count value among all diseases. This maximum value corresponds to the disease with the highest number of matching symptoms and is therefore the most likely diagnosis for the patient. This process of inner joining, counting, and finding the maximum value allows an accurate prediction of a disease based on the symptoms of the patient, as shown in Figure 4.

5. Results

In order to evaluate the performance of the application, various scenarios were employed.

5.1. Application Test with Real Subjects as Diagnosed by a Doctor

The first evaluation method involved seeking feedback from patients who currently have a CVD. These patients were asked to document the symptoms they had experienced before receiving a diagnosis, and a physician was also consulted to obtain information on symptoms reported by patients diagnosed with CVDs. Using this approach, three subjects were studied and the results are summarized in

Table 2. Comparison of diagnosis of CVDs conducted by physicians and an app.

Subject	Age [Years]	Gender	Reported Symptoms	D.D	A.D
1	70	F	Headache, fatigue, palpitations, blurred vision (which was attributed to headache), edema (which occurred after salt consumption), insomnia (when blood pressure was slightly elevated), dizziness	HTN and DM2	HTN
2	25	M	Dyspnea (difficulty breathing), cough, edema, elevated blood pressure, abdominal distension, paroxysmal nocturnal dyspnea and chest congestion	CHF	HF
3	50	M	Weakness on the right side, numbness on the arm, a tingling sensation and blurred vision accompanied by flashes of light, a feeling of heaviness and an inability to lift objects	CVA	Early CVA

Abbreviations: D.D—doctor diagnosis; A.D—app diagnosis; HTN—hypertension; DM2—diabetes mellitus 2; CHF—congenital heart failure; HF—heart failure; CVA—cerebrovascular accident, is synonymous with stroke.

.

5.2. Prototype Application vs. Existing Application

The second evaluation method involved comparing the performance of the prototype application with that of an existing application, specifically an AI-based Android application named “Symptom Checker”. Seven combinations of random symptoms were entered into both applications and the results were compared (see

Table 3. Comparison of diagnoses suggested by the prototype application and an existing application.

No.	Prototype Application	Existing Application
1	Hypertension, heart failure	hypertension, heart failure, heart disease
2	CHD, heart attack, heart failure	COVID-19, heart disease
3	COVID-19, heart failure	COVID-19, ischemic heart disease, pneumonia
4	Early stroke, hypertension, CBD	Stroke, pancreatic neoplasms, hypertension
5	Cirrhosis	Aortic dissection, anaphylaxis, pulmonary hypertension
6	Cirrhosis	COVID-19, iron deficiency
7	Heart attack, hypertension, heart failure, early stroke	Heart failure, lyme disease, COVID-19, stoke

).

• Symptom combination 1: chest pain, shortness of breath, back pain, nausea, feeling faint, cough, nose bleed.
• Symptom combination 2: chest pain, nausea, dizziness, cough, fever.
• Symptom combination 3: shortness of breath, body pain, nausea, fatigue, cough, fever, headache.
• Symptom combination 4: dizziness, disorientation, blurred vision, headache.
• Symptom combination 5: chest pain, pain (jaw/neck/back), disorientation, jaundice, lack of appetite.
• Symptom combination 6: shortness of breath, body pain, lack of appetite, dry mouth, bleeding gums.
• Symptom combination 7: chest pain, shortness of breath, weakness, headache, loose bladder.

5.3. Measuring Diagnostic Accuracy

Measuring the diagnostic accuracy of a new diagnostic application involves comparing its performance with established reference points provided by other applications. By evaluating the diagnostic results produced by the new application and comparing them to the results obtained from established apps, it is possible to assess the application’s ability to correctly identify and classify medical conditions. Key metrics such as sensitivity, specificity, positive predictive value, and negative predictive value can be calculated to quantitatively evaluate the performance of the new application [52].

According to [52], these parameters are ideally used to calculate the accuracy of diagnosis and to elucidate their respective meanings.

• True positive (TP): This refers to the number of individuals correctly identified as positive by the test. In medical terms, it represents the number of people with the condition being tested who were correctly identified as having the condition.
• True negative (TN): This represents the number of individuals correctly identified as negative by the test. It refers to individuals without the condition who were correctly identified as not having the condition.
• False positive (FP): These are the individuals who were incorrectly classified as positive by the test but do not have the condition. In medical terms, it represents the number of people who were mistakenly identified as having the condition when they were actually healthy.
• False negative (FN): These are the individuals who were incorrectly classified as negative by the test, but actually have the condition. It represents the number of individuals who were mistakenly identified as not having the condition when they actually did.

However, the representation is modified for each test to compare the diagnosis of the app with already proven or existing parameters. Test 1 is a comparison of the diagnosis provided by the designed app and a doctor and Test 2 is a comparison of the diagnosis provided by the designed application and an existing AI-based diagnostic Android application (see the confusion matrices in

Table 4. Confusion matrix of Test 1 (comparison of the suggested diagnosis by an app with the diagnosis conducted by a medical doctor.)

	Predicted Positive	Predicted Negative
Actual Positive	$4^{\mathrm{TP}}$	$0^{\mathrm{FP}}$
Actual Negative	$0^{\mathrm{FN}}$	$0^{\mathrm{TN}}$

and

Table 5. Confusion matrix of Test 2 (comparison of the suggested diagnosis by an app with the diagnosis suggested by an existing AI-based diagnosis Android app.)

	Predicted Positive	Predicted Negative
Actual Positive	$5^{\mathrm{TP}}$	$2^{\mathrm{FP}}$
Actual Negative	$0^{\mathrm{FN}}$	$1^{\mathrm{TN}}$

respectively). The performance metrics for both tests are provided in

Table 6. Performance indicators of the diagnosis made by the prototype application (App), medical doctor, and an existing application.

Parameters	App vs. Doctor Diagnosis	App vs. Existing Application
PPV	1.000	0.714
NPV	Undefined	1.000
Sensitivity	1.000	1.000
Specificity	Undefined	0.330
Accuracy	1.000	0.750

.

• True positive (TP): Number of CVD cases correctly identified by the app and that matches the preexisting diagnosis of the doctor (test 1)/AI-diagnosis android app (test 2).
• True negative (TN): Number of CVD cases identified as negative (that is, not identified as CVD) by the app and by the doctor (Test 1)/AI-diagnosis Android app (Test 2).
• False positive (FP): Number of cases of CVD that were inaccurately identified by the app, despite the absence of the actual condition, as proven by a doctor or AI-based diagnosis Android app.
• False negative (FN): Number of cases of CVD that went undetected by the app, but were actually identified and diagnosed by a doctor or an AI-based diagnosis Android app.
• Actual positive: Possible CVD diagnosis that was evaluated by a doctor or an existing application, with supporting evidence or validation.
• Actual negative: Possible diagnosis that was not identified as cardiovascular disease by a doctor or an existing application.
• Predicted positive: Possible CVD diagnosis made by the app, based on its analysis and algorithm.
• Predicted negative: This denotes other diagnoses that were not detected or identified as cardiovascular diseases by the app. These diagnoses may include conditions or diseases not related to CVD.

Parameters for Calculating Accuracy

• Positive predictive value (PPV), also known as precision, tells the proportion of correctly predicted positive instances out of all instances predicted as positive by the application. A higher PPV indicates that the application is more reliable for correctly identifying positive instances. It suggests that when the application predicts a positive result, there is a relatively higher chance that it is accurate.

  $$PPV=\frac{TP}{TP+FP}$$
• The negative predictive value (NPV) tells us the proportion of correctly predicted negative instances out of all instances predicted as negative by the application.

  $$NPV=\frac{TN}{TN+FN}$$
• Sensitivity, also known as the recall or true positive rate, indicates the proportion of correctly predicted positive instances out of all of the actual positive instances.

  $$Sensitivity=\frac{TP}{TP+FN}$$
• Specificity tells us the proportion of correctly predicted negative instances out of all actual negative instances.

  $$Specificity=\frac{TN}{TN+FP}$$
• Accuracy is a performance metric that tells us the overall correctness of the predictions made by the application or system being evaluated.

  $$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$

We observed that the accuracy of the application was 1.0 or 100%. This high accuracy was achieved because the three cases were accurately diagnosed with cardiovascular disease, matching the predictions made by the application. Furthermore, the positive predictive value was 100%, indicating that all positive diagnoses made by the application were true positives. However, the negative predictive value was undefined because there were no negative diagnoses; all cases were positively diagnosed as cardiovascular disease, according to the evaluations of medical doctors.

5.4. GUI

Figure 5, Figure 6 and Figure 7 show screenshots of the graphical user interface (GUI) of the designed application. Figure 5 presents the selection of symptoms and Figure 6 presents an example of the diagnosis results. An example of the report is shown in Figure 7 in the application and in the form of a PDF file, respectively.

The screenshots presented in Figure 5, Figure 6 and Figure 7 show the order of the potential diagnosis and show the user a possible diagnosis in a friendly and legible manner. The content of the generated PDF file with a report is consistent with the content shown in Figure 7.

6. Discussion

The use of an ontology-inspired database model and inner join or diagnosis demonstrates a reasonably good level of accuracy. Test 1 achieved a perfect accuracy rate of 100%, while Test 2 yielded an accuracy of 75%. Taking the average of these two tests, we can conclude that the overall accuracy of the diagnosis process is 85%, without considering any bias.

Other studies on the diagnosis of cardiovascular diseases based on ontologies reported an accuracy of 75.5% (10-fold cross-validation) in [8], sensitivity of 0.83 and precision of 0.87 in [7], and accuracy of more than 90% in [53].

The accuracy of the decision trees, SVM, or ANN models proposed in [54] was 89.6%, 92.1%, and 91.0%, respectively. In our study, we achieved a sensitivity of 1.0 (100%), indicating that it performed excellently in detecting positive cases of cardiovascular disease. Furthermore, the combinations of diseases used to test both evaluated and existing applications were similar to those associated with cardiovascular disease.

The reason for testing the application with real subjects diagnosed by physicians was to determine whether the application designed for the diagnosis of CVD can provide accurate diagnoses similar to those of physicians. However, only three patients with preexisting cardiovascular disease were tested, resulting in a small sample size that could introduce bias when evaluating the accuracy of the application.

Furthermore, when diagnosing cardiovascular disease, doctors can also consider other coexisting conditions, as demonstrated in the first subject where the doctor’s diagnosis included hypertension and diabetes mellitus. This highlights the complexity of the diagnosis of cardiovascular disease and the need to consider potential comorbidities.

The second test aimed to compare the diagnosis provided by the application under evaluation with that of an established and proven application. On the basis of the results of the second test, several observations can be made. In general, when testing the application against the existing one, it exhibited an accuracy rate of 0.75 (75%) and a positive predictive value (PPV) of 0.714 (71.4%). This indicates that the application was 75% accurate in its diagnoses. This result was lower than reported in [7,8,53] and suggests the need for improvement by changing the sources and design of the ontology.

However, the designed application has some limitations. First, the application was tested against the existing application only seven times. This limited sample size suggests a bias in the experiment and also implies that the accuracy values could be influenced by further testing. Second, only one architecture was considered because of the focus on the accuracy of diagnosis and the simplicity of the schema that implements the proposed ontology.

Second, this application concentrates only on the main cardiovascular diseases and on diseases that closely resemble cardiovascular disease. As a result, we were working with approximately 14 tables using the ontology-based database model. The efficiency of the diagnosis process may vary in terms of speed, but this aspect was not thoroughly explored due to the limited sample size of the diseases under investigation.

7. Conclusions and Future Work

The application has the potential to improve the efficiency and precision of the diagnosis of cardiovascular diseases, particularly in resource-limited settings where access to specialized medical personnel may be limited. By providing users with a tool to identify potential diagnoses and guide further testing, the application can help ensure that patients receive timely and appropriate care. Given that the application is based on novel research, accurate verification is needed.

There are many directions in which this work could be extended in the future:

• One potential avenue for future work is to improve the symptom checker by adding more detailed and specialized questions and a description of the database to create a more comprehensive and accurate diagnostic tool.
• The evaluation of the processing time and its influence on the accuracy.
• Another potential direction is to expand the range of diseases covered by the application. Although the current version focuses on major cardiovascular diseases, there is the potential to add support for other types of disease or health conditions.
• Another area of future work is the integration of the application with electronic medical record systems, allowing for the seamless exchange of data and a more efficient diagnosis and treatment.
• Developing a mobile version of the application that allows users to access it from their smartphones or tablets.
• Adding a feature that allows users to track their symptoms over time could aid in the diagnosis and management of chronic diseases. Integrating the application with wearable devices, such as fitness trackers or smartwatches, allows continuous monitoring of vital signs and other health metrics.