Encoding-Based Machine Learning Approach for Health Status Classification and Remote Monitoring of Cardiac Patients

Abstract

Remote monitoring of a patient’s vital activities has become increasingly important in dealing with various medical applications. In particular, machine learning (ML) techniques have been extensively utilized to analyze electrocardiogram (ECG) signals in cardiac patients to classify heart health status. This trend is largely driven by the growing interest in computer-aided diagnosis based on ML algorithms. However, there has been inadequate investigation into the impact of risk factors on heart health, which hinders the ability to identify heart-related issues and predict the conditions of cardiac patients. In this context, developing a GUI-based classification approach can significantly facilitate online monitoring and provide real-time warnings by predicting potential complications. In this paper, a general framework structure for medical real-time monitoring systems is proposed for modeling the vital signs of cardiac patients in order to predict the patient’s status. The proposed approach analyzes AI-driven interventions to provide a more accurate cardiac diagnosis and real-time monitoring system. To further demonstrate the validity of the presented approach, we employ it in a LabVIEW-based remote tracking system to predict three healthcare statuses (stable, unstable non-critical, and unstable critical). The developed monitoring system receives various information about patients’ vital signs, and then it leverages a novel encoding-based machine learning algorithm to pre-process, analyze, and classify patient status. The developed ANN classifier and proposed encoding-based ML model are compared to other conventional ML-based models, such as Naive Bayes, SVM, and KNN for model accuracy evaluation. The obtained outcomes demonstrate the efficacy of the presented ANN and encoding-based ML approaches by achieving an accuracy of $98.4\%$ and $98.8\%$ for the developed ANN classifier and the proposed encoding-based technique, respectively, whereas Naive Bayes and quadratic SVM algorithms realize $94.8\%$ and $96\%$, respectively. In short, this study aims to explore how ML algorithms can enhance diagnostic accuracy, improve real-time monitoring, and optimize treatment outcomes. Meanwhile, the proposed tracking system outperforms most existing monitoring systems by offering high classification accuracy of the heart health status and a user-friendly interactive interface. Therefore, it can potentially be utilized to improve the performance of remote healthcare monitoring for cardiac patients.

1. Introduction

Hundreds of preventable cardiac arrests occur every year as a result of a lack of real-time monitoring at home or in places apart from specialized healthcare facilities, i.e., the Intensive Care Unit (ICU) [1,2]. Fortunately, due to advances in healthcare technologies, the lifespan of chronic patients has been effectively extended [3]. People suffering from heart disease often require long-term daily monitoring of their vital signs, including blood pressure, heart beating rate, respiratory rate, chest pain, and other factors [4,5]. The available cardiac monitoring systems are very expensive since they rely on many complicated hardware devices that employ complex algorithms [6]. Additionally, medical care costs have been exponentially increased due to the growing number of people suffering from chronic diseases such as diabetes and cardiac disease [7,8,9].

Typically, complicated measurement devices are utilized in hospitals worldwide to diagnose serious heart diseases. Meanwhile, these types of healthcare equipment require physical in-person examination by medical specialists to identify any arrhythmia issues, i.e., irregular heartbeats [10]. One strategy for overcoming the healthcare delivery challenges of chronic diseases is to utilize a remote patient management system [11,12]. LabVIEW-based patient real-time monitoring systems have been sought to cope with unaffordable medical care. This is due to LabVIEW offering an excellent Graphical User Interface (GUI) for remote tracking that can be shared online via the LabVIEW webserver. In addition, LabVIEW provides a more user-friendly interactive interface, which allows quick understanding and cognitive medical care. On the other hand, to allow a practicing as an assistant tool for medical doctors, solid knowledge and skills are strongly required [13]. This is because any misinterpretation of the vital signs increases the diagnostic error, which is estimated to be 0∼15% wrong in different specialties such as general practice, emergency, and internal medicine [13]. Machine learning (ML) techniques have been widely implemented for data mining of medical applications since they have proven to be excellent tools for analyzing big data and revealing hidden patterns that can assist in making precise medical decisions [14,15]. Data mining in the medical field deals with patient outcomes, thereby applying machine learning techniques necessitates precise and careful decisions to achieve a high level of accuracy [16].

AI algorithms can accurately interpret and extract patterns from the vital signs and medical examinations of cardiac patients, particularly in identifying critical symptoms. These techniques enable the detection and precise identification of anomalies, facilitating early signs detection of cardiovascular diseases and timely intervention. This capability can significantly improve cardiac care and potentially save lives [17]. Moreover, identifying risk factors for heart-related issues is essential for effective patient risk stratification, which can direct management and preventive measures to decrease related morbidity and healthcare costs [18,19]. Due to the numerous and diverse predictors affecting health-related factors, machine learning has emerged and proved to be a leading technique [19]. Consequently, the accumulation of healthcare for personalized treatment can be facilitated through employing machine learning approaches to help analyze patients’ symptoms and predict whether or not a patient has developed risk factors that require urgent treatment. In this regard, only patients exhibiting severe symptoms will be transferred to the emergency room or ICU [20,21].

In response to these challenges, this paper presents a novel machine learning-based model that can be used to precisely anticipate the status of cardiac patients. The model employs a multi-class classification technique that encodes the most crucial risk factors and vital signs to reduce the computation complexity while increasing the prediction accuracy, achieving $97.3\%$ and $100\%$ minimum and maximum accuracy, respectively, based on K-fold cross-validation analysis. We have developed a real-time tracking system designed to monitor the health status of cardiac patients remotely, enabling comprehensive performance assessments that enhance the stability and efficiency of the healthcare system for cardiac patients. Below, we itemize the contributions of this manuscript as follows.

• We curated a balanced dataset of 750 patient records from three hospitals, categorized into three health statuses, namely stable, unstable non-critical, and unstable critical. This dataset serves as a valuable resource for training and evaluating machine learning models.
• Development of an efficient ANN classifier designed to achieve competitive accuracy compared to state-of-the-art ML models while optimizing the architecture by reducing the number of hidden-layer neurons.
• Introducing a novel encoding-based ML algorithm proposed that surpasses conventional methods, such as Naive Bayes, SVM, and KNN, in classification accuracy for predicting cardiac patient health, achieving up to 98.8% accuracy.
• The proposed encoding-based algorithm has demonstrated the efficacy of the encoding-based ML approach in enabling accurate predictions, supporting proactive monitoring and timely medical interventions.
• Implementing a LabVIEW-based monitoring system by integrating the proposed framework into a LabVIEW environment with an intuitive GUI, facilitating real-time remote monitoring for cardiac patients. This system enhances medical decision-making by allowing medical doctors to track patient status and provide immediate essential consultations and therapy.

The remainder of the article is organized as follows: Section 2 provides a literature review and discusses the state-of-the-art of various machine learning approaches employed at different abstraction levels for medical applications. Section 3 covers the description of the dataset and the preprocessing for feature extraction. The proposed models for cardiac patient status prediction are discussed in Section 4. Section 5 analyzes the performance of the presented model and evaluates the obtained results by comparing them with different machine learning techniques considering effective evaluation metrics such as accuracy, precision, and F1-score, while Section 6 concludes the paper.

2. Related Works

Applying machine learning concepts has been carried out in several studies [8,22,23,24,25], dealing with a variety of challenges that the healthcare industry encounters. These scholarly studies aim to support the medical field by thoroughly analyzing the medical information of a patient and providing excellent decisions that help improve healthcare systems to potentially open new industrial paths for different applications. Data analytics in medical diagnostics has been substantially employed to develop risk prediction models that involve machine learning approaches for decision support systems, intervention, and resuscitation [26,27] to lower healthcare costs.

The authors in [14] carried out a comparison between implementing a multilayer back-propagation perceptron network (MLPN) and Support Vector Machine (SVM) algorithms to diagnose five classes of heart disease with a dataset of 303 patients. The results revealed that applying SVM for online classification achieves higher prediction accuracy by delivering more accurate decisions. An economic model that also implements the SVM algorithm was proposed in [28]. It was simulated as an alternative to DNA testing to identify possible threats that form the foundation of diagnostic tests and disease prevention. This indicates that the presented model is able to predict unhealthy lifestyles arising from diseases that patients might suffer from. The developed model can provide a detailed report that tracks changes in an individual’s diet and weight to help maintain a healthy lifestyle. In [29,30], the authors applied machine learning techniques to conduct a survey that predicts heart diseases at the early stages. They employed various ML-based algorithms such as SVM, Naive Bayes, Decision Tree, K-Nearest Neighbor (KNN), and ANN to aware cardiac patients in case they develop complications. The survey concludes that higher performance and classification accuracy can be realized when the input features are selected efficiently by applying search methods first to specify only the most suitable features. Likewise, ML-based concepts have also been employed in [31] to improve the healthcare system through patient satisfaction by determining related characteristics of various parameters. The model concentrates on specific criteria to build a safe and supportive healthcare environment. However, the decision-making process is not adequately matured to provide a precise decision ($88\%$ regression accuracy).

Predicting hospitalization emergencies for cardiac patients was studied in [8,32]. The model developed in [8] takes several risk factors, such as age, disease types, disease conditions, etc., as input features and analyzes them to identify those associated with the high risk, whereas the authors in [32] employed data mining techniques to specify patients at high risk so that they could accurately and efficiently manage the admission of inpatients at the emergency department. Even though the predictive model facilitates emergency admissions and patient satisfaction, it only achieves a maximum classification accuracy of $80.31\%$. Therefore, clinical examination remains necessary. Additionally, IoT-based patient monitoring systems have been widely researched in the literature. The researchers in [22] proposed an IoT-based healthcare system that leverages remote monitoring. This would help in advancing the healthcare business by reducing consultations and in-person meetings with doctors. However, integrating the classification model into the monitoring system requires a real-time GUI-based display to indicate potential risks at early stages and then provide timely treatment and life care. In [33], the authors built an IoT-based architecture that can be employed in various types of online systems. Nevertheless, deep investigations are still in demand to achieve more customized models that provide an acceptable level of security and prediction outcomes.

Alternatively, cardiovascular diseases (CVDs) were investigated in [34] by applying ML techniques to predict whether people have heart disease or not, thereby providing a further diagnosis and awareness to the cardiac patients. The researchers combined different ML algorithms such as SVM, Naive Bayes classifier, Random Forest, and Logistic Regression to achieve hybrid ML-based algorithms. Pre-processing techniques were used to remove the noisy or missing data, auto-fill with default values if possible, and classify attributes for prediction and deciding different levels. The logistic regression model showed better data analysis and higher accuracy compared to the other aforementioned algorithms. However, the developed model fails to realize high-level accuracy, delivering $91.6\%$. Similarly, different supervised ML algorithms were applied in [35] to perform binary classification on 91 records and determine whether a patient has heart disease or not. It was shown that the Random Forest (RF) algorithm achieved the highest accuracy and F1-score ($83.52\%$ and $84.21\%$, respectively). This is consistent with the results in [36], which confirms that the RF algorithm provided the highest accuracy, achieving ($96.61\%$) accurate classification. To further improve the classification accuracy of heart diseases, a deep neural network was deployed in [1] to classify diseases of the coronary artery. The developed model delivers an accuracy of $98.15\%$. However, this was achieved at the expense of incurring complicated structure and massive computation operations. In [17], a scalable architecture for early cardiovascular disease prediction was proposed based on optimal feature selection approaches, such as FCBF and PSO optimization. Nevertheless, the developed system only concentrates on improving prediction performance while neglecting to offer real-time monitoring and warning outcomes.

The authors in [9] proposed a hybrid machine learning technique that employs different boosting approaches of machine learning to increase accuracy. The model involves a complicated feature extraction by combining Relief and the Least Absolute Shrinkage and Selection Operator (LASSO) to deliver accuracy of $99.05\%$. Similarly, an intelligent health surveillance system based on Fuzzy Logic for cardiac patients was introduced in [10]. The system is used to monitor patients with critical arrhythmia conditions via IoT-based health tracking to make an autonomous diagnosis. Likewise, the authors in [37] presented a Principal Component Heart Failure (PCHF) feature engineering approach, which identifies the most prevalent features to provide optimal hyperparameters. The proposed approach employs machine learning methods (PCHF with Decision Tree) to identify heart failure with a prediction accuracy of $100\%$. However, the proposed method fails to provide real-time remote monitoring of cardiac patients. In [19], machine learning algorithms were used for COVID-19 infection prediction, in which the Random Forest classification algorithm was demonstrated to deliver the highest classification performance of $97.4\%$.

Driver fatigue detection using electroencephalography (EEG) signals and deep neural networks was explored in [38], where five types of fatigue were classified. The study employed generative adversarial networks (GANs) and graph convolutional networks (GCNs) for preprocessing and feature selection. While the model demonstrated impressive classification performance, it exhibited higher computational complexity compared to other counterpart methods presented in the literature. Similarly, EEG signals were utilized in [39] to analyze human emotions triggered by different events. Following a similar approach to [38], a combination of graph theory and GAN was applied to achieve accurate recognition of various facial emotional categories. However, the proposed network architecture was computationally intensive, necessitating substantial resources to achieve real-time classification.

Prior approaches to monitoring cardiac patients often involve overly complex processing methods or rely on traditional machine learning techniques, which are suitable for making single decisions but struggle to adapt to complex applications, such as integrating new features and output labels. This presents a challenge in accurately modeling the conditions of cardiac patients. Consequently, there is an ongoing pursuit for alternative solutions that can reliably operate in real-time and achieve higher classification accuracy. In this paper, we concentrate on monitoring and quantifying various risk factors related to cardiac patients to improve the quality of healthcare. The developed system is designed to monitor the conditions of cardiac patients and share their health information with authorized medical facilities via a LabVIEW web server. Ultimately, this study aims to provide remote health tracking for cardiac patients, assisting medical doctors in online monitoring of their patients’ statuses through a GUI that displays the predicted health status of each patient.

3. Dataset Description

Our structured dataset has been collected from two to three hospitals located in Mosul city, and the output classes have been verified by more than three specialist medical doctors. The dataset consists of 750 information reports of cardiac patients, and each record has 13 features of which 10 are considered risk factors that represent the most important input features, listed in Table 1 under medical history and examination. The first five rows of Table 1 describe the selected features of our collected dataset. As shown, the type of these features can be either continuous or nominal. Features that deal with Boolean expressions (false/true) are considered nominal while the others are continuous variables. The dataset was classified as stable, unstable non-critical, and unstable critical cases. These classes were collected in a way that creates a balance among the output labels.

Table 1. A cardiac patient’s risk factors (features), health status, and required therapy.

Risk Factors and Extracted Features
Medical History		Medical Examination
Features	Type	Features	Type
Age	Continuous	Heart Rate	Continuous
Body Mass Index (BMI)	Continuous	Respiratory Rate	Continuous
Chest Pain (CP)	Nominal (0 → NO; 1 → YES)	SPO2	Continuous
Shortness of Breath (SoB)	Nominal (0 → NO; 1 → YES)	Blood Pressure	Continuous
Smoking	Nominal (0 → NO; 1 → YES)	Glasgow Coma Scale	Continuous
Possible Labels and Required Medical Therapy
Health Status (Labels)		Medical Therapy
Stable		Stay Home (SH)
Unstable Non-critical		Emergency Room (ER)
Unstable Critical		Intensive Care Unit (ICU)

To analyze the dataset, first, preprocessing is conducted to convert the ten most popular risk factors into five vital signs. This conversion process is carried out using a clinical decision-making comprehensive guide [13,40] to ensure excellent delivery of efficiency and safe medical care. These crucial vital signs can greatly aid medical doctors in rapidly delivering the necessary medical therapy. Hence the preprocessing for feature selection has been handled carefully. This is due to crucial information related to the modeling of the problem possibly being discarded, leading to degrading the system’s overall accuracy [41].

Table 2 outlines the necessary sensors and communication components for the proposed monitoring system for cardiac patients. These sensors are crucial for effective implementation of the system, while Table 1 presents the data volume from the sensors to illustrate the feasibility and realization of the system. It is important to note that the data types can be either continuous or categorical. Lastly, Table 3 lists the statistical summary of the collected dataset samples. These features include cardiac disease such as age, BMI, chest pain, shortness of breath, smoking, and so on, which mainly causes the hospitalization of cardiac patients.

Table 2. The required sensors and input data categories for practical cardiac system realization.

Factors	Applicable Sensors
Gender, Age, and Smoking	User Interface (input)
Body Mass Index (BMI)	Weight Sensor (e.g., load cell), Height Sensor (e.g., ultrasonic/infrared)
Chest Pain	User Interface (input)
Shortness of Breath	Piezoelectric Belt or Pulse Oximeter
Blood Pressure (Systolic and Diastolic)	Blood Pressure Monitor (Oscillo-metric)
Heart Rate	Heart Rate Sensor (e.g., photoplethysmography)
Respiratory Rate	Respiration Sensor (e.g., chest belt or nasal airflow sensor)
Glasgow Coma Scale	Clinician Assessment (e.g., EEG optionally for brain activity)
SPO2	Pulse Oximeter

Table 3. Statistical summary of the collected dataset for cardiac patients.

Factors	Min.	Max.	Mean ($\overline{\mathit{x}}$)	Std. Dev. ($\mathit{S}$)
Gender	0	1	$0.627$	$0.484$
Age	23	87	$54.159$	$12.267$
BMI	22	35	$28.86$	$2.375$
Chest pain	0	1	$0.665$	$0.472$
Shortness of breath	0	1	$0.536$	$0.5$
Smoking	0	1	$0.541$	$0.499$
Systolic	50	200	$133.33$	$32.421$
Diastolic	50	130	$80.536$	$18.821$
Heart rate	45	140	$87.181$	$17.2$
Respiratory rate	10	22	$14.231$	$2.808$
Glasgow coma scale	8	15	$13.067$	$1.926$
SPO2	52	98	$91.752$	$6.01$
Class	0	2	1	$0.817$

4. The Proposed Real-Time Monitoring System

To enable mobility of cardiac patients and improve their life quality, herein we propose a tracking and supervision real-time system that monitors the vital signs of a cardiac patient. The system analyzes the received data from the patients and then performs machine learning operations to classify the status of the patient whether it is stable, unstable non-critical, or unstable critical, illustrated in Figure 1. In case of a stable state, the patient should be advised to stay at home and maintain monitoring, while in the case of an unstable state, the patient should be taken to the nearest associated healthcare facility or hospital to receive the necessary medical therapy at the appropriate time. In the event of an emergency, the system sends out alert signals to indicate that medical assistance is urgently required. Patients in the ICU require proper monitoring of their blood circulatory and respiratory systems since these are time-critical processes that must be monitored regularly and intensively. Therefore, this article proposes two machine learning approaches that can be employed to predict the status of cardiac patients, listed in the lower portion of Table 1. Next, the proposed models are discussed.

Figure 1. State transition diagram for the cardiac patient’s status condition.

4.1. Developed ANN Model

Here, neural networks are employed to predict the status of cardiac patients, as depicted in Figure 1. An artificial neural network (ANN) with one fully connected hidden layer and an output layer of three classes was utilized. The hidden layer is a single layer since the inputs of the network are extracted features that do not require deep analysis to be trained for making correct decisions and therefore the structure of the developed ANN is not complicated. For model training and performance evaluation, Stochastic Gradient Decent with Momentum (SGDM) of $0.9$ provided the best fine-tuning, the learning rate was set to be 0.001, and maximum epochs of 60 were configured for the model to terminate the training. The training time per iteration is 14 s, and thus the performance was found to incur 829 s for completing all epochs, which confirms the straightforward training process of the developed ANN classifier. However, we could not determine how to construct an optimized number of neurons in the hidden layer, so modeling the dataset required intensive examinations of different design structures. Thus, a trial and error technique was used to determine the required number of neurons in the hidden layer. We started with from 10 to 70 neurons, and it turned out that the accurate solution can be realized when the number of neurons in the hidden layer equals 37, illustrated in Figure 2. As shown in Figure 3, we conducted various setups of the loss function for model fine-tuning via applying adaptive optimizers like Adam, SGD with momentum, etc., associated with increasing the number of neurons to more than 37; however, we found out that does not improve the accuracy ($98.4\%$ was the best achieved accuracy), and instead, it becomes worse.

Figure 2. The structure of the developed ANN classifier.

Figure 3. Accuracy vs. number of neurons in the hidden layer of the developed ANN model structure.

Additionally, the developed ANN model, containing 37 neurons, is specifically designed for a task-oriented dataset involving cardiac patients. Therefore, if the dataset type changes, the ANN classifier must be retrained to adapt and optimize for the new data. For this reason, we switched to come up with a more fine-grained model that can deliver higher prediction accuracy while incurring less intensive computational complexity, i.e., a lightweight model. Next, we discuss the proposed encoding-based machine learning model.

4.2. Feature Selection Based on PCA Algorithm

We primarily employed the Principal Component Analysis (PCA) algorithm to reduce the dimensionality of our dataset. PCA helps prioritize the most crucial features that contribute to the variance in the data. This process not only simplifies the dataset but also removes redundant or less significant features, allowing the model to learn more effectively from the available data. The statistical analysis of applying the PCA as a preprocessing step assisted in making the proposed encoding-based algorithm more robust, flexible, and applicable to various datasets. The analysis presented in Figure 4a revealed that the PCA retained 12 principal components, which together explained 80% of the variance in the dataset. Consequently, only the most informative components were considered, leading to significant improvements in classification outcomes for cardiac patients. Notably, the features contributing most significantly to the PCA components were found to be SPO2, GC, RR, HR, blood pressure (including diastolic and systolic (Dia. and Sys.)), CP, and SoB.

Figure 4. Applying the PCA algorithm for dimensional reduction, (a) variance level for PCA components, (b) components stability.

On the other hand, the clustering shown in Figure 4b demonstrates that the first two principal components (PC1 and PC2) effectively capture the key variations in the dataset concerning criticality and stability. Specifically, PC1 represents primary indicators of stability and criticality, such as HR and blood pressure (both systolic and diastolic). In contrast, PC2 reflects secondary factors, including SPO2, GC, and RR. As seen in Figure 4b, these five vital signs are insufficient to accurately differentiate the model classes. Therefore, we encoded the CP and SoB, along with the previously determined vital signs, to realize outstanding classification accuracy, and we discuss it further in the next section.

4.3. Proposed Encoding-Based Machine Learning Model

ML models evaluate a patient’s risk for heart attack, stroke, or sudden cardiac arrest by analyzing risk factors such as age, chest pain, and shortness of breath. This assessment helps in identifying heart-related issues in real-time and provides preventive care. In the proposed encoding-based model, the risk factors are encoded to calculate the risk score (RS) and vital score (VS), which represent the preprocessing for the feature extraction process. The calculation of the risk factor score is performed based on encoding two features (chest pain and shortness of breath) that have high priority for being important symptoms as listed in Table 4. However, this is due to applying the PCA algorithm for providing initial preprocessing, demonstrating that only the CP and SoB have a strong influence on impacting the model performance. Hence, we only considered these two crucial risk factors in the proposed encoding-based algorithm. It can be seen from Table 4, for instance, when $RS$ is equal to 0, this implies that the patient does not have an issue with these signs, whereas $RS=1$ indicates that the patient is suffering from shortness of breath. When $RS=3$, then the patient has both chest pain (CP) and shortness of breath (SoB). Therefore, they are encoded with two bits ($b_0$ and $b_1$), which are equivalent to SoB and CP, where CP represents the most significant bit (MSB = $b_1$) and SoB represents the least significant bit (LSB = $b_0$). The values of $RS$ are then encoded (steps 6 to 11 in the pseudo-algorithm) and stored in a lookup table (LUT1).

Table 4. Risk factor score.

Risk Factors
Risk Score (RS)	Chest Pain?	Shortness of Breath?
0	0	0
1	0	1
2	1	0
3	1	1

Vital score (VS), on the other hand, has been encoded based on the five most important vital signs, including Glasgow coma (GC), SPO2, systolic/diastolic blood pressure (BP), heart rate (HR), and respiratory rate (RR), which is equivalent to ($b_4$$b_3$$b_2$$b_1$$b_0$), respectively. These vital signs have been considered the highest priority factors to guide labeling the output classes for the classification process. We represent each factor with a single bit; thus, five bits are needed for encoding the vital signs and calculating the value of $VS$, as illustrated in Algorithm 1 from step 12 to step 26. Afterward, the values of the vital score are stored in a lookup table (LUT2). The encoded values of both RS and VS are then fed to 32 case statements to predict the status of a cardiac patient. The output condition of the comparison process, along with the vital sign score, is then fed to a lookup table (LUT3/LUT4) to provide the predicted status of the patient, as illustrated in Figure 5. Because the output of the condition is a Boolean (false/true), 64 possible statements are stored in two LUTs (LUT3 and LUT4) with a capacity of 32 locations. These LUTs work in parallel with a control signal for read enable. The VS value (five bits) is used as the address lines for both LUTs, whereas the condition is inverted and connected to the read enable (RE) signal of LUT3 and directly connected to the RE signal of LUT4. Thus, if the condition becomes false, then the status stored at the location indicated by the VS value of LUT3 goes to the output to be the predicted status. Otherwise, the statement in the equivalent location of LUT4 passes to the output.

Figure 5. The proposed encoding-based machine learning model.

To further understand the working mechanism of the encoding-based model, we provide a case study that takes an example of three different cardiac patients, listed in Table 5 as follows:

Algorithm 1 Pseudocode algorithm of the encoding-based model

1: procedure MyProcedure   2:     for $i\leftarrow 1$ to N do   3:         Read patient information   4:         Applay the PCA module to select the most important features (Risk Factors)   5:         Calculate: Risk Score (RS)=? (LUT1); As follows:   6:         if ($Chestpain==Yes$) then CP =1   7:            Else  CP = 0   8:         end if   9:         if ($Shortnessofbreath==Yes$) then SoB =1 10:            Else  SoB = 0 11:         end if Vital Score (VS)=? (LUT2); As follows: 12:         if ($ComaScale<13$) then GC =1 13:            Else  GC = 0 14:         end if 15:         if ($SPO2<94\%$) then SPO2 =1 16:            Else  SPO2 = 0 17:         end if 18:         if ($Systolic<100$ OR $Systolic>140$) OR ($Diastolic>90$ OR $Diastolic<60$) then Systolic/Diastolic (BP) = 1 19:            Else  Systolic/Diastolic (BP) = 0 20:         end if 21:         if ($HR>100$ OR $HR<60$) then HR =1 22:            Else  HR = 0 23:         end if 24:         if ($RR<12$ OR $RR>16$) then RR =1 25:            Else  RR = 0 26:         end if 27:         Case structure: 28:         Begin: 29:         Select the suitable case statement (case 0 to 31) 30:         if $(condition==true$) then Status = Non-critical OR Critical 31:         else Status = Stable OR Non-critical 32:         end if 33:         Index = VS 34:         if $(condition==true$) then 35:            Read LUT4(index) for decision making 36:         else 37:            Read LUT3(index) for decision making 38:         end if 39:         Select the predicted status (SH, ER, ICU) 40:         EndCase; 41:         goto step3 42:     end for 43: end procedure

Table 5. Case study: patients’ medical information for model testing.

#	Name	CP	SoB	Sys.	Dia.	HR	RR	GC	SPO2 (%)	Actual Status
1	Person1	YES	NO	90	50	93	12	10	87	Critical
2	Person2	YES	NO	100	80	83	13	14	97	Stable
3	Person3	NO	YES	150	70	95	13	15	92	Non-critical

Person1: the values for $RS$ and $VS$ are first calculated based on the risk factors. As listed in Table 5, the patient suffers only chest pain from the risk factors, so RS = 2, while VS = 28 because the patient has GC, SPO2, and BP lower or higher than that set of the threshold; therefore, the bit value of these vital signs becomes ‘1’ (refer to Algorithm 1 for the equations that determine the threshold values of these signs). The values of HR and RR are within the continuous range for normal operations of vital signs, and thus they are both set to ‘0’. Putting it all together, ($b_4$$b_3$$b_2$$b_1$$b_0$), it becomes (11100₂), which is equivalent to 28 in decimals, where the GC represents the MSB and RR represents LSB. The value of VS is then fed to the case structure to select the suitable statement, stable, non-critical, and critical. As seen in Figure 6a, the selected case is Case#28, so the status is either non-critical (emergency room) or critical (ICU). This can be determined based on both the value of the VS and the condition (false/true) of the selected case since they will then be used as the address and read enable (RE) to read one of the statuses stored in LUT3 when the condition is false or LUT4 when the condition is true (see Table 6). It is worth mentioning that each condition has two equivalent statuses. One is when the condition is false and stored in LUT3 at the location indexed by VS, while the others for the true condition are stored at the identical location in LUT4. Since the condition is true (see Figure 6a, row1 of Table 5 and Table 6), the predicted status is read from LUT4 and the model has classified it as critical, and thus both the actual and predicted status are the same. Note that Figure 7 illustrates the programming of Case#0 (default); however, there exist an additional 31 cases that are selected by the vital score.

Figure 6. Required case statements for the case study: (a) equivalent case structure for patient1, (b) equivalent case structure for patient2, (c) equivalent case structure for patient3.

Table 6. Encoding of conditions to enable either LUT3 or LUT4 for decision-making and classification.

Switch Case (VS)		Condition?
Vital Score	Case #	False (LUT3)	True (LUT4)
0	Case 0	Stable (SH)	Non-Critical (EM)
1	Case 1	Stable (SH)	Non-Critical (EM)
2	Case 2	Stable (EH)	Non-Critical (EM)
3,4	Case 3,4	Stable (SH)	Non-Critical (EM)
5	Case 5	Stable (SH)	Non-Critical (EM)
6	Case 6	Stable (SH)	Non-Critical (EM)
7	Case 7	Non-Critical (ER)	Critical (ICU)
8	Case 8	Non-Critical (ER)	Critical (ICU)
9	Case 9	Non-Critical (ER)	Critical (ICU)
10	Case 10	Non-Critical (ER)	Critical (ICU)
11	Case 11	Non-Critical (ER)	Critical (ICU)
12	Case 12	Non-Critical (ER)	Critical (ICU)
13	Case 13	Non-Critical (ER)	Critical (ICU)
14	Case 14	Non-Critical (ER)	Critical (ICU)
15	Case 15	Non-Critical (ER)	Critical (ICU)
16,17	Case 16,17	Non-Critical (ER)	Critical (ICU)
18,19	Case 18,19	Non-Critical (ER)	Critical (ICU)
20	Case 20	Non-Critical (ER)	Critical (ICU)
21	Case 21	Non-Critical (ER)	Critical (ICU)
22	Case 22	Non-Critical (ER)	Critical (ICU)
23	Case 23	Non-Critical (ER)	Critical (ICU)
24	Case 24	Non-Critical (ER)	Critical (ICU)
25	Case 25	Non-Critical (ER)	Critical (ICU)
26	Case 26	Non-Critical (ER)	Critical (ICU)
27	Case 27	Non-Critical (ER)	Critical (ICU)
28	Case 28	Non-Critical (ER)	Critical (ICU)
29	Case 29	Non-Critical (ER)	Critical (ICU)
30	Case 30	Non-Critical (ER)	Critical (ICU)
31	Case 31	Non-Critical (ER)	Critical (ICU)

Figure 7. Main part of the LabVIEW block diagram for the proposed encoding-based model.

Person2: the model first calculates the risk score and vital score, which are RS = 2 and VS = 0. The patient has chest pain, and thus CP = 1, while VS = 0 since all selected vital signs are normal. Thus, the proposed encoding-based model will forward this to Case#0. As can be found from Figure 6b, the condition is false because the RS is not equal to 3 and also none of the two inputs of the first OR gate is true. Hence, because the condition is false, then the status stored at location 0 is read from LUT3, which is stable. Notice that LUT3 and LUT4 store two states of status for each condition. This is due to the status of a patient rarely being a combination of stable or unstable critical at the same time; instead the possible combinations are (stable or unstable non-critical) and (unstable non-critical or unstable critical). This is mainly attributed to the state transition of a patient’s condition at any moment commonly moving from stable to unstable non-critical as it has to go through the state transition diagram, as illustrated in Figure 1, where the dashed line indicates the condition rarely happens.

Person3: the calculated values are RS = 1 and VS = 12, and thus Case#12 is selected and is depicted in Figure 6c. As can be seen, both comparisons must be true to drive the output condition to be true. Therefore, the condition is false and the status is classified according to the state stored in LUT3. From Figure 6, it can be noticed that the risk factor score has been involved in the processing of all case structures as a second factor. This is due to the fact that the medical information of a cardiac patient is correlated and dealing with it in a general processing method can lead to misclassifying the required medical therapy. Therefore, RS is encoded in all case statements to assist in breaking down the correlation, thereby increasing the correct prediction and thus improving the model’s overall accuracy. Notice that the considered observations (patients 1,2, and 3) in the case study are new data samples and the proposed approach did not see these samples in the training and testing, which confirms the model’s self-learning ability to provide correct predictions for unseen data samples.

Consequently, the proposed model can help medical doctors to be more confident about the decisions they make to provide exact medical therapy for cardiac patients. As depicted in Figure 1, earlier precise decision-making would strongly assist in preventing further complications that might lead to mortality. Furthermore, the proposed monitoring system can be applied to deliver seamless cardiac patient observation. For example, the status of a patient may move from unstable non-critical to unstable critical, and therefore, quick medical therapy becomes urgent to return the patient’s status to uncritical and then stable. Hence, approaches that deliver reliable and timely decision-making have been sought to accurately handle heart problems.

5. Experimental Setup and Results Evaluation

To build the ANN model, MATLAB R2022b is used, whereas the encoding-based model is constructed in LabVIEW 2017 since it provides a real-time and user-friendly GUI for monitoring. Additionally, the LabVIEW environment contains a comprehensive set of tools utilized for acquiring, analyzing, storing, and displaying data from real-time applications. Figure 7 depicts the main part of the designed graphical programming block diagram for the proposed encoding-based machine learning model. As can be seen, every feature is considered either as a nominal or continuous value. Then these values become encoded and compared to several threshold values to classify three classes of status and compute the accuracy accordingly.

To evaluate the performance of the proposed model, the dataset is divided into two groups, training and testing, with $70\%$ of the data used for training and the remaining $30\%$ utilized for testing (225 patients). Then, K distinct learning trials were executed by partitioning the data into K groups using the K-fold cross-validation, where K is set to 10 ($10\%$ of the testing data) in each testing cycle, resulting in the formation of 10 groups. Then, we run the presented encoding-based model and evaluate its accuracy. Figure 8 displays the testing window that shows the validation process of the encoding-based model. Clearly, it can be seen that both the actual (hospital) and predicted labels provide the same status, which indicates that the status of patient 136 (out of 225 for testing data) is unstable non-critical and the patient should be taken to the emergency room.

Figure 8. The LabVIEW front panel testing window of the proposed encoding-based model. Note that the yellow indicators signify the "ON" state for the corresponding risk factor.

5.1. Performance Evaluation

Here, we evaluate the performance of the proposed ANN and encoding-based model by comparing the obtained accuracy of each one with that of various algorithms including Fine Tree, Boosted Trees, Linear Discriminant, Naive Bayes, SVM, and KNN. As listed in Table 7, the proposed encoding-based model realizes the highest level of prediction accuracy ($98.8\%$) with a misclassification rate of $1.2\%$. The developed ANN model also provides a competitive accuracy; however, it incurs 37 neurons in the hidden layer to build a suitable structure. It can be seen that both the Gaussian Naive Bayes and the Quadratic SVM also achieve the third and fifth prediction accuracy ($94.8\%$ and $96\%$, respectively). To further evaluate the performance of the SVM algorithm, Naive Bayes classifier, proposed ANN, and encoding-based models in more detail, a confusion matrix is drawn. The confusion matrix is used to evaluate the effectiveness of classification models by categorizing data into various classes of true/false and positive and negative rates. Figure 9 depicts the confusion matrix of the presented ANN model. It can be noticed that the confusion happens between the first and second class (stay home and emergency room) and the second and third class (emergency room and ICU). This is because there are some correlated signs attributed to different classes and differentiating them requires more intensive computations by stacking more hidden layers. From Figure 10, it can be seen that the best validation performance (0.0263) was achieved at epoch 54, while the set goal was 0.01. This confirms that the model is capable of providing an accuracy of $97.37\%$ for the validation subset of data, but it utilizes a considerable number of neurons in the hidden layer (37 neurons).

Table 7. Accuracy of the proposed approaches compared to that of different ML-based multi-class classification algorithms.

ML-Based Algorithm	Accuracy (%)	Failure Prediction Rate (%)
Decision Tree (DT)	$92.8$	$7.2$
Medium Tree	$93.2$	$6.8$
XGBoost	$94.1$	$5.9$
RUSBoosted Trees	92	8
Linear Discriminant	$83.7$	$16.3$
Gaussian Naive Bayes	$94.8$	$5.2$
Kernel Naive Bayes	$92.8$	$7.2$
Linear SVM	$87.7$	$12.3$
Quadratic SVM	96	4
Medium Gaussian SVM	$95.9$	$4.1$
Fine KNN	90	10
Medium KNN	$85.1$	$14.9$
Weighted KNN	$89.6$	$10.4$
Developed ANN Model	$98.4$	$1.6$
Proposed Encoding-based Model	$98.8$	$1.2$

Figure 9. Confusion matrix of the developed ANN classifier: (a) training results, (b) validation results, (c) test results, (d) model’s overall performance evaluation.

Figure 10. Performance validation of the developed ANN model. Note that the green circle indicates the best performance intersection of validation and testing.

On the other hand, Figure 11 illustrates the confusion matrix of the Naive Bayes, Quadratic SVM, and proposed encoding-based algorithms. The encoding-based model achieves the lowest confusion between SH and ER and ER and ICU as it realizes the highest Positive Predictive Values (PPVs) and imposes the lowest False Discovery Rates (FDRs). Furthermore, the probability of the evaluated models, SVM, and Naive Bayes classifier is used to identify the status by computing the likelihood of each risk factor and then evaluating the prediction performance using precision, sensitivity, and F1-score as evaluation metrics. As listed in Table 8, the encoding model provides the highest prediction accuracy for all evaluation metrics. This can be explained as follows: the preprocessing of the proposed model concentrates on identifying the most crucial features to calculate RS and VS, then addressing them to specific lookup tables (LUTs). Therefore, unlike the probabilistic approaches, the prediction for the encoding-based method is performed based on a more mature analysis than the process of conventional ML approaches.

Figure 11. Confusion matrix: (a) number of observations of Naive Bayes, (b) evaluation of True Positive Rate (TPR) and False Negative Rate (FNR) of Naive Bayes, (c) evaluation of PPV and FDR of Naive Bayes, (d) number of observations of Quadratic SVM, (e) evaluation of TPR and FNR of Quadratic SVM, (f) evaluation of PPV and FDR of Quadratic SVM, (g) number of observations of the proposed encoding-based model, (h) evaluation of TPR and FNR of proposed encoding-based model, (i) evaluation of PPV and FDR of the encoding-based model.

Table 8. The classification performance of proposed and the employed competitive models.

Model	Precision (%)	Sensitivity (%)	F1-Score (%)
Gaussian Naive Bayes	$96.6$	91	$93.7$
Quadratic SVM	$96.8$	$95.2$	$95.8$
Developed ANN	$98.2$	96	97
Proposed Encoding-based Model	$98.2$	$98.2$	$98.3$

As can be observed in Table 9, the presented model achieves excellent accuracy in each iteration, realizing $100\%$ accuracy in six groups. The overall accuracy after K iterations of cross-validation for the testing data is $98.18\%$. K-fold cross-validation has also been performed on all 750 records of the dataset, and the proposed encoding-based model delivers slightly higher accuracy ($98.8\%$) since, in this case, each fold contains more data samples so it can learn better. The minimum and maximum accuracy for the proposed model are $97.3\%$ and $100\%$, respectively. This demonstrates that the proposed model can perform very well even when the dataset contains hundreds of samples (less than 1k, as in this scholarly research study) or when there are some underrepresented data samples. Furthermore, the two-stage encoding technique for breaking down the correlation among a patient’s data, which involves calculating RS and VS, allows code refactoring through coding the extended comparisons with fewer lines of instructions.

Table 9. Accuracy evaluation of the proposed ML encoding model using ($10\%$) K-fold cross validation.

	Testing Dataset (30%)		All Dataset (100%)
K	Range	Accuracy	Range	Accuracy
K1	1 → 22	$100\%$	1 → 75	$97.3\%$
K2	23 → 45	$95.45\%$	76 → 150	$100\%$
K3	46 → 68	$100\%$	151 → 225	$97.3\%$
K4	69 → 91	$95.45\%$	226 → 300	$100\%$
K5	92 → 114	$100\%$	301 → 375	$100\%$
K6	115 → 137	$100\%$	376 → 450	$98.67\%$
K7	138 → 160	$95.45\%$	451 → 525	$97.3\%$
K8	161 → 183	$100\%$	526 → 600	$100\%$
K9	184 → 206	$95.45\%$	601 → 675	$98.67\%$
K10	207 → 225	$100\%$	676 → 750	$98.67\%$
	AVE	$98.18\%$	AVE	$98.8\%$

5.2. Status Monitoring

To provide medical doctors with a user-friendly GUI, the LabVIEW webserver is used to share the designed front panel. As can be seen in Figure 12, medical doctors can easily track the status of their cardiac patients by analyzing the indicated processed data (information) in the front panel of the GUI. The proposed healthcare monitoring system can be particularly helpful for cardiac patients who are required to consult with their doctor for follow-up to prevent developing more risk factors from emerging. Additionally, it would work well to reduce the expenses of visiting a healthcare facility or remaining in a hospital for an extended period of time. For instance, if a patient has angina pectoris, she/he should receive immediate treatment to prevent the emergence of risk factors that may lead to myocardial infarction [40]. Consequently, the developed monitoring system can be used to reduce care costs while maintaining cardiac patients’ health and providing them with more confidence and securing their health status. Therefore, this would highly assist in improving the overall performance of the healthcare system by delivering intelligent and dependable decision-making techniques.

Figure 12. Designed monitoring window (GUI) for LabVIEW webserver. Note that the yellow indicators signify the “ON” state for the corresponding risk factor.

On the other hand, the designed GUI prioritizes simplicity and clarity to ensure usability, particularly for monitoring cardiac patients. It is tailored to display critical patient information effectively while supporting multiple-patient monitoring. Additionally, all displayed information is automatically stored in an Excel sheet on the server computer, providing a robust solution for maintaining patient records and enabling further analysis.

6. Conclusions

This study presents an encoding-based machine learning approach to analyze the risk factors associated with cardiac patients, thereby identifying their health status. We introduce a LabVIEW-based remote monitoring and classification system strategy for addressing the healthcare delivery challenges of cardiac patients. The developed monitoring system utilizes a novel encoding-based ML technique to enhance the prediction outcomes for cardiac patients. Thus, leveraging the encoding technique for heart health status classification achieves an impressive predictive outcome of $98.8\%$. Additionally, two stages of encoding were exploited to perform code refactoring, which allows for the use of a lower instruction count in the coding of extended comparisons. Furthermore, since the platform is based on LabVIEW, the system can easily be shared over the Internet with the relevant healthcare facilities for remote monitoring. Meanwhile, it offers a favorable decrease in the number of attached instruments, which is considered a burden that makes the daily life of patients more difficult. Thus, the presented approach for health status prediction can be an alternative low-cost monitoring system. It helps to bring mobility for cardiac patients at a lower expense and reduces diagnostic errors by consistently analyzing data patterns, achieving a low false prediction rate ($1.2\%$).