Weight Initialization‑Based Rectified Linear Algorithm for Accurate

Prediction of Chronic Heart Disease Compared with PCHF Feature

Engineering Technique

S. S. Deepak Senni and M. Krishnaraj

St Joseph’s Institute of Technology, Chennai, Tamil Nadu, India

Keywords: Cardiovascular Disease Prediction, Supervised Learning, Artificial Intelligence (AI), Weight

Initialization‑Based Rectified Linear Algorithm (WiReL), PCHF Algorithm, Medical Data Analysis, AI

Cardiologists, Patient Risk Assessment, Generative AI, Federated Learning.

Abstract: Cardiovascular disease continues to pose a significant challenge to global health, underscoring the critical

need for early and precise prediction to enable effective preventive strategies. This paper investigates the

promising role of supervised learning techniques within the realm of Artificial Intelligence (AI) for predicting

heart disease. We explore notable advancements in various algorithms, including logistic regression (LR),

support vector machines (SVM), and deep neural networks (DNN), emphasizing their ability to uncover

intricate patterns within extensive medical datasets. Moreover, the research highlights the capacity of AI-

enhanced cardiologists to analyze a wide array of patient data, encompassing demographics, medical histories,

lab test outcomes, and ECG readings. Such comprehensive evaluations promise to enhance the accuracy and

personalization of risk assessments, potentially facilitating earlier interventions and improving patient

outcomes. This study also addresses the significant challenges related to data quality, the mitigation of biases,

and the explainability of AI models, highlighting the need for ethical considerations in their design and

deployment. We classify ECG stages utilizing two models: a Cardiology model based on Machine Learning

techniques with a specific dataset and a Deep Learning Model focused on identifying cardiovascular disease

through ECG image classification. Additionally, the application of the Weight Initialization-Based Rectified

Linear Algorithm (WiReL) for heart disease prediction underscores the integration of optimized weight

initialization principles along with ReLU activation within a deep learning context. Our findings demonstrate

that the WiReL algorithm outperforms the Principal Component Heart Failure (PCHF) Feature Engineering

Technique in terms of predictive accuracy. Furthermore, this paper discusses potential future advancements

in AI-driven heart disease prediction, considering the implications of emerging methodologies such as

Generative AI and federated learning to further enhance this vital field. Our proposed research offers

meaningful contributions to medical science and its endeavors in combating cardiovascular disease.

1 INTRODUCTION

Heart failure represents a condition in which the heart

struggles to pump an adequate volume of blood to

meet the body's demands. Cardiovascular illnesses

have emerged as a major worldwide health problem.,

profoundly impacting public health across the world.

Heart failure, in particular, is a common and deadly

disorder affecting millions of people. The

incorporation of machine learning into medical

diagnostics and the larger healthcare sector has shown

much potential. Its application is many, including

drug discovery, diagnostic imaging, outbreak

forecasting, and heart failure prediction. By deriving

insights from extensive medical datasets, machine

learning techniques facilitate predictive analysis.

When compared to conventional medical

methodologies, machine learning offers noteworthy

advantages, including savings in time and costs,

thereby enhancing diagnostic efficiency (

M. Qadri et

al., 2023)

We are pleased to share our key research

contributions towards the detection of heart failure

using machine learning. Our proposed novel WiReL

(Weight Initialization-Based Rectified Linear)

algorithm represents an innovative strategy that

combines advanced weight initialization techniques

with the attributes of the Rectified Linear Unit

618

Senni, S. S. D. and Krishnaraj, M.

Weight Initialization-Based Rectiﬁed Linear Algorithm for Accurate Prediction of Chronic Heart Disease Compared with PCHF Feature Engineering Technique.

DOI: 10.5220/0013902900004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 3, pages

618-626

ISBN: 978-989-758-777-1

(ReLU) activation function. This approach is

specifically designed to predict the likelihood of heart

disease. By training a deep learning model with

comprehensive patient data, including demographics,

medical histories, laboratory results, and ECG

readings, we aim to analyze electrocardiograms

(ECGs) with precision. Our system adeptly identifies

complex patterns within the data, learns hierarchical

structures automatically—minimizing the need for

elaborate feature engineering—and enhances training

efficiency, thus improving performance in deep

architectures when compared to the existing PCHF

engineering technique. This robust system has the

potential to assess a patient’s risk of enabling

healthcare providers to implement tailored preventive

measures and initiate timely interventions, ultimately

enhancing patient outcomes. Nevertheless, it is

imperative to uphold data quality, mitigate bias, and

ensure interpretability (or explain ability) of AI-

driven decisions, all of which are critical for the

responsible adoption of these technologies.

We advocate for the utilization of our WiReL

algorithm to further refine performance by focusing

on the most pertinent features. This technique not

only identifies but also prioritizes essential dataset

characteristics with significant importance values,

thereby advancing the development of machine

learning models. By innovatively assembling a new

feature set, we have optimized the WiReL framework

to attain superior accuracy scores compared to

previously established methods. Additionally, we

conduct hyperparameter tuning for each machine

learning approach to pinpoint the optimal parameters,

resulting in enhanced accuracy outcomes.

2 RELATED WORKS

This section presents a comprehensive review of the

literature pertinent to our proposed research study,

examining previous investigations concerning heart

failure prediction. The findings and methodologies of

related research are systematically discussed and

compared.

In Study (

A. U. Hassan et al., 2022), the focus is on

the broader category of cardiovascular disease,

commonly known as heart disease. This includes

different conditions impacting the heart, which has

regrettably become the leading cause of mortality

worldwide in recent decades. Given the multitude of

risk factors associated with heart disease, there is a

pressing need for accurate and reliable methods of

early diagnosis to facilitate timely treatment. Within

healthcare, data analysis is crucial for managing and

understanding large datasets. Researchers employ a

range of statistical and machine learning approaches

to examine complex medical data, supporting

clinicians in anticipating cardiac problems.

The study (

A. U. Hassan et al., 2022) looks at several

aspects of heart illness and proposes a prediction

model based on supervised learning techniques

including Random Forest (RF), Decision Tree (DT),

and Logistic Regression. Using an existing dataset

from the UCI Cleveland database, which has 303

entries and 76 characteristics, the study evaluates 14

chosen features. The aim is to forecast the chance of

individuals developing heart disease, and the data

show that logistic regression has the highest accuracy

score of any approach studied.

Furthermore, dynamic CT image sequence

registration is a crucial preprocessing step for the

clinical evaluation of numerous cardiac physiological

parameters., including both global and regional

myocardial perfusion. In an innovative Approach,

we suggest a deep learning-based image registration

method that is specifically designed for quantitative

myocardial perfusion CT evaluations. This method is

capable of overcoming unique challenges, including

low image quality with limited anatomical landmarks,

fluctuating contrast agent concentrations within heart

chambers, and alignment difficulties that result from

cardiac stress, respiration, and patient movement.

(

Saboor et al., 2022).

To accommodate for temporal local contrast

variations, the proposed method utilizes a recursive

cascade network, a ventricular segmentation module,

and a unique loss function. The model is trained and

validated using data from 118 individuals with known

or suspected coronary artery disease and/or aortic

valve insufficiency. The results show that the

approach effectively registers dynamic cardiac

perfusion sequences, decreasing local tissue

displacements in the left ventricle (LV) while

maintaining contrast accuracy and CT (HU) values

throughout the series. Furthermore, the deep learning

technology has incredible processing speed, greatly

surpassing previous picture registration algorithms,

highlighting its potential for standardizing

quantitative cardiac perfusion CT in routine clinical

use.

The importance of accurate and reliable diagnoses

of cardiac disease cannot be overstated, particularly

as the incidence of fatalities from heart attacks

continues to escalate (

S. Sarah et al., 2021). Early

diagnosis is critical for prompt treatment of cardiac

problems. Using datasets from the University of

California, Irvine (UCI) repository, numerous

supervised machine learning approaches have been

Weight Initialization-Based Rectiﬁed Linear Algorithm for Accurate Prediction of Chronic Heart Disease Compared with PCHF Feature

Engineering Technique

619

examined to predict cardiac disease, including K-NN,

DT, LR, Naïve Bayes, and SVM.

The findings suggest that Logistic Regression

surpasses other classifiers across performance metrics

(

P. Rani et al., 2021), exhibiting a lower risk level with

fewer false negatives, as highlighted by confusion

matrix comparisons. The potential for enhancing

classifier. accuracy through ensemble methods is also

noted. To facilitate the implementation of these

models, Jupiter Notebook serves as an effective tool,

providing an array of libraries and modules

supportive of precise and accurate analyses.

In recent years, there has been a significant uptick

in interest toward auxiliary diagnostic technologies

for cardiovascular disease, particularly through the

detection of abnormal heart sounds (

G. O. Young.,

1964)

. Heart sound signals hold great promise for the

early diagnosis of cardiovascular conditions. Previous

study has mostly focused on the local aspects of heart

sounds., this work presents a unique approach for

mapping complicated heart sound patterns into fixed-

length feature embeddings, known as HS-Vectors, for

abnormality detection. To successfully capture the

comprehensive embedding of complicated heart

sounds, HS-Vectors (

W.-K. Chen., 1993) are developed

employing the compressed time and time delay

frequency expansion. A Dynamic Masked Attention

(DMA) module supplements the TCFE-TDNN neural

network (

A. U. Hassan et al., 2022). HS-Vectors are

intended to extract and highlight critical global heart

sound properties by filtering out irrelevant

information. The TCFE-TDNN module converts

cardiac sound signals over defined time periods into

fixed-length embeddings. The DMA module, which

includes a learnable masked attention matrix,

combines multiscale hidden features from

TCFETDNN layers to efficiently remove

inconsequential frame-level items.

The method was carefully validated using 10-fold

cross-validation (

H. Poor., 1985) on both the 2016

Physio Net Challenge dataset and a newly collected

pediatric heart sound dataset. The findings indicate

that the proposed technique has promising potential

for improving cardiac disease prediction when

compared to existing cutting-edge models (

Saboor et

al., 2021).

3 PROPOSEDMETHODOLODY

In this study, we examine a heart failure dataset

sourced from Kaggle, which comprises 1,025 patient

records, including both cases of heart failure and

healthy individuals. To enhance the dataset's quality,

we implement a range of data preprocessing

techniques, followed by exploratory data analysis that

provides valuable insights into patterns and variables

associated with heart failure. We employed the

Weight Initialization-Based Rectified Linear

(WiReL) Algorithm for predicting heart disease,

focusing on identifying the most pertinent features to

optimize the model's performance.

The WiReL algorithm is specifically designed to

improve machine learning model effectiveness by

selecting the most significant features from the

dataset. The dataset encompasses a variety of

attributes, such as age, sex, BP, cholesterol levels,

ECG results, and other relevant medical data. WiReL

processes inpatient data—consisting of 13 features

from the Kaggle dataset—utilizing layers with ReLU

as the activation function to accurately model

complex relationships. By assigning weights to

features based on their importance, the dataset is

divided into training and testing subsets. Ultimately,

the model produces probability scores or class

predictions regarding heart disease, demonstrating

enhanced accuracy. Our proposed model shows

promise in surpassing existing predictive methods,

positioning it as an efficient tool for heart failure

prediction.

3.1 Predictive Model

Machine learning fundamentally depends on the

careful collection of data, particularly a significant

amount of historical and raw information. However,

it's important to note that raw data requires

preprocessing before it can be effectively utilized for

analysis. (

P. Rani et al., 2021) This crucial step ensures

that the data is transformed into a suitable format.

Once this preprocessing is complete, appropriate

methods and models are selected for the analysis. The

chosen model undergoes a training and testing phase

to verify its performance and accuracy in making

predictions while minimizing errors. To maintain and

enhance the model's accuracy over time, it is

beneficial to periodically fine-tune it, as exemplified

in Figure 1.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

620

Figure 1: Predictive Model Process.

By thoughtful adapting Lent’s architecture and

training methodologies to better suit heart disease

data, it stands to serve as a highly effective predictive

model, particularly for image- based diagnostic tools.

This adaptability allows LeNet to automatically

extract relevant features from medical images,

thereby minimizing the necessity for extensive

manual feature engineering (

K. Chohan and D. C.

Dobhal., 2022)

. Its relatively shallow architecture

makes it particularly well-suited for the small datasets

that are often encountered in medical research. In the

case of ECG plots, LeNet is capable of processing

these images directly. If necessary, it is advisable to

convert medical imaging data into grayscale format.

Additionally, applying augmentation techniques,

such as rotation and flipping, can enhance the model's

generalization capabilities. It is also important to

normalize pixel intensity values to ensure they fall

within a specified range, such as [0, 1]. To align with

the specific requirements of heart disease prediction,

adjustments should be made to LeNet’s output layer.

This entails replacing the original 10-class output

layer, designed for digit classification, with a binary

output layer for predicting the presence or absence of

the disease, or alternatively, multiple nodes for multi-

class classification to reflect varying stages of disease

severity. Depending on the classification approach,

using a sigmoid activation function for binary

outcomes or softmax for multi- class scenarios is

recommended. Finally, a comparative analysis

between LeNet and conventional manual predictive

models is illustrated, providing insights into the

relative performance and efficacy of each approach as

demonstrated in Figures 2 and 3 for LeNet, and

Figures 4 and 5 for the manual predictive model (

Sarah et al., 2022).

Figure 2: Graph for Model Loss for Lanet Predictive Model.

Figure 3: Graph for Model Accuracy for Lanet Predictive

Model.

Figure 4: Graph for Model Accuracy for Manual Predictive

Model.

Weight Initialization-Based Rectiﬁed Linear Algorithm for Accurate Prediction of Chronic Heart Disease Compared with PCHF Feature

Engineering Technique

621

Figure 5: Graph for Model Loss for Manual Predictive

Model.

3.2 Data Pre-Processing

Validation techniques in machine learning estimating

the error rate of an ML model with the intent of

closely approximating the true error rate of the

dataset. While a large and representative dataset can

sometimes reduce the necessity for rigorous

validation, it is important to acknowledge that real-

world scenarios often present challenges where data

samples may not fully encapsulate the population

they aim to represent (

P. Rani et al., 2021).

In this context, validation becomes critical in

addressing potential issues such as missing values,

duplicate entries, and the accurate characterization of

data types, such as distinguishing between float and

integer variables. These techniques not only support

unbiased evaluations of the model's performance on

the training dataset but also assist in the careful fine-

tuning of hyperparameters (

G. O. Young., 1964),

ultimately leading to enhanced model performance.

It is essential to recognize that the evaluation

process can become increasingly influenced by the

model's performance on the validation dataset, which

may guide its configuration. Thus, while the

validation set serves to assess a model's effectiveness,

it is frequently utilized for ongoing evaluations,

enabling machine learning engineers to iteratively

refine the model’s hyperparameters.

Moreover, the tasks of data collection, analysis,

and addressing data quality, content, and structural

issues can require significant time and attention.

During the initial data identification phase, gaining a

deep understanding of the dataset and its

characteristics is imperative. This understanding not

only enhances preprocessing efforts but also

facilitates the selection of the most fitting algorithm

for model development.

3.3 Data Validation / Cleaning /

Preparing Process

To begin the process, it is essential to thoughtfully

import the necessary library packages and carefully

load the provided dataset. An initial analysis of the

dataset should involve a comprehensive variable

identification, with particular attention to its structure,

data types, and any potential issues such as missing

or duplicate values. It is advisable to hold back a

validation dataset from the training process, as this

can play a vital role in estimating the performance of

a model during the tuning of models and procedures.

The careful and appropriate use of validation and test

datasets is paramount for achieving reliable model

evaluation (

W.-K. Chen., 1993).

Moreover, the stages of data cleaning and

preparation include thoughtful actions such as

renaming the dataset, removing any superfluous

columns, and employing various analytical

techniques, including univariate, bivariate, and

multivariate analysis. The specific steps and

techniques may vary based on the unique

characteristics of the dataset at hand (

H. Poor., 1985).

The primary goal of data cleaning is to identify and

rectify any errors, inconsistencies, or anomalies,

thereby significantly enhancing the reliability and

applicability of the data in analytics and informed

decision-making.

3.4 Exploration Data Analysis of

Visualization

Data visualization plays an essential role in applied

statistics and machine learning, enriching the

quantitative aspects of statistics with valuable

qualitative insights. While statistical methods focus

primarily on numerical data and estimations, data

visualization provides effective tools for exploring

datasets, identifying underlying patterns, detecting

outliers, and revealing any inconsistencies in the data.

When combined with domain expertise,

visualizations can adeptly illustrate key relationships,

making complex information more intuitive and

impactful for stakeholders compared to traditional

numerical metrics like associations or significance

levels. By utilizing charts and plots, we can transform

intricate insights into accessible and actionable

knowledge, thereby bridging the gap between

technical analysis and practical decision-making (

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

622

Wang).

Preprocessing is a critical phase that entails

applying transformations to data before it is

introduced to a machine learning algorithm. This

significant step prepares raw data— often derived

from diverse sources—by transforming it into a clean

and structured format suitable for analysis. To ensure

that a machine learning model performs optimally, it

is imperative to appropriately prepare the data.

Specific algorithms, such as Random Forest, have

unique requirements; for instance, they don't

accommodate null values. Thus, effectively managing

null values is vital prior to applying such algorithms

to the dataset. Furthermore, preprocessing guarantees

that the dataset is formatted appropriately for various

machine learning and deep learning algorithms. This

adaptability opens the door for experimentation with

different models, helping to identify the most

effective one for the given dataset and ultimately

enhancing the overall efficiency and accuracy of the

analysis (

J. Kaufman., 1995).

4 COMPARISON ALGORITHM

Consistently comparing the performance of a

proposed algorithm with existing algorithms is of

utmost importance in the field of machine learning (

Yorozu et al.,)

. Python, in conjunction with libraries

such as scikit-learn, offers a robust framework that

facilitates these comparisons effectively. Each model

tends to exhibit distinct performance characteristics,

and employing resampling methods like cross-

validation can provide valuable estimates of a

model’s potential performance on unseen data. These

insights are instrumental in identifying the models

that stand out among a suite of developed options (

Wang., 1987)

When engaging with a new dataset, it can be

advantageous to utilize various visualization

techniques to explore the data from multiple angles.

This principle extends to model selection as

wellemploying diverse visualization methods can

greatly aid in examining metrics such as average

accuracy, variance, and other statistical attributes

related to model performance. Such visual insights

enrich our understanding and enhance the decision-

making process involved in selecting the most suitable

model. A fair comparison of machine learning

algorithms hinges on the premise of uniform

evaluation across all models, ensuring they are

assessed on the same data using a consistent test

harness. This method not only mitigates bias but also

guarantees that any performance differences stem

solely from the algorithms' inherent capabilities,

rather than external variances linked to data or

evaluation strategies.

Utilizing a standardized test harness, be it through

cross- validation or a fixed train-test split, ensures that

each algorithm operates under consistent training and

testing conditions. This level of consistency

contributes to a reliable assessment of their respective

strengths and weaknesses, ultimately leading to a

more objective selection of the best- performing

model

4.1 The K-Fold Cross Validation

To ensure a fair evaluation of machine learning

algorithms, it is essential to maintain consistent data

handling practices. This can be accomplished by

configuring the same random seed, which allows for

identical training data splits across different

algorithms. Before embarking on the comparison of

various algorithms, one can create a machine learning

model utilizing Scikit-learn. This involves installing

the library and performing important preprocessing

operations such as correcting missing values, scaling

features, and encoding categorical variables. Models

such as Logistic Regression from the linear model

module, Random Forest from the ensemble module,

and Decision Tree Classifier from the tree module are

excellent choices to consider. Employing cross-

validation methods, like K-Fold, will contribute to a

more robust performance evaluation. Furthermore,

train test split may effectively divide the dataset into

training and testing sets, allowing the model's

performance on previously unknown data to be

evaluated. Ultimately, each model can be trained and

tested, utilizing metrics such as accuracy to predict

outcomes and compare results. This structured

approach enables the identification of the best-

performing algorithm in a systematic manner while

ensuring fairness and consistency throughout the

evaluation process.

5 PCHF ENGINEERING

TECHNIQUE

The PCHF Technique encompasses the application of

Principal Component Analysis (PCA) and related

dimensionality reduction methods specifically

designed for predicting heart failure. This approach

aims to highlight the most pertinent features that

influence heart failure outcomes by transforming the

dataset into a more compact set of principal

Weight Initialization-Based Rectiﬁed Linear Algorithm for Accurate Prediction of Chronic Heart Disease Compared with PCHF Feature

Engineering Technique

623

components while preserving the majority of the

variance (

J. Wang., 1987). Data collection involves the

systematic gathering of clinical and diagnostic

information relevant to heart failure prediction.

Following this, data preprocessing and correlation

analysis are conducted, after which Principal

Component Analysis (PCA) is applied to minimize

dimensionality and identify the principal components

features the account for the most significant variance.

Finally, feature selection and model training are

undertaken to improve the predictive capabilities of

the model.

5.1 PCHF Process

Step 1: Data to standardize with zero mean and unit

variance.

Step 2: Apply PCA to find the most significant

components.

Step 3: Select components explaining at least 95%

of the variance.

Step 4: Train and evaluate models on the reduced

dataset.

The PCHF technique helps focus on the critical

variables influencing heart failure, enabling robust

and efficient prediction models that are easier to

interpret and apply in clinical settings. Feature

selection identifies the most relevant features for

heart disease prediction, improving the classifier's

performance and interpretability.

5.2 Bagging Classifier

Bagging, or bootstrap aggregation, serves as an

effective ensemble learning approach designed

primarily to mitigate variance in noisy datasets. This

approach selects random chunks of data with

replacement, enabling certain data points to be picked

more than once. Once these diverse samples are

generated, individual models are trained

independently on each one. In regression tasks, the

final prediction is generated by averaging the results,

whereas in classification tasks, a majority vote is used

(

J. Magn.Jpn., 1987). Notably, the Random Forest

algorithm can be viewed as a sophisticated extension

of bagging, as it incorporates feature randomness

along with the principles of bagging to construct an

uncorrelated ensemble of decision trees. This

combination not only enhances the robustness of the

model but also improves overall accuracy. Bagging is

particularly beneficial in addressing overfitting in

both classification and regression scenarios. Its major

goal is to improve the accuracy and performance of

machine learning algorithms by producing diverse

subgroups of the original dataset and fitting

appropriate classifiers or regressors to each subgroup.

By aggregating the predictions from multiple models,

bagging effectively reduces variance, thereby

enhancing the model’s stability. This technique plays

a crucial role in navigating the bias-variance trade-

off, as it diminishes variance without substantially

increasing bias. As such, bagging is widely

recognized for its effectiveness in managing

overfitting, especially when applied to decision tree

algorithms in various classification and regression

tasks (

Y. Yorozu et al.,).

5.3 BernouliNB

To effectively analyze heart disease, we utilize

datasets that encompass various features such as age,

cholesterol levels, BP and exercise-induced angina,

among others. A prudent approach is to transform

continuous features into binary values based on

established thresholds. For instance, we can designate

cholesterol levels as 1 if they exceed a threshold (e.g.,

200 mg/dL), and 0 otherwise. Similarly, we can

categorize age as 1 if it is above a specified age (e.g.,

50 years), and 0 otherwise. Regarding resting blood

pressure, we can label it as

1 if it is above 130 mmHg,

and 0 otherwise. In addition to these transformations,

it's essential to encode categorical features, such as

sex and chest pain type, using one-hot encoding for

clarity in analysis. For our target

variable, we

will define it as 1 to indicate the presence of heart

disease and 0 to denote its absence. With this

preprocessed binary dataset, we can proceed to train

the Bernoulli Naive Bayes (Bernoulli) model (

Magn.Jpn., 1987)

. This model effectively calculates the

likelihood of each feature being either 0 or 1 for the

two distinct classes: the presence or absence of heart

disease. For any new patient's binary feature set, the

model will compute the posterior probabilities for

each class in accordance with Bayes' theorem, thereby

assisting in informed decision- making.

𝑷(𝒀|𝑿) =

𝑷(𝑿|𝒀)𝑷(𝒀)

𝑷(𝑿)

(1)

Where:

• P(X|Y): Likelihood of the data given the

class.

• P(Y): Prior probability of the class.

• P(X): Evidence (normalizing constant).

The model predicts the class with the highest

posterior probability.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

624

6 RESULTS AND DISCUSSION

6.1 Prediction of Accuracy

WiReL: This technique is a neural network-based

model, which involves weight initialization and

rectified linear units (ReLU) for activation. After

training the model, use the test dataset to predict heart

disease outcomes (positive or negative). Calculate the

accuracy by comparing the predicted results to the

actual outcomes in the test set

J. Wang.

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =

   

   

(2)

The accuracy achieved by WiReL in heart disease

prediction depends on the quality of the dataset and

the hyperparameter tuning. Typically, WiReL can

outperform traditional algorithms like logistic

regression or decision trees, achieving higher

accuracy rates by minimizing errors during training

and testing.

6.2 Confusion Matrix

False Positives (FP): This term refers to instances

where an individual is mistakenly identified as having

heart disease. In such cases, the actual class indicates

"no" (the patient does not have heart disease), while

the predicted class indicates "yes" (the patient is

forecasted to have heart disease). For example, if it

turns out that the patient did not survive, but the

prediction suggested otherwise, this would represent

a false positive.

False Negatives (FN): This situation occurs when

an individual who genuinely has heart disease is

inaccurately predicted to be healthy. Here, the actual

class is "yes" (the patient does have heart disease),

whereas the predicted class is "no" (the patient is

forecasted to be healthy). For instance, if the actual

outcome reveals that the patient survived, but the

prediction indicated that they would not, this

exemplifies a false negative. True Positives (TP): This

concept describes cases where an individual with

heart disease is accurately identified as such. These

are the accurately predicted positive outcomes, with

both the actual and predicted classes indicating "yes"

(the patient has heart disease).

True Negatives (TN): This term pertains to

situations where an individual who does not have

heart disease is correctly recognized as not having it.

These represent correctly predicted negative

outcomes, where both the actual and predicted classes

indicate "no" (the patient does not have heart disease).

For example, if the actual outcome shows that the

there is no survival rate for patient, and the prediction

affirms this finding, it illustrates a true negative.

𝑇𝑟𝑢𝑒𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑅𝑎𝑡𝑒

(

𝑇𝑃𝑅

)

=𝑇𝑃/

(

𝑇𝑃 + 𝐹𝑁

)

(3)

𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑜𝑛 𝑟𝑎𝑡𝑒(𝐹𝑃𝑅) = 𝐹𝑃/(𝐹𝑃 + 𝑇𝑁) (4)

Accuracy: This statistic measures the fraction of total

predictions that are right, indicating how frequently

the model correctly detects accountable and non-

accountable. The computation of accuracy is as

follows:

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (𝑇𝑃 + 𝑇𝑁)/(𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁) (5)

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃/(𝑇𝑃 + 𝐹𝑃) (6)

While accuracy is a clear performance metric that

indicates the ratio of correctly predicted observations

to total observations, it is important to note that it

performs best when the dataset is balanced, which

implies that the false positive and false negative

values are equal.

Precision: This measure represents the proportion of

optimistic forecasts that are really correct. It

addresses the question: how many of the patients

identified as survivors actually did? An accuracy

score of 0.788 indicates a strong degree of precision,

with a decreased rate of false positives.

𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃/(𝑇𝑃 + 𝐹𝑁) (7)

Recall: This reflects the percentage of positive

observed values that are correctly anticipated. Recall

essentially identify genuine defaulters. Recall, also

known as sensitivity.

The F1 Score is a weighted average of

accuracy recall and recall classification is a

supervised learning technique used in machine

learning and statistics, where a computer program

examines input data and then uses what it has learned

to classify fresh observations.

General Formula:

𝐹− 𝑀𝑒𝑎𝑠𝑢𝑟𝑒=2𝑇𝑃/(2𝑇𝑃+𝐹𝑃+𝐹𝑁) (8)

𝐹1 − 𝑆𝑐𝑜𝑟𝑒 𝐹𝑜𝑟𝑚𝑢𝑙𝑎:

𝐹1 𝑆𝑐𝑜𝑟𝑒 = 2 ∗ (𝑅𝑒𝑐𝑎𝑙𝑙 ∗ 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)/ (𝑅𝑒𝑐𝑎𝑙𝑙 +

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛) (9)

This dataset may have numerous categories may be

Weight Initialization-Based Rectiﬁed Linear Algorithm for Accurate Prediction of Chronic Heart Disease Compared with PCHF Feature

Engineering Technique

625

binary in nature, such as identifying if a person is

male or female or whether an email is considered

spam. Speech recognition, handwriting analysis,

biometric identification, and document classification

are a few prominent examples of classification jobs.

In the context of supervised learning, algorithms use

labeled data to gain knowledge. Once the algorithms

have a thorough grasp of the data, they can find

underlying patterns and correlations to give new,

unlabeled data the proper classifications.

7 CONCLUSIONS

The proposed endeavor started with data cleaning and

processing, then moved on to missing value analysis,

model development, and evaluation. Finally, the

research compares the WiReL algorithm with the

Principal Component Heart Failure (PCHF) Feature

Engineering Technique, and established that

prediction of heart disease using WiReL proved to be

more accurate than the PCHF approach. WiReL

focuses on optimal weight initialization and enhanced

learning mechanisms using rectified linear activation.

It generally achieves high accuracy due to its ability

to handle complex, non-linear patterns in data

effectively. It excels in datasets with sufficient size

and feature richness. PCHF leverages dimensionality

reduction through principal component analysis

(PCA), which simplifies the dataset by focusing on its

most important components. While it performs well

in reducing overfitting and computational

complexity, it might lose some detailed relationships

in the data, slightly impacting predictive accuracy.

WiReL often achieves higher accuracy compared to

PCHF, especially in larger and more complex

datasets.

8 FUTURE WORK

Focus on integrating WiReL and PCHF with

other machine learning or deep learning

techniques, such as ensemble methods, to

leverage the strengths of both algorithms for

improved accuracy and robustness. Develop a

real-time heart disease prediction system that

incorporates WiReL or PCHF into wearable

health devices or cloud-based platforms for

continuous monitoring and early warning.

REFERENCES

A. Saboor, M. Usman, S. Ali, A. Samad, M. F. Abrar, and

N. Ullah, “A method for improving prediction of

human heart disease using machine learning

algorithms,” Mobile Inf. Syst., vol. 2022, pp. 1–9, Mar.

2022.

A. M. Qadri, A. Raza, K. Munir, and M. S. Almutairi,

“Effective feature engineering technique for heart

disease prediction with machine learning,” IEEE

Access, May 30, 2023.

B. Smith, “An approach to graphs of linear forms

(Unpublished work style),” unpublished.

C. J. Kaufman, Rocky Mountain Research Lab., Boulder,

CO, private communication, May 1995.

C. A. U. Hassan, J. Iqbal, R. Irfan, S. Hussain, A. D.

Algarni, S. S. H. Bukhari, N. Alturki, and S. S. Ullah,

“Effectively predicting the presence of coronary heart

disease using machine learning classifiers,” Sensors,

vol. 22, no. 19, p. 7227, Sep. 2022.

D. K. Chohan and D. C. Dobhal, “A comparison-based

study of supervised machine learning algorithms for

prediction of heart disease,” in Proc. Int. Conf. Comput.

Intell. Sustain. Eng. Solutions (CISES), May 2022, pp.

372–375

E. H. Miller, “A note on reflector arrays (Periodical style—

Accepted for publication),” IEEE Trans. Antennas

Propagat., to be published.

G. O. Young, “Synthetic structure of industrial plastics

(Book style with paper title and editor),” Plastics, 2nd

ed. vol. 3, J. Peters, Ed. New York: McGraw-Hill, 1964,

pp. 15–64.

H. Poor, An Introduction to Signal Detection and

Estimation. New York: Springer-Verlag, 1985, ch. 4.

J. Magn.Jpn., vol. 2, Aug. 1987, pp. 740–741 [Dig. 9th

Annu. Conf. Magnetics Japan, 1982, p. 301].

J. Wang, “Fundamentals of erbium-doped fiber amplifiers

arrays (Periodical style—Submitted for publication),”

IEEE J. Quantum Electron., submitted for publication.

M. Young, The Techincal Writers Handbook. Mill Valley,

CA: University Science,1989.

P. Rani, R. Kumar, N. M. O. S. Ahmed, and A. Jain, “A

decision support system for heart disease prediction

based upon machine learning,” J. Reliable Intell.

Environments, vol. 7, no. 3, pp. 263–275, Sep. 2021.

S. Sarah, M. K. Gourisaria, S. Khare, and H. Das, “Heart

disease prediction using core machine learning

techniques A comparative study,” in Advances in Data

and Information Sciences. Springer, 2022, pp. 247–260.

W.-K. Chen, Linear Networks and Systems (Book style).

Belmont, CA: Wadsworth, 1993, pp. 123–135.

Y. Yorozu, M. Hirano, K. Oka, and Y. Tagawa, “Electron

spectroscopy studies on magneto-optical media and

plastic substrate interfaces (Translation Journals

style),” IEEE Transl.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

626