A Genetic Algorithm Based Dynamic Expansion Framework for Medical Disease Prediction (https://doi.org/10.63386/619384)

Jinhui Shen1  Haojie Xie2 Zhao Liu1*
1.Xihua University
2.Shanghai Dianji University
Jinhui Shen ：Email：3120220871330@stu.xhu.edu.cn
Haojie Xie：Email：Xihua University
*corresponding author：Zhao Liu：Email：0720140040@ xhu.edu.cn

task. To address  this, we introduce a novel medical disease prediction approach, the Dynamic Ensemble  Framework (DEF), which leverages multiple predictive models as expert systems. This  design capitalizes on the strengths of diverse models, thereby enhancing generalization per– formance. Furthermore, we propose a novel Genetic Algorithm (GA)–based methodology to automatically optimize ensemble weight configurations, regulating model prediction  behavior and further improving performance. Empirical results, derived from experiments  on real–world medical datasets, demonstrate that the proposed approach outperforms  existing baselines.

Keywords: Medical Disease Prediction, SVM, Ensemble Framework

1. Introduction

The rapid advancement of science and technology is catalyzing significant transfor– mations within the medical domain. The confluence of an aging global population, the  escalating prevalence of chronic diseases, and the persistent increase in healthcare expendi– tures presents substantial challenges to conventional medical paradigms. Simultaneously, the emergence of advanced technologies, including big data analytics and artificial intel– ligence, offers novel opportunities for innovation within the healthcare sector. Medical  intelligent prediction, as a pivotal application, is gaining prominence as a critical instru– ment for enhancing the quality of medical services and augmenting the efficacy of disease  prevention and control strategies.

Medical intelligent prediction encompasses the application of cutting–edge technolo– gies, including artificial intelligence (AI), machine learning (ML), and deep learning (DL), to  perform comprehensive mining and analysis of extensive medical datasets. This approach  aims to forecast the onset, progression, and therapeutic outcomes of various diseases  with enhanced accuracy. The data utilized in these predictive models are derived from  diverse and multifaceted sources, such as electronic health records (EHRs), radiological  and pathological medical imaging, genomic sequencing and genetic testing results, lab– oratory and clinical test measurements, as well as detailed patient lifestyle factors and  environmental exposures. By integrating and systematically analyzing these voluminous  and heterogeneous datasets, intelligent prediction frameworks can identify latent patterns, correlations, and causal relationships that are not readily apparent through conventional  analysis. Consequently, these models provide robust, data–driven insights that significantly

augment clinical decision–making processes, enabling personalized medicine approaches     36

and improving patient prognosis and treatment efficacy.                                                         37

Recent advancements in medical intelligent prediction technologies have been substan–    38

tial [1,2]. In disease diagnosis, deep learning–driven image recognition methodologies have     39

been extensively integrated into medical image analysis, enhancing clinicians’ ability to     40

detect and diagnose pathologies with greater precision. For instance, in pulmonary disease     41

diagnostics, automated interpretation of chest radiographs and computed tomography     42

scans facilitates rapid identification of early–stage lung cancer indicators, thereby aug–    43

menting diagnostic accuracy and promptness. In the domain of disease risk stratification,    44 machine learning models leverage longitudinal patient data and biometric parameters      45

to forecast individual susceptibility to chronic conditions such as diabetes mellitus and      46 cardiovascular disorders, enabling proactive intervention and preventive strategies [3].         47

In this research, we aim to implement the medical disease prediction by proposing a     48

novel approach called Dynamic Ensemble Framework (DEF) that manages and optimizes     49

a series of models, where each one is implemented using different technologies. Such an     50

approach can improve the model’s generalization performance by exploring the advantages     51

of each model. Specifically, we consider to employ Support Vector Machine (SVM), DNN     52

and XGB (EXtreme Gradient Boosting) as the members of the proposed DEF. A simple way     53

for implementing the prediction process of the proposed DEF is to sum the outputs of all     54

models and calculate the average results. However, such an approach treats each model     55

as equally important, which would not achieve optimal generalization performance. In     56

order to address this issue, this paper proposes a novel Genetic–Algorithm (GA)–based      57

Ensemble Weight Optimization approach that automatically determines an optimal weight     58

configuration for the proposed DEF. By using such an approach, our model can further      59

improve the model’s generalization performance.  We build a series of experiments on      60

many real–world medical datasets and the results from the experiments demonstrate that     61

the proposed approach outperforms other baselines for the disease prediction tasks.

• In this paper, we propose a novel dynamic ensemble framework to deal with the 63 medical disease prediction, which explores advantages from different experts to     64 enhance the generalization performance.                                                                           65
• In this paper, we propose a novel Genetic Algorithm Ensemble Weight Optimization 66 (GAEWO) approach to automatically determine an optimal weight configuration for     67 the proposed framework, which further improves the model’s performance.                  68
• In this paper, we evaluate the effectiveness of the proposed framework on several real– 69 world medical datasets, and the results from the experiments show that the proposed     70 approach achieves the best performance compared to the baselines.                                 71

1. Background and Related Work

In this section, we provide a detailed background about several technologies that are     73

used in this research.                                                                                                                     74

75

1.1. Deep Neural Network

Deep Neural Network (DNN) has been a popular machine learning technology, which     76

achieves good performance in various applications [4–12]. Each hidden layer of a DNN     77

executes nonlinear transformations on the input data. For instance, in the context of image     78

recognition, the initial hidden layer may identify basic features like edges and textures; as     79

the layers progress, the intermediate hidden layer can discern more complex features such     80

as the local shape of the object; finally, the uppermost hidden layer is capable of recognizing     81

high–level features, including the overall semantic category of the object. This hierarchical     82

approach to feature learning, which transitions from low–level to high–level abstractions,    83

allows the DNN to gain a profound understanding of the data’s internal structure.                84

85

Due to its good performance, many studies have explored the DNN for disease predic–

87 88 89

tion [13,14]. Ramprakash et al propose to combine Deep Neural Network and X2–statistical     86 model to make the heart disease prediction.  The proposed approach can significantly

relieve over–fitting issues caused by the lack of training samples. The results from several

experiments indicate that the our approach can provide better classification accuracies

91

on all medical datasets when compared to other baselines. Reshan et al [15] implement      90 the heart disease prediction by proposing a new hybrid model that combines three neural

networks, including DNN, Long Short–Term Memory (LSTM) and Convolutional Neural     92

94

Networks (CNN). The proposed approach can learn robust representations from medical     93 data and thus can provide good performance in heart disease prediction.

1.2. Support Vector Machine                                                                                                              95

The Support Vector Machine (SVM) represents a supervised learning algorithm com–    96 monly employed in various applications [16], including classification, regression, and     97

99 100

anomaly detection. The fundamental concept behind SVM is improving the model’s gen–    98 eralization capability by identifying an appropriate hyperplane that effectively separates

data points from distinct categories and maximizing the margin between these classes.

103

In the linearly separable case, let us define D = {(x1, y1), (x2, y2), . . . , (xn, yn )} as the    101 training dataset, where xi  ∈ X = Rdyi  ∈ Y = {+1, -1}.  The goal of learning an SVM     102 model is try to determine a hyperplane, expressed as :

wTx + b = 0 ,                                                        (1)

where w and b represent the parameters belonging to the SVM model. Eq. (1) ensures that all    104

positive samples have wT xi + b ≥ 1 and all negative samples have wT xi + b ≤ -1. Support    105

vectors are the sample points closest to the hyperplane, and they satisfy | wT xi + b| = 1.        106

107 108

The optimization goal of the SVM model is maximizing the classification margin. This

optimization process is usually defined by :

(2)

109

As a result, we can transfer the above optimization problem as :

s.t.   yi(wT xi + b) ≥ 1,    i = 1, 2, . . . , n ,                          (3)

For linearly inseparable data, we introduce slack variables ξi   ≥ 0which allows some    110

samples not to meet the constraints.                                                                                           111

113 114 115

where C  >  0 denotes a penalty parameter, which controls the degree of penalty for    112 misclassified samples.

When the data samples are linearly inseparable in the original space, the SVM model

maps the data to a high–dimensional feature space through a kernel function, making the

data linearly separable in the new space. As a result, we can define the kernel function as :    116 K(xi, xj) = φ(xi)T φ(xj),                                                  (5)

117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135

where φ represents a mapping from the original space to the high–dimensional feature space. There are many choices for the kernel functions, including polynomial kernel K(xi, xj) =

(γxi(T)xj + r)d, linear kernel K(xi, xj) = xi(T)xj, Sigmoid kernel K(xi, xj) = tanh(γxi(T)xj + r)

and Gaussian radial basis function K(xi, xj) = exp(—γ∥xi — xj∥2 ).

Due to its good performance, the SVM has been widely employed for disease pre– diction [17]. Vijayarani et al [17] propose to employ the SVM and DNN to predict kidney diseases. The empirical results show that both SVM and DNN achieve good performance  in the kidney disease prediction task.  Mythili et al [18] propose to employ the SVM to  implement a rule–based model that is used for the heart disease prediction task. Hoque et  al [19] propose to predict cardiovascular diseases using an SVM model. Specifically, two  types of SVM models such as linear SVM and polynomial SVM are employed and show good performance in the cardiovascular disease prediction task. Although this SVM works  demonstrate promising performance in the disease prediction tasks, they have two primary limitations : (1) These works only focus on a specific disease type; (2) These works usually adopt a single model, which would not achieve optimal generalization performance; To  address these issues, this paper develops a new disease prediction approach that manages  and optimizes a series of machine learning models as experts and introduces to employ a Genetic Algorithm to find an optimal ensemble weight configuration for the proposed  framework, aiming to utilize advantages of all expert to improve the model’s performance.

136 137 138 139 140 141 142 143 144 145

2. Methodology

2.1. Problem definition.

In the medical disease prediction process, we usually have a training dataset Ds  =

respectively, where d represents the data dimension, respectively. Specifically, each class label yj is usually considered as a binary label. yj = —1 and yj = 1 denote the Illness and health for the j–th instance. During the training process, we optimize and update a model h on the whole training dataset Ds . The goal of the model h is to minimize the loss on all data samples, expressed as :

146 147 148 149 150 151 152 153 154 155 156 157

where floss(·, ·) denotes the loss function and |Ds | represents the total number of data samples for Ds . In the testing phase, we employ average classification accuracy calculated using the whole testing dataset as the final performance criterion.

2.2. The structure of the expert

The medical dataset usually has a few training data samples, and training a model on it would lead to overfitting issues. In order to address this issue, we propose a novel ensemble framework, which considers to manage several different expert models, where each one is implemented using different algorithms or models. Such a design can significantly enhance the model’s generalization performance.

Let Fθj : X  → Y denote the j–th expert network with the parameter set θj, which  receives a data sample x over the data space X and outputs a prediction y over the space Y. Let M represent the total number of expert networks for the proposed ensemble framework.

Features

Expert network

the data space

Weights

Patient             Database

the prediction space

• Initial ●crossover and mutation  the next generation    optimal weights

Figure 1. The network architecture of the proposed framework.

159

The prediction process of the proposed ensemble framework combines predictions from all    158 expert networks, expressed as :

(7)

where y/ is the final prediction made by the proposed ensemble framework and wj is the    160

component weight that determines the importance of the j–th expert network during the    161

prediction process. In practice, we should ensure that the sum of all ensemble weights is    162

equal to 1.                                                                                                                                       163

164

2.3. Optimizing the Ensemble Weights Using Genetic Algorithm

The prediction process defined in Eq. (7) usually assumes that each expert network has    165 the same influence on the prediction process. However, such a design would not achieve    166 optimal performance since each expert network has different prediction abilities. Finding    167 an appropriate ensemble weight configuration can significantly improve the model’s per–    168 formance. In order to achieve this goal, we propose to formulate the search of the optimal    169 ensemble weight configuration as an optimization problem, expressed as :                            170

(8)

where xc  denotes the c–th data sample and |Ds | represents the number of data samples     171

for Ds .  w1(*), · · · , wM(*) are the weights of the ensemble that allow the proposed ensemble     172

framework to achieve the best performance.  In order to find optimal weights of the    173 ensemble, we propose to employ the genetic algorithm, which is a popular approach to    174 find an optimal solution for the optimization problems. We provide the overall framework    175 in Fig. 1.                                                                                                                                           176

Specifically, each ensemble weight is between 0 and 1 and therefore we propose to     177 employ a ten–dimensional one–hot vector Vj   ∈ R10  to represent the ensemble weight.    178

Table 1. The feature information of the Obesity–Level dataset.

Variable Name	Role	Type	Demographic
Gender	Feature	Categorical	Gender
Age	Feature	Continuous	Age
Weight	Feature	Continuous
Height	Feature	Continuous
FAVC	Feature	Binary
FCVC	Feature	Integer
Family history with overweight	Feature	Binary
NCP	Feature	Continuous
SMOKE	Feature	Binary
CH2O CAEC	Feature Feature	Continuous Categorical
SCC	Feature	Binary
TUE	Feature	Integer
FAF	Feature	Continuous
MTRANS	Feature	Categorical
CALC	Feature	Categorical
NObeyesdad	Target	Categorical

Vj[i] = 1 indicates that the j–th ensemble weight is i/10. As a result, we can form a unified    179 vector V ∈ R10×M to represent all the weights of the ensemble model. First, we define a     180

fitness function, expressed as :                                                                                                     181

where Ft(·) is a function that transfers the vector Vi  to a set of weights {w1, · · · , wM}.   182 Ft(Vi)[j] denotes the j–th weight wj. The proposed GA–based ensemble weight approach    183 has several optimization steps:                                                                                                      184

Step 1: the algorithm initialization. At the beginning, the proposed approach randomly    185 generates 100 solutions {V1, · · · , V100 }.                                                                                      186

Step 2: the selection process. During each successive generation, we consider choosing a     187

small set of the existing solutions, aiming to produce a new generation. Specifically, we    188

propose to calculate the fitness value for each solution Vj  using Eq. (9) and then select     189

several good solutions for the next generation process.                                                            190

Step 3: the Genetic operations.  In this step, we propose to yield a second generation     191

192 193

population of solutions from those selected ones. Specifically, we consider a combination

of genetic operators, including mutation and crossover.

3. Experiment Results                      194

195 196

3.1. The Experiment Setting

We adopt the vscode as the development environment and python3.11 as the program–

198 199

ming language. For each experiment, we consider 10% of the entire dataset as the testing    197 samples and the remaining data samples are employed as the training samples.

In order to evaluate the performance of various models, we consider five popular

performance criteria, including True Positive Rate (TPR), Positive Predictive Value (PPV),    200 True Negative Rate (TNR), Accuracy and Negative Predictive Value (NPV). The TPR    201

203

denotes the proportion of positive samples that are correctly identified by the model and    202 its calculation is expressed as :

(10)

204 205 206 207

where TP denotes the examples that are actually positive are correctly predicted as positive

and FN represents the examples that are actually positive are mistakenly predicted as

negative.  The TNR evaluates how well a machine learning model correctly identifies

negative samples in a binary classification task, and its calculation is expressed as :

(11)

208 209 210 211

where FP denotes the examples that are actually negative and are incorrectly predicted as

positive. The PPV denotes a measure of how accurately a diagnostic test identifies those

who truly have a condition when the test result is positive.  The calculation of the PPV

criterion is defined by :

(12)

212

The NPV denotes the ratio of correctly predicted negative samples to all negative samples.

(13)

213

In addition, we also employ the classification accuracy as the performance criterion.

Table 2. The feature information of the Heart Disease dataset.

Variable Name	Role	Type	Demographic
age	Feature	Integer	Age
cp	Feature	Categorical
sex	Feature	Categorical	Sex
chol	Feature	Integer
fbs	Feature	Categorical
restecg	Feature	Categorical
trestbps	Feature	Integer
thalach	Feature	Integer
oldpeak	Feature	Integer
exang	Feature	Categorical
slope	Feature	Categorical
thal	Feature	Categorical
num	Target	Integer
ca	Feature	Integer

3.2. The data preprocessing                                                                                                               214

The first medical dataset employed in this study is the Obesity–Level dataset, which    215 encompasses data pertinent to assessing obesity prevalence across populations in Mexico,    216

217

Peru, and Colombia.  This dataset comprises variables related to dietary patterns and

physiological health metrics. It contains 17 attributes and 2,111 instances, each annotated    218 with the NObesity (Obesity Level) class label.  This label facilitates classification into    219

categories such as Insufficient Weight, Normal Weight, Overweight Level I, Overweight    220

222 223

Level II, Obesity Type I, Obesity Type II, and Obesity Type III. Notably, 77% of the dataset    221 was synthetically augmented using the Weka platform in conjunction with the SMOTE

algorithm, while the remaining 23% was sourced directly from participants via an online

Table 3. The feature information of the Breast Cancer Wisconsin dataset.

Variable Name	Role	Type
ID	ID	Categorical
radius1	Feature	Continuous
Diagnosis	Target	Categorical
texture1	Feature	Continuous
perimeter1	Feature	Continuous
smoothness1	Feature	Continuous
compactness1	Feature	Continuous
area1	Feature	Continuous
concavity1	Feature	Continuous
symmetry1	Feature	Continuous
fractal_dimension1	Feature	Continuous
concave_points1	Feature	Continuous
radius2	Feature	Continuous
perimeter2	Feature	Continuous
texture2	Feature	Continuous
area2	Feature	Continuous
compactness2	Feature	Continuous
smoothness2	Feature	Continuous
concavity2	Feature	Continuous
concave_points2	Feature	Continuous
fractal_dimension2	Feature	Continuous
symmetry2	Feature	Continuous
radius3	Feature	Continuous
perimeter3	Feature	Continuous
area3	Feature	Continuous
texture3	Feature	Continuous
smoothness3	Feature	Continuous
concavity3	Feature	Continuous
compactness3	Feature	Continuous
concave_points3	Feature	Continuous

interface.  Comprehensive feature details of the Obesity–Level dataset are presented in     224 Table 1. The dataset includes categorical variables incompatible with standard machine    225 learning algorithms; therefore, these were transformed into numerical representations.    226 For example, the ’Gender’ feature was binarized with 1 representing female and 0 male.   227 Similarly, binary encoding was applied to features such as family history of obesity, frequent    228 consumption of high–calorie meals, smoking status, and calorie monitoring, assigning 1 for    229 affirmative responses and 0 otherwise. Additionally, one–hot encoding was implemented    230 for multi–class categorical variables including eating habits, alcohol consumption frequency,    231 and transportation mode. Post–processing, each sample comprises 23 features alongside a    232 single class label.                                                                                                                           233

We employ the second clinical dataset, referred to as Heart Disease, which contains 76    234 variables; however, extant literature has predominantly analyzed a subset of 14 features.   235 Importantly, the Heart Disease database remains the exclusive dataset utilized by machine    236 learning practitioners to date. The ’goal’ attribute denotes the presence of cardiovascular    237 pathology, encoded as integer values from 0 (absence) to 4. Prior investigations leveraging    238 the Heart Disease dataset have primarily focused on binary classification distinguishing    239 disease presence (values 1-4) from absence (value 0). Comprehensive feature descriptions     240 of the Heart Disease dataset are detailed in Table 2.                                                                  241

We adopt the third dataset called Breast Cancer Wisconsin which contains 569 in–    242

stances, and each sample has 30 features. The detailed feature information of the Breast     243

(a) Heart Disease dataset                                             (b) Breast Cancer Wisconsin dataset

(c) Obesity–Level dataset                                            (d) Regensburg pediatric appendicitis dataset

(e) Myocardial infarction complications dataset

Figure 2. The values of the fitness functions over the optimization iterations.

(a) Heart Disease dataset                                     (b) Breast Cancer Wisconsin dataset

(c) Obesity–Level dataset                                   (d) Myocardial infarction complications dataset

Figure 3. The change in the ensemble weight configuration over time.

(a) Heart Disease dataset                                     (b) Breast Cancer Wisconsin dataset

(c) Obesity–Level dataset                                   (d) Myocardial infarction complications dataset

Figure 4. The heat map for change of the ensemble weight configuraiton over time.

Table 4. The performance of various models on the Heart Disease dataset.

Model	TPR	TNR	PPV	NPV	Accuracy
SVM	0.9286	0.7059	0.7222	0.9231	0.8065
DNN	0.8571	0.8235	0.8000	0.8750	0.8387
XGB	0.9286	0.7059	0.7222	0.9231	0.8065
Our	0.9286	0.8824	0.8667	0.9375	0.9032

249 250 251 252

Cancer Wisconsin dataset is provided in Tab. 3. For each sample, features are derived from    244 a digitized image of a fine needle aspirate (FNA) taken from a breast mass, detailing the    245 features of the cell nuclei observed in the image. Each instance contains 30 cell nucleus     246 features that are all numerical, describing the geometric characteristics of the cell nucleus    247 in the pathological image, including radius, texture, perimeter area, smoothness, concavity,    248 etc. Each feature contains three statistics such as mean, standard deviation and maximum

value.

In addition, we also consider evaluating the regensburg pediatric appendicitis dataset.

This dataset was obtained through a retrospective study involving a cohort of pediatric

255 256

patients who were admitted to Children’s Hospital St. Hedwig in Regensburg, Germany,    253 due to abdominal pain. For the majority of patients, multiple B–mode abdominal ultrasound    254 images were collected, with the number of views ranging from 1 to 15.  These images

illustrate various areas of interest, including the right lower quadrant of the abdomen, the

258 259 260

appendix, intestines, lymph nodes, and reproductive organs. In addition to the numerous    257 ultrasound images for each patient, the dataset encompasses information such as laboratory

test results, findings from physical examinations, clinical scores including the Alvarado

score and pediatric appendicitis scores, as well as ultrasonographic findings produced

by specialists. Ultimately, the subjects were categorized based on three target variables:    261

diagnosis (appendicitis versus no appendicitis), management (surgical versus conservative),    262

and severity (complicated versus uncomplicated or no appendicitis).                                     263

264 265

Furthermore, we also consider evaluating the myocardial infarction complications

dataset. This dataset is designed to tackle two major challenges: predicting complications

267

of Myocardial Infarction (MI) based on patient data (i) at the time of admission and (ii) on    266 the third day of hospitalization. Additionally, another vital category of tasks encompasses

disease phenotyping (cluster analysis), dynamic phenotyping (filament extraction and    268

270 271

identification of disease trajectories), and visualization (disease mapping).  Myocardial     269 Infarction is considered one of the most significant challenges in modern medicine. Acute

myocardial infarction is associated with a high mortality rate within the first year following

273

the incident. The incidence of MI remains high across all countries. This is particularly no–   272 ticeable among urban populations in highly developed nations, who face chronic stressors

and often consume irregular and unbalanced diets. In the United States, for example, over    274 a million people suffer from MI each year, with 200,000 to 300,000 dying from acute MI     275

276 277

before they can receive medical attention. The progression of the disease in MI patients

varies considerably. MI can occur without complications or with complications that do not

279 280 281 282

negatively impact the long–term prognosis. However, around half of the patients during    278 the acute and subacute stages experience complications that worsen their condition and

may even lead to death. Even experienced professionals cannot always foresee the onset of

these complications. Thus, the capacity to predict complications of myocardial infarction to

implement timely preventive measures is a crucial goal.

Table 5. The performance of various models on the Breast Cancer Wisconsin dataset.

Model	TPR	TNR	PPV	NPV	Accuracy
SVM	1.0000	0.9722	0.9545	1.0000	0.9825
DNN	1.0000	1.0000	1.0000	1.0000	1.0000
XGB	1.0000	1.0000	1.0000	1.0000	1.0000
Our	1.0000	1.0000	1.0000	1.0000	1.0000

3.3. The Experiment Results                                                                                                             283

291 292 293 294 295

We initially assess the performance of multiple models on the Heart Disease dataset,    284 with results summarized in Table 4. The SVM model attains TPR, TNR, PPV, NPV, and Ac–    285 curacy values of 0.92, 0.70, 0.72, 0.92, and 0.80, respectively. Conversely, the deep learning     286 model (DNN) demonstrates superior performance in several metrics, including TNR, PPV,    287 and Accuracy, relative to SVM. Furthermore, the XGB model achieves identical metrics to    288 SVM, with TPR, TNR, PPV, NPV, and Accuracy of 0.92, 0.70, 0.72, 0.92, and 0.80, respec–    289 tively. These findings indicate that each model excels in distinct performance measures.    290 Consequently, integrating the strengths of these models within a unified optimization

framework is expected to enhance overall performance, as empirically validated by the

proposed approach. As shown in Table 4, the proposed framework outperforms all baseline

models across all evaluated metrics, substantiating its efficacy.

We assess the efficacy of multiple models on the Breast Cancer Wisconsin dataset,

298 299

with the outcomes summarized in Table 5. The SVM model attains TPR, TNR, PPV, NPV,    296 and Accuracy values of 1.0, 0.97, 0.95, 1.0, and 0.98, respectively. Conversely, alternative     297 approaches demonstrate superior performance across all evaluation metrics. These findings

indicate that all methods achieve optimal predictive performance on the Breast Cancer

Wisconsin dataset.                                                                                                                        300

301

We assess the efficacy of multiple models on the Obesity–Level dataset, with results

304 305

summarized in Table 6.  The SVM model attains TPR, TNR, PPV, NPV, and Accuracy    302 values of 0.96, 0.62, 0.94, 0.72, and 0.91, respectively. Similarly, the DNN model exhibits     303 comparable performance metrics to the SVM. Notably, both the proposed methodology

and XGB demonstrate superior performance across all evaluation metrics relative to other

307 308 309 310 311 312

baselines. However, the proposed framework surpasses XGB on the Heart Disease dataset,    306 as evidenced by the results in Table 4. These findings indicate that the proposed framework

yields enhanced predictive accuracy over existing baselines, particularly when applied to

complex datasets.

For the regensburg pediatric appendicitis dataset, we train various models on the

training dataset and the classification accuracy of various models is reported in Table. 7. The

empirical results demonstrate that the SVM achieves similar performance results compared

314 315 316

to the DNN. In contrast, the XGB outperforms other baselines on all performance criteria.   313 Furthermore, the proposed approach achieves the best performance on the classification

accuracy as well as other performance criteria, when compared to other baselines. The

performance results of various models on the myocardial infarction complications dataset

318 319 320

are reported in Table.8. From the experiment results, we can find that each baseline achieves    317 good performance results on all performance criteria.  Again, the proposed approach

almost achieves the best performance on all performance criteria, when compared to other

baselines.

Table 6. The performance of various models on the Obesity–Level dataset.

Model	TPR	TNR	PPV	NPV	Accuracy
SVM	0.9615	0.6207	0.9409	0.7200	0.9147
DNN	0.9670	0.6207	0.9412	0.7500	0.9194
XGB	0.9835	1.0000	1.0000	0.9062	0.9858
Our	0.9835	1.0000	1.0000	0.9062	0.9858

Table 7. The performance of various models on the regensburg pediatric appendicitis dataset.

Model	TPR	TNR	PPV	NPV	Accuracy
SVM	0.9130	0.7500	0.8400	0.8571	0.8462
DNN	0.9130	0.7812	0.8571	0.8621	0.8590
XGB	0.9783	0.8438	0.9000	0.9643	0.9231
Our	0.9783	0.8750	0.9184	0.9655	0.9359

3.4. The Analysis Results                                                                                                                  321

The fitness analysis. To evaluate the efficacy of the proposed GA–based ensemble weight    322

optimization method, we computed the fitness function values at each iteration of the     323

optimization process, as illustrated in Fig. 2a. The results indicate that the method achieves    324

a stable fitness value after six iterations. Furthermore, the average fitness value exhibits    325

an increasing trend with the progression of optimization iterations, demonstrating the     326

capability of the GA–based approach to progressively converge toward an optimal solution.   327

Additional fitness outcomes for the Breast Cancer Wisconsin and Obesity–Level datasets    328

are presented in Fig. 2b and Fig. 2c, respectively, revealing that the optimal fitness value    329

is attained during the early optimization stages. Collectively, these findings substantiate    330

the suitability of the proposed GA–based optimization framework for handling complex    331

datasets such as the Heart Disease dataset.                                                                                   332

In addition, we also report the values of the fitness function of the regensburg pedi–   333 atric appendicitis and myocardial infarction complications datasets in Fig. 2d and Fig. 2e,    334

respectively. These empirical results show that the proposed GA–based approach can find    335

the best solution at the initial optimization process. Furthermore, the average fitness values    336

increase as the number of optimization iterations increases.                                                    337

The ensemble weight analysis. The proposed GA–based optimization approach aims to    338

find an optimal ensemble configuration for the proposed DEF. In order to analyse how the    339

proposed GA–based approach finds the optimal solution, we record the ensemble weights    340

at each training time. The line chart and heat map are presented in Fig.  and Fig. 3 and     341

Fig. 4, respectively. we can observe that the proposed GA–based optimization approach    342

provides different ensemble configurations for different datasets. For example, the SVM–    343

based component of the proposed DEF is assigned by a large weight while the XGB–based    344

component is less important on the Heart Disease dataset.  Similar results can also be    345

observed for the Breast Cancer Wisconsin dataset.  In contrast, the proposed GA–based     346

optimizaiton approach gives a large weight for the XGB–based component while SVM–    347

and DNN–based components are less important. These results indicate that the proposed    348

GA–based optimizaiton approach can find the most suitable ensemble weights for the    349

Table 8. The performance of various models on the myocardial infarction complications dataset.

Model	TPR	TNR	PPV	NPV	Accuracy
SVM	0.0000	1.0000	0.0000	0.9059	0.9059
DNN	0.0000	0.9935	0.0000	0.9053	0.9000
XGB	0.0000	0.9805	0.0000	0.9042	0.8882
Our	0.0000	1.0000	0.0000	0.9059	0.9059

4. Conclusions                                                                               351

This paper introduces a novel disease diagnosis system, termed the dynamic ensemble    352

framework, designed for enhanced predictive accuracy on medical datasets. The framework    353

employs an ensemble of expert models, each instantiated with a distinct machine-learning    354

algorithm. This architectural choice leverages the strengths of individual machine-learning    355

methodologies to improve the model’s generalization capabilities. Furthermore, we present    356

a novel genetic algorithm (GA)-based ensemble weight optimization strategy to determine    357

the optimal weight configuration for the proposed framework, thereby facilitating expert    358

interaction and further enhancing model performance. Empirical validation, conducted on    359

a suite of real-world medical datasets, demonstrates that the proposed framework exhibits    360

superior performance compared to established baselines.                                                        361

362

1. Barghouthi, E.D.; Owda, A.Y.; Owda, M.; Asia, M. A Fused Multi-Channel Prediction Model 363 of Pressure Injury for Adult Hospitalized Patients—The “EADB” Model.  AI 2025, 6.  https:    364 //doi.org/10.3390/ai6020039.                                                                                                                365
2. Karkas, A.Y.; Durak, G.; Babacan, O.; Cebeci, T.; Uysal, E.; Aktas, H.E.; Ilhan, M.; Medetalibeyo- 366 glu, A.; Bagci, U.; Cakir, M.S.; et al. Radiomics-Based Machine Learning Models Improve Acute     367 Pancreatitis Severity Prediction. AI 2025, 6. https://doi.org/10.3390/ai6040080.                          368
3. Ding, N.; Möller, K. Minimally Distorted Adversarial Images with a Step-Adaptive Iterative 369 Fast Gradient Sign Method. AI 2024, 5, 922–937. https://doi.org/10.3390/ai5020046.                  370
4. Bang, J.; Koh, H.; Park, S.; Song, H.; Ha, J.W.; Choi, J. Online Continual Learning on a     371 Contaminated Data Stream With Blurry Task Boundaries. In Proceedings of the Proceedings of    372 the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp.    373 9275–9284.                                                                                                                                                374
5. Cha, H.; Lee, J.; Shin, J. Co2l: Contrastive continual learning. In Proceedings of the Proceedings 375 of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 9516–9525.                   376
6. Gu, Y.; Yang, X.; Wei, K.; Deng, C. Not Just Selection, but Exploration: Online Class-Incremental 377 Continual Learning via Dual View Consistency.   In Proceedings of the Proceedings of the     378 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp.    379 7442–7451.                                                                                                                                                380
7. Guo, Y.; Liu, B.; Zhao, D. Online continual learning through mutual information maximization.    381 In Proceedings of the International Conference on Machine Learning. PMLR, 2022, pp. 8109–    382
8. 383
9. Polikar, R.; Upda, L.; Upda, S.S.; Honavar, V. Learn++: An incremental learning algorithm     384

for supervised neural networks.   IEEE Trans.  on Systems Man and Cybernetics, Part C 2001,     385

31, 497–508.                                                                                                                                              386

9. Ren, B.; Wang, H.; Li, J.; Gao, H. Life-long learning based on dynamic combination model.     387

Applied Soft Computing 2017, 56, 398–404.                                                                                              388

10. Ritter, H.; Botev, A.; Barber, D. Online Structured Laplace Approximations for Overcoming     389

Catastrophic Forgetting.  In Proceedings of the Advances in Neural Information Processing     390

Systems (NeurIPS), 2018, Vol. 31, pp. 3742–3752.                                                                                  391

11. Tiwari, R.; Killamsetty, K.; Iyer, R.; Shenoy, P. GCR: Gradient Coreset Based Replay Buffer Selec– 392 tion for Continual Learning. In Proceedings of the Proceedings of the IEEE/CVF Conference on    393 Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 99-108.                                     394
12. Huo, F.; Xu, W.; Guo, J.; Wang, H.; Fan, Y. Non–exemplar Online Class–Incremental Continual 395 Learning via Dual–Prototype Self–Augment and Refinement. In Proceedings of the Proceedings     396 of the AAAI Conference on Artificial Intelligence, 2024, Vol. 38, pp. 12698-12707.                          397
13. Pan, Y.; Fu, M.; Cheng, B.; Tao, X.; Guo, J. Enhanced deep learning assisted convolutional neural 398 network for heart disease prediction on the internet of medical things platform. Ieee Access 2020,    399

8, 189503-189512.                                                                                                                                    400

14. Ramprakash, P.; Sarumathi, R.; Mowriya, R.; Nithyavishnupriya, S. Heart Disease Prediction 401

Using Deep Neural Network. In Proceedings of the 2020 International Conference on Inventive    402

Computation Technologies (ICICT), 2020, pp. 666-670. https://doi.org/10.1109/ICICT48043.2     403

020.9112443.                                                                                                                                             404

15. Reshan, M.S.A.; Amin, S.; Zeb, M.A.; Sulaiman, A.; Alshahrani, H.; Shaikh, A. A Robust     405

406

Heart Disease Prediction System Using Hybrid Deep Neural Networks.   IEEE Access 2023,

11, 121574-121591. https://doi.org/10.1109/ACCESS.2023.3328909.                                              407

16. Suthaharan, S. Support vector machine. In Machine learning models and algorithms for big data 408

classification: thinking with examples for effective learning; Springer, 2016; pp. 207-235.                        409

17. Vijayarani, S.; Dhayanand, S.; Phil, M. Kidney disease prediction using SVM and ANN    410 algorithms. International Journal of Computing and Business Research (IJCBR) 2015, 6, 1-12.                411
18. Mythili, T.; Mukherji, D.; Padalia, N.; Naidu, A. A heart disease prediction model using    412 SVM–decision trees–logistic regression (SDL). International Journal of Computer Applications 2013,     413

414

68.

19. Hoque, R.; Billah, M.; Debnath, A.; Hossain, S.; Sharif, N.B.; et al. Heart disease prediction     415 using SVM. International Journal of Science and Research Archive 2024, 11, 412-420.                             416