An Explainable Fuzzy Pooling CNN Model for Multi-Chronic Disease Prediction with

Scrutiny Boosted Graph LSTM Monitoring

Abstract: The global rise in chronic diseases and associated mortality has underscored the importance of early diagnosis to
improve patient outcomes. While many studies have proposed classification models, most suffer from poor feature extraction,
negatively affecting prediction accuracy. This study introduces a novel fuzzy-based deep learning framework to predict seven
major chronic diseases: diabetes, liver, thyroid, heart, kidney, Alzheimer’s, and breast cancer. For robust feature extraction, an
ensemble of LeViT and SincNet is applied. A Fuzzy Pooling Convolutional Neural Network (FPCNN) is then used for disease
prediction. To continuously monitor patient health records, the Scrutiny Boosted Graph Convolutional LSTM (SBGC-LSTM)
is employed. To interpret the impact of each feature on prediction, Shapley Additive Explanations (SHAP) are used. The
proposed method achieved a high accuracy of 99.28%. Comparative evaluations across multiple chronic disease datasets
revealed that our approach outperforms benchmark models such as SVM, Random Forest, and standard deep learning methods.
Keywords: chronic diseases; LeViT approach; SincNet approach; Fuzzy Pooling Convolutional Neural Network (FPCNN);
SBGC-LSTM; and Shapley Additive Explanations.

1 Introduction high gradient datasets or by training XGBoost models based

on RF models to attain particular DT (Asif et al., 2024; Asif
In the field of healthcare, the prevalence of chronic diseases et al., 2024). A number of traditional approaches fail to use
around the world is a significant concern. Chronic illnesses ensemble methods, which combine several models to
still have a big influence on world health and provide increase the prediction rate. It may be challenging to
enormous problems for healthcare systems everywhere understand the rationale behind the forecasts if a method
(Islam et al., 2024). These chronic illnesses, which include lacks interpretability (Kalra et al., 2024). This might restrict
diabetes, cancer, kidney disease, and cardiovascular diseases, its use in medical practice, as decision-making depends on
not only cause significant morbidity and mortality but also interpretability. Therefore, in this research, we introduced a
place a significant financial strain on societies (Kerth et al., novel framework.
2024). There is an urgency to develop novel and precise
predictive techniques for early identification and tailored The major key contributions of this research are as follows,
interventions as the frequency of chronic diseases keeps
 For resolving the class imbalance issue, the
increasing as a result of aging populations and changes in
Synthetic Minority Over-sampling Technique
lifestyle. In order to stop the advancement of chronic
(SMOTE) is employed, improving the learning
diseases, lessen related consequences, and improve the
process for minority-class instances.
general prognosis for those who are impacted, early detection
is essential. To reduce the potential severity of such  Feature extraction is enhanced through an ensemble
disorders, it is crucial to identify them as soon as they appear of LeViT and SincNet, ensuring rich and
(Huang et al., 2024; Chunduru et al., 2024). discriminative representations from input data.
Many chronic diseases, such as diabetes, heart disease,  FPCNN is introduced for robust disease
and some types of cancer, are defined by their asymptomatic classification, effectively capturing both local and
early stages, which postpones important discovery and global patterns with fuzzy pooling mechanisms.
intervention until they have progressed significantly and may  To monitor the patient records SBGC-LSTM
be irreversible (Patterson et al., 2024; Ram et al., 2024; supports continuous monitoring, enabling long-term
Smokovski et al., 2024). Recent advances in artificial temporal analysis in dynamic healthcare
intelligence (AI) have had a significant impact on healthcare environments.
and medical research. Earlier intervention and prevention  The interpretability of model outcomes is improved
methods can be implemented due to AI's ability to identify using Shapley Additive Explanations (SHAP),
people who are more likely to develop particular diseases allowing clinicians to understand feature impact on
(Ma et al., 2024). This revolutionary potential enables predictions.
medical practitioners to detect chronic illnesses early, hence
reducing the strain on individuals and healthcare systems. 2 Related Works
The benefits of early detection, especially in melanoma, are This section outlines the relevant research conducted in order
demonstrated in the following influential illustration of the to build the suggested model for chronic disease prediction.
benefits of early diagnosis (Kumar et al., 2024; Abhin et al.,
2025; Mohamed et al., 2024). (Rajeashwari et al., 2024) suggested using two Deep
In medical data processing, it can be difficult to predict CNNs (Deep Convolutional Neural Networks) to extract
and diagnose diseases simply based on the underlying features. In this instance, pertinent characteristics are
symptoms. The primary goal of the proposed work has been extracted using the optimal, maximum, and minimal hidden
to cluster and classify disease datasets (Akour et al., 2024). layers. Moreover, categorization is done using ME-RF
To improve classification accuracy, disease datasets with (Modified Extreme-Random Forest). The XGBoost method,
high dimensionality can be pre-processed using appropriate which has some inherent benefits like high convergence and
feature selection techniques. The suggested model extracts low calculations, is taken into consideration in this procedure.
features using a particular architecture that is built on two On the other hand, this model functions well with many
Deep CNN networks. Optimum models for prediction with leaves in DT (Decision Tree) when its predictability is low.
high specificity, sensitivity, and effectiveness can be obtained Concurrently, RF consists of many trees with equally
by training RF algorithms based on XGBoost algorithms for
weighted leaves, allowing for the flexible attainment of to classify it as either benign or malignant. Using the
maximum precision and accuracy using the available data. Crayfish optimization method (CFO), the RRNN weight
In order to accurately anticipate a patient's chronic parameter is tuned for precise classification. RRNN's
health status (heart or kidney disease), (Malibari et al., 2023) Semantic Web Language Rules (SWRL) are used to construct
presented the EO-LWAMCNet model, an EO-optimized the ontology approach. At least, the ontology classifier
Lightweight Automatic modulation categorization network. classifies the data as either benign or malignant.
The EO-LWAMCNet framework initiates the categorizing (Ayus et al., 2024) introduced two hybrid approaches
process to forecast chronic disease based on the obtained for the prediction of Alzheimer's disease: CNN-Conv1D-
sensor data. CKD and HD databases are used to predict the LSTM and ensemble recurrent ensemble network
disease. In this case, the training phase uses the preprocessed (HReENet). CNN is used as an extractor of features and
data for categorization. Following training, sensor data from Conv1D-LSTM is used as a classifier in the hybrid CNN-
the Cloud server (CS) is examined and classified as normal Conv1D-LSTM model. On the other hand, HReENet
or abnormal (heart or renal problems). provides an ensemble model that makes use of CNN-
The integration of fuzzy logic (FL), a mathematical Conv1D-LSTM, CNN, and LSTM predicted values.
approach, with quantum machine learning (QML) and its (Aytac et al. 2025) proposed a model called COOT-
application to a chronic disease dataset was proposed by CNN for thyroid cancer diagnosis, which uses the COOT
(Khushal et al., 2025). The high-dimensional space becomes Metaheuristic Optimization Algorithm to optimize CNN
low-dimensional when two variables are combined into one parameters. In order to identify the best layers and
variable, which decreases the number of features. When parameters of the CNN model for thyroid cancer diagnosis,
applied to the dataset under consideration, machine learning this study uses the COOT method. This technique uses a
has demonstrated low accuracy and a high computation time. carefully thought-out coding strategy to enable effective layer
This problem has been resolved by integrating FL with ML parameter tuning.
(FML), which has improved computation accuracy and time. (Amin et al. 2023) suggested a combined feature
Since various studies have shown that QML is inaccurate and extraction method for hepatic patient classification. The
requires a lot of computations when data quantities increase, suggested approach first imputes the outliers and missing
fuzzy concepts, specifically SVM and KNN are combined values for pre-treatment in the pipeline. The essential
with QML. features for categorization are then extracted using integrated
A machine learning-based kidney disease prediction feature extraction on the pre-processed data. Additionally, a
(ML-CKDP) model was developed by (Halder et al. 2024) to simulation investigation is being carried out to support the
enhance dataset preparation for CKD classification and offer suggested approach. A number of machine learning (ML)
a web-based application for CKD prediction. The proposed techniques are included in the suggested method, such as the
model includes missing data imputation, normalization using ensemble voting classifier, logistic regression (LR), random
Min-Max scaling, numerical value conversion for categorical forest (RF), K- KNN, SVM, and MLP.
variables, and a comprehensive data pretreatment process.
The approach employs seven classifiers to forecast CKDs. 2.1 Research Gap
(Mandava et al., 2024) presented an ensemble deep-
learning intelligent system for effective CVD prediction. The In recent years, there has been a surge in interest in deep
UCI repository's five benchmark datasets for heart disease learning-based disease prediction because of its improved
have been used for tests and evaluations. In the pre- accuracy and effective feature extraction capabilities.
processing phase, three data processing methods are initially Because of their simpler architectures than the conventional
applied to enhance the quality of the dataset by avoiding convolutional neural network (CNN), several authors have
undesirable distortions: eliminating outliers, substituting employed different architectures, such as ResNet, DenseNet,
missing values, and addressing issues with data imbalance. etc., for disease prediction. The number of parameters
The disease-related features are then extracted using utilized for training frequently rises due to the size of deep
Modified DenseNet201 (MDenseNet201), a deep learning- learning architectures. Because of its low latency and modest
based algorithm. The features are chosen using the Least computational performance, its applicability in lightweight
Absolute Shrinkage and Selection Operator (LASSO) and devices is commonly limited. The current methods
relief techniques. Lastly, to predict cardiovascular illness, an frequently concentrate on a specific disease that suffers from
enhanced deep residual shrinkage network (IDRSNet) based class imbalance and vanishing gradients. When used on
on deep learning was used. lightweight devices, the majority of current methods are
(Singh et al. 2024) introduced an approach that uses ineffective. Therefore in this research we introduced a novel
three soft-computing algorithms ETO, GSA, and their framework for predicting multiple chronic diseases.
suggested hybrid hGSAEPO to effectively find important
3 Proposed Methodology
features while also reducing the frequency of irrelevant ones,
reducing total complexity and improving reliability. A Using suitable data mining methods, the current study aims
workable structure for prognostic analysis through data to detect the four primary chronic diseases: diabetes, heart
instance categorization is presented by the combination of disease, renal disease, and breast cancer. Despite their best
these soft computing techniques and six machine learning efforts, traditional research has failed to achieve this in terms
classifiers. of accuracy. The current work aims to improve the prediction
(Rajeswari et al., 2025) suggested using Ontological rate based on a stepwise procedure, as shown in Figure 1,
Modeling with Recursive Recurrent Neural Network and since the accuracy rate appears to be the most critical metric
Crayfish Optimization for Reliable Breast Cancer Prediction in disease diagnosis. The below subsection explains the
(OM-RRNN-CFO-RBCP) to create an ontological predicate proposed methodology in detail.
for breast cancer. RRNN is provided with pre-processed data
 Figure 1 Overall framework of proposed methodology

3.1 Preprocessing To preserve the integrity of the datasets, issues that arose
during the data pre-processing stage, such as differences in
It is impossible to overestimate the importance of data quality data quality and the existence of anomalies, were carefully
in the field of predictive analytics, especially when resolved. In addition to improving the precision of disease
considering actual medical datasets. These datasets present prediction, our thorough data pre-processing efforts ensured
difficulties due to their varied sources and nature, such as the validity of our conclusions.
irrelevant features, missing data points, and outliers. These Following preprocessing, the datasets were randomly
variables may significantly impact the accuracy of later selected and methodically split, with 20% being used for
analysis and model training. Given the inherent complexity, testing and 80% proceeding to training. The 20% testing set
our work focused heavily on thorough data pre-processing to allows for a comprehensive assessment of the model's
ensure the accuracy and resilience of our prediction models. generalization ability on unknown data, while the large
A number of essential phases were included in the pre- training set ensures strong model learning.
processing pipeline, each specifically designed to handle the
unique difficulties presented by actual medical data. Normalization: The magnitudes, ranges, and units of some
Important features that were relevant to illness prediction characteristics of the dataset vary significantly. Feature
were carefully extracted. In accordance with ethical scaling is commonly used to normalize the range of
principles, steps were taken to protect patient confidentiality. independent variables and guarantee that they contribute
Finding and managing outliers was part of the data cleaning equally to the prediction because the datasets are so diverse.
procedure, which is essential for precise model training. The data in the range (0, 1) is normalized using a Min–Max
A robust replacement method was used to handle normalization. The mathematical formula required to
outliers, which were identified as possible sources of bias in determine the Min–Max normalization is displayed in
the analysis. To preserve data integrity, typical values, more Equation (1).
especially, the mean, were used to substitute outliers for 𝑥−𝑥𝑚𝑖𝑛
numerical attributes. A similar strategy was used for 𝑋̅ = (1)
𝑥𝑚𝑎𝑥 −𝑥𝑚𝑖𝑛
categorical characteristics, replacing the mode for outliers in
order to maintain the categorical distribution. 𝑥 represents the initial feature value, 𝑋̅represents the
Another crucial component of our data pre-processing normalized value, and 𝑥𝑚𝑎𝑥 and 𝑥𝑚𝑖𝑛 are the feature's
approach was handling missing data. Depending on the type maximum and minimum values.
of data, different imputation techniques were used, such as
mode imputation for categorical characteristics and mean 3.2 Data imbalancing using SMOTE
imputation for numerical values. By ensuring a thorough Class imbalance can become a problem due to the nature of
handling of missing entries, this method reduced the network traffic. Because the hacker may target the network
influence on ensuing analyses and forecasts. on a particular day and time, normal network traffic always
outpaces abnormal network traffic. Additionally, they could For every minor category 𝑥𝑖 , the k-nearest neighbor is
employ a variety of attacks to undermine network determined using the Euclidean distance approach.
functionality. Every network type has a corresponding class
label, and the network dataset includes a wide range of Next, choose the closest and a random sample 𝑥𝑗 from a
attributes associated with different kinds of assaults. It group of 𝑥𝑖 closest neighbors. A new instance is now created
significantly affects the detection accuracy and evaluation using formula (2).
criteria of the suggested IDS-INT. The deep learning method
could focus on the primary class during training while 𝑥𝑛𝑒𝑤 = 𝑥𝑖 + |𝑥𝑖 + 𝑥𝑗 | + 𝛿 (2)
disregarding the minor. SMOTE technique is utilized to
A random factor, represented by 𝛿 (0,1) determines the
oversample minority classes in order to solve this problem.
distribution of the newly produced samples. Figure 2 shows
Depending on how similar they are to the earlier samples,
the data imbalance outcomes.
they present different minority class samples. As a result, the
effect of smaller classes becomes more comparable to that of
primary classes. The way SMOTE functions is as follows:

Figure 2 Data Imbalance outcome: Before and after data instances

3.3 Feature Extraction

In order to improve prediction accuracy, our suggested 𝑔[𝑛, 𝑓1 , 𝑓2 ] = 2𝑓2 𝑠𝑖𝑛𝑐(2𝜋𝑓2 𝑛) − 2𝑓1 𝑠𝑖𝑛𝑐(2𝜋𝑓1 𝑛) (7)
framework's feature extraction phase seeks to extract and
preserve the most essential features from datasets related to While the definition of the 𝑠𝑖𝑛𝑐 function is:
chronic diseases. This is accomplished by using an ensemble sin(𝑥)
of LeViT and SincNet, which successfully combines 𝑠𝑖𝑛𝑐(𝑥) = (8)
𝑥
frequency-based and spatial feature representations to extract
highly discriminative features from intricate medical data. A set of trainable parameters 𝑔 is formed by the cutoff
frequencies 𝑓1 and 𝑓2 , which are established in the training
3.3.1 LeViT phase:

LEVIT is a low-power ViT that uses less processing power 𝜃 = {(𝑓𝑖,1 , 𝑓𝑖,2 )|𝑖𝜖𝐶 + ∩ 𝑖 ≤ 𝑙} (9)
to extract both local and global characteristics. With the help
of its attention system, it may analyze Mel-Spectrograms at A block of fully connected layers is fed a single feature vector
various scales. The attention system dynamically switches that has been concatenated from the features that the
between several visual regions to focus on the main speech subnetwork extracted.
problem patterns. The LEVIT model is supported by the
parallelization capacity to produce features faster. The
3.3.3 Feature Fusion Layer
LEVIT model, in contrast to the CNN models, requires a Using an element-wise addition method, the author fused the
significant amount of memory and processing power. features by combining a fusion layer with LeViT. Features
Equation 3 represents the feature map extraction procedures with varying dimensions were found using a dimension-
utilizing the LEVIT model. matching procedure. The feature maps were reshaped into
distinct dimensions using a reshape tool. The elements of
𝐹𝑖 = 𝐿(𝑀), 𝑖 = 1,2, … , 𝑛 (3)
LeViT and SincNet are then combined using element-wise
LeViT represents the model used to extract characteristics addition. The mathematical representation of feature fusion
from the Mel-Spectrogram (M), and 𝐹𝑖 represents the is seen in equation (10).
characteristic map. In the LeViT approach, the authors
∑𝑛𝑖=1 𝑓𝑓𝑢𝑠𝑒𝑑 = ∑𝑛𝑖=1 𝑓𝑆𝑖𝑛𝑐𝑁𝑒𝑡 + ∑𝑛𝑖=1 𝑓𝐿𝑒𝑉𝑖𝑇 (10)
substituted the Linformer transformer mechanism for the
self-attention mechanism. The attention matrix is Where n is the number of features, 𝑓𝑓𝑢𝑠𝑒𝑑 is the fused
approximated by the Linformer self-attention mechanism,
features, 𝑓𝑆𝑖𝑛𝑐𝑁𝑒𝑡 is the SincNet approach extracted features,
which lowers the quadratic complexity.
and 𝑓𝐿𝑒𝑉𝑖𝑇 is the LeViT approach extracted features.
The investigators developed a multi-scale feature
pyramid by combining FPN with the LeViT model. Then, 3.4 Classification using Fuzzy Pooling
the feature extraction procedure from the particular LeViT
layer is started using a top-down pathway and lateral
Convolutional Neural Network (FPCNN)
connection, and iteratively, the features are combined with The process of classification utilizes a robust and adaptable
earlier features to upsample them. Equation 4 describes the learning approach to forecast a number of chronic diseases
procedures for FPN. accurately. This is achieved by using a Fuzzy Pooled
Convolutional Neural Network (FPCNN), which
𝑃𝑛 = 𝐶𝑜𝑛𝑣1×1 (𝑓) + 𝑈(𝑃𝑖+1 ), 𝑛 = 𝑖 − 1, 𝑖 − 2, … . ,1 (4)
incorporates fuzzy logic into the CNN architecture to
Wherein 𝑈() represents the upsample function and 𝑃𝑛 is the strengthen the performance of classification and decision
feature pyramid. To lessen the accuracy of the model weights boundaries across a variety of chronic disease datasets.
and activation, the authors implement quantization-conscious In order to classify data, the Convolution Neural
training during the training process. Furthermore, a loss Network (CNN) carries out a number of operations to extract
function is included to ensure the feature extraction model's intuitive elements from it. In general, a CNN network can be
resilience. The proposed quantization procedure is shown in divided into two main parts: the neural network section and
Eqn. 5. the convolution segment. Convolution, activation, and
pooling are merely some of the procedures that may be
𝐹
𝑄(𝐹) = 𝑟𝑜𝑢𝑛𝑑 ( ) × ∆ (5) included in the CNN's convolution segment. The pooling
∆
procedure decreases the input feature map's height and width.
The quantization function represents 𝑄(𝐹), the accuracy The reduction operation is carried out by the pooling
reduction function is round(), the quantization step size is Δ, procedures known as max-pooling, min-pooling, and
and the features are denoted by 𝐹. average-pooling. We have used an effective fuzzy-based
pooling operation in this work.
3.3.2 SincNet
3.4.1 Fuzzy pooling
The purpose of SincNet layers is to extract low-level
characteristics from data samples of a raw signal. SincNet The fuzzy pooling operation in cutting-edge CNN
layers use convolution on the input signal to train "wavelets" architectures has been used in this study. The fuzzy-based
for feature extraction: pooling approach has two stages. In the initial stage, a fuzzy
membership function receives the input's crisp value.
𝑦[𝑛] = 𝑥(𝑛) ∙ 𝑔(𝑛, 𝜃) (6) Defuzzification is then used to transform the fuzzified values
into clear ones. A map of characteristics with values inside a
While the characteristics of the wavelets, 𝑔 are established range is the input of a pooling layer. Following convolution
during training, and n represents the probe's index. The operations by the convolution layer, the output is subjected to
equation defines the wavelet function 𝑔: an activation function the Rectified Linear Unit (ReLU). The
membership functions' value is 𝑟𝑚𝑎𝑥 = 6. As shown in with great certainty using Equation (14). This choice marks
Figure 3, this study transforms fuzzification using three type- the end of the fuzzification phase.
1 fuzzy membership functions (𝑚𝑓𝑣, where v = 1, 2, 3).
Equations (11) through (13) include the function definitions. 𝑠𝜋𝑍𝑣 = ∑𝑘𝑖=1 ∑𝑘𝑗=1 𝜋𝑣𝑍𝑖,𝑗 (14)

0 𝑍
𝐹𝑀𝑖,𝑗 >𝑑 𝜋́ = {𝜋́ 𝑣𝑍 = 𝜋𝑣𝑍 |𝑣 = 𝑎𝑟𝑔𝑚𝑎𝑥(𝑠𝜋𝑍𝑣 ), 𝑍 = 1,2,3, … , 𝑧} (15)
𝑍
𝑍 𝑑−𝐹𝑀𝑖,𝑗
𝑚𝑓1 (𝐹𝑀𝑖,𝑗 )= 𝑍
𝑐 ≼ 𝐹𝑀𝑖,𝑗 ≼𝑑 (11) The dimensionality of patches is decreased during the
𝑑−𝑐
𝑍 defuzzification phase. We employed a variety of
{ 1 𝐹𝑀𝑖,𝑗 <𝑐 defuzzification techniques in this investigation and
𝑟𝑚𝑎𝑥 𝑑 discovered that the Center of Gravity (COG) method worked
Where 𝑑 = and 𝑐 = well. Equation (16).
2 3

0 𝑍
𝐹𝑀𝑖,𝑗 ≼𝑎 ∑𝑘 𝑘 𝑍 𝑍
𝑖=1 ∑𝑗=1(𝜋́𝑖,𝑗 ∙𝑃𝑖,𝑗 )
𝑧
𝐹𝑀𝑖,𝑗 −𝑎
𝑧
𝑃̇ 𝑧 = ∑𝑘 𝑘 𝑍 (16)
𝑎 ≼ 𝐹𝑀𝑖,𝑗 ≼𝑚 𝑖=1 ∑𝑗=1 𝜋́𝑖,𝑗
𝑚−𝑎
𝑍
𝑚𝑓1 (𝐹𝑀𝑖,𝑗 )= 𝑍 (12)
𝑏−𝐹𝑀𝑖,𝑗 𝑍
𝑚 ≼ 𝐹𝑀𝑖,𝑗 ≼𝑏 Where 𝑝́ = {𝑝́ 𝑍 |𝑍 = 1,2,3, … , 𝑧}
𝑏−𝑚
𝑍
{ 0 𝐹𝑀𝑖,𝑗 ≽𝑏 3.4.2 FP-CNN Framework
𝑟𝑚𝑎𝑥 𝑟𝑚𝑎𝑥 The results of the analysis show that the Xception design with
𝑎= ,𝑚= and 𝑏 = 𝑚 + 𝑎
4 2 fuzzy-based pooling and all internal pooling layers is better
0 𝑍
𝐹𝑀𝑖,𝑗 <𝑟 on performance metrics. The fuzzy pooling layer in the
𝑍
Xception architecture does not take the role of the global
𝑍 𝐹𝑀𝑖,𝑗 −𝑟 𝑍
𝑚𝑓3 (𝐹𝑀𝑖,𝑗 )= 𝑟 ≼ 𝐹𝑀𝑖,𝑗 ≼𝑞 (13) average pooling layer. Fuzzy-based pooling layers have
𝑞−𝑟
𝑍
replaced all of the internal max-pooling layers. Two dense
{ 1 𝐹𝑀𝑖,𝑗 >𝑞 layers with 1024 and 256 neurons were included in every
𝑟𝑚𝑎𝑥 𝑟𝑚𝑎𝑥 model, and the initial dense layer was replaced with a global
Where 𝑟 = and 𝑞 = 𝑟 + pooling layer. For the resultant layer, we have used softmax,
2 4
and for the internal layers, we have used ReLU as an
Each patch's total value (𝑠𝜋𝑍𝑣 )
is its score for being eligible for activation function. The stride was 2×2, and the filters' kernel
the patch's overall membership (𝑝𝑍). The patches are chosen size was 3×3. Three neurons make up the output layer, which
is connected to the global average pooling layer.

Figure 3 Framework of FPCNN method

3.5 Scrutiny Boosted Graph Convolutional LSTM features in temporal dynamics and spatial configuration, as
(SBGC-LSTM) for Monitoring well as assessing the relationship involving the co-occurrence
of the temporal and spatial domains. The three gates of
SBGC-LSTM are an input 𝑖𝑔𝑡 , forget 𝑓𝑔𝑡 , and output 𝑜𝑔𝑡
Continuous tracking and analysis of patients' chronic health
gate, just like those of LSTM. In contrast, a graph
conditions over time is ensured by the monitoring phase. This
convolution operator is used in the construction of these
is accomplished through the use of the Scrutiny Boosted ̂ 𝑡 , and input Xt are in the
Graph Convolutional LSTM (SBGC-LSTM), which gates. Cell memory, hidden state 𝐻𝑆
successfully models dynamic changes in health records and SBGC-LSTM framework. The graph-structured data is
patient development by fusing temporal LSTM learning with termed 𝑐𝑚𝑡 . The spatial and temporal dynamics of the cell
memory (CMt) and hidden state (𝐻𝑆 ̂ 𝑡 ) in SBGC-LSTM are
graph-based spatial insights.
Numerous research have demonstrated that long-term made possible via the graph convolution operator. The inner
temporal connections for sequence modeling may be handled operators of SBGC-LSTM graph convolution computations
well by LSTM, a form of recurrent neural network (RNN). differs from LSTM. By using this inspection approach,
The temporal dynamics of skeletal sequences have been critical nodes' qualities are enhanced to display more
studied using a variety of LSTM-based models. SBGC- discriminative data. The SBGC-LSTM units perform the
LSTM outperforms LSTM in capturing discriminative following tasks:
 |𝑆|!(|𝐹|−|𝑆|−1)!
𝑖𝑔𝑡 = 𝜎(𝕎𝑥𝑖 ∙ 𝔾𝑋𝑡 + 𝕎ℎ𝑖𝑔 ∙ 𝔾 × 𝐻𝑡−1 + 𝑏𝑖𝑔) (17) ∅𝑡 = ∑𝑆⊆𝐹\{𝑡} [𝑓𝑥 (𝑆⋃{𝑖}) − 𝑓𝑥 (𝑆)] (27)
|𝐹|!

𝑓𝑔𝑡 = 𝜎(𝕎𝑥𝑓 ∙ 𝔾𝑋𝑡 + 𝕎ℎ𝑓𝑔 ∙ 𝔾 × 𝐻𝑡−1 + 𝑏𝑓𝑔) (18) Whereas F represents the characteristics, S denotes the
characteristics subset that does not include i, |𝑆| represents the
set of cardinality S, |𝐹| denotes the overall number of
𝑜𝑔𝑡 = 𝜎(𝕎𝑥𝑜 ∙ 𝔾𝑋𝑡 + 𝕎ℎ𝑜𝑔 ∙ 𝔾 × 𝐻𝑡−1 + 𝑏0𝑔) (19) characteristics. Thus, characteristic I's average marginal
contribution across all feature combinations is represented by
𝑐𝑚𝑡 = 𝑓𝑔𝑡 ⨀𝑐𝑚𝑡−1 + 𝑖𝑡 ⨀𝑚𝑖𝑡 (20) the SHAP value ∅𝑡 .

̂ 𝑡 = 𝑜𝑔𝑡 ⊙ 𝑡𝑎𝑛ℎ(𝐶𝑚𝑡 )
𝑐𝑚𝑡 = 𝑓𝑔𝑡 ⨀𝑐𝑚𝑡−1 + 𝑖𝑡 ⊙ 𝑚𝑖𝑡 𝐻𝑆 4 Results and discussions
(21) Predictive algorithms were used in this study to forecast the
development of a variety of illnesses, including diabetes,
̂ 𝑡 = 𝑜𝑔𝑡 ⊙ 𝑡𝑎𝑛ℎ(𝐶𝑚𝑡 )
𝐻𝑆 (22) kidney disease, breast cancer, and heart attacks. In order to
support the predictive models' robustness and applicability,
̂ ) + (𝐻𝑆
𝐻𝑆𝑡 = 𝑓𝑎𝑡𝑡 (𝐻𝑆 ̂) (23) two datasets were used to assess each disease. The inherent
𝑡 𝑡
variety in healthcare data is taken into consideration by using
various datasets for each chronic disease. This method
Equation (17) − (23) above, where ⊙ denotes Hadamard improves the effectiveness of tactics meant to manage and
products and 𝑔 denotes graph convolution operators. A prevent chronic disease by providing a more thorough basis
sigmoid activation function is shown by σ(·). 𝑚𝑖𝑡 stands for for accurate forecasts.
inputs that are modulated. The intermediate hidden states are
denoted by 𝐻𝑡 . We first aggregate all of the nodes' data into 4.1 Experimental Setup
a query feature:
Using PyTorch version 1.11.0, the framework for predicting
̂ chronic diseases was created and implemented using the
𝑞𝑡 = 𝑅𝑒𝐿𝑈(∑𝑁
𝑖=1 𝕎𝐻𝑆𝑡𝑖 ) (24)
PyCharm programming environment's Streamlit-based
interface. The server used for all of the studies was running
Learnable parameter matrices are represented by the 𝕎 in the Windows 10 and had a 16-core Intel(R) CPU running at 2.60
equation (25), and the scrutiny scores of nodes may be GHz, an NVIDIA RTX 3090Ti GPU with support for CUDA
calculated using 11.3, and 24 GB of RAM.

̂𝑡 + 𝕎𝑞 𝑞𝑡 + 𝑏𝑠 ) + 𝑏𝑢 ) (25)
𝑎 = 𝑠𝑖𝑚𝑜𝑖𝑑(𝑉𝑠 𝑡𝑎𝑛ℎ(𝕎ℎ 𝐻 4.2 Dataset Description
Seven different chronic disease datasets are used in this
Biases include 𝑏𝑠 and 𝑏𝑢 . Due to the presence of numerous research. The details are explained below
key joints, nonlinear sigmoid functions, Local features are
created and the weighed sum of focused nodes. Framingham Heart Study dataset (26): The information
for the Framingham Heart Study comes from a significant
𝑔
𝐿𝐹𝑡 = ∑𝑁 ̂ cardiovascular study that was started in 1948 in Framingham,
𝑖=1 𝛼𝑡𝑖 ∙ 𝐻𝑡𝑖 (26)
Massachusetts. Its goal was to pinpoint the traits or common
elements that contribute to cardiovascular disease. The
3.6 Shapley Additive Explanations project began with 5209 participants between the ages of 30
and 62, and it has now been extended to include later
By determining the most significant factors influencing the
generations, yielding a plethora of data covering many
prediction outcome, the interpretation phase guarantees
decades. Variables like age, sex, blood pressure, cholesterol,
transparency. The use of Shapley Additive Explanations body mass index (BMI), smoking status, and diabetes status
(SHAP) improves the predictability and interpretability of the are examples of the dataset commonly used for
chronic illness forecasts by offering transparent, model-
cardiovascular disease prediction.
agnostic insights into feature contributions. Game theory
provides a coherent, theoretically valid, and consistent Thyroid disease datasets (27): The UCI thyroid illness
measure of feature importance, which is provided by Shapley datasets were used to construct the datasets for our
Additive Explanations values. The prediction of a scenario investigation. Numerous datasets related to thyroid disorders
can be explained by using SHAP values, which compute the are kept up to date in the UCI machine learning repository.
impact of every characteristic to the difference between the There are 9172 observation samples in the dataset, with 31
present scenario and the mean prediction for all occurrences. characteristics per sample.
The output of a model may be decomposed in an interpretable
and comprehensive manner using this method, which makes Liver Disease Classification (28): Liver disease
it an effective way to comprehend complex models. The categorization is done using the Indian Liver Patient Dataset
Shapley value, a notion from cooperative game theory that (ILPD) from the UCI machine learning repository. Ten
distributes compensation to players according to their impact attributes and a target variable are spread across 11 columns.
to the overall payoff, serves as the basis for SHAP values. The characteristics include SGPT, SGOT, Alkphos, ALB,
Features are referred to as "players" in machine learning, and albumin and globulin ratio (A/G), TB, DB, TP, and age.
the model's prediction output is known as the "payout." The Patients with and without liver disease are represented by the
following formula is used to determine feature I's Shapley two categories in the output variable. The dataset includes
value: 583 patient records that were gathered from the northeastern
Indian state of Andhra Pradesh.
Pima Indian Dataset (29): The Pima Indians Diabetes Breast cancer dataset (32): Each data point in the Wisconsin
Dataset is a publicly accessible dataset that was used in this Breast Cancer (Diagnostic) dataset is a digital image of a
study. Eight independent parameters pregnancies, BP, ST, breast mass obtained by fine needle aspiration (FNA). Using
insulin, BMI, diabetes pedigree function, and age as well as 30 attributes that were taken from these photos, the dataset is
one dependent parameter outcome are included in this utilized to categorize tumors as either benign (non-cancerous)
dataset, which includes data from 668 female patients in total. or malignant (cancerous). Of the 569 instances (samples) in
it, 357 are benign and 212 are malignant.
Chronic Kidney disease (30): It has 400 samples in two
classes. Age, blood pressure, anemia, diabetes mellitus, and 4.3 Evaluation criteria
hypertension are the characteristics at play. Edema of the
pedals, Pus cells, clusters of Pus cells, red blood cells, The proposed illness prediction framework is assessed using
Bacteria urine's density to volume ratio, The number of four performance evaluation indicators. True positives (TP),
white blood cells, red blood cells, hemoglobin, Volume of which are accurate predictions of the target as a patient with
packed cells, blood's albumin level, serum creatinine levels, a chronic illness; true negatives (TN), which are accurate
Blood sugar, blood urea, and Patient class, potassium, predictions of people without illnesses; false positives (FP),
sodium, and sugar level. which are inaccurate predictions of healthy people as ill; and
false negatives (FN), which are inaccurate predictions of
Alzheimer disease dataset (31): With IDs ranging from healthy people, make up the confusion matrix. The four
4751 to 6900, each of the 2,149 patients in this dataset has parameters used for performance evaluation are described as
comprehensive health information. Demographic follows.
information, lifestyle choices, medical history, clinical
𝑇𝑝+𝑇𝑛
measurements, cognitive and functional evaluations, 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (28)
𝑇𝑝+𝑇𝑛+𝐹𝑝+𝐹𝑛
symptoms, and an Alzheimer's disease diagnosis are all
included in the dataset. Researchers and data scientists who 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑝
(29)
want to investigate risk factors for Alzheimer's, create 𝑇𝑝+𝐹𝑝
predictive models, and do statistical analysis would find the 𝑇𝑝
data suitable. 𝑅𝑒𝑐𝑎𝑙𝑙 = (30)
𝑇𝑝+𝐹𝑛

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑅𝑒𝑐𝑎𝑙𝑙
𝐹 − 𝑀𝑒𝑎𝑠𝑢𝑟𝑒 = 2 ∗ (31)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙

Figure 4 Evaluation of training and testing accuracy and their loss

Evaluation of training and testing accuracy and their losses is illustrated in Figure 4. The learning rate was set as 0.01 and we
trained over 200 epochs.

Figure 5 Home page

Figure 5 shows the home page for the prediction here we can fill the patients details.

Figure 6 Prediction Report page

The chronic prediction report page is shown in Figure 6.

Figure 7 Sample No chronic disease prediction page

No chronic disease sample page is shown in Figure 7.

Table 1 Performance differentiation of Framingham dataset

Approaches Accuracy Precision Recall F-Measure

Random Forest 83.5% 84.5% 85.47% 84.09%
AdaBoost 86% 84.2% 79% 81.4%
XGBoost 91.7% 89.9% 92% 90%
Optimized Random forest 89% 92% 90.5% 89.8%
CNN-LSTM 74.2% 77.1% 77% 77%
BiLSTM 98.8% 98.8% 98.8% 98.8%
DNN 98.1% - 96.7% -
Proposed 99.12% 99.08% 99.06% 99.07%
 Figure 8 Comparison of proposed with existing approaches using Framingham dataset

Table 1 and Figure 8 shows the performance analysis over differentiating with existing approaches, proposed approach
proposed with existing approaches. The existing approaches obtain 99.12%, 99.08% precision, 99.06% recall and 99.07%
like Random Forest, AdaBoost, XGBoost, Optimized F-Measure. The second best result obtain in BiLSTM which
Random forest, CNN-LSTM, BiLSTM and DNN was is 98.8% accuracy, 98.8% precision, 98.8% recall and 98.8%
utilized to compare with proposed approach. While F-Measure.

Figure 9 ROC curve of Framingham dataset

Figure 9 shows the ROC curve of Framingham dataset.

Figure 10 Summary Plot and mean values for SHAP on Framingham dataset

Summary plot and mean absolute values for SHAP on and smoking patterns are important predictors of heart
Framingham dataset is shown in Figure 10. The "Age" feature disease risk that the model employs.
in this case has the greatest absolute mean SHAP, indicating
that it is the most significant feature in terms of mean effect Table 2 Performance differentiation of Thyroid disease dataset
on the output of the categorizer. "sysBP" and "cigsPerDay" Approaches Accuracy Precision Recall F-Measure
are two more noteworthy variables that imply blood pressure AlexNet 89.4% 84.89% 92% 79%
 ResNet50 81.4% 76.9% 81.4% 81.6% compare with proposed approach. While contrasting from
Vgg16 LSTM 85.9% 78.62% 89.4% 90%
existing approaches, proposed approach obtain a better
DNN 98.8% 99.2% 96.56% 93.24%
Proposed 99.08% 99.02% 99% 99.01% outcomes (99.08% accuracy, 99.02% precision, 99% recall
and 99.01% F-Measure.

Figure 11 Comparison of proposed with existing approaches using

Thyroid disease dataset

The performance differentiation of thyroid disease dataset is

shown in table 2 and Figure 11. The existing approaches like Figure 13 ROC curve of thyroid disease dataset
AlexNet, ResNet50, Vgg16-LSTM and DNN are used to
Figure 13 shows the ROC curve of Thyroid disease.

Figure 14 Summary plot and mean absolute values for SHAP on Thyroid disease dataset

The summary plot and mean absolute values for SHAP on Thyroid disease dataset is shown in Figure 14.
Table 3 Performance differentiation of Pima Indians Diabetes Database

Approaches Accuracy Precision Recall F-Measure

Decision Tree 70.77% 67% 69% 67.5%
Random Forest 78.57% 75% 72.5% 73.5%
Kernel SVM 79.22% 76% 72.5% 74%
Naïve Bayes 79.22% 75.5% 74.5% 74.5%
KNN 79.87% 76.5% 75.5% 75.5%
Logistic Regression 82.46% 80% 77% 78%
SVM-K means Clustering 98.7% 98.6% 96.8% 97.5%
Proposed 99.15% 99.10% 99.05% 99.08%
 Figure 15 Comparison of proposed with existing approaches using
Pima Indians Diabetes Database

The performance differentiation of Thyroid disease dataset is

shown in table 3 and Figure 15. The existing approaches like
Decision tree, Random forest, Kernel SVM, Naïve Bayes and
KVM, Logistic Regression and SVM-K Means clustering are
used to compared with proposed approach. While contrasting
from existing approaches, proposed approach obtain a better
outcomes (99.15% accuracy, 99.10% precision, 99.05%
recall and 99.08% F-Measure.

Figure 16 Summary plot and Mean absolute values for Pima Indians Diabetes Database

Figure 16 shows the summary plot and Mean absolute values for Pima Indians Diabetes Database.

Figure 17 ROC curve of diabetes disease dataset

Figure 17 shows the ROC curve of Diabetes disease dataset.

Table 4 Performance differentiation of Chronic Kidney disease dataset

Approaches Accuracy Precision Recall F-Measure

ANN 99.33% 95.14% 94% 94.57%
KNN 99.5% 95.98% 95.61% 94.75%
Gradient Boosting 99% 92.33% 91.61% 91.97%
FuDNN-FOSMO 99.7% 98.85% 99.6% 99.7%
EOAEDL-CKDD 98.12% 98.16% 97.83% 97.99%
1D CorrNN-LSTM 96% 97.82% 98.43% 98.60%
Proposed 99.20% 99.06% 99.06% 99.06%

The performance differentiation of Chronic Kidney disease

dataset is shown in table 4 and Figure 18. While contrasting
from existing approaches, proposed approach obtain a better
outcomes (99.20% accuracy, 99.06% precision, 99.06%
recall and 99.06% F-Measure.
 Figure 18 Performance comparison

Figure 19 ROC curve of kidney disease

Figure 19 shows the ROC curve of Kidney disease dataset.

Figure 20 Summary plot and Mean absolute values for Chronic Kidney disease dataset

Figure 20 shows the summary plot and Mean absolute values Figure 21 Performance comparison using Indian Liver dataset
for Chronic Kidney disease dataset.
The performance differentiation of Indian Liver dataset is
Table 5 Performance differentiation of Indian Liver Patient shown in table 5 and Figure 21. While contrasting from
Records existing approaches, proposed approach obtain a better
outcomes (99.18% accuracy, 99.10% precision, 99.07%
Methods Accuracy Precision Recall F1-score
(%) (%) (%) (%) recall and 99.09% F-Measure).
DT 74.78 70.86 88.43 78.68
RF 79.57 89.40 81.33 85.17
NB 78.67 81.88 86.75 84.24
Proposed 99.18 99.10 99.07 99.09
Method

Figure 22 ROC curve of Liver disease

Figure 22 shows the ROC curve of Liver disease.

 Figure 23 Summary plot and Mean absolute values for Indian Liver dataset.

Figure 23 shows the summary plot and Mean absolute values Figure 24 Performance comparison using Alzheimer's disease and
for Indian Liver dataset. Healthy Aging Data

Table 6 Performance differentiation of Alzheimer's Disease and The performance differentiation of Alzheimer's disease and
Healthy Aging Data Healthy Aging Data is shown in table 6 and Figure 24. While
contrasting from existing approaches, proposed approach
Model Accuracy Precision Recall F1-Score
(%) (%) (%) (%)
obtain a better outcomes (99.02% accuracy, 98.91%
GCN 89.23 71.0 56.0 62.6 precision, 98.83% recall and 98.85% F-Measure).
GAT 89.23 80.0 44.0 57.1
Multi-Level 91.30 89.5 88.2 88.8
GCN
Bi-GRU + 88.50 86.0 84.0 84.9
LSTM
Proposed 99.02 98.91 98.83 98.85

Figure 25 ROC curve of Alzheimer disease

Figure 25 shows the ROC curve of Alzheimer disease.

Figure 26 Summary plot and Mean absolute values for Alzheimer's disease and Healthy Aging Data

Figure 26 shows the summary plot and Mean absolute values Table 7 Performance differentiation of Breast Cancer Wisconsin
for Alzheimer's disease and Healthy Aging Data (Diagnostic) Data Set
 Model Accuracy Precision Recall F1-Score Figure 27 Performance comparison using Breast Cancer
(%) (%) (%) (%) Wisconsin (Diagnostic) Data Set
MLP + RFE 99.12 98.91 98.42 98.66
RFE + 98.25 98.13 96.63 98.59 The performance of proposed with existing approaches using
GWO +
Breast Cancer Wisconsin (Diagnostic) Data Set is shown in
Ensemble
CNN 98.25 97.52 98.21 97.86 table 7 and Figure 27. While differentiating with existing
LSTM 95.61 94.85 94.73 94.79 approaches proposed approach obtain superior performances.
Proposed 99.25 99.12 99.08 99.10

Figure 28 ROC curve of breast cancer

Figure 28 shows the ROC curve of breast cancer disease

dataset.

Figure 29 Summary plot and Mean absolute values for Breast Cancer Wisconsin (Diagnostic) Data Set

Figure 29 shows the summary plot and Mean absolute values for Breast Cancer Wisconsin (Diagnostic) Data Set
Table 8 K-Fold Cross validation

Fold (k) Accuracy (%) Precision (%) Recall (%) F1-Score (%)
Fold 1 99.30 99.10 99.06 99.08
Fold 2 99.27 99.11 99.09 99.10
Fold 3 99.29 99.15 99.08 99.11
Fold 4 99.32 99.14 99.12 99.13
Fold 5 99.25 99.10 99.04 99.07
Fold 6 99.28 99.12 99.10 99.11
Fold 7 99.26 99.08 99.06 99.07
Fold 8 99.29 99.13 99.10 99.12
Fold 9 99.31 99.16 99.11 99.13
Fold 10 99.28 99.12 99.08 99.10
Mean ± Std 99.28 ± 0.02 99.12 ± 0.02 99.08 ± 0.02 99.10 ± 0.02

K-fold cross-validation is a method for splitting data into In order to assess a model's performance, predictive analysis
subsets for model training and testing. This method is typically uses a dataset that has the necessary class labels for
primarily utilized with predictive models to assess how well training. Once the model is constructed, it is tested on a fresh
the model will function when exposed to new, unseen data. dataset with undetermined output labels. Split data and cross-
validation are the two methods for dataset splitting. Since K- ANN-PSO 0.48
CNN and KNN 0.29
fold cross-validation performs well on biased datasets, it has
RF 0.22
the benefit over the split data technique. The number of times Proposed 0.12
the data is trained using k-1 training subsets and 1 testing
subset is denoted by K. In cross-validation, every occurrence
is examined for both testing and training purposes. To train Table 9 shows the comparison of computational time. While
and evaluate categorization models, ten-fold cross-validation comparing with existing approaches proposed approach takes
is employed. The k-fold cross validation is shown in table 8. less time.
Table 9 Comparison of Computational time 4.4 Ablation Study
Method Running time (ms)
In this section we analyze the performances of ablation study.
MLP 0.45
L-BOA 0.34 Table 10 shows the comparison of ablation analysis.
Table 10 Ablation analysis

Phase Accuracy Precision Recall F-Measure

Preprocessing 91.45% 90.80% 89.72% 90.25%
Preprocessing+ Feature Extraction 95.78% 95.10% 94.85% 94.97%
Preprocessing+ Feature Extraction+ Prediction 97.63% 97.18% 96.90% 97.04%
Preprocessing+ Feature Extraction+ Prediction+ AI 98.72% 98.35% 98.18% 98.26%
Monitoring
Preprocessing + Extraction+ Prediction+ AI 99.28% 99.12% 99.08% 99.10%
Monitoring+ Explainable AI

The effects of each stage in the suggested framework are 4.5 ANOVA Statistical Analysis
shown in Table 10. The accuracy is 91.45% when
preprocessing alone is used. Accuracy increases to 95.78% In order to determine if the distinction among the mean of any
when feature extraction is added, and to 97.63% when the number of groups is significant, the ANOVA test uses sample
prediction model is added. The accuracy of AI-based data. Since the results of the right-tailed ANOVA test and
monitoring is 98.72%. Finally, including explainable AI the two-tailed pooled variance t-test are identical for two
techniques like SHAP raises the overall performance to groups, ANOVA is typically utilized when there are three or
99.28%, confirming the efficiency of each component in more groups.
enhancing chronic disease prediction.
Table 11 Statistical analysis

Source DF Sum of Square Mean Square F Statistic P-Value

Groups ( Between 5 6500.2835 1300.0567 15.2111 9.397e-12
Groups)
Error ( Within Groups) 130 11110.8183 85.4678
Total 135 17611.1018 130.4526

H0 is rejected since the p-value is less than α. Several of the 15.211064, which is below the 95% area of acceptance, (0:
groupings' means are thought to be uneven. In other words, 2.2839). The effect size f that was observed is large (0.76).
the representative averages of certain categories differ in a That suggests a significant magnitude of the averages'
way that is statistically significant. The (p(x ≤ F) = 1 P-value discrepancy. The value of η2 is 0.37. It indicates that 36.9%
is 9.39659e-12. 9.397e-12 (9.4e-10%) shows that the of the variance from the average can be explained by the
likelihood of a type 1 error is low. The p-value decreases group. Table 11 shows the statistical analysis
with increasing support for H1. The test statistic F is

Table 12 Turkey test analysis

Pair Difference SE Q Lower CI Upper CI Critical Mean p-value

x1-x2 8.7182 1.8634 4.6787 1.0951 16.341 7.6228 0.01507
x1-x3 7.4565 1.9139 3.896 -0.3729 15.286 7.8294 0.017159
x1-x4 3.7055 2.1926 1.6914 -5.2612 12.672 8.9697 0.8381
x1-x5 7.4225 2.0672 3.5906 -0.0324 14.877 8.4567 0.1206
x1-x6 8.033 2.0063 4.0021 0.7452 15.321 8.2042 0.0273
x2-x3 16.1747 1.6916 9.5615 9.2544 23.095 6.9203 6.337e-9
x2-x4 12.4237 2.0162 6.1605 3.1795 21.669 8.1882 0.4888
x2-x5 16.1407 1.8946 8.5195 8.1987 24.083 7.8297 0.00003247
x2-x6 16.7512 1.8634 8.9897 9.1242 24.378 7.6228 4.794e-8
x3-x4 -3.751 2.0487 1.5499 -7.693 8.3809 8.2034 0.002463
x3-x5 -0.034 2.0672 -0.016 -7.489 7.4206 7.8294 0.000002902
x3-x6 0.5765 2.0063 0.2874 -6.711 7.8644 7.8294 0.9999
x4-x5 3.717 2.1926 1.6963 -5.2637 8.9697 8.9697 0.8372
x4-x6 11.7415 2.0956 5.535 2.7718 20.711 8.9697 0.003119
x5-x6 15.4555 2.0672 7.4765 6.9988 23.912 8.4567 0.000007529
Table 12 shows the Turkey test analysis.
Table 13 Summary of related work with proposed

Reference Advantage Drawback How Proposed Approach Accuracy (%)

Overcomes
(Rajeashwari et al., 2024) Dual Deep CNNs extract ME-RF and XGBoost have Uses ensemble of LeViT 96.85
relevant features effectively limited generalization and and SincNet for richer
suffer from low feature learning and
predictability with many FPCNN for generalization
DT leaves
(Malibari et al., 2023) EO-optimized lightweight Restricted to heart and Proposed model handles 7 95.91
CNN for sensor-based kidney disease only chronic diseases with
prediction higher flexibility and
scalability
(Khushal et al., 2025) Fuzzy + QML improves QML becomes inaccurate Uses fuzzy logic with deep 97.25
accuracy and computation with large-scale data learning and LSTM for
scalable and accurate
monitoring
(Halder et al., 2024) Web-based CKD prediction Focused only on CKD, Framework handles multi- 94.70
with smart pre-processing lacks interpretability disease prediction with
SHAP for explainability
(Mandava et al., 2024) Deep ensemble for CVD Focused on CVD; high Uses fuzzy pooling and 96.02
prediction with LASSO & preprocessing dependency hybrid feature extraction,
Relief reducing preprocessing cost
(Singh et al., 2024) Hybrid soft computing + Limited to breast cancer; Proposed model extracts 96.50
ML classifiers improves high dimensionality optimal features with fuzzy
feature selection filtering + GCN-LSTM for
broader use
(Rajeswari et al., 2025) Ontological RRNN + Focused only on breast Our framework is 97.11
Crayfish Optimization cancer; high tuning lightweight and less reliant
gives strong classification complexity on tuning; supports
multiple diseases
(Ayus et al., 2024) Hybrid CNN-Conv1D- Needs large training data SBGC-LSTM handles 98.00
LSTM improves Alzheimer and suffers from sequence temporal learning with
detection overfitting regularization, improving
generalization
(Aytaç et al., 2025) COOT-based CNN Only applicable to thyroid Our model generalizes 95.84
parameter tuning improves cancer; lacks multi-disease across 7 diseases and needs
thyroid detection design fewer parameter
optimization cycles
(Amin et al., 2023) Integrated feature Accuracy suffers due to Our deep learning-based 94.20
extraction with multiple ML limitations on complex approach captures complex
ML classifiers patterns patterns better with higher
accuracy
Proposed approach 99.28

Table 13 shows the performance and summary overview of Convolutional Neural Network (FPCNN) for precise
the proposed with existing approach comparison. classification, hybrid feature extraction utilizing LeViT and
SincNet. SHAP for interpretability and a Scrutiny Boosted
4.6 Proposed limitations Graph Convolutional LSTM (SBGC-LSTM) for efficient
illness monitoring further improve the model. For diabetes,
Even though the proposed fuzzy-based deep learning liver, thyroid, heart disease, kidney, Alzheimer and breast
framework showed excellent accuracy on several datasets cancer, our model achieved prediction accuracy of 99.25%,
related to chronic diseases, there are still some drawbacks. 99.18%, 99.08%, 99.12%, 99.20%, 99.02% and 99.25%.
The benchmark datasets used to test the model might not Although our suggested model yielded the best results, it was
accurately reflect the unpredictability and inconsistencies discovered that the data features employed for outcome
prevalent in actual clinical data. Additionally, it works prediction had a substantial impact on the accuracy of
effectively for the chosen chronic illnesses, but it has not yet classification. As a result, our suggested model is applied to
been proven to be beneficial for other uncommon or less various features on up to two datasets for every chronic
prevalent illnesses. Finally, the model relies on preprocessed illness. The outcomes demonstrated that our suggested
and structured input data; more research is needed to see how strategy worked better than alternative approaches. Our
well it performs with unprocessed or partial data. methods took less time to process than random forest, deep
learning, and SVM, according to a comparison of classifier
5 Conclusion and Future Scope
processing times. Our research can be extremely helpful in
Chronic disease early diagnosis is a significant research hospitals and medical labs that use fuzzy-based prediction
challenge. Numerous artificial intelligence methods are algorithms to forecast a variety of ailments. Reducing patient
employed in the literature to classify medical data and data attributes for medical reasons may benefit from further
forecast diseases. With a small collection of characteristics, optimizing key aspects.
these methods can frequently be used to diagnose particular For improved generalizability, future studies can
diseases in selected datasets. In order to overcome these concentrate on expanding the suggested fuzzy-based deep
obstacles, this research offers a thorough fuzzy-based deep learning framework to bigger and more varied clinical
learning architecture that combines a Fuzzy Pooling datasets. Optimizing important feature subsets and
integrating real-time hospital data could significantly Kumar, Y., Kaur, I., & Mishra, S. (2024) ‘Foodborne disease
symptoms, diagnostics, and predictions using artificial
increase diagnostic accuracy while lowering computational
intelligence-based learning approaches: A systematic review’,
costs. Archives of Computational Methods in Engineering, Vol. 31,
no.2, pp.553-578. [Link]
Declaration 0
Abhin, B., Kumar M, H., Kamath S, S., & Sugumaran, V. (2025, July).
Conflict of interests: The authors declare that they have no Leveraging Generative AI for Chronic Kidney Disease
Prediction: A Comparative Study of Open-Source LLMs. In
known competing financial interests or personal relationships
International Conference on Applications of Natural Language
that could have appeared to influence the work reported in to Information Systems (pp. 216-228). Cham: Springer Nature
this paper. Switzerland.
Mohamed, N., Almutairi, R. L., Abdelrahim, S., Alharbi, R.,
Acknowledgements Alhomayani, F. M., & Elhag, A. A. (2024) ‘Towards precise
chronic disease management: A combined approach with
We declare that this manuscript is original, has not been binary metaheuristics and ensemble deep learning’, Journal of
Radiation Research and Applied Sciences, Vol. 17, no.4,
published before and is not currently being considered for pp.101092. [Link]
publication elsewhere. Akour, I., Nuseir, M. T., Alshurideh, M. T., Alzoubi, H. M., Al Kurdi,
B., & AlHamad, A. Q. M. (2024) Explainable artificial
Funding: No funding was received to assist with the intelligence (EAI) based disease prediction model. In Cyber
preparation of this manuscript. Security Impact on Digitalization and Business Intelligence:
Big Cyber Security for Information Management:
Availability of data and material: Data will be available Opportunities and Challenges (pp. 207-221). Cham: Springer
International Publishing. [Link]
when requested 31801-6_12
Asif, S., Zhao, M., Li, Y., Tang, F., Ur Rehman Khan, S., & Zhu, Y.
Authors' contributions: The author confirms sole (2024) ‘AI-based approaches for the diagnosis of Mpox:
responsibility for the following: study conception and design, challenges and future prospects’, Archives of Computational
data collection, analysis and interpretation of results, and Methods in Engineering, Vol. 31, no.6, pp.3585-3617.
manuscript preparation. [Link]
Asif, S., Zhao, M., Li, Y., Tang, F., Ur Rehman Khan, S., & Zhu, Y.
(2024) ‘AI-based approaches for the diagnosis of Mpox:
Ethics approval: This article does not contain any studies challenges and future prospects’, Archives of Computational
with human participants or animals performed by any of the Methods in Engineering, Vol. 31, no.6, pp.3585-3617.
authors. [Link]
Kalra, N., Verma, P., & Verma, S. (2024) ‘Advancements in AI based
References healthcare techniques with FOCUS ON diagnostic
techniques’, Computers in Biology and Medicine, Vol. 179,
Islam, R., Sultana, A., & Islam, M. R. (2024) ‘A comprehensive review pp.108917.
for chronic disease prediction using machine learning [Link]
algorithms’, Journal of Electrical Systems and Information Rajeashwari, S., & Arunesh, K. (2024) ‘Chronic disease prediction
Technology, Vol. 11, no.1, pp.27. with deep convolution based modified extreme-random forest
[Link] classifier’, Biomedical Signal Processing and Control, Vol.
Kerth, J. L., Hagemeister, M., Bischops, A. C., Reinhart, L., Dukart, J., 87, pp.105425. [Link]
Heinrichs, B., & Meissner, T. (2024) ‘Artificial intelligence in Malibari, A. A. (2023) ‘An efficient IoT-Artificial intelligence-based
the care of children and adolescents with chronic diseases: a disease prediction using lightweight CNN in healthcare
systematic review’, European journal of pediatrics, Vol. 184, system’, Measurement: Sensors, Vol. 26, pp.100695.
no.1, pp.83. [Link] [Link]
Huang, M., Zhang, X. S., Bhatti, U. A., Wu, Y., Zhang, Y., & Ghadi, Khushal, R., & Fatima, U. (2025) ‘Fuzzy quantum machine learning
Y. Y. (2024) ‘An interpretable approach using hybrid graph (FQML) logic for optimized disease prediction’, Computers in
networks and explainable AI for intelligent diagnosis Biology and Medicine, Vol. 192, pp.110315.
recommendations in chronic disease care’, Biomedical Signal [Link]
Processing and Control, Vol. 91, pp.105913. Halder, R. K., Uddin, M. N., Uddin, M. A., Aryal, S., Saha, S., Hossen,
[Link] R., & Akter, M. F. (2024) ‘ML-CKDP: Machine learning-
Chunduru, A., Kishore, A. R., Sasapu, B. K., & Seepana, K. (2024) based chronic kidney disease prediction with smart web
‘Multi chronic disease prediction system using CNN and application’, Journal of Pathology Informatics, Vol. 15,
random forest’, SN Computer Science, Vol. 5, no.1, pp.157. pp.100371. [Link]
[Link] Mandava, M. (2024) ‘MDensNet201-IDRSRNet: Efficient
Patterson, E. J., Bounds, A. D., Wagner, S. K., Kadri-Langford, R., cardiovascular disease prediction system using hybrid deep
Taylor, R., & Daly, D. (2024) ‘Oculomics: a crusade against learning’, Biomedical Signal Processing and Control, Vol. 93,
the four horsemen of chronic disease’, Ophthalmology and pp.106147. [Link]
Therapy, Vol. 13, no.6, pp.1427-1451. Singh, L. K., Khanna, M., & Singh, R. (2024) ‘An enhanced soft-
[Link] computing based strategy for efficient feature selection for
Ram, M., Afrash, M. R., Moulaei, K., Parvin, M., Esmaeeli, E., timely breast cancer prediction: Wisconsin Diagnostic Breast
Karbasi, Z., & Sabahi, A. (2024) ‘Application of artificial Cancer dataset case’, Multimedia Tools and Applications, Vol.
intelligence in chronic myeloid leukemia (CML) disease 83, no.31, pp.76607-76672. [Link]
prediction and management: a scoping review’, BMC cancer, 024-18473-9
Vol. 24, no.1, pp.1026. [Link] Rajeswari, V., & Priya, K. S. (2025) ‘Ontological modeling with
12764-y recursive recurrent neural network and crayfish optimization
Smokovski, I., Steinle, N., Behnke, A., Bhaskar, S. M., Grech, G., for reliable breast cancer prediction’, Biomedical Signal
Richter, K., & Golubnitschaja, O. (2024) ‘Digital biomarkers: Processing and Control, Vol. 99, 106810.
3PM approach revolutionizing chronic disease management— [Link]
EPMA 2024 position’, EPMA Journal, Vol. 15, no.2, pp.149- Ayus, I., & Gupta, D. (2024) ‘A novel hybrid ensemble based
162. [Link] Alzheimer’s identification system using deep learning
Ma, C. Y., Luo, Y. M., Zhang, T. Y., Hao, Y. D., Xie, X. Q., Liu, X. technique’, Biomedical Signal Processing and Control, Vol.
W., & Lin, H. (2024) ‘Predicting coronary heart disease in 92, pp.106079. [Link]
Chinese diabetics using machine learning’, Computers in Aytaç, Z. İ., İşeri, İ., & Dandil, B. (2025) ‘A hybrid coot based CNN
Biology and Medicine, Vol. 169, pp.107952. model for thyroid cancer detection’, Computer Standards &
[Link] Interfaces, pp.104018.
[Link]
Amin, R., Yasmin, R., Ruhi, S., Rahman, M. H., & Reza, M. S. (2023)
‘Prediction of chronic liver disease patients using integrated
projection based statistical feature extraction with machine
learning algorithms’, Informatics in Medicine Unlocked, Vol.
36, pp.101155. [Link]
[Link]
study-dataset
[Link]
[Link]
[Link]
database
[Link]
[Link]
healthy-aging-data
[Link]
data?select=[Link]