(PDF) Secure and Efficient IoT–Cloud Healthcare Framework for Disease Prediction Using an Optimized ProbSparse Graph Network

Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6170 E-ISSN: 2349 5359; P-ISSN: 2454-9967 Secure and Efficient IoT–Cloud Healthcare Framework for Disease Prediction Using an Optimized ProbSparse Graph Network R. Reka1*, S. Anuradha2, K. Gayathri Devi3, M. Sathiya4 1Department of AI & DS, Mahendra College of Engineering, Mahendra - Salem Campus, Salem - Chennai Highway, Minnampalli Post, Salem, Tamilnadu-636106, India 2Department of Computer Science and Applications, SRM Institute of Science and Technology (FSH), Ramapuram Campus, Chennai, Tamilnadu – 600089, India 3Department of Electronics and Communication Engineering, Nandha College of Technology, Erode - Perundurai Main Road, Vaikkalmedu, Pitchandampalayam, Erode, Tamilnadu- 638052, India 4Department of Computer Science and Engineering, Karpagam Institute of Technology, Bodipalayam Post, Salem - Kochi Hwy, Seerapalayam, Coimbatore, Tamilnadu-641021, India ABSTRACT: The rapid integration of cloud computing and Internet of Things (IoT) technologies has transformed modern healthcare into an intelligent, data-driven ecosystem capable of continuous monitoring and real-time analysis of patient information. Cardiovascular disease remains the leading global cause of mortality, demanding early, accurate, and secure diagnostic solutions. This study introduces a secure and efficient disease classification framework, the Multi-Head ProbSparse Cascaded Graph Network optimized with Draco Lizard Optimizer (MH-PSCGN-DLO), designed to enhance prediction accuracy while ensuring strong data protection. The system incorporates MultiLevel Encryption (MLE) to safeguard sensitive medical data, followed by an advanced preprocessing pipeline—Adjusted Min-Max, Decimal Scaling, and Statistical Column Normalization (AMM-DS-SCN)— to reduce noise and standardize heterogeneous IoT data. Parrot Optimization (PO) selects the most informative features, while MH-PSCGN combines Disentangled Cascaded Graph Convolution and MultiHead ProbSparse Self-Attention for robust classification. The DLO optimizer further enhances model stability and convergence. Experiments on the cardiovascular disease dataset demonstrate 98.4% accuracy, 0.01 error rate, and significantly lower encryption/decryption times compared to existing methods. These results indicate that MH-PSCGN-DLO provides a reliable, secure, and computationally efficient foundation for real-time clinical decision support, with strong potential for future extension to other chronic disease domains. KEYWORDS: Cardiovascular Disease Classification, Draco Lizard Optimizer, Multi-Head ProbSparse Self-attention Network, Multi-Level Encryption, and Parrot Optimization. https://doi.org/10.29294/IJASE.12.3.2026.6170-6192 ©2026 Mahendrapublications.com, All rights reserved 1. INTRODUCTION The development of digital technologies with an accelerated pace has turned healthcare into a networked, smart, and data-driven ecosystem. One such innovation has been the creation of Internet of things (IoT) in combination with cloud computing that has been shown to be an effective and secure solution to efficient and scaled healthcare deliveries [1]. The IoT network can be integrated with learning equipment, such as wearable sensors and smart watches that, in turn, can constantly monitor the vital signs of the patient, including heart rate, blood pressure, and electrocardiogram (ECG) signals, among others [2]. Together with cloud infrastructure, this kind of data streams may be stored and analyzed in real-time so that physicians were able to diagnose patients early, provide them with treatment, and manage patients remotely [3]. It is known that this crossroads does not leave indifferently the diagnosis and classification of *Corresponding Author:

Received: 15.12.2025 Accepted: 19.02.2026 Published on: 25.03.2026 Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6171 cardiovascular illnesses (CVDs) that represent the major cause of mortality globally [4]. Among the conditions, which are covered by the umbrella term of cardiovascular disease, there is heart failure, arrhythmia, coronary artery disease, and stroke [5]. The World Health Organization estimates that CVDs cause 17.9 million deaths annually or 31 percent of all deaths in the world [6]. The disease is especially widespread in poorly accessible and underdeveloped countries where the access to specialist medical centers is poor [7]. Timely intervention and early identification of the risk factors of CVDs would be very important in reducing mortality and enhancing quality of life [8]. Nevertheless, classical diagnostic techniques are based on face-to-face meetings and paperbased processing of medical records, which could not be useful during emergency or rural medical conditions [9]. IoT-enabled healthcare solutions based on cloud computing assist in eliminating these limitations by monitoring the patient in real-time, providing a remote diagnosis, and making predictive analytics [10]. The devices of the IoT will be capable of ongoing gathering of cardiovascular-related information of the patients in a real environment, thus complying with the minimum number of visits to hospitals [11]. The cloud can also be used to offer scalable storage and the highest level of computer infrastructure to accommodate the processing of enormous amounts of health data by means of machine-learning and neural network models. [12]. It is through these systems that healthcare professionals are able to group diseases correctly and produce an immediate alert in the case of anomalies and is also enabling patients take an active role in the management of its own health [13]. Despite these advantages, some issues related to the development of reliable IoT-cloud medical solutions still persist. One of the largest concerns are security and privacy of sensitive information about patients. [14]. It has a high likelihood of healthcare data being exposed to cyber-attacks and illegal access along with abuse. One violation has the potential to damage patient confidence and clinical care. It is therefore urgent to incorporate involving safe encryption and authentication activities in the IoT-cloud systems. The other problem lies in the way to ameliorate the condition of the disease classification models. Noise, inconsistencies, or even unrecorded values can be present in the health data collected by IoT devices and may decrease predictive performance. Such issues demand sophisticated preprocessing and E-ISSN: 2349 5359; P-ISSN: 2454-9967 powerful machine approaches, including functional neural networks [15]. In this study, the secure cardiovascular disease classification IoTcloud framework is suggested. The assurance of data confidentiality is provided with the use of encryption, the cleaning of raw health data is performed with the help of preprocessing, significant properties are determined with the help of feature selection, classification is provided by neural networks, and optimization increases the accuracy and efficiency. To sum up, the combination of cloud and Internet of Things is a major innovation in the struggle against cardiovascular diseases. With the integration of secure data management and ingenious techniques of classifying data, healthcare systems will be able to initiate improved patient results, mitigate the strain of heart disease, and make contributions to millions of lives saved across the world. This paper has novelty and contributions as follows:  The proposed MH-PSCGN-DLO system has a synthesis of features of selection, graph-based learning, attention optimization, to achieve high accuracy classification at the cost of interpretability and structural information.  Multi-Level Encryption (MLE) system is offered, which is premised on stratified security mechanisms to guarantee a feeling of confidence, prevent unauthorized access, and provide an increased level of protection of patient data.  Adjusted Min‐Max with Decimal Scaling and Statistical Column Normalization (AMM-DSSCN) as preprocessing techniques to provide both efficiency in controlling the different attributes and enhance homogeneity of the data.  The special Parrot Optimization Algorithm is the model of the feature selection which is able to minimize the redundancy, improve relevance and maximize computational efficiency in process of medical data.  The model combines Disentangled Cascaded Graph Convolution Networks (DCGNN), MultiHead ProbSparse Attention, which makes sure that it classifies cardiovascular diseases more effectively with high interpretability and robustness.  The novel Algoritive Draco Lizard Optimizer (DLO) improves the capabilities of the classification by adjusting the balance between exploration and exploitation in real Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6172 time, thus reducing the error rates and improving convergence time. In 2021, Adewole, et al. [16] introduced a cardiovascular disease prediction cloud-based IoMT framework. The strategy is a combination of cloud computing and Internet of Things to improve healthcare service provision through efficient utilization of resources, energy conservation, storage, and processor power. This framework takes advantage of paper-based data to solve problems in hospitals with particular interests in evaluating, treating, and prognosing heart diseases. In 2022, Nancy, et al. [17] designed an IoTCloud-based smart-healthcare heart disease prediction system based on deep learning. The system predicts the diseases using the risk of heart diseases by collecting IoT and electronic clinical data and using Bi-LSTM (bidirectional long short-term memory) to perform predictive analytics and accurately monitor the diseases. In 2025, Verma et al. [18] introduced an adaptive, secure IoT-cloud disease classification framework. The method utilizes MATLAB for computational assessment and combines ant lion optimization and generalized fuzzy intelligence for predicting illness and calculating severity. It also uses new Elapid encryption for safe cloud storage. In 2023, Pati et al. [19] developed an IoT-FogCloud integrated framework for real-time diagnosis of cardiovascular disease. The method combines deep neural networks with ensemble methods and integrates Fog computing to support instantaneous remote diagnosis, low latency, and efficient energy use, trained on multiple datasets related to heart disease. In 2023, Raheja et al. [20] presented an IoTenabled secure healthcare framework for heart disease diagnosis. The technique uses deep CNN classification after preprocessing ECG signals with Savitzky-Golay filtering & MOWPT. Security and authentication are ensured by triple DES encryption & water cycle optimization, while ThingSpeak is utilized to integrate IoT devices for remote monitoring. In 2021, Verma, et al. [21] proposed a hybrid intelligent cloud IoT system in the field of disease prediction. The technique employs a generalizedfuzzy-intelligence-based gray wolf ant lion optimization (GMI-GWALO) technique that E-ISSN: 2349 5359; P-ISSN: 2454-9967 predicts illness and selects severity of the illness, which is developed within the MATLAB to take a correct diagnosis of the disease. Moreover, a secure cloud transmission is used with the help of hybrid Elapid encryption. In 2024, Janarthanan et al. [22] created the secure e-health system in predicting heart disease. In order to predict the disease, it employs Hybrid Binary Particle Firefly Optimized Extreme Learning Machine (HybBPF-ELM) with an intelligent encryption and decryption system (IEDF), which applies AES, DES, RSA, MBF, and Automatic Sequence Encryption to ensure safety in cloud applications. In 2022, Singh et al. [23], developed an IoTbased smart healthcare cloud consisting of fog computing and AI applications to identify thyroid diseases. To identify the patient, the method will use ensemble-based classifier. Information security would be provided through encryption and decryption algorithms data confidentiality and the system performance would be measured based on latency, network usage, RAM usage and energy usage. The increasing volume of sensitive medical data in IoT-cloud medical chip systems poses grave issues in attaining precise disease classification, besides maintaining data security. The existing methods are normally described as costly calculations, time wastage and vulnerability to attacks. To address such problems, this study introduces the MH-PSCGN-DLO model that is a combination of Multi-Head ProbSparse Cascaded Graph Network and Draco Lizard Optimization to find the optimal games to predict diseases in a way that avoids threats and inaccuracies. The general aim of this study is to design a secure, precise, and scalable IoT-cloud-based cardiovascular disease classification paradigm by incorporating state-of-the-art preprocessing, feature selection, graph-based learning, attention, and optimization methods. Particularly, the study will also focus on: (i) make sensitive medical data confidential and intact through a multi-level encryption scheme; (ii) enhance the quality of data by using powerful data normalization methods; (iii) choose the most suitable clinical characteristics by employing metaheuristic optimization; and Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6173 (iv) improve classification accuracy and reliability with an MH-PSCGN model optimized on DLO. The proposed study hypothesizes that probabilistic sparse attention integration with graph-based learning and bio-inspired optimization can significantly enhance cardiovascular disease classification accuracy while preserving data security in the IoT-cloud healthcare setting. This work is essential because it provides a holistic approach to security, feature optimization, and intelligent disease classification on a single IoT-cloud platform. In contrast to current methods that can be applied to ensure security or prediction performance, the proposed MH-PSCGN-DLO model can preserve data secrecy while guaranteeing high classification accuracy. This innovation lies in the fact that, using Multi-Head ProbSparse Cascaded Graph Networks and the Draco Lizard Optimizer, both patient-feature interactions and network E-ISSN: 2349 5359; P-ISSN: 2454-9967 parameter optimization can be performed efficiently. This hybrid method is more robust, interpretable, and scalable to real-time prediction of cardiovascular disease. MATERIALS AND METHODS The proposed cardiovascular disease predictive framework (Fig. 1) will combine datasets on cardiovascular disease, medical check-up records, and IoT-enabled health monitoring devices. The preprocessing of the collected data is carried out through modified Min–Max normalization, statistical Normalization, decimal scaling, and multi-level encryption to guarantee data security. This encrypted information is then sent to the AWS IoT cloud, where Parrot Optimization is used to identify the most useful clinical features [24]. Lastly, a Multi-Head ProbSparse Cascaded Graph Network, optimized by the Draco Lizard Optimizer, is adopted to achieve high accuracy and reliability in cardiovascular disease classification. Figure 1: Overall Architecture of Proposed MH-PSCGN-DLO Framework Input Acquisition: Secure Disease Classification starts by receiving input in cardiovascular disease dataset and complemented by real-time data, provided by an IoT-enabled medical device [25]. The devices probe the most crucial patient information such as heart rate, blood pressure, and cholesterol levels. The data gathered is first sent to be encrypted to provide protection whereby it is ready to be subjected to appropriate analysis and classification as discussed below. Multi‑Level Encryption (MLE) Encryption is a necessary point of Secure Disease Classification in IoT-Cloud healthcare systems [26]. It can avoid the unauthorized access to medical information and guarantee confidentiality and safety of the privacy of sensitive information in the course of transmission and storing. Multi-Level Encryption (MLE) provides enhanced security by applying the layered protection, incorporating three fundamental encryption techniques. Each layer enhances the confidentiality where authorized keys can retrieve the confidentiality. MLE is also effective to provide a high level of protection to Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6174 sensitive health information due to its hardening to attack and makes the process more difficult which allows to protect patient information at every level of storage and transmission. The encryption process of the Secure Disease Classification starts with the collection of medical data by IoT-enabled healthcare equipment and the cardiovascular disease dataset [28]. The data is coded into a format that cannot be read and encrypted three times in a row to create a secure message, which is then encrypted and stored or transmitted safely in the IoT-Cloud environment [29]. The algorithm 1 for Decryption is shown below. Algorithm 1: Encryption: E-ISSN: 2349 5359; P-ISSN: 2454-9967 every stage, preventing unauthorized access. The encrypted output is then input into the preprocessing stage as follows. Preprocessing using Adjusted Min‑Max with Decimal Scaling and Statistical Column Normalization (AMM-DS-SCN) Secure Disease Classification in IoT-cloud healthcare systems is one of the necessary steps in cleaning, scaling, and transforming data to enhance quality. The preprocessing method used is the Adjusted Min-Max with Decimal Scaling and Statistical Column Normalization (AMM-DSSCN) technique to bring the feature ranges to a consistent level and control the magnitudes of the attributes.  Statistical Column Normalization Input: Encrypted Medical Information H enc Output: Original Medical Information H Steps: 1. Begin 2. Receive the encrypted medical information H enc from IoT-enabled healthcare devices and the dataset. 3. Use the decryption key for Layer 3 to reverse the last encryption stage. 4. Apply Decryption for Layer 2 to remove intermediate protection. 5. Apply Decryption for Layer 1 to restore baseline readability. 6. Convert the decrypted data back into the original medical information H . 7. Verify data integrity to ensure no tampering occurred during storage or transmission. 8. End Output: Original Medical Information H The algorithm for the decryption process involves systematically processing encrypted medical information from IoT-enabled devices and datasets through three decryption layers. The first stage eliminates a level of protection, and the data is restored to its original format. Lastly, data integrity is ensured to prevent the recovery of medical information from being tampered with or compromised. Multi-Level Encryption (MLE) provides protection for sensitive medical information of IoT-enabled devices and datasets through layered security. The encryption step reinforces confidentiality at Statistical Column normalization individually scales each column by the difference between the column mean and the column mean, divided by the column mean, and then multiplies it by a small bias (0.1). This provides similar scales on features and minimizes the impact of huge variations. The given transformation can be described as in the following equation (1): Y   Y  b f m  / b f m  0.1 (1)  where, Y is the original column value, Y is the normalized value, b f m  is the column mean, and 0.1 is the small bias factor.  Decimal Scaling Normalization The preprocessing phase entails the application of Decimal Scaling Normalization to counter the differences in the level of attributes values. This will scale the number of places allocated to a row of attributes depending on the highest number of digits in a column. The scaling can give a proportional transformation without distorting the distribution of data. This normalization equation (2) is: Y   Y / 10 n (2)  where, Y is the original attribute value, Y is the standardized value, and n is the highest number of digits in the column used to scale the data.  Adjusted Decimal Scaling Normalization [30] Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6175 Decimal Scaling Normalization is used to obtain a finer form of the decimal scaling, by substituting the power n by f  1 , where f is a calculation of the highest value of the column. This leads to a more controlled scaling, especially when dealing with large attribute values, resulting in more balanced normalized data. The equations (3-4) representing this method are: Y   Y / 10  f 1 (3) f  log 10 max Ym  (4) E-ISSN: 2349 5359; P-ISSN: 2454-9967 enhance model efficiency, decrease complexity, and improve interpretation. The Parrot Optimization (PO) algorithm is used in this method to select the best features. By removing redundant or unnecessary characteristics, this met heuristic approach, which is based on the social as well as communication behavior of parrots, is efficient and effective in exploring the research space, improving classification accuracy while lowering computation costs. Initialization log base 10 of the maximum column value. The suggested Parrot Optimization (PO) initialization formulation, defined with a swarm size R , maximum iterations Maxiter , and search space boundaries represented by the lower bound y and upper bound vy , is expressed through the following equation (6):  Adjusted Min-max normalization Ym0  d  rand 0,1 . vy  y  where, Y is the initial value of the characteristic, Y  is the value that has been normalized, Ym represents each value in the column, and f is the Min-Max Normalization balances the data using the conventional min-max Normalization and modifies it with a new scale. It is a value distribution method that allocates values in the range [0, 1.5], thereby enhancing the comparability of features and retaining relative differences. The equation (5) of this method is: Y   Y  min Y  / rangeY   Y  min Y  /max Y   min Y  (5)  where, Y is the regularized value, Y is the   regulated value, min Y is the minimum column   value, max Y is the extreme column value, and rangeY  is the change between extreme and minimum values. In Secure Disease Classification in IoT-Cloud Healthcare Systems, the Adjusted Min-Max with Decimal Scaling and Statistical Column Normalization are used to scale and transform the data, ensuring that the ranges of features are even and that the magnitude of the attributes is regulated. The normalized output is then used as input to the feature selection stage as follows. Feature Selection using Parrot optimization (PO) [31]     (6) where, Ym0 is the initial position, y and vy are   the lower and upper bounds, and rand 0,1 is a random number between 0 and 1.  Fitness Function The fitness function measures the quality of a solution or a possible solution by examining the performance or objective of the solution. A fitness object is typically defined based on the problem that needs to be solved, as stated in Equation (7). Fitness Function K Ym   Objec E Ym  (7) where, K Ym  is the fitness of the m solution, th ObjecE Ym  is its objective function, and Ym is the candidate feature subset. The fitness function evaluates how well Ym meets the optimization goal, guiding the algorithm toward the best features.  Foraging behavior-Exploration [31] In PO foraging behavior, the rough position of food is determined either by watching the food itself or the position of the owner, after which they proceed to the rough position of the food. The positional movement is therefore represented by the equation (8): Feature selection is an operation that extracts the most pertinent attributes from a set of data to Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6176 E-ISSN: 2349 5359; P-ISSN: 2454-9967 2d  d  Maxiter d  Ymd 1  Ymd  YBest  Levytma  rand 0,1.1  .Ymean (8) Max iter     th where, Ymd 1 represents the updated position of the m solution at iteration d  1, Ymd is its current position, YBest is the best clarification found so far, and Ymean is the mean position of all solutions. The update incorporates a Levy flight Levytma in tma dimensions, a random factor rand 0,1 , the current iteration d , and the extreme number of repetitions Maxiter , guiding the search toward optimal solutions.  Staying behavior-Exploitation [31] Pyrrhura Molinae is a very social bird, and the behavior of staying consists of running across the body of the owner and staying in that part of the body for some time. The mathematical representation is the following equation (9): Ymd 1  Ymd  YBest  Levytma  rand 0,1  ones1, tma (9) th where, Ymd 1 is the updated position of the m solution at iteration d  1, Ymd is its current position, and YBest is the best solution found so far. The update uses a Levy flight Levytma in tma dimensions, a     random factor rand 0,1 , and a vector of ones ones 1, tma to guide the solution toward optimal positions.[31]  Termination The Parrot Optimization algorithm stops when it becomes stagnant because of either reaching an extreme number of iterations or achieving a satisfactory fit value. The algorithm in the termination stage then uses the most suitable set of features to classify the data, aiming to maximize the incorrect classifications and minimize the irrelevant and redundant features present in the best set of features. This will guarantee that there is effective calculation, reduced time of processes, and the selected features are effective in balancing exploration and exploitation, violating the model to improved performance and consistent results of the classification task. [31] Classification using Multi-Head ProbSparse Cascaded Graph Network (MH-PSCGN) The process of classifying medical data is through grouping of the medical data into predefined categories depending on characterized traits as well as patterns that have been established. Two approaches are combined in this framework: Disentangled Cascaded Graph Convolution Networks (DCGNN) [32] that learn factor-specific representations that define complex relationships between patients and their features and the Multi-Head ProbSparse Self-Attention Network (MHPSN) [33] that prioritizes the utilization of key clinical features with the help of probabilistic sparsity. It is a combination of these two approaches that forms the Multi-Head ProbSparse Cascaded Graph Network (MH-PSCGN) that can perform precise, effective, and robust predictions of various diseases, as further explained below. Disentangled Cascaded Graph Convolution Networks (DCGNN) The DCGNN is a deep learning network implemented on graphs, which captures the complicated interaction between patients and medical features by disentangling independent variables and stacking more than one graph convolutional layer, in sequence. This makes possible correct prediction of disease since it can be represented with factors and can also focus on factors. [33]  Embedding Initialization In the DCGNN model, each patient k  K and feature z  Z is represented with ID embeddings. These embeddings are initialized to Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6177 provide a starting point for learning in the cascading GCN blocks. Formally, the embeddings of a patient and a feature are given by equation (10): f  K  ID , f 0 va V a  G  ID  b 0 mb the variables include f (initial IDb (ID vector of feature $n$), and K G (learnable projection matrices).   mRm 1 Rv f m y ,  , Rm  y, p m y,P m  Prediction with Attention Lastly, the model assesses disease risk by summarizing patient-feature embeddings across determinants and actions, utilizing an attention mechanism. Equation (13) gives the attentionweighted prediction: P  D   D  bˆvm     mv y , p  f vb, p     f m y , p   p 1  y 1   y 1  (11)  y ,  fv (13) Rv 1 features m  Rv normalized by Rv and Rm multiplied by the feature embedding f m y ,  from the previous layer; similarly, f m y , 1 is the updated embedding of feature m at layer   1, calculated as the sum over neighboring patients normalized  T Rm 1 f v y  is the concatenated embedding vector of the patient v under behavior y , consisting of factor-specific embeddings f v y ,1 , , f v y , p  , , f v y , P  , similarly, f m y  is the concatenated embedding vector of the feature m under behavior y , composed of y ,1 where, f v y , 1 is the updated embedding of the patient v at layer   1under behavior y , computed as the sum over its neighboring v  Rm , y,P m where, m To capture nuanced patient-feature interactions across multiple stages of disease progression, DCGNN applies a cascading GCN. For patient k and feature z at layer   1 under behavior y , the embeddings are updated using equation (11): fm y, p m  f   , , f   , , f    .  Disentangled Cascading GCN Blocks  y , 1 y ,1 m (12) 0 va of feature mb ), IDaV (ID vector of patient $m$), mRv   f     f   ,  , f   ,  , f   , f v y   f v y ,1 ,  , f v y , p  ,  , f v y , P  , y embedding of patient va ), f m0b (initial embedding f v y , 1   into P independent blocks, which guarantee factor-specific representations. This is expressed as in equation (12): m (10) where, E-ISSN: 2349 5359; P-ISSN: 2454-9967 by 1 Rm and Rv multiplied by the patient embedding f v y ,  from layer  .  Disentangled Representation Learning The patient outcomes are affected differently by various factors (e.g., blood pressure, cholesterol, heart rate). DCGNN decomposes the embeddings where, b̂vm is the predicted outcome for patient v and feature m , P is the number of independent disentangled blocks, D is the number of behaviors, m v y , p  is the attention weight for patient v in behavior y and block p , f vb , p  and f m y , p  are the embeddings of patient v and feature m for behavior y and block p , respectively. Multi-Head ProbSparse Self-Attention Network (MHPSN) The Multi-Head ProbSparse Self-Attention Network (MHPSN) is a network that enhances the use of secure cardiovascular disease classification by considering the complex interaction between the patient and features. It relies on several attention heads to focus on the key clinical characteristics, utilizing a probabilistic sparsity mechanism to emphasize the most important data, thereby enhancing efficiency, robustness, and predictive quality. Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6178 MHPSN identifies the features that are most important and need to be highlighted during the patient classification process, as expressed in equation (14):  PH D  G AttentionP, H , G   soft max   t  p   (14) where, P , H , and G are the query, key, and value matrices, respectively, and t p is the key D dimension for scaling, P H is the dot product of the query matrix 𝑄 soft max normalizes attention scores to highlight important patientfeature interactions. MHPSN utilizes ProbSparse attention to enhance efficiency and resiliency in the presence of noisy or irrelevant patient data. Active queries are only computed based on relevant features of the cardiovascular system. Equation (15) is used to state the measure of sparsity:  hm p nD  1 H P hm p nD (15) R hm , P   max    n H t t  1 n P   where, R hm , P  represents the sparsity measure of the query hm , hm is the query vector, p nD is the key vector, t is the feature dimension, and H P is the total number of keys used for attention calculation. MHPSN is a model that employs a multiattention head to simultaneously model various interactions between patients and features. The distinctive patterns of each head across the clinical characteristics are captured, and the final feature representation is the concatenated output that is used in the classification. This can be expressed as shown in equations (16-17): Multihead P, H , G   Concat q1 , q 2 , , qb  (16)   q m  Attention P   mp , H   mH , G   mG (17)   where, Multihead P, H , G represents the output of the multi-head ProbSparse self- E-ISSN: 2349 5359; P-ISSN: 2454-9967 value matrices, and  mp ,  mH ,  mG are the corresponding weight matrices. MHPSN will be utilized to enhance classification in cardiovascular diseases by efficiently incorporating multi-head features, focusing on relevant clinical variables, and optimizing the accuracy, robustness, and efficiency of the classification process. The results of this classification are then fed into the DLO optimization process, which is then optimized and refined to add to the final finish and performance as follows: Disease Identification and Decision-Making Using DLO-Optimized MH-PSCGN In the presented system, cardiovascular disease identification is performed by the joint work of the Multi-Head ProbSparse Cascaded Graph Network (MH-PSCGN) and the Draco Lizard Optimizer (DLO). Where MH-PSCGN needs to learn discriminative representations from patient data, DLO is a metaheuristic optimizer that designs and optimizes the internal parameters of the classification network to achieve better predictive accuracy and stability. The MH-PSCGN architecture processes and filters patient feature vectors using Disentangled Cascaded Graph Convolution layers, which learn patient-feature interactions, and Multi-Head ProbSparse Self-Attention layers, which focus on clinically meaningful cardiovascular features. The ProbSparse system selectively preserves high-impact query-key interactions, avoiding unnecessary Computation and concentrating on major risk factors such as blood pressure, cholesterol levels, and heart rate. The Draco Lizard Optimizer (DLO) is applied after the original MH-PSCGN training stage. It optimizes the trainable parameters of the classification model, including graph convolutional weights, attention weight matrices, and fully connected output-layer parameters. DLO sequentially samples parameter sets using a fitness function that estimates classification loss and accuracy, and updates the parameters using its exploration-exploitation algorithm until convergence. DLO is therefore additionally involved in parameter tuning and classification refinement, but not feature selection, which is a separate concern of the Parrot Optimization algorithm. th attention mechanism, q1 is the output of the m attention head, P , H , G are the query, key, and To make classification decisions, the optimized MH-PSCGN would generate a two-dimensional Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6179 probability score vector associated with Cardiovascular Disease Present (CDP) and No Cardiovascular Disease (NCD). These scores are normalized to probabilities using a softmax activation function at the output layer. A maximum probability criterion is used to arrive at the final disease label, where a sample is considered a CDP when the predicted probability of a CDP class is greater than the probability of an NCD class, and a sample is regarded as an NCD when the likelihood of the NCD class is greater. E-ISSN: 2349 5359; P-ISSN: 2454-9967 6. Grant the end cardiovascular disease label on the maximum score of likelihood. It is a comprehensive combination of ProbSparse attention and Draco Lizard Optimization to ensure accurate, reliable, and computationally efficient cardiovascular disease classification in the proposed IoT–cloud healthcare ecosystem. Draco Lizard Optimizer (DLO) The Draco lizards have inspired a natureinspired metaheuristic algorithm known as the Draco Lizard Optimizer (DLO), which mimics their hunting and gliding behaviors. It is also extremely efficient in finding equilibrium between exploration and exploitation to solve nontrivial optimization problems, being dynamically adaptive to search spaces comprising variables. Having the capacity to learn heuristic and cooperative behavior, and simulate cooperative strategies and strategy adaptation, DLO yields better quality solutions and is also robust to engineering, computational, and real-world optimization problems. The Flow chart of the Draco Lizard Optimizer is shown in Figure 2 [34]. The process of cardiovascular disease identification is step-by-step, as shown below: 1. Inserted preprocessed optimized feature vectors into the MH-PSCGN model. 2. Disentangled Cascaded Graph Convolution layers were used to learn patient-feature representations. 3. Multi-Head ProbSparse Self-Attention is applied to emphasize important cardiovascular features. 4. Use the Draco Lizard Optimizer to maximize the network parameters by minimizing classification loss. 5. Produce scores of class probability using the softmax output layer. Initialization Random Process Initialization of DLO parameters for optimizing the weight parameters of MH-PSCGN hh1 Fitness Function Yes The Draco Lizard Optimizer (DLO) updates and optimizes the weight parameters of MHPSCGN Halting criteria No The Draco Lizard Optimizer s greedy selection strategy updates the best positions to ensure convergence toward the optimal solution. An improvement in accuracy is achieved, accompanied by a reduction in both error rate and computational complexity Termination Figure 2: Flow chart for Draco Lizard Optimizer Step 1: Initialization: The Draco Lizard Optimizer (DLO) starts with the initiation stage, which involves the population size, iteration constraints, and weight factors in the optimization of MH-PSCGN. This move defines the starting search space, ensuring a balanced starting point for the next search and optimization processes. Step 2: Random Generation: Candidate solutions at this stage are seeded randomly in the search space that has been Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6180 specified to model a range of starting positions. The solutions pertain to a collection of possible weight parameters of MH-PSCGN. This randomness guarantees variation, avoidance of early convergence, and makes the DLO efficient in searching at a wide range of locations in the search space that can have global optima. terms of classification. It determines accuracy, errors, and performance. Better-fit solutions are used to continue the iterations, which guide the DLO to weight parameter settings that maximize the performance of MH-PSCGN, as shown in equation (18). Fitness Function  Min  (18) Step 3: Fitness Function: The all the candidate solutions are compared on the basis of fitness value which considers the performance of the secure disease predictor in E-ISSN: 2349 5359; P-ISSN: 2454-9967 where,  denotes the weight matrix in the classification process. Step 4: The Draco Lizard Optimizer’s greedy selection strategy ensures the attainment of the optimal result: After iteration, the DLO algorithm applies a greedy selection strategy to update both the individual best positions and the global best position. The corresponding mathematical formulations for these updates are given in equations (19) and (20).    m d  1, iff ma d  1  f mbest a ) (19) mbest a    a  mbest otherwise , a   m d  1, iff ma d  1  f qbest ) qbest   a qbest a , otherwise  (20)   where, mbesta denotes the individual best position of particle a , ma d  1 is the updated position of  particle a at iteration d  1, and f . is the fitness function used to evaluate solution quality; the update rule states that mbest is replaced by ma d  1 if it yields a better fitness value, otherwise it remains unchanged. Similarly, qbest represents the global best position among all particles, updated ma d  1 if the new position achieves a better fitness value than the current global best, otherwise it is retained. Step 5: Termination: Output: The termination stage takes place when the termination condition is met which is usually when a maximum number of iterations are got or when convergence is attained. In this step, the DLO calculates the best weight parameters of MH-PSCGN, which gives the best accuracy of classification. The process outcome is greater strength, low error rate and low cost of Computation, which provides a stable and capture cardiovascular disease classification. Classification of predicted disease (cardiovascular disease/No Cardiovascular Disease) and performance measures. Begin 1. 2. 3. Algorithm 2: MH-PSCGN-DLO-Based Secure Cardiovascular Disease Classification Input: IoT sensor data, cardiovascular disease dataset Data Acquisition: Gather patient physiological information from IoT-enabled medical devices and organize cardiovascular disease data. Multi-Level Encryption (MLE): To maintain the confidentiality of the collected medical data, encrypt it using a layered encryption mechanism before transmission to the cloud. Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6181 E-ISSN: 2349 5359; P-ISSN: 2454-9967 4. Cloud Reception and Decryption: Get the encrypted data to the cloud computing environment and perform multiple levels of Decryption to recover the original medical data. 5. Data Preprocessing (AMM-DS-SCN): Normalize the distributions of features using Statistical Column normalization. Apply Decimal Scaling normalization to normalize attributes. Use Adjusted Min-Max Normalization to get a normalized dataset. 6. Feature Selection using Parrot Optimization (PO): Starting population of candidate feature subsets. Factors assess fitness through the classification performance. Find the solutions through exploration and exploitation. Choose the best set of features to make the classification. 7. Classification using MH-PSCGN: Initialize patient and feature embeddings. Apply Disentangled Cascaded Graph Convolution layers to learn patient–feature relationships. Apply Multi-Head ProbSparse Self-Attention to emphasize critical clinical features. Generate preliminary disease predictions. Parameter Operating System Programming Language Cloud Platform Encryption Algorithm Datasets Used Data Type Train–Test Split Classification Model Optimization Algorithm Performance Metrics 8. Model Optimization using Draco Lizard Optimizer (DLO): Optimize MH-PSCGN parameters using exploration–exploitation balancing. Update global and local best solutions iteratively until convergence. 9. Output Generation: Produce the final cardiovascular disease classification results and compute evaluation metrics. End 2.RESULTS AND DISCUSSION The implementation of the proposed Secure Disease Classification framework in the IoTCloud Healthcare Systems for the concept of cardiovascular disease data classification is given in Table 2. The model is implemented on the Windows 10 platform using Python 3.7.14, where the input is based on IoT-enabled medical sensors and the cardiovascular disease dataset. The Draco Lizard Optimizer (DLO) is utilized to optimize the proposed Multi-Head ProbSparse Cascaded Graph Network (MH-PSCGN). Accuracy, Precision, Recall, Specificity, and F1score are employed to evaluate the network's performance. The simulation parameters used in this investigation are compiled in Table 2. Table 2: Simulation Parameter Value Windows 10 Python 3.7.14 AWS IoT (simulated environment) Multi‐Level Encryption (MLE) Cardiovascular Disease dataset IoT Sensor Data, Text Data 80% training, 20% testing Multi-Head ProbSparse Cascaded Graph Network (MH-PSCGN) Draco Lizard Optimizer (DLO) Accuracy, Precision, Recall, F1-Score, Specificity Dataset Description The proposed framework is evaluated using a publicly available cardiovascular disease dataset obtained from the Kaggle repository (Cardiovascular Disease Dataset: https://www.kaggle.com/datasets/sulianova/ca rdiovascular-disease-dataset). This dataset is an offline hospital dataset that is based on regular check-ups. It was not obtained in real time by wearable sensors. Still, it is applied in the present research as the simulated IoTproduced healthcare information, i.e., the physiological measurements that are usually acquired by IoT-enabled healthcare and wearable tracking sensors in realistic smart healthcare settings. Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6182 The dataset contains 56,372 patient records, each with several clinical and physiological features. The primary factors are demographic (age), physiological (systolic and diastolic blood pressure, cholesterol levels, body mass index, and heart rate), and lifestyle-related. All these heterogeneous attributes help predict cardiovascular disease reliably and are very similar to the data streams produced in IoTbased healthcare. The dataset is classified into two groups: Cardiovascular Disease Present (CDP), which includes patients diagnosed with heart-related conditions, and No Cardiovascular Disease (NCD), which consists of healthy people. The sample size in each class is 28,186, ensuring balanced class distribution and reducing bias during model training and evaluation. E-ISSN: 2349 5359; P-ISSN: 2454-9967 The raw data are then subjected to a preprocessing pipeline based on Statistical Column Normalization (SCN), Decimal Scaling Normalization (DSN), and Adjusted Min-Max Normalization (AMM) to address scale variation, noise, and feature heterogeneity. After preprocessing, the data is split into an 80:20 train:test split, with 80% (45098 records) assigned to training and optimization, and 20% (11274 records) to independent testing. This division is also confirmed with the k-fold crossvalidation to make the proposed MH-PSCGN-DLO model robust and generalized. The distribution of the cardiovascular disease dataset in Table 3 indicates the sample size by class, separating it into training (80%) and testing (20%) subsets, with equal representation of the No Cardiovascular Disease (NCD) classification and the Cardiovascular Disease Present (CDP) classification. Table 3: Class-wise Sample Distribution for Training and Testing in the Cardiovascular Disease Dataset Cardiovascular Disease dataset Class NCD CDP Total Total Samples Training (80%) Testing (20%) 28,186 28,186 56,372 22,549 22,549 45,098 5,637 5,637 11,274 Evaluation metrics: The evaluation criteria, which include recall, accuracy, specificity, precision, F1-score, Error rate, Encryption time, Decryption time, and Security level, will also be considered in the recommended method [35]. The proposed MHPSCGN-DLO compares its performance to earlier methods, such as C-IoMT, Bi-LSTM, GF-ALO, DNN, CNN, GWALO, HybBPF-ELM, EC-FIoT, for the cardiovascular disease dataset, and Hyb-AESRSA, PKE, 3-DES, PAEKS, AES for the existing Encryption methods. Comparison of Cardiovascular Disease dataset: The performance analysis of a cardiovascular disease (CVD) prediction model, based on a confusion matrix and class correlation analysis, is presented in Figure 3. Subfigure (a) shows perfect accuracy of classification with equal separation of No CVD and CVD Present cases, whereas subfigure (b) indicates the existence of strong positive and negative correlation between respective classes. The model's validation and training performance across 100 epochs is shown in Figure 4. The plot of accuracy shows a gradual increase in accuracy during training, reaching almost 1.0, and in validation, at about 0.85. The loss plot shows that the training loss has decreased considerably, and the validation loss has remained at a low level. Figure 5 illustrates the distribution of samples for cardiovascular disease (CVD) and non-CVD conditions in various representations. Subfigure (a) shows the sample class balance in blocks, (b) class frequency depending on indicators, and (c) almost equal numbers of classes guaranteeing training and evaluation without bias. Figure 6 illustrates the relative performance of various encryption schemes, basing them on encryption time, decryption time, privacy, and error rate. This is shown by Subfigures (a) and (b) that the proposed method has low encryption and decryption times, and under Subfigure (c), it can sustain high privacy with low error rates. Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6183 E-ISSN: 2349 5359; P-ISSN: 2454-9967 (b) (a) Figure 3: Confusion Matrix and Class Correlation for CVD Prediction Figure 4: Training and Validation Accuracy and Loss Curves (a) (b) (c) Figure 5: Distribution of CVD and Non-CVD Samples Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6184 (a) E-ISSN: 2349 5359; P-ISSN: 2454-9967 (b) (c) Figure 6: Performance Comparison of Encryption Schemes Figure 7: Accuracy Comparison Across Cross-Folds Figure 7 illustrates the comparison of accuracy among different models across various crossvalidation folds. It is clear that the proposed approach consistently demonstrates higher accuracy compared to all other methods, and its accuracy values are the highest in all folds, whereas traditional models, such as CNN, DNN, and Bi-LSTM, show relatively low accuracy values. The Figure 8 illustrates ROC curve analysis of various models in terms of their true positive and false positive rates, along with the corresponding AUC values. Most of the models have high AUC scores of 0.97 to 0.98; however, the proposed method achieves an ideal AUC of 1.00. Figure 9 presents the outcomes of the classification of cardiovascular disease (CVD) Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6185 detection, the prediction results, and the actual outcomes. The dominance of triangles implies the proper classification of the CVD present and absent cases, whereas the out of triangle points E-ISSN: 2349 5359; P-ISSN: 2454-9967 represent the misclassifications. These points to the effectiveness of the model in separating the cases of CVD and non-CVD. Figure 8: ROC Curve Comparison of Models Figure 9: Classification Results for CVD Detection Figure 10 shows the distribution of indicators of the presence and absence of cardiovascular disease (CVD). The cut points of the two groups (No CVD and CVD Present), may be characterized as rather distinct, there is no significant overlap between the two categories, showing the effectiveness of the model in discriminating the two states with a minor overlap. The classification analysis of cardiovascular disease (CVD) with different visualization has been demonstrated in Figure 11. Figure (a) indicates the sample-wise distribution of classes, Figure (b) empirical cumulative distribution function (ECDF) of CVD and non-CVDs and Figure (c) actual and predicted values confirm that the model is very sound. Figure 10: Density Distribution of CVD Classification Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6186 E-ISSN: 2349 5359; P-ISSN: 2454-9967 (a) (b) (c) Figure 11: Classification Analysis of CVD Detection Figure 12: Visualization of No CVD Indicator Across Samples The visual representation of the No CVD indicators based on a color gradient scale with a sample index is presented in Figure 12. The fringe areas depict the concentration of spots, whereas the focus area is thin, which points to the definite segmentation of CVD and non-CVD cases. This means that it is well classified. Table 4 provides the comparison of the performance of the proposed MH-PSCGN-DLO model of categorizing secure healthcare data with the current approaches. In order to prove that the proposed solution is more superior to the traditional models in all the board considerations, it concentrates on the most important metrics like F1-score, sensitivity, specificity, recall, accuracy and precision. The error and computation time of the currently available methods is given in Table 5 relative to the proposed MH-PSCGN-DLO model in classifying secure healthcare data. It stresses that the proposed solution is much more productive (error rate: 0.01), much rapid (0.2 seconds) and more precise, which makes it applicable to the real-life healthcare case. Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6187 E-ISSN: 2349 5359; P-ISSN: 2454-9967 Table 4: Performance Comparison of the cardiovascular disease dataset Method / Metrics C-IoMT [16] Bi-LSTM [17] GF-ALO [18] DNN [19] CNN [20] GWALO [21] HybBPF-ELM [22] EC-FIoT [23] MH-PSCGN-DLO (Proposed) Accuracy 85.5 82.3 82.7 83.1 83.4 83.6 83.9 84.1 98.4 Precision 84.2 82.0 82.5 82.9 83.2 83.5 83.7 83.9 98.2 Recall 83.7 82.4 82.8 83.2 83.5 83.7 84.0 84.2 98.6 Sensitivity 81.9 82.6 83.0 83.4 83.7 83.9 84.2 84.4 98.8 Specificity 82.0 82.8 83.2 83.5 83.8 84.0 84.3 84.5 98.0 F1-score 81.5 82.3 82.7 83.1 83.4 83.7 84.0 84.2 98.7 Table 5: Error rate/computational time Comparison Method / Metrics C-IoMT [16] Bi-LSTM [17] GF-ALO [18] DNN [19] CNN [20] GWALO [21] HybBPF-ELM [22] EC-FIoT [23] MH-PSCGN-DLO (Proposed) Error Rate (%) 0.50 0.48 0.46 0.45 0.44 0.43 0.42 0.41 0.01 CT (sec) 0.95 0.93 0.91 0.90 0.89 0.88 0.87 0.86 0.2 Table 6: Comparison of Encryption and Decryption Time for Different Methods Encryption Method Hyb-AES-RSA [24] PKE [25] 3-DES [26] PAEKS [27] AES [28] MH-PSCGN-DLO (Proposed) Encryption Time (ms) 15.3 17.1 20.8 11.6 23.4 5.1 Decryption Time (ms) 12.7 14.5 18.2 9.9 20.8 4.3 Table 6 represents the encryption and decryption time of the currently used methods and the proposed method, MH-PSCGN-DLO. The best results of the proposed method are 5.1 ms and 4.3 ms in the encryption and decryption stages, respectively, which are significantly superior to traditional methods. This indicates that the approach is highly efficient and can be applied in practice to the IoT-Cloud healthcare context for real-time secure data processing [36]. The statistical significance of the existing methods compared to the proposed MH-PSCGNDLO model is presented using the Kruskal-Wallis and Wilcoxon Signed-Rank tests in Table 7. The lowest p-values are obtained with the proposed method (0.020 and 0.018), indicating that its performance gains are statistically significant and can be reliably better compared to traditional methods for secure healthcare data classification. Statistical Analysis of the Proposed Method/Existing Methods Table 8 presents the performance measures of the proposed model, as evaluated by eight-fold cross-validation, in the cardiovascular disease dataset. There are metrics such as IJ, MCC, k, and CSI, which exhibit high values across all folds. This means that the model exhibits consistent, high, and robust classification performance, making it suitable for use in real-life healthcare applications. A statistical analysis of the suggested approach, which serves as the foundation for comparing it with the current ones, is shown in Table 7. Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6188 E-ISSN: 2349 5359; P-ISSN: 2454-9967 Table 7: Statistical Validation Method / Test [References] C-IoMT [16] Bi-LSTM [17] GF-ALO [18] DNN [19] CNN [20] GWALO [21] HybBPF-ELM [22] EC-FIoT [23] MH-PSCGN-DLO (Proposed) Kruskal–Wallis Test (p-value) 0.058 0.056 0.054 0.052 0.050 0.048 0.046 0.044 0.020 Wilcoxon Signed-Rank Test (p-value) 0.054 0.052 0.050 0.048 0.046 0.044 0.042 0.040 0.018 Table 8: Performance Metrics across Eight-Fold Cross-Validation for the Cardiovascular Disease Dataset Scenario Cardiovascular (k1) Cardiovascular (k2) Cardiovascular (k3) Cardiovascular (k4) Cardiovascular (k5) Cardiovascular (k6) Cardiovascular (k7) Cardiovascular (k8) IJ 0.9758 0.9770 0.9745 0.9775 0.9757 0.9655 0.9670 0.9648 3. DISCUSSION The developed Secure Disease Classification model demonstrates significant advancements over conventional IoT-cloud medical systems, particularly in cardiovascular disease prediction and data protection. Multi-Level Encryption (MLE) is integrated to ensure high data confidentiality, thereby minimizing the risks of unauthorized access during data transmission and storage. Further, adjusted normalization methods in preprocessing are useful because maximizing the consistency of the data is attained, but Parrot Optimization is useful in identifying significant features and minimizing redundancy and cost of Computation. MH-PSCGN model is grounded on the idea of integrating Disentangled Cascaded Graph Convolution Network (DCGNN) and Multi-Head ProbSparse Self-Attention (MHPSA) that allows representing patient data correctly and factor-specifically, thus contributing to a significant improvement in the quality of predictions. The Draco Lizard Optimizer (DLO) has been defined as being more accurate in classification, 98.4% according to existing models like Bi-LSTM, CNN, and GWALO with offerings ranging between 82% and 84%. Moreover, the system boasts of low error (0.01) and quicker processing speed (0.2 sec) which MCC 0.9885 0.9892 0.9878 0.9897 0.9886 0.9792 0.9805 0.9780 κ 0.9830 0.9835 0.9820 0.9842 0.9829 0.9738 0.9745 0.9728 CSI 0.9752 0.9760 0.9748 0.9768 0.9753 0.9648 0.9662 0.9643 makes it so convenient in real-time healthcare. The encryption and decryption times are also superior to those of conventional algorithms, which attest to the framework's efficiency in scaling to a large scale. Overall, the results demonstrate that the proposed model is clinically significant, reliable, and has the potential to revolutionize disease prediction in the IoT-based cloud computing environment. Interpretability, XAI, and Clinical Relevance The interpretability, Explainable Artificial Intelligence (XAI), and clinical relevance play a major role in rolling out the advanced healthcare models. To make the inner workings of predictive models interpretable, such that clinicians can trace the impact of input features, e.g., patient vitals/lab result, on the final outcome, interpretability is required. This has been extended with XAI, which offers a transparency and accountability by using characteristics like visualization of attention, feature importance ranking and the explanation of the decision path. Such transparency makes trust grow, yet it also aids the physicians to advocate the reasons behind the model using the available medical expertise. Clinical relevance in its turn is concerned with the way the outputs of the model Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6189 can be transformed into actionable information that is to be applied to real-life scenarios, e.g., forecasting the development of a disease, determining risk factors, or prescribing timely intervention. Even the best-fit models may be disputed within the medical practice because of the unclear interpretation and applicability. Hence, the inclusion of XAI practices in healthcare systems ensures that models are comprehensible, stable, and morally oriented, thereby eliminating the gap between artificial intelligence systems and clinical decision support in patient care. The authors of this manuscript have no conflict of interest to declare. REFERENCES [1] [2] CONCLUSION The proposed Secure Disease Classification framework integrating IoT and cloud computing has been successfully implemented and demonstrated highly effective performance in detecting cardiovascular diseases. Multi-Level Encryption, state-of-the-art preprocessing, Parrot Optimization, and MH-PSCGN classifier, optimized with Draco Lizard Optimizer, provide the system with the ability to provide not only a safe data transfer but also a perfect classifier. The results of the experiments ascertained high accuracy, precision, recall, and efficiency as opposed to the currently used methods, with exceptionally low errors and computation time. The framework has therefore attained effective detection of diseases in real time in a healthcare setting and preserved information confidentiality and privacy. Further research will involve applying this model to other chronic conditions, combining multimodal medical data (e.g., ECG signals or images) with it, and improving their interpretability based on explainable AI to be able to make informed clinical decisions. Moreover, large-scale deployment on real-world IoT-enabled hospital infrastructures will be explored to validate scalability, robustness, and practical applicability in diverse healthcare scenarios. Funding No specific grants or funding organizations from the public, commercial, or non-profit sectors provided financial support for this study. Declaration The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. E-ISSN: 2349 5359; P-ISSN: 2454-9967 [3] [4] [5] [6] [7] Deepika, J., Rajan, C., Senthil, T., 2021. Security and privacy of cloud‐ and IoT‐ based medical image diagnosis using fuzzy convolutional neural network. Computational Intelligence and Neuroscience, 2021, Article ID 6615411. https://doi.org/10.1155/2021/661541 1 Patel, S. K., 2023. Improving intrusion detection in cloud-based healthcare using neural network. Biomedical Signal Processing and Control, 83, 104680. https://doi.org/10.1016/j.bspc.2023.10 4680 Ali, A., Almaiah, M. A., Hajjej, F., Pasha, M. F., Fang, O. H., Khan, R., Teo, J. Zakarya, M., 2022. An industrial IoT-based blockchain-enabled secure searchable encryption approach for healthcare systems using neural network. Sensors, 22(2)572. https://doi.org/10.3390/s22020572 Kamalam, G. K., Anitha, S., 2022. CloudIoT secured prediction system for processing and analysis of healthcare data using machine learning techniques. In: Advanced Healthcare Systems: Empowering Physicians with IoTEnabled Technologies, pp.137-172. https://doi.org/10.1002/97811197692 93.ch10 Ho, E. S., 2021. Data security challenges in deep neural network for healthcare IoT systems. In: Security and Privacy Preserving for IoT and 5G Networks: Techniques, Challenges, and New Directions, pp. 19–37. Cham: Springer International Publishing.https://doi.org/10.1007/978 -3-030-76002-6_2 Refaee, E. A., Shamsudheen, S., 2022. A computing system that integrates deep learning and the internet of things for effective disease diagnosis in smart health care systems. The Journal of Supercomputing, 78(7) 9285–9306. https://doi.org/10.1007/s11227-02104153-7 Bhatia, M. & Kumari, S., 2022. A novel IoT-fog-cloud-based healthcare system for monitoring and preventing Conflict of interest Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6190 [8] [9] [10] [11] [12] [13] encephalitis. Cognitive Computation, 14(5),16091626. https://doi.org/10.1007/s12559-02109944-2 Mansour, R. F., El Amraoui, A., Nouaouri, I., Díaz, V. G., Gupta, D. Kumar, S., 2021. Artificial intelligence and internet of things enabled disease diagnosis model for smart healthcare systems. IEEE Access, 9, 45137–45146. https://doi.org/10.1109/ACCESS.2021. 3066297 Kathamuthu, N. D., Chinnamuthu, A., Iruthayanathan, N., Ramachandran, M. Gandomi, A. H., 2022. Deep Q-learningbased neural network with privacy preservation method for secure data transmission in IoT healthcare application. Electronics, 11(1) 157. https://doi.org/10.3390/electronics110 10157 Mohiyuddin, A., Javed, A. R., Chakraborty, C., Rizwan, M., Shabbir, M., Nebhen, J., 2022. RETRACTED ARTICLE: Secure cloud storage for medical IoT data using adaptive neuro-fuzzy inference system. International Journal of Fuzzy Systems, 24(2) 1203–1215. https://doi.org/10.1007/s40815-02101178-7 Thilagam, K., Beno, A., Lakshmi, M. V., Wilfred, C. B., George, S. M., Karthikeyan, M., Peroumal, V., Ramesh, C. Karunakaran, P., 2022. Secure IoT healthcare architecture with deep learning-based access control system. Journal of Nanomaterials, 2022, Article ID 2638613. https://doi.org/10.1155/2022/263861 3 Padinjappurathu Gopalan, S., Chowdhary, C. L., Iwendi, C., Farid, M. A., Ramasamy, L. K., 2022. An efficient and privacy-preserving scheme for disease prediction in modern healthcare systems. Sensors, 22(15) 5574. https://doi.org/10.3390/s22155574 Ramanan, M., Singh, L., Kumar, A. S., Suresh, A., Sampathkumar, A., Jain, V., Bacanin, N., 2022. Secure blockchain enabled cyber-physical health systems using ensemble convolution neural network classification. Computers and Electrical Engineering, 101, p. 108058. https://doi.org/10.1016/j.compeleceng. 2022.108058 [14] [15] [16] [17] [18] [19] [20] E-ISSN: 2349 5359; P-ISSN: 2454-9967 Nasser, A. R., Hasan, A. M., Humaidi, A. J., Alkhayyat, A., Alzubaidi, L., Fadhel, M. A., Santamaría, J., Duan, Y., 2021. IoT and cloud computing in healthcare: A new wearable device and cloud-based deep learning algorithm for monitoring of diabetes. Electronics, 10(21) 2719. https://doi.org/10.3390/electronics102 12719 Kumar, A., Satyanarayana Reddy, S. S., Mahommad, G. B., Khan, B., Sharma, R., 2022. Smart healthcare: Disease prediction using the cuckoo-enabled deep classifier in IoT framework. Scientific Programming, 2022, Article ID 2090681. https://doi.org/10.1155/2022/209068 1 Adewole, K. S., Akintola, A. G., Jimoh, R. G., Mabayoje, M. A., Jimoh, M. K., UsmanHamza, F. E., Balogun, A. O., Sangaiah, A. K., Ameen, A. O., 2021. Cloud-based IoMT framework for cardiovascular disease prediction and diagnosis in personalized e-health care. In: Intelligent IoT Systems in Personalized Health Care, pp. 105– 145. Academic Press. https://doi.org/10.1016/B978-0-12821641-2.00006-2 Nancy, A. A., Ravindran, D., Raj Vincent, P. D., Srinivasan, K. & Gutierrez Reina, D., 2022. IoT-cloud-based smart healthcare monitoring system for heart disease prediction via deep learning. Electronics, 11(15),2292. https://doi.org/10.3390/electronics111 52292 Verma, A., Agarwal, G., Gupta, A. K., Kumar, V., Singh, S., 2025. An adaptive secure internet of things and cloud-based disease classification strategy for smart healthcare industry. Wireless Networks, 31(1), 879–897. https://doi.org/10.1007/s11276-02303405-6 Pati, A., Parhi, M., Alnabhan, M., Pattanayak, B. K., Habboush, A. K., Al Nawayseh, M. K., 2023. An IoT-fog-cloud integrated framework for real-time remote cardiovascular disease diagnosis. Informatics, 10(1)21. https://doi.org/10.3390/informatics10 010021 Raheja, N. & Manocha, A. K., 2023. An IoTenabled secured clinical healthcare framework for diagnosis of heart Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6191 [21] [22] [23] [24] [25] [26] [27] diseases. Biomedical Signal Processing and Control, 80, 104368. https://doi.org/10.1016/j.bspc.2022.10 4368 Verma, A., Agarwal, G., Gupta, A. K., Sain, M., 2021. Novel hybrid intelligent secure cloud internet of things-based disease prediction and diagnosis. Electronics, 10(23) 3013. https://doi.org/10.3390/electronics102 33013 Janarthanan, V., Annamalai, T., Arumugam, M., 2024. Enhancing healthcare in the digital era: A secure ehealth system for heart disease prediction and cloud security. Expert Systems with Applications, 255, 124479. https://doi.org/10.1016/j.eswa.2024.12 4479 Singh, P. D., Dhiman, G., Sharma, R., 2022. Internet of things for sustaining a smart and secure healthcare system. Sustainable Computing: Informatics and Systems, 33, 100622. https://doi.org/10.1016/j.suscom.2021. 100622 Denis, R., Madhubala, P., 2021. Hybrid data encryption model integrating multiobjective adaptive genetic algorithm for secure medical data communication over cloud-based healthcare systems. Multimedia Tools and Applications, 80(14)21165–21202. https://doi.org/10.1007/s11042-02010457-6 Zhao, Z. Z., Guo, F., Wu, G., Susilo, W., Wang, B., 2022. Secure infectious diseases detection system with IoTbased e-health platforms. IEEE Internet of Things Journal, 9(22)22595–22607. https://doi.org/10.1109/JIOT.2022.316 3483 Raheja, N., Manocha, A. K., 2022. IoTbased ECG monitoring system with encryption and authentication in secure data transmission for clinical healthcare approach. Biomedical Signal Processing and Control, 74,103481. https://doi.org/10.1016/j.bspc.2022.10 3481 Qamar, S., 2022. Healthcare data analysis by feature extraction and classification using deep learning with cloud-based cybersecurity. Computers and Electrical Engineering, 104,108406. [28] [29] [30] [31] [32] [33] [34] E-ISSN: 2349 5359; P-ISSN: 2454-9967 https://doi.org/10.1016/j.compeleceng. 2022.108406 Das, S., Namasudra, S., 2022. A novel hybrid encryption method to secure healthcare data in IoT-enabled healthcare infrastructure. Computers and Electrical Engineering, 101, 107991. https://doi.org/10.1016/j.compeleceng. 2022.107991 Al-Kateeb, Z. N., Jader, M., 2025. Multilevel of encryption and steganography depending on Rabinovich hyperchaotic system and DNA. Multimedia Tools and Applications, 84(3) 1211–1237. https://doi.org/10.1007/s11042-02418567-0 Sinsomboonthong, S., 2022. Performance comparison of new adjusted min-max with decimal scaling and statistical column normalization methods for artificial neural network classification. International Journal of Mathematics and Mathematical Sciences, 2022, Article ID 3584406. https://doi.org/10.1155/2022/358440 6 Lian, J., Hui, G., Ma, L., Zhu, T., Wu, X., Heidari, A. A., Chen, Y., Chen, H., 2024. Parrot optimizer: Algorithm and applications to medical problems. Computers in Biology and Medicine, 172, 108064. https://doi.org/10.1016/j.compbiomed. 2024.108064 Gunavathie, M. A., Arivarasi, A., Aggarwal, P. K., Arunachalam, K. P., 2025. Blockchain-enabled 6G wireless network securities using multi-relational graph attention disentangled cascaded graph convolution networks. Computers and Electrical Engineering, 127, 110617. https://doi.org/10.1016/j.compeleceng. 2024.110617 Liu, W., Gu, Y., Ge, Y., 2024. Multi-factor stock trading strategy based on DQN with multi-BiGRU and multi-head ProbSparse self-attention. Applied Intelligence, 54(7) 5417–5440. https://doi.org/10.1007/s10489-02305184-1 48. Wang, X., 2025. Draco lizard optimizer: A novel metaheuristic algorithm for global optimization Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com Int. J. Adv. Sci. Eng. Vol.12 No.3 6170-6192 (2026) 6192 [35] [36] problems. Evolutionary Intelligence, 18(1)10. https://doi.org/10.1007/s12065-02400910-5 https://www.kaggle.com/datasets/sulia nova/cardiovascular-disease-dataset Raja, L., Periasamy, P. S., 2022. A trusted distributed routing scheme for wireless sensor networks using blockchain and jellyfish search optimizer based deep generative adversarial neural network (Deep-GANN) technique. Wireless [37] E-ISSN: 2349 5359; P-ISSN: 2454-9967 Personal Communications, 126, 1101– 1128. https://doi.org/10.1007/s11277-02209634-6 Mathivanan, M., Balamurugan, M., Harish, L., Nandini., Reddy, M., 2018. IoT-based continuous monitoring of cardiac patients using Raspberry Pi. AIP Conference Proceedings, 2039(1) 020025. https://doi.org/10.1063/1.5078987 All © 2026 are reserved by International Journal of Advanced Science and Engineering. This Journal is licensed under a Creative Commons Attribution-Non-Commercial-ShareAlike 3.0 Unported License. Reka et al., International Journal of Advanced Science and Engineering www.mahendrapublications.com