Journal of Computing & Biomedical Informatics

Comparative Analysis of Machine Learning Classifiers for Detecting Backdoor Malware in IoT Networks

Adeeb Alsaaidah — Wed, 20 May 2026 00:00:00 +0000

The growth of Internet of Things (IoT) networks has led to the fact that it can now be more exposed to cyber-attacks, including those based on backdoor malware that makes it possible to gain unauthorized and persistent access to the compromised devices. Proper and effective identification of such attacks is hence critical in ensuring the safety of the IoT environments. This paper provides a comparative overview of classical machine-learning (ML) models of binary backdoor malware detection in IoT networks, in which the network traffic samples are either labeled Backdoor malware or Benign. The experimental workflow starts with an unbalanced backdoor of CICIoT2023 dataset, which is then preprocessed and balanced to create the final benchmark dataset to be used in model evaluation. This dataset is composed of 39,160 samples and 47 features, and the distribution of classes are the equal (19,580 benign and 19,580 backdoor malware samples). The ML pipeline was applied with a single evaluation pipeline based on data cleaning, label encoding, and feature standardization, and then 10-fold cross-validation. Twenty classical ML classifiers were benchmarked, including ensemble, linear, probabilistic, kernel-based, instance-based, and neural-network-based ML methods. Accuracy, Precision, Recall, F1-score and execution time were used to measure performance indicates . The findings indicate that tree-based ensemble models always perform better than other models with ExtraTrees having the highest overall performance (Accuracy = 0.999872, Precision = 0.999872, Recall = 0.999872, F1-score = 0.999872), then followed by RandomForest (0.999719) Among the best performing models, XGBoost also produced very competitive results (Accuracy = 0.999489) with desirable computational efficiency. These results emphasize the use of ensemble-based ML methods as highly efficient and feasible baselines in detecting backdoor malware in IoT and the evolution of effective, deployment-friendly IoT security tools.

Intelligent Healthcare Chatbot to Enhance Patient Satisfaction and Engagement with Implementation of Advance HCI Techniques

Sat, 18 Apr 2026 00:00:00 +0000

Effective communication is crucial to the quality of healthcare delivery, whereas communication breakdowns between patients and healthcare providers are an ongoing issue. AI Chatbot have proven to be a potential solution to enhance patient interaction and satisfaction through availability and customization. The purpose of this research was to develop and evaluate an AI-based healthcare Chatbot, created in accordance with the principles of advanced HCI, to improve patient interdependence and satisfaction and to compare the performance of the developed Chatbot with the existing healthcare Chatbot. The Chatbot has been created based on publicly available healthcare datasets in Kaggle, and preprocessing, including cleaning and parting of the data and correction of spelling mistakes, have been done. Principles of HCI like usability, accessibility, error handling, and user feedback were taken into consideration and performance determined by usability testing of 150 users, engagement metrics evaluation, and deep-learning testing using accuracy, precision, and recall. It was found that usability was very high (88% of tasks in 12.5 seconds on average, 8% error rate), user satisfaction was 4.6/5 on average, and 85% were of the opinion that the interface was easy to use. The engagement metrics showed that the average session time was 4.5 minutes, 78% of tasks completed and 62% retention rate. The deep learning model was 91.3 %, 89.7 %, and 87.5 % accurate, precise, and recalled, respectively, in terms of the ability to interpret patient requests. Altogether, the HCI-based AI Chatbot enhanced patient engagement and satisfaction considerably through usability, accessibility, and receptive interactions, the error management, individualized communication, and multimodal interface, which supported effective patient-providing communication. The next idea that should be incorporated in the future work is to combine multilingual support, voice-based interaction, and connectivity with the electronic health record in order to increase its efficacy.

A Hybrid HHO-BA Feature Selection Framework for High-Accuracy Malicious URL Detection Using LightGBM

Sun, 01 Mar 2026 00:00:00 +0000

Malicious URLs are frequently used as delivery channels for malware and continue to represent a challenge in cybersecurity. The following paper propose URL detection framework based on using a combination of the Harris Hawks Optimizer (HHO) and Bat Algorithm (BA) using a union feature selection strategy. The goal is to build an informative and diverse subset of features by using the features that were chosen using two complementary metaheuristic search methods. LightGBM and Naive Bayes classifiers are used to evaluate the selected features on ISCX-URL2016 dataset. As it has been experimentally found, LightGBM has a higher accuracy of 99.52 % and Naive Bayes has an accuracy of 81.12 % indicating a distinct difference in the capability of modelling structured URL features. The proposed framework is competitive or better accurate when compared to some of the past studies. The results demonstrate that the HHOUBA union approach is successful in minimizing feature redundancy and discriminative information is maintained, which results in higher learning performance and a steady classification performance. The suggested solution is a solid solution to malicious URL detection.

Cybersecurity Architecture for Medical Live Monitoring Systems Using HTTP Protocols and SHA-256 Encryption

J. Dafni Rose, Sunitha T, Suma T, Cinthuja K, Mohanaprakash T A, Justindhas Y — Sun, 01 Mar 2026 00:00:00 +0000

Live health monitoring systems are revolutionizing patient care, yet the security and integrity of sensitive medical data remain a critical challenge. To address the vulnerabilities of centralized databases and basic authentication in traditional systems, we propose a secure framework that leverages a decentralized architecture. This proposed system uses standard HTTPS protocols for communication, with each client request authenticated using JSON Web Tokens (JWT) signed with HMAC-SHA256. Patient records are encrypted using AES-256-GCM before storage, with cryptographic hashes of these records written to a private Hyperledger Fabric blockchain for tamper-proof auditability. A smart contract enforces decentralized, role-based access control to ensure only authorized personnel can view sensitive information. A predictive analytics module processes secured vital data to forecast potential health deterioration, triggering automated alerts to the responsible physician. In a simulation handling 1,000 patient records over 30 days, the system demonstrated 99.98% data integrity, processed an average of 4,820 authenticated requests per minute with a mean latency of 185ms, and achieved 87.3% accuracy in predicting critical health events with a 30-minute lead time, confirming its efficacy as a robust and secure solution for remote patient monitoring.

Iot-Based Rain Detector System Using Arduino ESP-32

Sun, 01 Mar 2026 00:00:00 +0000

Is it feasible to automate the rain detector using an Arduino ESP-32 IoT device to enhance its performance? The Internet of Things [IOT] is changing weather practices and human life. Researchers are implementing IoT in weather practices. The development and statistical testing of an IoT-based rain detection module using the Arduino ESP32 microcontroller was the main focus of this study to enhance weather-based automation in smart environments. It was hypothesized that digital sensors and controllers could be interconnected to automate real-life systems, such as detecting rainfall and responding accordingly. This research aimed to provide a solution for urban areas where real-time rain monitoring is essential, especially for smart irrigation or safety systems. By implementing a rain detector system based on the IoT approach using the ESP32, the study offers a reliable, low-cost, and automated method to detect rain, trigger alerts, and integrate with other smart devices. An ESP-32 microcontroller was adapted and tested to collect real-time environmental data, specifically focused on detecting rainfall, humidity, temperature, and atmospheric conditions. Through its integrated Android application, the ESP-32 microcontroller enables users to monitor weather conditions remotely, providing real-time environmental updates and data analysis. This system supports better decision-making in weather-sensitive areas, automating responses such as halting irrigation or activating covers during rainfall. It has the potential to reduce water wastage by 30% and minimize weather-related damage by providing timely alerts through automatic rain detection. Using this IoT-based project, users, particularly in urban and rural areas, can manage outdoor systems more efficiently without relying on manual monitoring. It is recommended to implement this IoT-based rain-detection solution across both urban and rural agricultural environments to enhance weather management and disaster preparedness.

Hybrid Attention-Enhanced Deep Learning for Adaptive Digital Literacy Prediction in College Learners

Yuhan Chen — Sat, 23 May 2026 00:00:00 +0000

Digital literacy has become a pillar of rural revitalization efforts across the globe, and the creation of smart, adaptive systems to forecast and tailor learner outcomes constitutes a major area of research gap. This paper presents a new Hybrid Attention-Enhanced Deep Learning Network (HAE-DLN) a multi-task deep learning model that predicts overall scores on digital literacy (regression), classifies learners into proficiency levels (3-class), and classifies above/below-median learners (binary classification). The HAE-DLN is built upon a Feature Attention Module (FAM), Residual Dense Blocks (RDB), as well as a Multi-Task Learning Head (MTL), whilst being trained on a complete synthetic dataset of 1,000 rural learners, covering 23 demographic, behavioral, engagement, and outcome variables, in one end-to-end trainable framework. The first comparison on the baseline with the Random Forest and the Gradient Boosting regressors in their tuning with the help of the GridSearchCV proves that the proposed model has a better predictive power. Findings indicate that post-training scores, quiz performance and engagement efficiency are most effective predictors of digital literacy outcomes. The results have a direct implication on the design of adaptive, personalized digital literacy initiatives in rural and underserved people.

AI-Driven Deep Learning Discourse Analysis: Class Narrative and the Shifting Subjectivity of Women in the Film Adaptations of Jane Eyre

Rui Zhao — Mon, 25 May 2026 00:00:00 +0000

Gender representation in cinema has long reflected and reproduced broader societal inequalities, yet the mechanisms through which canonical literary adaptations negotiate female subjectivity across historical moments remain underexplored by computational methods. This study presents an AI-driven deep learning discourse analysis of female representation across film adaptations of Charlotte Brontë's Jane Eyre (1847), one of the most frequently adapted texts in the English literary tradition. We propose AttentionFusionNet, a novel deep learning architecture combining multi-head self-attention over tabular gender feature tokens with a residual multilayer perceptron (MLP) branch, fused through a gated mechanism and trained under a multi-task learning objective that jointly predicts Bechdel Test scores operationalised as a quantitative measure of female narrative subjectivity and Academy Award recognition. Based on information on The Movie Database, Bechdel Test Movie List, and historical data on the Academy Awards, the model combines cast and crew gender representation scores, commercial performance indicators and prestige indicators to generate interpretable, data driven evaluations of gender dynamics in the cinema industry. Synthetic Minority Oversampling Technique (SMOTE) and class-weighted loss functions are used to overcome ordinal imbalance in classes of Bechdel scoring schema of four classes. The analysis of explainability through GradientExplainer-based SHapley Additive exPlanations (SHAP) attribution demonstrates that the female representation of the cast and crew members is the major predictive signal, and the commercial and prestige indicators represent the secondary predictive ability, especially in intermediate Bechdel classes that represent partial narrative conformity. The results are discussed in the discourse analytical convention of feminist studies of adaptation with references to the theoretical frameworks of Mulvey, de Lauretis, Fairclough and Bourdieu to place the computational outputs into the greater context of cinematic gender construction politics. The publication provides the first methodologically comprehensive computational account of gendered literary adaptation at the scale of single canonical text, an integrated framework of modellable results, interpretable feature analysis, and evidence-based implications to feminist film theory, computation humanities, and production that is inclusive of all.

Quantum-Safe Wireless Sensor Networks: A Post-Quantum Cryptography Framework with Adaptive Security Optimization

Sun, 01 Mar 2026 00:00:00 +0000

Quantum computing poses a significant threat to Wireless Sensor Networks (WSNs) by undermining traditional cryptographic algorithms such as RSA and Elliptic Curve Cryptography (ECC). This work proposes a quantum-safe architecture for WSNs that integrates post-quantum cryptography (PQC) with lightweight IoT protocols to ensure long-term confidentiality, authenticity, and resilience. The framework leverages ML-KEM for key encapsulation and ML-DSA/SLH-DSA for signatures, and seamlessly integrates with EDHOC, OSCORE, and COSE. A novel Efficient Adaptive Parameter Selection (EAPS) mechanism dynamically adjusts cryptographic strength to balance security, energy consumption, and latency under varying network conditions. Experimental evaluation demonstrates that ciphertext fragmentation (8–27 pieces) results in manageable completion times of 4–10 seconds, while join operations consume only 0.005–0.03 J per cryptographic handshake event, while system-level energy consumption including network overhead is higher (~1.0–1.4 J per join cycle depending on hop count and retransmissions). Battery lifetime projections under steady-state sensing workloads range from approximately 115–280 months for low-duty-cycle operation, whereas realistic deployment conditions with periodic communication yield effective lifetimes of 11.5–24 months. Security analysis confirms robust resistance against downgrade attacks and Harvest-Now-Decrypt-Later (HNDL) threats. Moreover, EAPS reduces risk scores by over 50% compared to fixed schemes with less than 10% additional energy overhead, and batch signature verification improves scalability by increasing throughput from ~1,600 to ~2,800 verifications per second. Overall, the proposed framework demonstrates that WSNs can achieve quantum-resistant security with minimal performance trade-offs, ensuring readiness for the post-quantum era.

An Effective Rapid Threat Detection Framework for Detecting Malicious Domains Using Supervised Machine Learning Approach

Sun, 24 May 2026 00:00:00 +0000

The Domain Name System (DNS) that translates human readable domain names into IP addresses is important in Internet communication. The critical importance of DNS and its widespread trust has made it a typical target of cybercrimes such as cache poisoning and DNS spoofing, which redirect users to dangerous websites where they lose a lot of money and data. Conventional blacklist-based detection methods are becoming less and less effective against threats that are changing quickly. This research investigates the effectiveness of the supervised machine learning techniques to identify malicious domains using DNS data. Three models (RF TFIDF Lex, SAE Stacking, and CharCNN BiLSTM) are developed and evaluated using the CAIDA DNS dataset. The standard performance measures of accuracy, precision, recall, and F1-score are used to evaluate the performance of the models in a multi-class classification scenario. The experimental data indicate that CharCNN_BiLSTM model is superior to the rest of the approaches, particularly in detecting more complex patterns in the domain, such as DGA and phishing domains. These findings show that deep learning models constructed using sequences are better at reflecting the structural characteristics of malicious domains. The research indicates that in the case of DNS-based threat identification, machine learning-based methods will provide a more credible alternative to traditional methods.

A Hybrid Trust Evaluation Framework for Cloud Service Providers in Cybersecurity-Critical Environments

Sun, 01 Mar 2026 00:00:00 +0000

The rapid expansion of cloud computing has transformed digital service delivery, but establishing the trustworthiness of Cloud Service Providers (CSPs) remains a significant challenge, especially for sensitive workloads. This study presents a novel hybrid evaluation framework aimed at quantifying CSP trustworthiness from multiple perspectives. The framework combines a deterministic Multi-Attribute Trust Model (MATM) with a sophisticated Fuzzy Inference System (FIS). The MATM assesses CSPs based on three key attributes: Cost Efficiency (CE), Performance (P), and Reputation & Trustworthiness (RT). It utilizes a weighted normalization function to compute a transparent Trustworthiness Level (TL), offering a clear quantitative benchmark. To handle the inherent uncertainty and subjectivity in trust evaluation, the framework incorporates an FIS that uses fuzzy logic to interpret linguistic variables and capture the complex relationships among attributes. This dual approach provides both a straightforward, scalable assessment method and a more nuanced, human-centric evaluation. The framework’s effectiveness was validated through scenario analysis. As an initial proof of concept, the evaluation utilized constructed provider scenarios representing typical market profiles. The results demonstrate the model’s ability to clearly differentiate between service offerings, confirming its utility as a comprehensive decision-support tool.

Reinforcement Squeeze-and-Excitation Learning for Automated Gleason Grading in Prostate Cancer Histopathology

Maulika Patel, Parag Sanghani, Niraj Shah — Sun, 01 Mar 2026 00:00:00 +0000

Proper diagnosis and treatment plans of prostate cancer can only be based on accurate Gleason grading of the prostate cancer using histopathological whole-slide images and manual assessment is likely to be subject to inter-observer variability. The present paper will suggest a new Reinforcement Squeeze-and-Excitation Learning-based Convolutional Neural Network (RIL-SE-CNN) to perform automated Gleason grading. The framework combines reinforcement learning and squeeze-and-excitation blocks to establish dynamically recalibrated channel-wise features and provides an ability to focus diagnostically relevant glandular patterns. Reinforcement feedback is an optimal method of reinforcing the attention mechanism by means of rewarding the discriminative feature representations in heterogeneous tissue regions. The proposed model is tested on two benchmark datasets, PANDA and DiagSet-A, with the help of the typical performance indicators. On the PANDA dataset, the model has an accuracy of 0.9646, precision of 0.9642, recall of 0.9657, F1-score of 0.9642 and Matthew correlation coefficient of 0.9634. The suggested method achieves better results on the DiagSet-A dataset with the following accuracy of 0.9958, precision of 0.9954, recall of 0.9957, F1-score of 0.9955, and MCC of 0.9965. These findings indicate the strength and generalization power of the suggested system in the automated Gleason grading.

Computational Assessment of nsSNPs Associated with FGFR2 Gene

Hira Saleem, Anum Munir — Sat, 18 Apr 2026 00:00:00 +0000

The fibroblast growth factor receptor 2 (FGFR2) controls cell proliferation, differentiation, angiogenesis, and wound healing. Single nucleotide polymorphisms (SNPs) of the FGFR2 gene are linked with Pfeiffer syndrome, Crouzon syndrome, and Jackson-Weiss syndrome. Missense mutations in FGFR2 have also been reported in breast, gastric, and lung cancer. Objective: This study aimed to systematically analyze missense SNPs (nsSNPs) in FGFR2 using an integrative computational pipeline to identify variants with the strongest predicted pathogenic impact. Methods: FGFR2 gene data and protein sequence were retrieved from the NCBI dbSNP [build 153 (March 2020)] and UniProt database [release 2020_02; UniProt ID: P21802]. A total of eleven bioinformatics tools—SIFT, PolyPhen-2, PROVEAN, SNAP2, SNPs&GO, PANTHER, PhD-SNP, PMut, I-Mutant3.0, ConSurf server, and Project HOPE—were used to examine the deleterious potential of nsSNPs. The interaction of FGFR2 with different genes was analyzed using GeneMANIA. Results: We investigated 28,027 total SNPs from dbSNP, of which 701 were coding variants: 294 synonymous and 407 non-synonymous. Among the latter, 393 were missense mutations, 7 frameshift mutations, and 7 nonsense mutations. Only missense nsSNPs were retained for further analysis. Stepwise filtering identified 90 consensus-deleterious variants (≥ 6/8 predictors), 54 extremely damaging variants (all 8 predictors), and 38 unstable variants (ΔΔG ≤ −0.5 kcal/mol). A final set of 24 highly conserved and damaging variants (ConSurf score ≥7) was prioritized. Conclusion: A total of 24 nsSNPs (L757S, G690R, P666S, R664W, D644N, Y616C, L572F, L550P, L550V, I547M, V514M, G502E, G502R, E489K, R450C, P286S, C278Y, G271R, P263L, P256S, D225E, S224P, E219G, and Y105C) were predicted to have the most damaging and disease-causing effects on FGFR2 protein function and structure. Thus, the early prediction of FGFR2 gene functions could aid in disease prognosis. The results of our study provide beneficial information for devising early diagnostic and therapeutic measures.

IoT Based Garden Managing System with Wireless Control Using ESP-8266 Controller

Sun, 01 Mar 2026 00:00:00 +0000

Internet of Things (IoT) is changing agricultural practices and human life. The main goal of this study was to develop and do statistical testing of an IoT-based module for home gardening for the main purpose of increasing the production. It was hypothesized that digital devices may be interconnected to automate real-life systems in agriculture. This research further, found out the answer to fix the issue of lack of expertise for maintaining homegrown gardens in urban areas by developing a project-based on IoT approach for smart garden monitoring system. An ESP-8266 microcontroller was modified /extended and test statistics such as temperature, moisture, humidity, intensity of light, plants growth, and levels in gardens. Through its Android software, the ESP-8266 microcontroller provided data analysis to verify real-time updates and display environmental profiles. This system helps users optimize gardening practices, promoting plant growth, and achieve 65% reduction of water usage and 45% increases in plant growth. Using this IoT-Base project, the home gardeners of urban areas tackle with the challenges of maintaining gardens, eliminating the need for external gardening services. It is recommended to implement this IoT-based project in urban and rural areas.

A Machine Learning-Based Firewall Model for Effective Attack Detection Using Dragonfly and Bat Algorithms

Hani Al-Mimi, Ali Al Dahoud, Ahmad Al Dahoud — Sun, 01 Mar 2026 00:00:00 +0000

This work proposes a new machine learning (ML)-based firewall model for attack detection in modern networks. In this regard, the proposed ML-based firewall model is integrated with an advanced feature selection method to optimize the significant features that will improve the accuracy of detection by using the Dragonfly Algorithm (DA) and Bat Algorithm (BA). Logistic Regression (LR) and Gradient Boosting Trees (GBT) were used for the classification in this model. The model was tested and validated on the UNSW-NB15 dataset for the experimentation process because this dataset represents a comprehensive modern network activity. The GBT classifier achieved an accuracy of 100%, demonstrating its great capability in handling selected features and finding the attacks. The LR also attained a very high of accuracy of 99.84%. These outcomes highlight the efficiency of the proposed model in detecting attacks with minimal false positives. The integration of DA and BA for feature selection and the use of robust classifiers make the proposed ML-based firewall a promising solution for safeguarding modern networks.

LLM-Pep: Targeted Modeling for Anti-Parasitic Peptide Detection Using Large Language Models

Aqsa Amjad, Faria Nazir, Tayyaba Anees, Nosheen Qamar, Wajeeha Khalil — Mon, 25 May 2026 00:00:00 +0000

Parasites pose serious threats to host organisms, and anti-parasitic peptides (APPs) have shown potential in inhibiting parasite growth and reproduction. However, traditional biological screening methods such as nanomedicine-based assays and organism-based approaches are costly and time-consuming, highlighting the need for efficient computational prediction methods. In our study, we introduce a two-stage machine learning framework for accurate APP identification. To handle the class imbalance in the training data, we apply a random under sampling strategy to construct a balanced training set. Next, peptide sequences are encoded using pre-trained large language model-based embed dings and classified using a multi-layer perceptron (MLP) model. Unlike existing approaches that suffer from limited feature representation and poor generalization on independent datasets, our method leverages deep contextual sequence embedding combined with data balancing to improve robustness. Experimental results demonstrate that our model achieves an accuracy of 91.7% and an AUC of 0.939 on independent test sets, surpasses the existing approaches in APP prediction.

Analyzing the Impact of Pretrained Language Models on Low-Resource Languages

Muhammad Irshad Hussain, Shafiq Hussain, Aleena Jamil, Adeen Amjad, Sajid Iqbal — Sun, 01 Mar 2026 00:00:00 +0000

The rapid advancement of natural language processing has predominantly benefited high-resource languages such as English, Chinese, and Spanish, leaving thousands of languages underserved. This digital language divide limits equitable access to technology and threatens global linguistic diversity. This paper presents a systematic evaluation of eight pretrained language models across seven low-resource languages representing five distinct language families. Through extensive experiments on sentiment analysis, named entity recognition, and machine translation tasks, we demonstrate that multilingual BERT achieves the highest average accuracy of 74.5%. We further propose a novel adaptation framework combining vocabulary augmentation, continual pretraining, task-adaptive fine-tuning, and knowledge distillation that improves performance by up to 18.7%. Our analysis identifies vocabulary overlap as the strongest predictor of cross-lingual transfer success, explaining 76.3% of performance variance. These findings provide evidence-based guidelines for researchers and practitioners developing inclusive NLP technologies for underserved language communities. Limitations of this study include the focus on seven languages (generalizability to other low-resource languages requires further validation), computational constraints that prevented evaluation of models exceeding 300M parameters, and potential biases introduced by dataset availability and quality across languages.

An ANOVA-SMOTE-Based Framework for Multi-Class Autoimmune Disease Classification

Mon, 25 May 2026 00:00:00 +0000

Diagnosis of autoimmune disorders is prone to significant difficulties as the clinical picture is overlapped with complicated pathophysiology. The proposed study is the multi-class classification of autoimmune diseases based on a machine learning framework and taking into account four disease categories, namely, Graves’ Disease, Sjögren Syndrome, Rheumatoid Arthritis, and Systemic Lupus Erythematosus as well as a healthy control group, which leads to a five-class classification problem. The framework takes advantage of an integrated dataset, which includes demographic, hematological, biochemical, as well as immunological characteristics. K-Nearest Neighbors (KNN), Linear Support Vector Classifier (SVC), Decision Tree, Naive Bayes, Logistic Regression, and Ridge Classifier are six classification algorithms that were systematically tested under the same conditions of experiment. In order to improve the performance of the models and solve the problem of the class imbalance, an ANOVA-based feature selection strategy was utilized to be combined with the Synthetic Minority over Sampling Technique (SMOTE). The findings of experimental studies confirm that nonlinear classifiers, especially, Decision Tree and KNN, are much better than the linear models in terms of F1-score, recall, and precision. The added feature of the framework is the incorporation of a healthy control group, which enhances the capacity of the framework to differentiate between a diseased and a non-diseased person and increase the clinical applicability. In general, the suggested ANOVA-SMOTE-based framework can be considered as a powerful and scalable method of non-invasive autoimmune diseases classification, which can be implemented in the real clinical diagnostics and decision support.

Multi-Class Classification of Alzheimers Impairment and Dementia Stages using an Efficientnet-B0 Deep Learning Framework

Sat, 18 Apr 2026 00:00:00 +0000

Alzheimer’s disease is a progressive neurodegenerative condition that leads to deterioration of memory and cognitive function, where early detection is critical as therapeutic interventions are most effective prior to extensive neuronal loss. Many existing deep learning approaches focus on binary or ternary classification schemes, which inadequately reflect the full clinical spectrum of cognitive impairment and dementia, thereby limiting their practical diagnostic relevance.This study proposes a lightweight deep learning framework based on EfficientNet-B0 for fine-grained eight-class Alzheimer’s staging, encompassing four levels of cognitive impairment and four levels of dementia. A structured preprocessing pipeline, including normalization, contrast enhancement, resizing, and targeted data augmentation, is employed to improve MRI consistency and enhance discriminative feature learning.Under a strict patient-level evaluation protocol, the proposed model achieves a peak testing accuracy of 99.78% and demonstrates strong and stable performance when compared with commonly used 2D convolutional neural network baselines. Owing to its low computational complexity and consistent performance across classes, the framework represents a promising research-stage tool for automated Alzheimer’s disease staging and warrants further validation, including comparison with transformer-based and volumetric approaches, before clinical deployment.