Assistant Professor, Assistant Head Research- Department of Information Technology, Vishwakarma Institute of Technology Pune (Affiliated to Savitribai Phule Pune University, Maharashtra, India
Authors - Shruti Thakur, Shilpa Nikhil Bhosale, Priti Prakash Jorvekar, Sandeep Muktinath Chitalkar, Harshala Shingne, Rupali Vairagade Abstract - This study examines the effectiveness of ensemble learning models for detecting fraud in e-wallet transactions under extreme class imbalance and temporal dependence. Using the PaySim bench-mark dataset, a time-aware experimental framework is developed that incorporates forward-chaining evaluation, imbalance-aware resampling, hyperparameter optimisation, probability calibration, and cost-sensitive threshold tuning to reflect real-world deployment conditions. RF and XGBoost are systematically compared across multiple dataset scales and train–test splitting strategies. Empirical findings show that XGBoost consistently outperforms RF, achieving the highest F1-score, maintaining PR-AUC above 0.88, and demonstrating near-perfect ROC-AUC, indicating strong discriminative capability. Following isotonic calibration, XGBoost also produces the lowest Brier score, highlighting superior probability reliability for risk-based decisions. Performance gains plateau beyond a 75% training share, while XGBoost preserves stable performance as the test window expands, unlike RF. Overall, the results support prioritising gradient boosting models, adopting time-aware validation, and integrating calibrated risk scoring in operational e-wallet fraud detection systems.
Authors - Shoh-Jakhon Khamdаmov, Muazzam Akramova, Rano Abdullaevna Sadikova, Azamat Kasimov, Jasurbek Pozilovich Kurbonov, Alisher Bakberganovich Sherov, Dilshoda Akramova Abstract - Attention Deficit Hyperactivity Disorder (ADHD) is one of the most common neurodevelopmental disorders in children, characterized by inattention, hyperactivity, and impulsivity that impair academic and social functioning. Due to its heterogeneous presentation and symptom overlap with other cognitive disorders, early and accurate diagnosis remains challenging. This study proposes a multimodal machine learning framework integrating behavioral, neuroimaging, and physiological data to predict ADHD in children. Convolutional Neural Networks (CNNs) are used to extract features from brain MRI scans, Long Short-Term Memory (LSTM) networks model temporal patterns in physiological signals such as EEG and heart rate variability, and ensemble learning methods incorporate behavioral and clinical attributes. Both feature-level and decision-level fusion strategies are evaluated. Results on benchmark datasets show that the multimodal model consistently outperforms unimodal approaches in accuracy, sensitivity, and F1- score, demonstrating the potential of AI-driven multimodal systems for early, objective, and interpretable ADHD diagnosis.
Authors - Matjere Matsebe, Nobubele Angel Shozi Abstract - It is possible to increase the acceptability of small wind turbines for wind regions with low wind velocities for rural as well as urban sectors by placing them inside diffusers. The research on development of various diffusers is a major re-search area nowadays. Curved flanged diffusers can deliver better performance by adding a cylindrical throat section between converging and diverging sections. This research paper presents a systematic study on short curved flanged diffusers with converging-diverging sections and extended uniform throat between them. Twenty-five diffuser models are studied using Computational Fluid Dynamics using ANSYS Fluent. These models are finalized using the design of experiments for six variables at five levels. The throat diameter for all diffuser models is fixed. The investigation is performed by considering radial average velocity and percentage velocity variation along the radial planes. The global velocities are observed as 1.18 to 1.47 times that of the radial average velocities. The diffuser dimensions are optimized to maximize radial average velocity and to minimize the velocity variation along the radial planes. The diffuser with optimized dimensions is manufactured and tested experimentally in a wind tunnel. Good matching is seen between the predicted results and experimental results. The optimized diffuser has the ability to produce more than two times the power that of the turbine without a diffuser.
Authors - Murodov Gayrat Nekovich, Kholmuhamedov Bakhtiyor Farkhodovich, Avezov Sukhrob Sobirovich, Khudayberganov Nizomaddin Uktambay ogli, Yunusova Maftuna Shokirovna, Mansurova Shahinabonu Najmiddin qizi Abstract - The classification of ECG signals continues to be a major focus in intelligent healthcare systems, especially for the early identification of cardiac arrhythmias. In this work, we propose a hybrid probabilistic neural strategy that integrates Bayesian Networks with Artificial Neural Networks (ANNs) to enhance the reliability of ECG classification. The approach begins by extracting informative ECG features, such as crosscorrelation and phase-based characteristics. A Bayesian Network is then applied to model the probabilistic dependencies among these features and identify those most relevant to classification. At the same time, an ANN is trained on the refined feature set to learn complex non-linear patterns present in the signals. The two models are subsequently combined through a weighted voting mechanism to form an ensemble classifier. Experimental evaluation using an ECG dataset indicates that the proposed ensemble achieves higher accuracy and stability compared to its individual components. Notably, the method demonstrates strong capability in distinguishing multiple arrhythmia categories, which are typically difficult to classify. Overall, the results highlight the promise of hybrid probabilistic–neural models for improving automated ECG interpretation and supporting more accurate diagnosis of cardiac abnormalities.
Authors - K S Shubham, Uma Mudengudi, Ujwala Patil Abstract - Secure, compliant, and interoperable data sharing remains a core bottleneck for cross-organizational analytics and AI, particularly under evolving privacy regulations, contractual obligations, and adversarial threats. This paper introduces HARMONIA, a pluggable, risk-aware data sharing framework that integrates policy-as-code enforcement, continuous compliance monitoring, provenance-grade evidence, and revocation with machine unlearning. HARMONIA is inspired by the iterative Analyzer–Mechanic and Conductor–Observer operational pattern described in the HARMONIA strategic perspective, generalizing its quality-gate-and-repair loop to a policyand- risk-gated release lifecycle. We formalize an architecture that separates governance, control, and data planes; define a release-mode lattice that enables explainable fallbacks among raw export, masking, kanonymity, differential privacy, synthetic data, query-only access, and federated compute; and propose an evidence model aligned with W3C PROV. We provide a proof-of-concept (POC) blueprint implemented with commodity components (OPA, OAuth2/OIDC, PostgreSQL, and object storage) and specify interfaces that support end-to-end request-to-release-to-revocation workflows, including batch-scoped unlearning for model derivatives. The paper concludes with an evaluation methodology and a standards-aligned roadmap for deployment in sovereign data spaces.
Authors - Mehzabul Hoque Nahid, Fatema Tuz Zahra, Mubashshir Bin Mahbub, Saleh Ahmed Jalal Siam Abstract - Personalizing learning in higher education presents a significant challenge due to the difficulty of providing individual feedback to large student cohorts. This study proposes an intelligent tutoring system based on a multi-agent architecture utilizing Large Language Models (LLMs) to address scalability and adaptability issues. The proposed architecture integrates two complementary subsystems: a reactive module that answers student queries using Retrieval-Augmented Generation (RAG) to ensure accuracy based on course materials, and a proactive module that autonomously analyzes student profiles to generate personalized study plans without direct instructor intervention. The system was implemented using Lang- Graph for agent orchestration and MongoDB for state persistence. Experimental validation was conducted using a curated golden dataset from a university course. Results demonstrate a retrieval precision of 94.2% and a faithfulness score of 87.8%, significantly mitigating hallucinations common in monolithic models. Furthermore, the operational cost analysis indicates high financial viability for mass implementation. This dual approach offers a robust solution for automated, highquality educational support, effectively bridging the gap between standardized teaching and personalized learning needs.
Authors - Vemuri Bharath Kumar, Anjan Babu G Abstract - Healthcare data scarcity poses significant challenges for machine learning applications in clinical settings, particularly for conditions with limited patient populations. This paper presents a novel quantumenhanced data augmentation framework that addresses this challenge through a three-pillar architecture: Quantum Random Number Generation (QRNG) for true randomness, Statistical AI for intelligent parameter optimization, and Generative AI for clinical interpretability. Our implementation utilizes Bell state quantum circuits to generate genuinely random perturbations, ensuring higher entropy than classical pseudorandom methods. The framework incorporates medical domain knowledge through constraint-aware augmentation, maintaining clinical validity while generating synthetic patient records. Experimental evaluation on the Pima Indians Diabetes dataset (768 samples, 8 features) demonstrates that our quantum-enhanced approach achieves 100% medical constraint compliance while generating high-quality synthetic data. The system provides both command-line and web interfaces, with automatic fallback to classical methods when quantum resources are unavailable. Our contributions include: the first practical application of quantum computing to healthcare data augmentation, an AI-driven optimization system that automatically determines augmentation parameters, integration with large language models for non-technical summarization of validation reports, and a production-ready implementation with comprehensive validation mechanisms. The framework represents a significant advancement in synthetic medical data generation, offering a scalable solution for addressing data scarcity in healthcare AI applications.
Authors - Sandhya Awate, Vipin Kumar Gupta Abstract - Rural communities face significant challenges in accessing essential healthcare services due to language barriers, limited health literacy, and insufficient medical support. Difficulties in understanding medical information, communicating symptoms, and interpreting diagnostic reports further hinder effective healthcare delivery. Additionally, unreliable internet connectivity restricts the reach of conventional digital health platforms. To address these challenges, this paper presents a Multilingual AI Health Assistant designed to operate on low-cost edge devices, enabling offline functionality to ensure continuous access and data privacy in low-connectivity areas. The proposed system integrates Artificial Intelligence (AI), Machine Learning (ML), Natural Language Processing (NLP), Optical Character Recognition (OCR), and speech recognition, allowing users to interact in their native languages via text or voice. It analyzes user-reported symptoms to predict probable health conditions, translates complex medical reports and prescriptions into simplified, localized explanations, and provides recommendations for nearby healthcare facilities. Unlike internetdependent telemedicine systems, this edge-based solution processes data directly on the device, safeguarding sensitive health information while maintaining reliability. By bridging linguistic and literacy gaps, the proposed assistant empowers rural populations with accessible and actionable healthcare insights, ultimately improving health outcomes in underserved regions.
Authors - P. Sivaperumal, R. Naresh, S. Prawin, B. E. Viruthatchanan Abstract - The food portion estimation is a critical component of automated dietary assessment systems, enabling better monitoring of nutritional intake and supporting healthcare, weight management, and public health applications. Traditional self-reporting methods are often inaccurate and time-consuming, motivating the need for computer vision–based approaches that can reliably estimate food portions from images captured in real-world conditions. This paper presents deep learning pipeline for food portion estimation that integrates image preprocessing, deep learning–based segmentation, and geometric volume computation. The data preprocessing with Mask R-CNN used for precise food seg-mentation, providing pixel-level masks and bounding boxes that isolate individual food items from complex backgrounds. The segmented mask is used to estimate the pixel area of the food region. Experimental evaluation demonstrates that the proposed method achieves high segmentation accuracy, with a segmentation IoU of 87.6%, precision of 90.3%, recall of 88.9%, and an F1-score of 89.6%. The pixel area estimation error is limited to 6.8%, resulting in an overall portion estimation accuracy of 89.1%, indicating reliable and consistent performance across different food images. The proposed framework highlights the effectiveness of combining deep instance segmentation with geometric volume estimation for accurate food portion assessment. Future work will focus on multi-view image integration and real-time deployment in mobile dietary monitoring systems to enhance robustness and scalability.
Authors - Chalani Dinitha, Saadh Jawwadh Abstract - Automated Image Enhancement from CCTV surveillance relies heavily on accurate image segmentation; however, real-world footage is often degraded by low illumination, motion blur, occlusion, and background clutter, causing conventional segmentation models to lose boundary precision and small object details. This paper proposes EdgeLite-CrimSegNet, a novel lightweight boundary-aware segmentation network designed specifically for crime scene analysis. Unlike existing fast segmentation models that prioritize global context, the proposed architecture adopts a boundary-first learning strategy, where crime-relevant contours are explicitly extracted and refined before region-level segmentation. A compact edge-aware encoder, boundary-guided feature refinement module, and progressive region filling strategy are introduced to improve segmentation accuracy while maintaining real-time performance. Experiments on CCTV frames derived from the UCF-Crime dataset demonstrate improved boundary preservation, higher IOU, and better segmentation of overlapping and small objects compared to conventional lightweight segmentation networks, confirming the suitability of EdgeLite-CrimSegNet for real-time surveillance applications.
Assistant Professor, Assistant Head Research- Department of Information Technology, Vishwakarma Institute of Technology Pune (Affiliated to Savitribai Phule Pune University, Maharashtra, India
