Associate Professor and Head, Department of Artificial Intelligence and Data Science, Vidyavardhini's College of Engineering and Technology, Maharashtra, India
Authors - Aiswarya Rajan K K, K Nattar Kannan Abstract - This study presents a systematic literature review on the emergence, adoption, and challenges of AI-driven Human Resource Management (AI-HRM). Thematic synthesis and bibliometric insights were used to analyze eighteen Scopus-indexed studies published between 2019 and 2024 using the PRISMA framework. Using the Technology Acceptance Model (TAM/UTAUT), Socio-Technical Systems (STS) Theory, and Responsible AI principles, the review shows how AI improves HRM by automating repetitive tasks, facilitating data-driven decision-making, and allowing for individualized employee development. However, ethical risks like algorithmic bias, lack of transparency, privacy issues, and employee resistance continue to be major obstacles. The results imply that only when technological capabilities are in line with human judgment, organizational culture, and ethical governance can AI pro-vide long-term value in HRM.
Authors - Padma Lakshmi G, Swetha V, Monik Raj Murugan S, Srinivasa Perumal R, Lakshmi Priya G G Abstract - We have proposed ”Haze to vision: Pipeline for Underwater Image Restoration, Enhancement and Object detection”.The images captured underwater suffer from bluish tint,greenish tint,haze,color distortion. As light travels in water it will undergo scattering, refraction and absorption, the higher the wavelength will be observed first, and the lower wavelength will be absorbed later. This phenomenon affects the bluish/greenish color in the captured images . To study underwater species, underwater environments, we need good quality images and videos. The images captured underwater are poor quality. There have been several researches yet they have many drawacks.We have proposed pipeline.Our model consists of restoration,enhancement,object detection. Restoration process built from deep convolutional neural network called autoencoder .Which has been trained by 5000 synthetic images. The second model is the self-supervised enhancement model. The selfsupervised model is trained for 10,000 epochs of 5,000 datasets.We have used the customized gan model to obtain the best results.We have also used transfer learning and residual network for the improvement of the model.We have reached the PSNR value of 38.33 . CIQUE value 0.82 and UIQM 0.5.Our third model is object detection model. We have used the latest version of YOLOv5 for the betterment and the best object detection model.
Authors - Kishore S, Jeganathan L, Janaki Meena M, Ummity Srinivasa Rao, Jayaram Balabaskaran Abstract - Finding movies from an enormous number of movies that fit our interests and preferences becomes a challenging endeavor. Because recommendation systems address information overload by recommending the most appropriate products to users, they have become widely used in today’s world. The majority of recommendation systems disregard the constraints of the user such as not suggesting certain exceptional movies to them because they aren’t as popular as others. Furthermore, the lack of transparency about how these recommendation algorithms operate creates concerns regarding accountability. In this work, we propose an improved ALS-based recommendation framework that is implemented on Apache Spark and uses HDFS for processing and storing data. In order to address the long tail bias problem, we utilize the ALSbased framework that enhances exposure to low-frequency items through strong interaction filtering. This study employs SHAP to improve transparency and facilitate fairness analysis by explaining the elements generating recommendations to overcome this limitation. Root Mean Square Error (RMSE) and Top-K long-tail exposure metrics are used to assess the model’s performance on a large movie interaction dataset.
Authors - Pradnya Gotmare, Aryan Halkude, Manish Potey Abstract - The high pace of the data-driven applications growth in the distributed settings has enhanced the pressure to ensure that the data sharing infrastructure remains secure, efficient, and privacy-sensitive. The classic centralized data sharing architectures have the intrinsic limitations of being single-point-of-failure, untransparent, and unauthorized access to data, and prone to data corruption. To curb these hurdles, this paper proposes a decentralized approach of sharing secure data with the use of blockchain technology. The suggested system also uses the decentralized and unalterable features of blockchain to provide data integrity, transparency, and confidence among the involved parties without involving third-party intermediaries. Access control policies are the policies implemented using smart contracts to allow only trusted users to access the shared data. The solution is to keep sensitive information in off-chain repositories, where blockchain limitations of storage and scalability do not exist, yet cryptographic hash values and access control measures (ACMs) are stored in the blockchain registry. This design makes sure that the data transactions are confidential and data verifiability and auditability maintained.
Authors - Mutiara Ayu Mawaddah, Norhalina Senan, Mohd Norasri Ismail, Larisang, Muchlis Almubaraq Abstract - With the growing use of smart meters, massive amounts of electricity consumption data are being generated every day. Managing and analyzing this data efficiently is a big challenge. In this study, we generated a smart meter dataset of 10 million records, adding realistic anomalies such as missing values, noise, and unusual spikes to reflect real-world conditions. The data was stored in Hadoop Distributed File System (HDFS) on a single-node virtual machine running on Kali Linux for distributed processing . Using Apache PySpark, we cleaned the data, filled in missing values, identified outliers, and normalized features. For predicting electricity consumption, we trained a linear regression model which achieved a Root Mean Squared Error (RMSE) of 0.0141 and a R2 score of 0.9891, showing that the model predicts consumption very accurately. Overall, this study demonstrates a practical end-to-end approach that combines big data tools and machine learning for smart meter analytics. In the future, this workflow could be extended to multi-node clusters to improve fault tolerance and handle even larger datasets.
Authors - Shilpa Dhopte, Lalit Damahe 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 - Md. Shahidul Islam, Md. Murad Hossain, Omar Faruck Ansari Abstract - Time series prediction plays a critical role in monitoring and control of electrical power systems, particularly for detecting frequency fluctuations caused by imbalances between generation and demand. This study proposes an early warning framework for frequency fluctuation events using a hybrid k-Nearest Neighbour (KNN) and Dynamic Time Warping (DTW) approach combined with a global confidence interval based decision mechanism. Electricity frequency data collected from the New Zealand power grid over a six-month period were segmented into training, validation, and testing sequences. Alignment distances between historical and incoming sequences were used to identify precursor patterns indicative of impending frequency disturbances. Experimental results show that the proposed method achieves high warning accuracy with a very low false negative rate, outperforming baseline models such as ARIMA and LSTM. The findings demonstrate that KNN–DTW provides an effective and practical solution for early warning of frequency fluctuations, supporting improved operational reliability in modern power systems.
Authors - Gina Gallegos-Garcıa, Nidia A. Cortez Duarte, Jose A. Arellano Munguıa, Humberto A. Ortega Alcocer Abstract - "Communication has been a topic as ancient as man and at the same time so important that, over time, various forms have been cre- ated to facilitate it, among which stand out: mail, telephony, telegrams, and fax, to name a few. Nowadays many people use instant messaging applications to communicate with each other by feeling that their con- versations are protected. However, that feeling could not be further from reality and should not be taken lightly, since there are always groups focused on taking advantage of the vulnerability of this kind of applica- tions, resulting in users’ privacy being compromised. In this paper, we present the development of an instant messaging application that inte- grates a novel key establishment protocol based on a quantum-resistant algorithm. Our application employs cutting-edge lattice-based crypto- graphic techniques, ensuring robust security against quantum attacks while maintaining operational efficiency. Obtained results show the ap- plication’s viability by offering a practical solution to safeguard mobile communication in the impending quantum era."
Authors - Soumen Halder, Subhamoy Bhaduri, Binayak Mukherjee Abstract - Paraphrasing is significant in applications that require controlled lexical variation to original text with semantic equivalence, especially in educational assessment systems where student answers should be scored on more than surface level matching. Recent transformer-based paraphrasing models do not exhibit regulated structural changes but instead generate uncontrolled changes, are costly in terms of computation, and are not feasible in low-resource or real-time implementations. These limitations are overcome by this work with a lightweight synonymreplacement paraphrasing framework on the basis of exclusive embedding clustering. The proposed EEC-SRP model groups semantically similar words into local embedding clouds and limits the search of synonyms to the tiny areas, which lowers the complexity of search considerably. An embedding augmentation algorithm involves perturbation to form embedding clusters and a neural network is trained to output contextually favorable synonym embeddings in those clusters. Strict semantic fidelity and controlled lexical substitution is ensured by the model by maintaining word count and sentence structure. Experimental analysis of standard paraphrasing tasks show that the suggested methodology attains high levels of semantic similarity, competitive levels of BLEU and ROUGE, and significantly quicker inference than conventional embedding-based and transformer-based models. The proposed model can be effectively implemented in automated assessment systems, controlled text rewriting and resource-constrained applications of natural language processing due to its low memory footprint and computational efficiency.
Authors - Pablo Ramon, Josue Piedra, Nelson Piedra Abstract - The increasing numbers of deepfakes and AI tools have made it difficult to trust digital images these days. Images can be altered and ownership can be established without revealing private information. Current systems have many limitations, and systems that either rely on easyto change metadata or on cryptographic methods that are too costly like ZKSNARKs. To overcome these limitations, an authentication verification model has been presented named ZKP-Guard based on a Dual- Lock architecture framework. The detection system verifies an image is a real image by using ECDSA signatures and a custom ownership in the Schnorr-based Zero-Knowledge Proof for the protocol. This framework was tested on a dataset with significant number of images and produced desired results.
Associate Professor and Head, Department of Artificial Intelligence and Data Science, Vidyavardhini's College of Engineering and Technology, Maharashtra, India
Authors - Pablo Figueroa, Valeria Yunga, Pablo Ramon, Nelson Piedra Abstract - Traditional airport meet-and-greet operations are often characterized by a sea of physical placards and manual, paper-based logging systems. This manual approach not only creates logistical clutter in arrival halls but also leads to significant information lag and frequent data entry errors during the administrative reconciliation process. This paper presents the design and implementation of a centralized digital platform developed to streamline the coordination be-tween airport authorities, hotel representatives, and arriving passengers. Utilizing a responsive web-based architecture, the system eliminates the requirement for native application installations, thereby ensuring immediate accessibility for international travelers and hotel staff through their mobile devices. The platform integrates a multi-tier interface that facilitates real-time booking, automated digital check-ins, and instantaneous data synchronization. By replacing error-prone manual key-in tasks with an automated data pipeline, the system provides airport management with real-time operational visibility and analytics. Preliminary results from the implementation demonstrate a substantial reduction in guest waiting times and a marked improvement in data accuracy. Ultimately, this digital transition enhances terminal space management and provides a more seamless, professional experience for international arrivals, establishing a scalable model for modern airport ground handling services.
Authors - Monali Deshmukh, Payal Shete, Tanvi Pakhale, Pranjal Alhat, Krutika Salve Abstract - Because of their expensive price, large size, and reliance on lab settings, conventional oscilloscopes are inconvenient tools for signal analysis. They have made it necessary to have small, inexpensive, portable devices that can see waveforms outside of typical lab settings. The creation of a portable digital oscilloscope utilizing a 2.8-inch TFT display and an ESP32 microprocessor is detailed in this paper. Because of its autonomous operation, the gadget can record data in real time and display analog signals. Because it runs on batteries, the oscilloscope is affordable, lightweight, and portable. The ESP32 samples analog signals and displays them with user-controlled time-base settings. This oscilloscope has features including a grid display, waveform zooming, and freeze for convenience and readability. Both AC and DC signals can be monitored with an oscilloscope. According to tests, the device accurately displays common waveforms including sine, square, and sawtooth signals, which makes it ideal for embedded system development, simple troubleshooting, and instructional purposes.
Authors - Luis Anthony Hidalgo Ponce, Maricela Pinargote-Ortega Abstract - Technical support management in university environments often faces a high manual operational load due to the constant increase in digital service requests. This paper presents a multi-agent system based on Large Language Models (LLMs) designed to automate the ticket lifecycle, including classification, urgency-based prioritization, and intelligent routing. The proposed solution is built upon a modular architecture coordinated by an orchestrator agent and integrated with Retrieval-Augmented Generation (RAG) techniques to resolve frequent queries without human intervention. The system’s performance was evaluated through a controlled dataset, achieving a classification accuracy of 85.7% and a 100% effectiveness rate in user intent detection. The results demonstrate a significant reduction in response times compared to manual processes, validating the efficacy of generative artificial intelligence to optimize efficiency and user experience within university technology service desks.
Authors - Madhuri Surwase, Trupti Bansode, Jyoti Pawar, Smita Katkar, Vaishali Kalsgonda, Prakash Bansode, Namdev Falake Abstract - Automatic Speech Recognition (ASR) systems have achieved remarkable progress through deep learning and Transformer-based architectures, demonstrating near-human accuracy on clean audio. However, their performance degrades significantly under challenging conditions and specialized domains. This comprehensive study evaluates leading commercial ASR APIs—Google Cloud Speech-to-Text, Microsoft Azure Speech Service, AssemblyAI, Deepgram, OpenAI Whisper, Speechmatics, and others—across multiple dimensions: general speech recognition, low-quality forensic-like audio, domain-specific mathematical notation, and personalized speaker adaptation. Results demonstrate 100% accuracy on clean audio for leading systems (Deepgram, Speechmatics, Webkit SpeechRecognition), but dramatic performance degradation to 10− 81% word error rates on forensic-like audio. Analysis of domain-specific challenges reveals that none of the tested commercial ASR systems natively support direct transcription of mathematical symbols and Greek letters into structured symbolic output (e.g., LaTeX). The study identifies critical limitations in robustness, modularity, and domain adaptation, while highlighting promising customization mechanisms including custom vocabularies, language models, and post-processing integration. Performance improvements through speaker personalization ranged from 3% for natural voices to 10% for synthetic voices. Despite notable advances in end-to-end and Transformer-based approaches, ASR systems remain unsuitable for forensic applications and specialized domains without substantial customization and post-processing. Future research must address low-resource performance, linguistic diversity, robustness in extreme noise, and the integration of Large Language Models for semantic understanding. This paper synthesizes recent advances and critical gaps, providing a roadmap for advancing ASR technology in specialized and challenging acoustic environments.
Authors - Nurul Istiq faroh, Nur Asitah, Amiruddin Hadi Wibowo, Ricky Setiawan, Abdur-Razaq Aliyy Abolaji, Hendratno Abstract - Detecting structural breaks and anticipating volatility regimes in foreign exchange markets remain challenging due to the non-stationary and nonlinear nature of exchange rate dynamics. This study proposes a non-parametric framework for identifying structural breaks in the NZD/ USD exchange rate by integrating sliding-window volatility estimation, concentration bound based change point detection, and wavelet-based time frequency analysis. Volatility is first quantified using a movingwindow approach and compared against a Hoeffding bound to detect extraordinary events. The resulting change points are used to segment the exchange rate series into statistically reliable sequences, which are subsequently analyzed using wavelet scalograms. Empirical results reveal a consistent three-regime structure in the wavelet domain, comprising post-event reaction, stable market behavior, and pre-event escalation phases. Non-parametric statistical tests confirm significant differences in volatility distributions across these regimes, with the pre-event regime exhibiting markedly higher variability and acting as a precursor to structural breaks. The findings demonstrate that wavelet coefficients contain informative signatures of impending market instability. Overall, the proposed framework provides an interpretable and robust approach for analyzing regime-dependent volatility dynamics and offers valuable insights for early warning and risk management in currency markets.
Authors - Syeda Zaina Rohana Sneha, Mohammad Shamsul Arefin, M. M. Musharaf Hussain Abstract - This study details the development and evaluation of a web-based digital health platform that uses Optical Character Recognition (OCR) and Artificial Intelligence (AI) to automate the reading of medication labels and manage appointments. Users photograph medication labels and appointment slips, and the system automatically extracts and organizes relevant data to generate medication schedules, appointment calendars, and reminders with minimal manual effort. Designed with a user-centered approach to lessen cognitive load, the platform was tested with 35 users. Three experts verified the content validity of the assessment tool via the Item Objective Congruence (IOC) index. User satisfaction analysis indicates high approval, particularly for reducing the memory burden associated with medication routines and appointments. The results indicate that integrating OCR and AI can support continuous care, enhance usability, and increase patient engagement in the sustainable management of chronic diseases.
Authors - Tirupathi Rao Dockara, Manisha Malhotra Abstract - The prediction of cardiovascular disease (CVD) risk by machine learning is frequently impeded by duplicated and associated clinical characteristics, leading to complex and less robust models. Feature selection is therefore essential to improve model compactness while maintaining predictive performance. This study presents a systematic evaluation of meta-heuristic-based feature selection for CVD risk modeling under a standardized experimental setting. Feature selection is formulated as a wrapper-based optimization problem and evaluated using representative population-based meta-heuristic algorithms from multiple families. All methods are assessed using the XGBoost Histogram classifier on a public cardiovascular dataset comprising approximately 70,000 records with 13 clinical features. Experimental results show that meta-heuristic feature selection consistently reduces the number of input features by more than 60% while achieving comparable predictive performance across different algorithmic families. In addition, SHAP analysis is employed to examine the contributions of the selected features and support model interpretability.
Authors - Md. Shahidul Islam, Ronobir Chandra Sarker Abstract - The widespread adoption of artificial intelligence (AI) and automation is emerging as a central driver of productivity growth in European firms. Yet identifying the causal impact of AI adoption on firm productivity is complicated by endogeneity, selection bias, and heterogeneous treatment effects. This paper analyzes the productivity effects of AI and automation adoption using a unified framework that combines traditional econometric techniques with causal machine learning methods. Using firm-level data from Orbis merged with industry-level productivity and ICT capital measures from EU KLEMS for the period 2010–2023, we estimate both average and heterogeneous treatment effects. Double Machine Learning yields a robust average productivity gain of approximately 4.5 percent, while Causal Forests reveal substantial heterogeneity across industries, firm size, human capital, and digital maturity. The results provide credible causal evidence that AI adoption enhances firm productivity and highlight the importance of complementary capabilities in realizing its economic benefits.
Authors - Sonia Kuwelkar, Veena Gauns, Rohit Sopan, Sonia Shetkar, Dinanath Usgaonkar Abstract - Prompt engineering has emerged as an essential paradigm in leveraging desired behaviors from large language models (LLMs) without altering their parameters. Although the majority of the current literature has revolved around the introduction of novel prompt engineering strategies, there has been comparatively less emphasis on the contribution of the evaluation and optimization of prompts in concrete systems. In this paper, we offer a specialized review of prompt engineering from an evaluation/optimization centric viewpoint with a larger nod to conceptual developments and illumination rather than detailing the comparisons of approaches. Furthermore, we attempt to establish the concrete importance of prompt engineering via a real-life application, which resulted in improved performances in tasks through the process of prompt refinement and informal evaluations without the need to change the architecture and weights of the models. The paper will also introduce the deficiencies in prompt engineering in the realms of re-producibility, robustness, and the unavailability of standardized approaches in the aspect of concrete evaluations.
Authors - Domenico Vito, Carol Maione, Gabriela Fernandez, Catia Algieri, Sudip Chakraborty Abstract - The demand for long-endurance, intelligent drone systems is growing across diverse domains including defense, sports analytics, and industrial inspection. This paper presents the design and implementation of a solar-powered drone platform equipped with an autonomous, image-based range scoring system. Leveraging high-efficiency monocrystalline photovoltaic panels and Silicon- Carbide (SiC)-based lithium-ion batteries, the drone achieves extended flight durations while maintaining energy reliability. A centralized Energy Management System (EMS), featuring Maximum Power Point Tracking (MPPT) control, optimizes real-time energy harvesting and distribution. The platform also integrates an AI-enhanced thermal imaging module for precise target impact detection and scoring, with results computed using a multi-parameter range scoring model. An interactive Ground Control Station (GCS) interface enables intuitive mission planning, telemetry visualization, and data export. Experimental evaluations demonstrate significant gains in energy efficiency and scoring precision, underscoring the system’s potential for sustainable, autonomous aerial operations in real-world conditions.
Authors - Felix Kabwe, Jackson Phiri Abstract - The growth of Open Educational Resources (OER) has created a paradox of abundance, causing “academic infoxication” where students struggle to find content aligned with their competency levels. Traditional recommender systems often fail to interpret pedagogical context effectively. This paper presents the implementation and empirical validation of OPMAS, a multi-agent architecture orchestrated with LangGraph that utilizes Large Language Models (LLMs) to automate the curation and adaptation of educational resources. Unlike linear chatbots, OPMAS employs a state-graph of specialized agents (Router, Query, Search, Adaptation) to map user queries to European competency frameworks like DigComp. The system, built using Gemini 2.5 Flash and a hybrid retrieval strategy, was validated through a Minimum Viable Product (MVP). Results demonstrate a functional success rate of 95% in complex reasoning flows and a semantic precision of 0.77. Although the deep reasoning process introduces an average latency of 96 seconds, the system successfully prioritizes pedagogical relevance and content adaptation over immediate retrieval, proving the technical viability of agentic architectures for personalized education.
Authors - Minal Deshmukh, Aakash Dabhade, Daksh Jethwa, Siddhi Jadhav, Ketki Khirsagar Abstract - In this paper, we outline the design and implementation of a novel electronic voting kiosk, dubbed BlockVote, which helps counter identity-related fraud and data tampering via biometric and blockchainbased approaches. The proposed system is a standalone embedded system running on an ESP32-S3 SoC-based microcontroller. The system includes a touchscreen display for user input and an optical fingerprint sensor for identity checking. This collected bio-data and voting selection are then integrated in such a manner that a secure transaction is created through cryptography. This is then sent through the Node.js gateway, which leads it to the secure Ethereum-based blockchain network. Such an application of physical verification technologies with blockchain technology ensures that the proposed voting system is more secure than the traditional e-voting machines or e-voting websites. Block-vote is a hybrid security system in which hardware-based verification techniques are combined with blockchain-based data management in a power-saving, compact format. The prototype has shown proof of its functional viability, its module-based construction, and its reliability, particularly in the field of embedded systems. The experimental results demonstrate the system’s high precision, low latency, and robustness against illegitimate use. The suggested framework demonstrates the practical feasibility of blockchain and biometric technology in the creation of trustworthy electronic voting systems that can be used in both urban and rural areas.
Authors - S.D.P. Abeysekara, J.A.D.N. Jayakody, K.A. Dilini T. Kulawansa Abstract - Breast cancer is the second most prevalent cancer globally and a leading cause of death among women. According to the World Health Organization, over 2.3 million new cases are diagnosed annu ally, emphasizing the need for early and accurate detection.In this work, Wavelet-Driven Intelligent Model for Multi-Class Breast Cancer Diagno sis is proposed. In this proposed work, three level wavelet decomposition is used on BreakHis data to extract wavelet based features. These fea tures were fed to Artificial Neural Network Classifiers such as Multi-Layer Perceptron (MLP), Radial Basis Function (RBF) and Machine Learning Classifier Random Forest (RF). Multi-class classification (binary , be nign sub-types, 4 malignant sub-types) of breast tumour has been done. The experimental results show that RF achieved high accuracy of 94% for benign and malignant, 97% for benign sub- type and 92% for malig nant subtype classification compared to RBF and MLP. Deep learning models such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) are more effective when trained on large-scale datasets but for small datasets and limited resource environments, the proposed framework ensures efficient and consistent diagnostic approach. In future, a prototype breast cancer alert system can be developed using raspberry pie for real time application.
Authors - Md. Shahidul Islam, Atiqur Rahman, Md. Murad Hossain Abstract - This study examines the influence of both demographic and natural factors on climate change risk perception in New Zealand. Using data from a nationally representative survey, the analysis applies exploratory factor analysis to construct a composite measure of risk perception, followed by correlation and regression modeling to evaluate the relative contribution of environmental exposure and human characteristics. The findings indicate that while natural factors such as temperature anomalies and extreme weather exposure significantly shape perceived risk, demographic variables including prior disaster experience, trust in scientific institutions, and media exposure exert a stronger overall influence. These results underscore the importance of incorporating social and behavioral dimensions into climate risk assessments and policy development to enhance public engagement and adaptive capacity.
Authors - Shreyas M S, Kumar P K, Venkateswara Rao Kolli Abstract - The Newborns mostly use infant crying as their main form of communication and it represents a great variety of physiological and emotional conditions. Despite the high potential of automated infant cry analysis in early diagnosis and support of caregivers, the application in real-life still has low usage rates because of environmental noise, imbalance of classes, low interpretability, and high computational cost. This paper is a compilation of an effective, interpretable, and real-time infant cry classification system using a two-step hierarchical methodology. The first stage involves a distinction of cry and non-cry sounds to reduce the rate of false alarms due to background noise. The second stage involves categorizing detected cries into a particular intent. An adaptive feature fusion strategy based on reinforcement learning, gives the cepstral and prosodic and qualitative acoustic features dynamic weights, and SHAP-based explainability offers explicit feature interpretations. Data augmentation, SMOTE-Audio, and model pruning are used to find solutions to the issues of class imbalance, noise robustness, and deployment constraints. Experimental evidence shows that the proposed approach outperforms single feature base-lines, it is also stable in noisy environments and also attains significant parameter reduction without significant loss in performance, making it possible to run in resource-constrained devices in real time. The system is tested on a publicly available infant cry dataset which contains 889 audio samples of cry and non-cry signals in five categories of cry intent and was recorded in realistic conditions.
Authors - Md. Shahidul Islam, Md. Raihan Habib Abstract - Detecting structural breaks and anticipating volatility regimes in foreign exchange markets remain challenging due to the non-stationary and nonlinear nature of exchange rate dynamics. This study proposes a non-parametric framework for identifying structural breaks in the NZD/ USD exchange rate by integrating sliding-window volatility estimation, concentration bound based change point detection, and wavelet-based time frequency analysis. Volatility is first quantified using a movingwindow approach and compared against a Hoeffding bound to detect extraordinary events. The resulting change points are used to segment the exchange rate series into statistically reliable sequences, which are subsequently analyzed using wavelet scalograms. Empirical results reveal a consistent three-regime structure in the wavelet domain, comprising post-event reaction, stable market behavior, and pre-event escalation phases. Non-parametric statistical tests confirm significant differences in volatility distributions across these regimes, with the pre-event regime exhibiting markedly higher variability and acting as a precursor to structural breaks. The findings demonstrate that wavelet coefficients contain informative signatures of impending market instability. Overall, the proposed framework provides an interpretable and robust approach for analyzing regime-dependent volatility dynamics and offers valuable insights for early warning and risk management in currency markets.
Authors - Diego Perez-Lopez, Rodolfo Bojorque, Jorge Duenas-Lerin, Raul Lara-Cabrera Abstract - Accurate early detection of liver cancer remains a significant clinical challenge, primarily due to scarce annotated imaging data, inconsistencies in radiological interpretation, and the inherent opacity of deep learning models. To address these limitations, this study proposes a clinically informed, explainable deep learning framework designed specifically for low-annotation settings. The framework combines transfer learning with advanced visualization techniques, enabling both high diagnostic accuracy and medically meaningful outputs that integrate seamlessly into clinical workflows. Three pre-trained CNN architectures — ResNet-50, DenseNet-121, and EfficientNet-B4 — were adapted to liver cancer imaging through domain-specific fine-tuning. Model generalizability was reinforced by combining geometric data transformations with StyleGAN2-derived synthetic lesion generation. Model transparency was facilitated through Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP), while clinical trustworthiness was evaluated via predictive uncertainty quantification, subgroup bias analysis, and resistance to adversarial perturbations. The proposed framework was evaluated on the LiTS and TCGA-LIHC datasets, demonstrating a 15–20% improvement in accuracy over baseline models that consisted of standard convolutional neural networks trained from scratch without transfer learning or data augmentation. EfficientNet-B4 achieved 94.2% accuracy, 0.96 specificity, and an AUC-ROC of 0.978. Grad-CAM accurately highlighted tumor regions in 89.4% of cases, and Bayesian dropout identified 7.3% of predictions as uncertain. These findings demonstrate the framework’s potential for clinical deployment by balancing performance, transparency, and reliability.
Authors - Jutika Borah, Debarun Chakraborty, Bhabesh Deka, Rosy Sarmah, Siddeswara Bargur Linganna, Diptadhi Mukherjee, Ram Bilas Pachori, Mohit Khamele Abstract - Electroencephalogram (EEG) signal modeling for downstream tasks, such as classifying neurological states and identifying biomarkers, is essential for designing effective brain-computer interfaces. Conventional methods often treat EEG channels independently, overlooking inter-channel dependencies, while existing graph-based approaches address this limitation either through fixed electrode geometry or entirely data-driven connectivity. In this paper, we propose a graph representation framework that combines coherence-based spectral connectivity with domain-informed priors, such as anatomical structure and regional proximity, based on graph signal processing (GSP). The resulting representation embeds multichannel EEG signals as attributed graphs through graph convolutional networks (GCNN) to learn discriminative embeddings. Experimental results demonstrate that the hybrid framework enhances classification performance, with the proposed GCNN-deep model achieving the highest area under the receiver operating characteristic curve (AUC) across all datasets and reaching 93% on Dataset 1. These EEG datasets correspond to three independent populations and include recordings from both healthy individuals and patients with neurological disorders such as major depressive disorder (MDD) and epilepsy.
Authors - Samiksha Chougule, Kirti Satpute, Krishnraj Patil, Om Kumbhardare, Sumedha Patil 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 restrict effective healthcare delivery. Moreover, unreliable internet connectivity limits the reach of conventional digital health platforms. 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 AI, ML, NLP, OCR, and speech recognition to allow users to interact in their native languages through 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 internet-dependent 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 - Noor, Soumya Mukherjee, Shivraj Singh Yadav Abstract - The increasing numbers of deepfakes and AI tools have made it difficult to trust digital images these days. Images can be altered and ownership can be established without revealing private information. Current systems have many limitations, and systems that either rely on easyto change metadata or on cryptographic methods that are too costly like ZKSNARKs. To overcome these limitations, an authentication verification model has been presented named ZKP-Guard based on a Dual- Lock architecture framework. The detection system verifies an image is a real image by using ECDSA signatures and a custom ownership in the Schnorr-based Zero-Knowledge Proof for the protocol. This framework was tested on a dataset with significant number of images and produced desired results.
