Browsing by Author "Sowmya Kamath, S."

Now showing 1 - 14 of 14

A Comprehensive Analysis ofÂ Classification Techniques forÂ Effective Multi-class Research Article Categorization onÂ anÂ Imbalanced Dataset
(Springer Science and Business Media Deutschland GmbH, 2025) Gowhar, S.; Kempaiah, P.; Sowmya Kamath, S.; Sugumaran, V.
Categorizing scientific articles into specific research fields is a challenging problem, affected by the volume and variety of literature published. However, existing classification systems often suffer from limitations regarding taxonomy or the models used for classification. This article explores a comprehensive analysis of approaches built on Sentence Transformer embeddings combined with Machine Learning algorithms, Neural Networks, and Transformers to classify articles into 123 predefined classes, with the dataset being heavily imbalanced. The effectiveness of Large Language Models (LLMs) for generating synthetic data is also experimented with, along with synonym augmentation SMOTE and employing 1D CNNs for text classification. The best-performing model is a hierarchical classification model trained on MP-Net sentence embeddings that achieved an accuracy of 78%, outperforming all other models. Â© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
A Multimodal Approach Integrating Convolutional and Recurrent Neural Networks for Alzheimer's Disease Temporal Progression Prediction
(IEEE Computer Society, 2024) Durga Supriya, H.L.; Thomas, S.M.; Sowmya Kamath, S.
Alzheimer's Disease (AD) poses a substantial healthcare challenge marked by cognitive decline and a lack of definitive treatments. As the global population ages, the prevalence of AD escalates, underscoring the need for more advanced diagnostic techniques. Current single-modality methods have limitations, emphasizing the critical need for early detection and precise diagnosis to facilitate timely interventions and the development of effective therapies. We propose a novel multimodal medical diagnostic framework for AD employing a hybrid deep learning model. This framework integrates a 3D Convolutional Neural Network (CNN) to extract detailed intra-slice features from MRI volumes and a Long Short-Term Memory (LSTM) network to capture intricate inter-sequence patterns indicative of AD progression. By leveraging longitudinal MRI data alongside spatial, temporal, biomarkers, and cognitive scores, our framework significantly enhances diagnostic accuracy, particularly in the early stages of the disease. We incorporate Grad-CAM to enhance the interpretability of MRI-based diagnoses by highlighting relevant brain regions. This multimodal approach achieves a promising accuracy of 92.65%, outperforming state-of-the-art works by a margin of 6%. Â© 2024 IEEE.
Automated hard exudate segmentation using neural encoders and attention mechanisms for diabetic retinopathy diagnosis
(Inderscience Publishers, 2023) Gawas, P.; Sowmya Kamath, S.
Diabetic retinopathy (DR) is a complication caused by increased blood glucose levels, which causes retinal damage in diabetic patients’ eyes. If not discovered and treated early, it can lead to vision loss. Hard exudates (HE) are one of its characteristic signs. Identification of HE is a paramount step in early diagnosis of DR. In this work, the suitability of U-Net-based deep CNN with different encoder configurations and attention gates (AG) is experimented, for HE segmentation. The proposed models were benchmarked on the standard IDRiD dataset. To overcome the challenges related to the limited dataset, data augmentation techniques were also applied to generate image patches and used for model training. Extensive experiments on the dataset revealed that U-Net with AG achieved an accuracy of 98.8%. The U-Net with ResNet50 as the encoder backbone achieved an accuracy of 98.64%. The findings show that the presented models are effective and suitable for early-stage clinical diagnosis. © © 2023 Inderscience Enterprises Ltd.
Automated Health Insurance Management Framework withÂ Intelligent Fraud Detection, Premium Prediction, andÂ Risk Prediction
(Springer Science and Business Media Deutschland GmbH, 2024) Devaguptam, D.; Gorti, S.S.; Leela Akshaya, T.; Sowmya Kamath, S.
Private insurance is already one of the sectors with the greatest growth potential. For the majority of high-value assets today, including houses, jewelry, cars, and other valuable items, there are insurance solutions available. To maximize profits while handling client claims, insurance firms are leading have adopted cutting-edge operations, procedures, and mathematical models for estimating risks and serving customer best interests, while also maximizing profits. In this work, we aim to develop a machine learning-based automated framework that minimizes human involvement, protects insurance operations, identifies high-risk consumers, uncovers false claims, and lowers financial loss for the insurance industry. This framework consists of fraud detection followed by risk prediction and premium prediction. We trained and tested different machine learning approaches for each of the three insurance processing tasks; the observations are presented in this article. Â© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.
Automated Marine Debris Detection from Sentinel-2 Satellite Imagery
(Institute of Electrical and Electronics Engineers Inc., 2024) Priyadarshini, R.; Arya, V.; Sowmya Kamath, S.
Marine debris present a severe, escalating threat to oceans and coastal ecosystems, requiring effective monitoring and detection. This work proposes an automated marine debris detection system utilizing satellite imagery data from the MARIDA dataset, sourced from Sentinel-2. Advanced AI techniques are leveraged to analyze high-resolution satellite imagery, and the models are trained to facilitate the identification/tracking of marine debris across various water bodies. Experiments reveal that the machine learning models form a robust baseline, while the UNet model achieves improved precision. The proposed Attention-activated UNet model achieved the best performance, particularly in challenging conditions. Â© 2024 IEEE.
Content Based Surgical Video Retrieval using Scene and Motion Feature Embeddings
(Institute of Electrical and Electronics Engineers Inc., 2024) Parwal, V.; Sowmya Kamath, S.
Effective clinical care increasingly relies on the ability to identify relevant details within extensive heterogeneous medical datasets, a task that becomes particularly challenging with complex media formats. Traditional keyword-based retrieval methods have limitations, and content-based retrieval techniques have gained popularity. Surgical video retrieval, a novel and largely unexplored research area, offers significant utility, especially in real-time clinical scenarios. As surgical procedures are predominantly recorded via video, these recordings are often stored on servers for subsequent access, necessitating efficient retrieval of specific footage segments by surgeons. In this work, we explore various models to extract scene and motion feature embeddings, accounting for motion intensity in videos. Accurate retrieval of similar videos or images requires precise embeddings of objects within the frames. For experimental validation, we utilized the public Cholec80 dataset, which includes cholecystectomy surgery videos. The combined scene and motion features are hashed to generate frame-level binary codes, facilitating rapid search and retrieval. Â© 2024 IEEE.
Ensemble Machine Learning Approaches for Automated Fungal Keratitis Diagnosis Using In Vivo Confocal Microscopy Images
(John Wiley and Sons Inc, 2025) Sowmya Kamath, S.; Reji, S.; Vaibhava Lakshmi, V.; Supreetha, S.; Gawas, P.; Mayya, V.; Hazarika, M.
Fungal keratitis (FK) is a severe ocular infection that can lead to significant vision problems or blindness if not diagnosed and treated promptly. Early and accurate detection of FK is essential for effective management. Traditional diagnostic methods are often time-consuming and require specialized laboratory resources. Recently, advances in artificial intelligence and computer vision have enabled automated diagnosis of FK using slit-lamp images. In this article, a comprehensive evaluation of state-of-the-art techniques adopted for classifying FK using in vivo confocal microscopy (IVCM) images is presented. Detailed experiments and performance evaluation of various machine learning models are systematically performed, with a focus on evaluating the effect of diverse techniques for image processing, data augmentation, hyperparameters and model finetuning to assess each model's strengths and limitations. Experiments revealed that applying green channel preprocessing with a 12-feature set achieved 99% accuracy using Random Forest, highlighting its effectiveness in FK detection, while complex techniques like histogram modelling reduced accuracy to 64%. Robust models like AdaBoost and RUSBoost maintained high F1-scores, demonstrating adaptability to imbalanced medical datasets, and to real-world clinical scenarios. © 2025 The Author(s). Healthcare Technology Letters published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology.
HALE Lab NITK at TouchÃ© 2024: A Hybrid Approach for Identifying Political Ideology and Power in Multilingual Parliamentary Speeches
(CEUR-WS, 2024) Simhadri, S.; Patel, M.M.; Sowmya Kamath, S.
In this article, an approach to determine the political views and stances of speakers for identifying whether they support or oppose the government in parliamentary discussions is presented. The work was carried out as part of the TouchÃ© 2024 Task 2, â€œIdeology and Power Identification in Parliamentary Debatesâ€ . Towards this, two systems were developed, the first employs traditional machine learning methods with TF-IDF embeddings, while the second utilizes advanced NLP techniques with the LASER encoder for multilingual embeddings. Both systems incorporate standard preprocessing techniques and also integrates a variety of models, after which a voting classifier is used to combine the predictions from both approaches. Experiments revealed that this comprehensive framework effectively addresses the complexities and nuances of political discourse, providing valuable insights into speakers' ideologies and governing statuses within parliamentary debates. Â© 2024 Copyright for this paper by its authors.
ICU Patients’ Pattern Recognition and Correlation Identification of Vital Parameters Using Optimized Machine Learning Models
(J.J. Strossmayer University of Osijek , Faculty of Electrical Engineering, Computer Science and Information Technology, 2023) Yallabandi, G.; Jeganathan, J.; Mayya, V.; Sowmya Kamath, S.
– Early detection of patient deterioration in the Intensive Care Unit (ICU) can play a crucial role in improving patient outcomes. Conventional severity scales currently used to predict patient deterioration are based on a number of factors, the majority of which consist of multiple investigations. Recent advancements in machine learning (ML) within the healthcare domain offer the potential to alleviate the burden of continuous patient monitoring. In this study, we propose an optimized ML model designed to leverage variations in vital signs observed during the final 24 hours of an ICU stay for outcome predictions. Further, we elucidate the relative contributions of distinct vital parameters to these outcomes The dataset compiled in real-time encompasses six pivotal vital parameters: systolic (0) and diastolic (1) blood pressure, pulse rate (2), respiratory rate (3), oxygen saturation (SpO2) (4), and temperature (5). Of these vital parameters, systolic blood pressure emerges as the most significant predictor associated with mortality prediction. Using a fivefold cross-validation method, several ML classifiers are used to categorize the last 24 hours of time series data after ICU admission into three groups: recovery, death, and intubation. Notably, the optimized Gradient Boosting classifier exhibited the highest performance in detecting mortality, achieving an area under the receiver-operator curve (AUC) of 0.95. Through the integration of electronic health records with this ML software, there is the promise of early notifications regarding adverse outcomes, potentially several hours before the onset of hemodynamic instability. © 2023, J.J. Strossmayer University of Osijek, Faculty of Electrical Engineering, Computer Science and Information Technology. All rights reserved.
NBA MVP Prediction and Historical Analysis Using Cross-Era Comparison Approaches
(Institute of Electrical and Electronics Engineers Inc., 2024) Godbole, I.; Murali, S.S.; Sowmya Kamath, S.
In order to understand the crucial player statistics that decide the Most Valuable Player (MVP) Trophy, this research study dives into a substantial 32-year dataset of the National Basketball Association (NBA). We build a predictive framework trained on historical player statistics and MVP voting results using a sophisticated ensemble of machine learning models, including Support Vector Machines (SVM), ElasticNet, AdaBoost, Random Forest and Back-propagation Neural Network (BP). We determine the key elements influencing this renowned award by evaluating connections between player stats and MVP picks. Our research provides insights into the MVP selection process by utilising the models' ability to capture complex patterns and nonlinear interactions, providing stakeholders with a reliable tool for assessing player performances.This work advances the discourse surrounding the NBA MVP Trophy and enriches our comprehension of player value assessment. Also, the prediction models are used to conduct various historical analysis experiments, by finding an objective method to compare performances of players from different eras. Â© 2024 IEEE.
Ocular Region Segmentation Model for Diagnosis of Microbial Keratitis Using Slit-Lamp Photography
(Institute of Electrical and Electronics Engineers Inc., 2023) Supreetha, R.; Sowmya Kamath, S.; Mayya, V.
Corneal disease, a prevalent cause of global blindness, can lead to severe complications such as Microbial Keratitis, an inflammatory condition of the cornea often caused by bacterial or fungal infections. Early detection and timely treatment are crucial to prevent vision loss associated with this condition. Slit-lamp photography, a standard tool for ocular examination, is commonly employed for diagnosis. To address the growing demand for ophthalmology specialists, numerous studies have explored the use of Deep Learning (DL) algorithms to achieve precise and accurate segmentation of ocular structures, including the cornea, from slit-lamp photography images. In this study, an ocular region segmentation model trained on heterogeneous slit-lamp image datasets for improving learning performance is presented. Various data augmentation strategies are experimented with, and optimization techniques are incorporated. Experiments revealed that the model outperformed several state-of-the-art works concerning Dice score. Furthermore, the model can also be utilized for the unsupervised learning task of mask generation, as the segmentation findings are on par with the ground truth. Â© 2023 IEEE.
Predicting Air Quality Index with Recurrent Neural Networks and Meta-heuristic Algorithms
(Institute of Electrical and Electronics Engineers Inc., 2024) Jayanth, P.; Sowmya Kamath, S.
Millions of people worldwide suffer from the impacts of air pollution, a significant health risk. The metric Air Quality Index (AQI) serves as a crucial tool, providing valuable insights into current air quality conditions and potential health risks. This study utilizes two datasets: one from Wuhan City and the other from Shanghai. The features utilized for forecasting the AQI include PM2.5, PM10, SO2, NO2, O3, CO, l-temp, h-temp, temp, wet, wind, Hecto-pascal Pressure Unit (hpa), visibility, precipitation, and cloud content. This work focuses on developing models to predict AQI for a given data by comparing Long Short Term Memory (LSTM) and its variants, including Bidirectional LSTM (BiLSTM), Stacked LSTM, and Gated Recurrent Unit (GRU) models. Additionally, Particle Swarm Optimization is utilized as an evolutionary feature selection method. Â© 2024 IEEE.
Predictive Model for Enhancing Water Quality Monitoring leveraging Satellite Data
(Institute of Electrical and Electronics Engineers Inc., 2024) Prakash, P.; Sowmya Kamath, S.; Bhattacharjee, S.; Umesh, P.; Gangadharan, K.V.
Remote sensing data can be used instead of conventional methods to collect image data from multiple satellites with acceptable spatial and temporal coverage. The proposed study makes use of Landsat 8 Operational Land Imager (OLI) data. The relationship between reflectance retrieved from Landsat 8 OLI data and in-situ data is established through the application of machine learning model. The dataset is made up of Landsat8 band extractions for water quality features. Water with high turbidity is predicted and verified using in-situ data that was gathered within the chosen temporal and spatial limits. Â© 2024 IEEE.
RIVER: A Bio Inspired Routing Protocol For High Data Rate Wireless Sensor Networks
(IEEE Computer Society, 2024) Geetha, V.; Malik, P.; Sowmya Kamath, S.
Wireless sensor networks are deployed to observe the environment such as disaster management facilities and some industrial applications to observe temperature, pressure and humidity etc. The network is dynamic, battery operated and as a result it phases many issues compared to other networks like MANETs. Compared to MANETs, the issue or challenges are more as the nodes in the network has to send data to a common node called sink node. The challenges increases if the data type is other than scalar data. In future WSN gets extended to high data rate applications to sense and transmit multimedia, image or audio files. Normal routing algorithms, which are currently suitable for low data rate WSNs (LWSNs) are not sufficient for HWSNs. As a result, a new type of routing protocol which can solve various issues of HWSN is essential. It has been found that bio inspired algorithms perform very well for optimization and combinatorial problems. Bio-inspired algorithms are inspired from nature and nature has found very good solutions to very complex problems. In this regard, we are proposing and developing a new bio inspired routing protocol for high data rate wireless sensor networks called River which is based on the Intelligent Water Drops algorithm related to swarm intelligent techniques. The proposed method is implemented in ns-2, and the results are compared with AODV. The results shows on an average 75% improvement with respect to various parameters. Â© 2024 IEEE.