Modern mobile devices often struggle with limited computational capabilities, hindering their ability to efficiently process data-intensive applications such as augmented reality, mobile healthcare, and intelligent navigation. Multi-access Edge Computing (MEC) presents a viable solution by enabling the offloading of complex computational tasks to geographically proximate edge servers. This offloading approach alleviates the processing burden on user devices and significantly reduces end-to-end latency, thereby enabling real-time responsiveness. In this work, we propose GRU-OptiCom, a novel task offloading framework that leverages meta-reinforcement learning (MRL) to dynamically optimize offloading decisions across varying environments. The model incorporates a GRU-based sequence-to-sequence neural architecture for capturing task dependencies, and employs Proximal Policy Optimization (PPO) to ensure stable and efficient training. We evaluate GRU-OptiCom using latency as the primary performance metric and compare its performance with baseline methods such as MRLCO and HEFTbased greedy algorithms. Experimental results demonstrate that GRU-OptiCom consistently achieves lower latency and improved task distribution, setting a new benchmark for adaptive and intelligent task offloading in MEC environments.

Agriculture remains the backbone of India's economy, employing over two-thirds of the population and contributing approximately 19% to the national GDP. This paper introduces the Generative Agricultural Transformer (GAT), a novel transformer-based architecture tailored for the Indian agricultural domain. GAT powers Kisan-Mitra, a multilingual generative AI chatbot that delivers context-aware, region-specific guidance aligned with government schemes and agronomic best practices. Unlike generic language models such as BERT or GPT, GAT integrates domain-specific attention mechanisms, crosslingual embeddings for 12 regional languages, and real-time connectivity with government databases, including the Kisan Call Centre (KCC) and agricultural subsidy portals. As a result, the model is fine-tuned for tasks such as crop advisory, disease diagnosis, seasonal planning, fertilizer optimization, and market price forecasting. GAT achieves a response accuracy of 89.6% and covers over 85% of key agricultural topics, significantly outperforming baseline transformer models. The proposed system is scalable, fully trainable end-to-end, and effectively bridges the digital divide by providing accessible, accurate agricultural assistance to rural communities. This study presents the complete development pipeline of Kisan-Mitra, including dataset construction, architectural innovations, training strategies, and deployment approach. Results indicate that task-specific transformer architectures—when grounded in local context and institutional integration—can substantially enhance the impact and inclusivity of AI-driven agricultural services.

Our study focuses on detecting confusion from EEG data in higher education. Using BiLSTM models, we enhance EEG analysis efficiency and precision. Starting with extensive feature extraction and preprocessing, we improve data quality. We then apply diverse deep learning methods, including BiLSTM models, to predict perplexity-associated EEG signals. Unique data augmentation techniques like random noise injection and synthetic data synthesis bolster model resilience and generalization. Rigorously evaluating our approach with real-world EEG datasets, we achieve a significant accuracy improvement of 99.89%. The evaluation includes analyzing the confusion matrix and classification report, validating our methods' efficacy. Our research offers a comprehensive framework with state-of-the-art deep learning algorithms, particularly BiLSTM models, advancing EEG pattern identification and classification in educational contexts.

Congestive Heart Failure (CHF) is a prevalent cardiovascular disorder that requires early detection for effective management. This paper presents a comprehensive framework leveraging Bidirectional Long Short-Term Memory (BiLSTM) networks for the early detection of CHF using Electrocardiogram (ECG) data. The methodology involves preprocessing ECG signals, extracting relevant features, normalizing the data, and training a BiLSTM model. Our experimental results demonstrate high accuracy and robustness, highlighting the potential of BiLSTM networks in clinical applications for CHF prediction. The proposed method achieves an accuracy of 98.50%, sensitivity of 98.36%, and specificity of 98.57%, outperforming existing techniques. The model's ability to capture temporal dependencies in ECG signals through bidirectional learning contributes to its superior performance in identifying early signs of CHF. The findings suggest that BiLSTM networks hold promise for enhancing CHF detection accuracy, which could significantly impact healthcare outcomes by enabling early intervention and reducing medical costs.

Chronic Obstructive Pulmonary Disease is a progressive health condition characterized by restricted air-flow in the lungs, leading to life-threatening risks including loss of life. Detecting and diagnosing COPD early can significantly enhance disease management and patient well-being. However, existing diagnostic approaches for COPD are costly, time-intensive, and demand specialized equipment. Consequently, there is a pressing demand for an automated, accessible, and cost-effective diagnostic tool for diagnosing COPD. This work proposes a novel TriSpectraLSTM model that listens to the melody of lung sounds through multiple audio features, such as: Mel-frequency cepstral coefficients, chromagram, and Mel spectrograms. Each sub-model analyzes its assigned instrument, extracting distinct sonic signatures, before the hybrid network merges them, revealing the hidden COPD features. The model achieves a remarkable 93% accuracy on test data, surpassing existing methods. This multimodal approach by capturing the entire range of lung sounds, it produces a more precise and detailed diagnosis. This work employed the power of machine learning with healthcare, paving the way for a future where lung sounds become the key to diagnosing COPD diseases.

Diabetic Retinopathy (DR) is a serious eye condition that impacts individuals with diabetes, leading to retinal damage and potentially resulting in gradual vision loss. In this paper, we proposed a novel framework for diabetic retinopathy detection using deep learning techniques. The proposed framework employs an ensemble learning model named WENN-4 to classify fundus images into five classes and measure the severity of blindness of individuals. The proposed framework consists of following steps: customized preprocessing method that includes Gaussian Blurring and augmentation techniques, subsequently followed by inputting the processed images into diverse CNN architectures to extract discriminatory feature vectors for classification and early diagnoses and categorizes blindness severity levels, including no-DR, mild DR, severe, moderate, or Proliferative Diabetic Retinopathy (PDR). To enhance the overall performance of our proposed framework, we utilized an ensemble of models, WENN-4. This ensemble comprises ResNet-50, DenseNet-121, InceptionV3, and EfficientNetV2. This model achieves a remarkable accuracy of 93.7% accuracy on test data of APTOS 2019, surpassing existing methods. This ensemble of models lead to increased accuracy by leveraging complementary strengths, correcting individual misclassifications, and ultimately producing superior outcomes.