Sanjivani College of Engineering

This research aims to develop a comprehensive mental health prediction system based on Internet of Things (IoT) and deep learning methods. Specifically, this prediction system is especially sensitive to changing stress levels, to allow prompt intervention and personalized support, in response to the critical demand related to precise, convenient, and stigma-free mental health observation. The proposed system utilizes Interval type-2 Fuzzy Gradient recurrent mixed multimodal Convolutional Neural Network (IFG-CNN) that has three major modules. (i) Stress detector component performs a question–answer based evaluation with the help of Robustly Optimized BERT Approach (RoBERTa). (ii) and A facial emotion recognition module takes visual data and uses Scale-Invariant Feature Transform (SIFT) to extract features and then uses emotion classification using a Rotation-Invariant Surface Attention Radial basis function neural Network (RISAR-Net) trained by the white-faced capuchin optimizer. (iii) The classification module is a mental health status module that classifies them into highly stressed, moderately stressed and normal. The proposed model has a better prediction accuracy with accuracy of 97.85

Skin cancer has recently become the fifth most common cancer worldwide, burdening both the economy and global health. Industrialization, genetic modification, and the rapidly changing environment have all contributed to an increase in the incidence of skin cancer. To overcome this challenge, this research suggests employing Fuzzy Self-Guided Structure Convolution Retention Generative Adversarial Network with Greater Cane Rat Algorithm (FSGSCR-GAN-GCRA) approaches to improve the identification accuracy of skin cancer. Initially, dermoscopic images are collected from the ISIC 2017 and ISIC 2018 datasets for analysis. Before this, preprocessing is performed using the Dual Bilateral Least Squares Hybrid Filter (DBLSF), which helps protect edges and remove outliers effectively. This is followed by the Feedback DenseNet201 Network (FDNet201), which enhances feature extraction and accentuates clinically significant regions by jointly using feedback attention strategies with the dense connections of DenseNet201. The Fuzzy Self-Guided Structure Convolution Retention GAN (FSGSCR-GAN) then applies Fuzzy logic and convolutional transformers to the extracted features to retain the structural integrity of lesions, local textures, and global patterns simultaneously. To overcome structural information loss, feature under-representation, and training instability, the GAN enables structure-preserving adversarial learning; the Feedback attention-enhanced DenseNet201 is more effective at lesion-specific feature discrimination; and optimization with the GCRA ensures efficient convergence of hyperparameters. The FSGSCR-GAN-GCRA shows better skin cancer identification than the current methodology. The proposed model has excellent Accuracy (99.71

A decision-centric workflow was developed to link a QbD-defined design space with interpretable surrogate modeling for rapid selection of ultrasound-triggered doxorubicin-loaded chitosan microbubbles. A 15-run Box–Behnken design spanning chitosan, palmitic acid, and Pluronic F68 was used to quantify three critical quality attributes (CQAs)—mean size, encapsulation efficiency (EE), and burst release at 40 s (Burst40)—and to train response-specific surrogates evaluated by leave-one-out cross-validation (LOOCV). The DoE produced a broad performance envelope (size 2.35–4.85 µm; EE 60–82%; Burst40 55.6–95%) with clear composition-driven trade-offs, notably between EE and acoustic responsiveness. A regularized polynomial surrogate (PolyRidge) provided strong LOOCV performance for size (R²=0.885; RMSE = 0.264 µm) and EE (R²=0.871; RMSE = 2.18%), whereas Burst40 was only moderately predictable (R²=0.501; RMSE = 8.61%), consistent with threshold-like cavitation and microstructure sensitivity. Digital screening with multi-objective desirability and explainability (SHAP, permutation importance, partial dependence, LIME) shortlisted manufacturable candidates; experimental validation showed QbD/RSM was better calibrated for absolute size and EE, while both modeling approaches remained similarly limited for Burst40. Importantly, shortlisted formulations preserved suppressed baseline release and pronounced ultrasound-enhanced release, yielding trigger-dependent cytotoxicity and increased apoptotic commitment in MCF-7 cells.

Global agriculture is seriously threatened by plant diseases, which drastically reduce crop yields, economic productivity, and food security. Conventional methods of disease detection rely on manual inspections, which are time-consuming, error-prone, and often inconsistent under various field conditions. To overcome these limitations, the current study proposes an explainable hybrid CNN (Convolutional neural networks) method for precise plant disease identification and yield enhancement. The proposed architecture combines CNNs for dependable feature extraction with classifiers that make complex decisions using machine learning. A comprehensive preprocessing pipeline that incorporates noise reduction, augmentation, and normalization improves the model's predictive power in a variety of lighting scenarios, crop types, and occlusions. Explainable AI (XAI) methods, like Grad-CAM, demonstrate the important elements affecting classification results, improving interpretability and boosting user confidence. Even in the face of challenging lighting and background changes, experimental evaluations show superior efficacy over traditional models, achieving high F1-scores, recall, accuracy, and precision. Furthermore, by including yield-related insights that support data-driven agricultural decision-making, the methodology extends beyond simply identifying illnesses. A comprehensive strategy reduces crop losses, promotes sustainable farming methods, and allows for proactive disease management. For modern precision agriculture, the suggested system provides an open, scalable, and useful solution.

Gait recognition has emerged as a powerful biometric technique thanks to its capability to identify individuals from afar, eliminating the need for physical interaction or high-resolution imagery. However, the performance of gait recognition models largely depends on the quality and discriminative power of the features extracted from gait patterns. This paper presents a comprehensive study aimed at identifying the prominent features that most substantially enhance accurate gait recognition. Both traditional handcrafted features and modern deep-learning-based representations are examined across appearance-based, model-based, and spatio-temporal approaches. Through a systematic review and comparative analysis of state-of-the-art methods, this work highlights key gait attributes such as silhouette shape cues, joint–angle trajectories, limb motion dynamics, periodicity of gait cycles, and deep spatio-temporal embeddings. The findings reveal that robust gait recognition is typically achieved by combining multi-level features—capturing both structural and temporal characteristics—while ensuring invariance to covariates such as view angle, clothing, and walking speed. This study provides a consolidated understanding of the most effective feature categories and offers insights for developing next-generation gait recognition systems with improved accuracy and robustness.