Delay prediction in urban public transport systems is a critical task for improving operational efficiency and service reliability. While numerous predictive models exist, understanding the relative importance of contributing factors remains a challenge, with traditional approaches often overestimating the impact of stochastic weather conditions. This study proposes an approach that combines predictive modelling and factor analysis based on interpretable machine learning. An eXtreme Gradient Boosting model was developed using a large dataset of operational and meteorological data from a city with approximately one million inhabitants. The model demonstrated high predictive accuracy, explaining 72% of the variance in delays (Coefficient of Determination R²=0.72). Analysis of the model’s feature importance revealed that operational cycles (seasonal, weekly, daily) and spatial context (routes, stops) are the dominant predictors, collectively accounting for over 52% of the model’s total feature importance. Contrary to common assumptions, weather conditions were identified as a powerful secondary, rather than primary, factor. While their cumulative feature importance was substantial (contributing nearly 45%), the model revealed their impact to be highly contextual: the negative effects of adverse weather were significantly amplified during predictable peak operational hours but were minimal otherwise. This research demonstrates how Explainable Artificial Intelligence methods can transform complex predictive models into practical tools, providing a data-driven basis for shifting from reactive management to proactive, evidence-based planning.

In Artificial Intelligence, voice categorization is important for various applications. Tamil, being one of the oldest languages in the world, comprises rich regional slang differing in tone, pronunciation, and emotive expression. These slang words are difficult to categorize because they are informal and there is limited annotated audio data. This study proposes an enhanced deep learning framework for Tamil slang classification using a balanced audio corpus. The framework integrates data-specific pre-processing techniques, including Mel spectrograms, Chroma features and spectral contrast, to capture the nuanced characteristics of Tamil speech. A DenseNet backbone, combined with LSTM and GRU layers, models both temporal and spectral information. The suggested FRAE-PSA module is an innovative application of the Pyramid Split Attention (PSA) mechanism adapted to support regional and affective variations of speech. Different from current PSA or Transformer-based approaches, FRAE-PSA splits the audio frequency spectrum and adapts attention weights dynamically based on auxiliary tasks. A multi-branch architecture is employed to fuse temporal and spectral features effectively and multi-task learning is used to enhance regional accent and emotion detection. Custom loss functions and lightweight networks optimize model efficiency. Experimental results show up to a 15% improvement in classification accuracy over baseline models, demonstrating the framework's effectiveness for real-world Tamil slang classification tasks.

The deep learning models are being used in the agricultural sector, ushering in a new age of decision support for crop pest and disease control. In light of the difficulties faced by farmers, this literature study seeks to identify the critical components of deep learning models that affect decision support. All the way from model design to data input to training approaches and their effects on efficient decision-making. Examining the deep learning model factors influencing decision support in crop diseases and pest management was the primary goal. The researcher looked at articles and journals published by IEEE access, ACM, Springer, Google Scholar, Wiley online library, Taylor and Francis, and Springer from 2014 to 2024. From the search results, sixty-three publications were selected according to their titles. The paper provides a synopsis of deep learning models used for crop health management based on a careful evaluation of scholarly literature. In order to shed light on the merits and shortcomings of different models, the article conducts a thorough literature review and literature synthesis. Future studies might be guided by the identification of methodological issues and gaps in present research. Applying deep learning to the problem of agricultural diseases and pest control has real-world consequences, as shown by several case studies and applications. Insightful for academics, practitioners, and legislators in the field of precision agriculture, this extensive study adds to our knowledge of the complex relationship between elements in deep learning models and their impact on decision support.

Stress and anxiety are some of the most public mental health illnesses that people in the current society face. It is important to determine these conditions early to be able to effectively promote the well-being of individuals. This research work presents the possibility of identifying stress and anxiety through social media (SM) data and an anonymous survey, by machine learning (ML) and natural language processing (NLP). The paper starts with data collection, using the DASS-21 questionnaire and a sample of tweets obtained from Twitter users from India, aimed at determining which language is associated with stress and anxiety. The gathered data is pre-processed in some of the steps, such as URL removal, lower casing, punctuation removal, stop words removal, and lemmatization. After data preprocessing, the textual content is transformed into numerical form through Word2Vec to facilitate pattern analysis. To enrich the analysis of the main topics in the dataset, the Latent Dirichlet Allocation (LDA) and the Non-Negative Matrix Factorization (NMF) techniques are applied. For the classification, the work uses ML algorithms such as Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM) networks. Lastly, the project involves an application created with Streamlit to allow the user to interact with the model. 

Cardiovascular Disease (CVD) is a hazardous condition for humans that is rapidly expanding around the globe in developed as well as developing nations, eventually resulting in death. In this disease, the heart often fails to supply sufficient oxygen to other parts of the body so that they are cannot able to perform their normal activities. It is critical to identify this problem immediately and precisely in order to save patients lives and avoid additional damage.Henceforth, this work proposes an efficient image processing strategy based on a hybrid algorithm optimised Convolutional Neural Network (CNN) classifier, which is used in this present research for precise identification of cardiac vascular disease.In the beginning, the Electrocardiogram (ECG) images are obtained and processed by removing noise using an adaptive median filter. The pre-processed ECG image is then divided into different regions using the Fuzzy C-Means (FCM) algorithm, which improves the accuracy of heart illness detection.Following segmentation, the Grey level Co-occurrence Matrix (GLCM) is employed to efficiently extract high-ranked features. Subsequently, the characteristics are considerably identified using a novel hybrid Crow Search Optimization (CSO) Dragon Fly Optimisation (DFO) algorithm-based CNN classifier for optimally categorising the cardiacvascular illness.The entire work is validated in Python software, and the results show that the novel method produces the best possible outcomes with a maximum precision of 98.12%.

Safe and energy-aware navigation for unmanned aerial vehicles (UAVs) requires the simultaneous optimization of path length, curvature, obstacle clearance, altitude, energy expenditure, and mission time—within the tight computational limits of on-board processors. This study proposes a two-phase hybrid optimizer that couples the global search capability of Differential Evolution (DE) with an Enhanced Whale Optimization Algorithm (E-WOA) specialized for local refinement. E-WOA improves on the canonical WOA through three principled modifications: real-time boundary repair to ensure path feasibility, quasi-oppositional learning to restore population diversity, and an adaptive stagnation trigger that re-initiates exploration when progress stalls. When the population’s improvement plateaus, control transfers from DE to E-WOA, combining broad exploration with focused exploitation. Comparative experiments conducted in 3D environments with static obstacles that block direct line-of-sight routes demonstrate that the hybrid achieves lower composite cost—normalized over path length, curvature, risk, altitude, energy and time—shorter and smoother trajectories, and faster convergence than standard metaheuristics while preserving obstacle clearances and curvature limits. Averaged over 30 independent trials, our hybrid framework reduced the normalized composite cost by 14.5% relative to the next-best algorithm (Grey Wolf Optimizer) and produced feasible paths in an average of 2.35 seconds on commodity hardware—adequate for strategic re-planning, though further optimization is needed for sub-second control loops. Blending DE’s global reach with a diversity-aware, adaptively stalled WOA provides a practical foundation for strategic, near-real-time replanning in 3D airspaces.

Pancreatic cancer, characterized by its high mortality rate and scarce treatment options, poses a formidable challenge in the field of oncology. Now, we live in a reality that requires immediate progress in diagnostic and prognostic methodologies to find pancreatic cancer early and understand its stage. This study deals with the pressing requirement for better diagnostic tools by evaluating and deciding the suitable machine learning (ML) algorithms for detecting pancreatic cancer at an early stage. This work uses a publicly available dataset with 590 urine samples which included control, benign hepatobiliary disease as well as Pancreatic Ductal Adenocarcinoma (PDAC) samples. The primary objectives of the research included developing a predictive model based on clinical data, examining various machine learning (ML) algorithms for their diagnostic precision, and improving the early detection rates for pancreatic cancer. The study assessed the efficacy of a broad array of ML algorithms in forecasting outcomes associated with pancreatic cancer. This analysis systematically explored Random Forest, Support Vector Machine, Decision Trees, K-Nearest Neighbours, XGBoost, ADABoost, CatBoost, and GradientBoost. The assessment focused on standard performance metrics such as accuracy, precision (also known as positive predicted value or PPV), recall (sometimes called sensitivity or true positive rate), F1-score, and support. Notably, CatBoost achieved the highest accuracy of 75%, outperforming other models such as Random Forest (74%) and XGBoost (74%), demonstrating its superior classification performance in distinguishing between pancreatic cancer, benign conditions, and non-cancerous cases. In addition to performance evaluation, this study integrates SHAP (Shapley Additive Explanations) analysis to enhance model interpretability, ensuring transparency in feature contributions. SHAP analysis revealed that Plasma CA19-9, LYVE1, and TFF1 were the most influential biomarkers across all classifications, reinforcing their diagnostic significance. This research emphasizes the critical importance of early detection, model interpretability, and clinical applicability, demonstrating that ML algorithms, particularly CatBoost, not only enhance diagnostic precision but also provide explainable predictions that support real-world medical decision-making.

Machine learning model training, which ultimately optimizes a model’s cost function is usually a time- consuming and computationally intensive process on classical computers. This has been more intense due to the in- creased demand for large-scale data analysis, requiring unconventional computing paradigms like quantum computing to enhance training efficiency. Adiabatic quantum computers have excelled at solving optimization problems, which require the quadratic unconstrained binary optimization (QUBO) format of the problem of interest. In this study, the squared error minimization in the multiple linear regression model is reformulated as a QUBO problem enabling it to be solved using D-wave adiabatic quantum computers. Same formulation was used to obtain a solution using gate-based algorithms such as quantum approximate optimization algorithm (QAOA) and sampling variational quantum eigensolver (VQE) im- plemented via IBM Qiskit. The results obtained through these approaches in the context of runtime and mean squared error(MSE) were analyzed and compared to the classical approaches. Our experimental results indicate a runtime ad- vantage in the D-wave annealing approach over the classical Scikit learn regression approach. The time advantage can be observed when N>524288 compared to Sklearn Linear Regression and when N>65536  compared to Sklearn SGDRegressor. Support vector machine induced neural networks, where the margin-based entropy loss is converted into a QUBO with Lagrangian approach is also focused in this study concerning the applicability for nonlinear models.

According to anticipation, Diabetic Retinopathy (DR) is one among the most potential causes of visual disability and even blindness across the globe. Prompt diagnosis and treatment will stall the development towards irreversible damage. Hence, detection and staging of DR must play a huge role in early medical intervention and treatment planning. There exist tremendous challenges in accuracy, ability to capture retinal features, overfitting from poorly represented features, and inefficiencies in optimisation of parameters. These are likely to provide a very challenging situation for the clinical application of these methods, especially when dealing with huge heterogeneous datasets. The paper is discussing a new hybrid framework for detection and classification of DR that integrates the Swin-Transformer Integrated radial Residual Network (Swin-RadialNet) with a Butterfly-Mayfly Radial Optimizer (BMRO). This framework is intended to address those challenges. Hierarchical extraction of features from a model trained with radial residual connections atop the Swin Transformer architecture is carried out by Swin-RadialNet to guarantee the best learning of the complex k structures in retina on-the-fly through BMRO, which would act as an optimizer hybridizing to estimate radial spread parameters, thereby supporting acceleration of models' performance and converging rates. The core novelties of the proposed method shall lie in merging advanced transformer-based feature extraction with nature-inspired hybrid optimization to tackle efficiently critical issues regarding feature abstraction, parameter fine-tuning, and classification reliability. The method will be evaluated on several benchmark datasets such as Kaggle's Diabetic Retinopathy and APTOS 2019, and will expect to show state-of-the-art results for all DR stages in terms of accuracy, precision, recall, and F1-score.

Over the past two decades, the automatic recognition of human activities has been a prominent research field. This task becomes more challenging when dealing with multiple modalities, different activities, and various scenarios. Therefore, this paper addresses activity recognition task by fusion of two modalities such as RGB and depth maps. To achieve this, two distinct lightweight 3D Convolutional Neural Network (3DCNN) are employed to extract space time features from both RGB and depth sequences separately. Subsequently, a Bidirectional LSTM (Bi-LSTM) network is trained using the extracted spatial temporal features, generating activity score corresponding to each sequence in both RGB and depth maps. Then, a decision level fusion is applied to combine the score obtained in the previous step. The novelty of our proposed work is to introduce a lightweight 3DCNN feature extractor, designed to capture both spatial and temporal features form the RGBD video sequences. This improves overall efficiency while simultaneously reducing the computational complexity. Finally, the activities are recognized based the fusion scores. To assess the overall efficiency of our proposed lightweight-3DCNN and BiLSTM method, it is validated on the 3D benchmark dataset UTKinectAction3D, achieving an accuracy of 96.72%. The experimental findings confirm the effectiveness of the proposed representation over existing methods.