Mid Sweden University

The disparity in the data from intensive care units, where stroke victims and heart attack patients make up a minority, makes this effort extremely difficult. A well-known difficulty in data mining is handling unbalanced data. The main contribution of this work is a method that accurately identifies and categorises minority-class data, even in highly imbalanced datasets with small class sizes. This work predicts stroke from the balanced and compressed data from MIMIC III dataset. The Convolutional Neural Network-Gated Recurrent Unit with Imbalanced Data Handling (CNN-GRU-IDH) is proposed. Additionally, it reduces the amount of data transferred by compressing healthcare data using the Lempel Ziv Markov Chain Algorithm (LZMA). Class imbalance problems are addressed with the Synthetic Minority Over-sampling Technique (SMOTE). Notably, this study adds a novel element by employing the Improved Multi-Objective Wolf Pack Algorithm (IMOWPA) to choose the appropriate K nearest neighbour value for SMOTE. The suggested model surpasses existing models when used on the dataset, obtaining a remarkable accuracy rate of 87.66% and 85.63% of F1 score for 70% of training and 30% of testing data. The CNN-GRU-IDH approach, which tries to forecast the incidence of strokes, is used as the major data classification technique. This study makes a substantial advancement to improving patient-specific early stroke prediction, which might save lives and lower death rates. 

Early and precise recognition of developmental disabilities (DDs) is essential for the timely intervention of preschool-age children. Although various predictive models have been developed in the past, they cannot provide accurate detections and face challenges like high false prediction, limited generalization, etc. To address these issues, a novel artificial intelligence (AI)-based algorithm was proposed for accurate DD detection. The proposed system integrates digital biomarkers from speech and behavioral domains for predicting DDs. Consequently, an Enhanced Gaussian-Based Noise Filtering (EGNF) is applied to speech data, and Z-score normalization is applied to behavioral data for preprocessing, which makes the raw data reliable for subsequent analysis. Further, feature extraction is performed using Convolutional-Bidirectional Long Short-Term Memory (Conv-BiLSTM) for speech data and Graph Attention Network (GAT) for behavioral data. Subsequently, a Hybrid Feature Selection (HFS) method was developed by combining the Harmony Search Algorithm (HSA) and ReliefF ranking for selecting informative and relevant attributes for model training. Finally, a hybrid classification model named DeepSTNet has been developed by integrating MnasNet and Independently Recurrent Neural Networks (IndRNN), which allows the capture of spatial and temporal long-term dependencies within the data, resulting in high detection performances. The proposed framework was implemented in Python and validated using comprehensive data formed by combining the Dysarthria Detection dataset and the Autism Dataset for Toddlers. The execution outcomes highlighted that the developed framework achieved an accuracy of 99.46%, a precision of 97.74%, and a recall of 98.05%, illustrating that this AI-centric system can be equipped by end users, including healthcare professionals and caregivers, with tools for early screening and detection of DDs. 

The high prevalence of eye diseases these days is a critical health problem. With everyday use of digital products, the diseases are becoming increasingly common, and underlines the urgent demand for early diagnosis and timely treatment. In that context, multi-modality image fusion has recently attracted much interest for automated detection of ocular disorders like glaucoma, cataracts, diabetic retinopathy (DR), high myopia and macular degeneration. An innovative and fully automated deep learning framework, H-CapsNet, has been designed in this study for robust yet accurate eye disease classification. The suggested framework consists of a structured pipeline that involves preprocessing, characteristic extraction, attribute selection and categorization. For instance, a combined deep neural network architecture called H-Net and an Attention Block is used to initially extract a wide range of distinguishing features from the retinal images. A SENet Block further enhances the robustness of these representations so that multi-modality retinal image features can be fused into a unified description. The new Hybrid RemoPel algorithm is then applied for an optimum feature selection, whereby the most informative attributes are preserved. Both starts and ends are different. Finally, the most advanced clustering-based binary grey wolf optimizer (KCBGWO) soft capsule model through KCBGWO optimization process is utilized for final classification. The appropriate means of categorizing different kinds of eye diseases, the suggested system can help improve the accuracy of diagnosis and drive forward a new trend in ophthalmic disease detection using deep learning.

The rapid advancement of technology and the exponential increase in data volumes have emphasized the critical significance of effective feature selection (FS) methods in data mining. FS aims to eliminate redundant and irrelevant features from datasets while maintaining or improving the accuracy of classification models. All past studies on the FS problem have tried to improve parameters such as classification accuracy and the size of chosen feature subsets. Despite their success, there is still opportunity to improve these parameters further by selecting newer and more efficient algorithms. In this context, this paper proposes an innovative approach to FS utilizing the Binary Human Mental Search (BHMS), a novel algorithm that has not previously been used in FS. The method introduced in this paper adapts the HMS algorithm into a binary framework specifically tailored to identify optimal feature subsets for classification tasks. The K-Nearest Neighbors (KNN) classifier is also employed as the classification technique in data mining. The experimental evaluations are conducted on standard benchmark datasets from the UCI Dataset Collection. The performance of the proposed BHMS method is compared against nine advanced metaheuristic methods. Results indicate that the BHMS variant achieves competitive performance, demonstrating its effectiveness in feature selection for classification tasks. Parameters such as the average classification accuracy, the average size of the chosen feature subsets, the average fitness, and the value of the Wilcoxon test has improved significantly compared to comparative algorithms. In summary, this paper emphasizes the critical necessity for effective feature selection methods in the age of big data and technological progress. The proposed BHMS algorithm presents a promising approach to identifying optimal feature subsets, contributing to improved classification accuracy in data mining applications. 

Accurate forecasting of real estate taxes is essential for property owners and government agencies. It affects financial strategies and public revenue generation. Conventional deep learning approaches for real estate tax often struggle to select appropriate features and are highly sensitive to changes in hyperparameters. To address these challenges, this paper proposes a real estate tax model that uses reinforcement learning to guide feature selection. The reinforcement learning component constantly updates the selection of important features based on data interactions. It helps focus on the most relevant features and prevents overfitting. The reinforcement learning model incorporates a scope loss function that balances learning from current data with exploring new patterns. This balance helps preserve the accuracy and generalizability of the model. Additionally, the artificial bee colony algorithm enhances the flexibility and efficiency of the model by optimizing hyperparameter settings. The model was evaluated using several real estate datasets from Kaggle. These datasets cover Bucharest, California, Helsinki, King County, and Saudi Arabia. The results demonstrate that the proposed model outperforms traditional prediction tools. Across all datasets, it achieves mean absolute percentage error values ranging from 0.216 to 1.081. This strong performance demonstrates the potential of the model for use across various markets. It can improve economic assessments and planning. Code is publicly available at https://github.com/Mahyar-ramezankhani/Tax.

Recognising the growth and transformation of plants over time is of great significance for both agriculture and environmental preservation. It aids stakeholders in efficiently utilising resources and providing optimal growth conditions. This paper investigates the automated Plant Growth Temporal Labelling (PGTempLab) problem, which involves plant growth classification based on visual imagery—specifically focusing on pine seedlings, a key forest species in Scandinavia. A central challenge of PGTempLab lies in the Class-Based Variability (CBV) Paradox, where significant intra-class variability and inter-class similarity occur due to temporal categorisation, complicating traditional classification tasks. To address this issue, a transfer learning approach is adopted using 12 well-known convolutional neural networks (CNNs) as pre-trained models. Additionally, three novel evaluation criteria are introduced, incorporating temporal aspects: time-dependent accuracy (TDA), temporal boundary deviation (TBD), and a temporal confusion matrix. Experiments, conducted on a custom dataset collected specifically for this study, consisting of over 38,000 time-lapse images of pine seedlings captured from two camera perspectives over 44 days, demonstrate that MobileNet achieves the highest performance in the top-view setting with 86.85% TDA, while SqueezeNetV2 performs best in the angle-view setting with 86.94% TDA. These results highlight the effectiveness of deep transfer learning in addressing the CBV paradox in plant growth classification. 

Detecting advertisements in digitized newspapers is a key step in large-scale media analytics and digital archiving. However, variations in layout, typography, and advertisement design across publishers and time periods cause significant domain shifts that reduce the generalization ability of supervised detectors. This paper presents AdAPT, a confidence-guided pseudo-labeling pipeline for unsupervised domain adaptation in advertisement detection. The proposed method leverages both advertisement-free (Null) and advertisement-containing pages from unlabeled target domains to generate reliable pseudo-labels. By retraining a YOLO-based detector using labeled source data combined with filtered pseudo-labeled target samples, AdAPT achieves robust adaptation without requiring manual annotation. Experiments conducted on two unseen newspapers (Adresseavisen and iTromsø) demonstrate that Null-based pseudo-labeling provides the most stable and accurate adaptation, yielding up to 38% error reduction compared to the baseline. The results highlight AdAPT as a simple, scalable, and annotation-efficient solution for maintaining high-performance advertisement detection across diverse newspaper collections. 

Purpose: Artificial Intelligence (AI) is increasingly recognized for its potential in improving the detection, classification, prediction, and segmentation of colon cancer. Yet, the reliability of these applications depends on the quality and completeness of the underlying studies. This systematic review evaluates the current state of AI applications in colon cancer research, focusing on their impact on diagnostic accuracy, treatment planning, and patient outcomes. Methods: A comprehensive search was conducted in PubMed, Scopus, and Web of Science for articles published between 2020 and 2024. The quality of the included studies was assessed using standardized criteria. A meta-analysis was performed where applicable, and a subgroup analysis was conducted based on the type of AI technology (e.g., deep learning, machine learning) and its application (detection, classification, etc.). Additionally, we recorded whether each study incorporated Explainable AI (XAI) techniques or Generative AI (e.g., GANs) as part of its methodology. Results: In 80 articles, AI models showed significant improvements in diagnostic accuracy, particularly in polyp detection during colonoscopies and histopathological analysis. Deep learning approaches often outperformed traditional methods. However, clinical integration remains challenging due to data and validation gaps. Conclusion: AI holds great promise in colon cancer diagnosis and treatment. Future work should focus on integrating AI tools into clinical workflows through explainable models and standardized validation. 

Effective intervention and detection of developmental disabilities (DDs) should be carried out early and accurately. This paper discusses a novel AI-based approach for detecting DDs precisely, by incorporating both speech and behavioral data in a unique manner. The new architecture comprises several key stages, with innovations attributed to the system: Enhanced Gaussian-Based Noise Filtering (EGNF) is used to enhance speech; Conv-BiLSTM and Transformer Embeddings for performing strong feature extraction from speech and behavior; Harmony-ReliefF Optimization (HRO) for feature selection, and BioNeuroFusionNet (BNFN) as a novel classification network integrating IndRNN and MnasNet, with GNA selected as the class superiors. The Harmony-ReliefF Optimization (HRO) strategy combines the exploration capabilities of the Harmony Search Algorithm (HSA) with the feature ranking ability of ReliefF for optimizing the feature space. The features selected and optimally extracted are then fed into the novel BioNeuroFusionNet (BNFN) intended for final classification. The BNFN is another significant contribution of this work, where Independently Recurrent Neural Networks (IndRNN) are incorporated for fast processing of sequential data set and MnasNet (MN) for extracting high-level abstract features. The experimental result indicates that this integrated architecture works effectively with high accuracy (0.98942 for voice; 0.99632 for behavior) and very high scores in precision, recall, specificity, and other relevant metrics. Indeed, these results indicate that the system proposed in this work represents a considerable improvement in the area of early diagnosis and intervention of developmental disorders. 

Machine learning techniques are utilized for early detection of diseases, which can significantly enhance probabilities of positive treatment and existence. The traditional machine learning algorithms may be unable to predict outcomes with sufficient accuracy. In this work, an Effective ECOLASSO with Black Widow Optimization for Feature Selection and Stagewise Adaptive Learning Rate (ELBWOSALR) classifier is proposed for feature selection and prediction. The proposed work comprises two phases, in the first phase, Ecological similarity Least Absolute Shrinkage and Selection Operator (ECOLASSO) model is utilized to predict the best features from the dataset by removing the feature with smallest absolute regression coefficient from the feature set. A Black Widow Optimizer (BWO) is used to choose the subset of optimal features and to reduce local optima. In the second phase, Stagewise Adaptive Learning Rate (SALR) involves combining several weak learner classifiers into a strong ensemble classifier by adaptive learning rate. The key contribution of this work is the integration of ECOLASSO model with BWO for robust feature selection, combined with a SALR classifier. This hybridization addresses two critical challenges simultaneously: (i) ECOLASSO ensures sparsity and ecological similarity-driven selection of relevant features, while (ii) BWO prevents premature convergence and enhances global search efficiency. By coupling these with SALR, our model achieves superior accuracy and generalization compared to conventional classifiers. Lung cancer, breast cancer and heart disease datasets are used for experimentation. The ELBWOSALR classifier is compared with various classifier models such as Support Vector Classifier, Decision Tree Classifier, Random Forest Classifier, Logistic Regression, Extreme Gradient Boost Classifier, Gradient Boosting Classifier, K-Nearest Neighbors Classifier and CatBoost Classifier and the results are observed. The proposed ELBWOSALR classifier achieves accuracies of 98%, 97% and 91% with AUC values of 92%, 99% and 94% for lung cancer, breast cancer and heart disease datasets respectively.