Orlando Belo

Introduction: The synergistic relationship between Artificial Intelligence and Medicine has the potential to revolutionize our understanding of how genomics, transcriptomics, and other omics influence susceptibility to specific diseases. Finding the most appropriate biomarkers from omics is still a challenge for most diseases, but it can bring tremendous benefits for disease diagnosis, prognosis, and treatment. Methods: This scoping review provides a comprehensive analysis of all existing research on omics biomarkers for skin conditions discovered by machine learning algorithms, following the Joanna Briggs Institute methodology. Results: After screening, 73 papers met the eligibility criteria. These articles correspond to studies conducted mainly in China (49.3%) and the United States (20.5%), and most are relatively recent (2020–2022). The studies mainly focus on cancers and systemic lupus erythematosus, as well as the discovery of diagnostic (49.3%) and prognostic biomarkers (45.2%). Only 11.0% of the publications explore predictive biomarkers. LASSO, random forest, and support vector machines are among the top three most frequently employed machine learning algorithms. Finally, transcriptomic data (80.8%) are the most commonly used type of omics, and HLA-DQB1 and IFI27 emerged as promising biomarkers. Discussion: The findings highlight the growing number of machine learning-based omics studies in dermatology. However, key gaps remain, including geographic concentration and limited exploration of non-cancer dermatologic conditions. Conclusions: This scoping review enabled the identification and comparison of potential biomarkers associated with skin diseases discovered through machine learning techniques. Despite their promising role in disease management and advancing precision medicine, such studies remain scarce in dermatology.

The evaluation of bank branch performance remains a critical strategic imperative, particularly in an era characterized by the coexistence of physical branches and digital transformation. Traditional performance evaluation methods, such as Data Envelopment Analysis (DEA), often rely on static, two-dimensional matrix representations that fail to capture the inherent multi-way and temporal heterogeneity of branch operations. Furthermore, while advanced machine learning techniques offer predictive power, their black-box nature limits their adoption in highly regulated banking environments where explainability is paramount. This study proposes a novel, integrated analytical framework that combines DEA, Non-negative Tucker Tensor Decomposition, K-Means clustering, Extreme Gradient Boosting (XGBoost), and SHapley Additive exPlanations (SHAP) to uncover latent performance factors and provide explainable managerial intelligence. Utilizing a comprehensive dataset of 563 bank branches observed over 12 months (6,756 observations), the proposed framework constructs a third-order tensor (Branch × Feature × Time) to preserve multidimensional interactions. Non-negative Tucker decomposition extracts latent factors, which are subsequently clustered into seven distinct branch archetypes using K-Means. XGBoost is employed to model cluster membership, and SHAP is utilized to unpack the black box, revealing the specific financial and operational drivers influencing branch behavior. The results reveal distinct archetypes, ranging from “Elite Branches” to “Critical Underperformers,” and identify Foreign Exchange Earnings and Low-Cost Deposits as the most critical drivers of branch archetype membership. This study contributes to the literature by introducing a unified Tensor Mining-XAI framework, offering a mathematically rigorous and interpretable decision support system for bank branch optimization.