Dementia is a progressive neurological condition characterized by cognitive decline and structural brain changes that evolve. Longitudinal modeling of these changes is important for improving disease monitoring, identifying progression patterns, and supporting early risk stratification. This study developed an explainable longitudinal machine-learning framework for dementia progression, using cognitive and Magnetic Resonance Imaging (MRI)-derived biomarkers from the Open Access Series of Imaging Studies (OASIS-2) longitudinal dataset. The dataset included 150 subjects and 373 repeated observations classified as Non-demented, Demented, or Converted. Current-visit features, previous-visit features, and slope-based temporal features were constructed from Mini-Mental State Examination, Clinical Dementia Rating, normalized whole-brain volume, estimated total intracranial volume, atlas scaling factor, Age, and MRI delay. Baseline models were compared with a longitudinal gradient-boosted model, using patient-level splitting to reduce data leakage across repeated visits. The proposed longiGradient Gradient boosting model achieved the best held-out test performance, with an accuracy of 88.16%, a macro F1-score of 0.776, and a weighted F1-score of 0.860. The model showed strong classification performance for Demented and Non-demented individuals, while converted cases remained more difficult to identify. A regularized gradient boosting model was also evaluated as an overfitting sensitivity analysis; although it reduced the perfect training fit, it did not improve held-out test performance. Feature importance, permutation importance, and SHapley Additive exPlanations identified Clinical Dementia Rating as the dominant predictor, with slope-based Clinical Dementia Rating providing additional longitudinal information. These findings suggest that combining cognitive measures, MRI-derived biomarkers, and temporal feature engineering can improve dementia progression modeling, although external validation in larger longitudinal cohorts is needed.

The Gauss-Markov Theorem is central to linear statistical inference, assuring that Ordinary Least Squares (OLS) is the Best Linear Unbiased Estimator (BLUE) under the classical assumptions. But textbook proofs typically involve two assumptions - column full rank of the design matrix and spherical error covariance - that are often violated in practice. In this paper, we provide a single unified proof encompassing both rank deficiency and non-spherical errors using linear algebra: orthogonal projections, the Moore-Penrose pseudoinverse, and the Loewner positive semi-definite matrix order. We sequentially address both of these conditions, culminating in a unified theorem allowing for rank deficient design matrices and arbitrary positive definite error covariances. Our proof takes the form of three lemmas: (1) OLS fitted values are the unique orthogonal projection onto the column space of the design matrix; (2) any alternative linear unbiased estimator is dominated by OLS in the positive semi-definite order; and (3) Generalized Least Squares is BLUE under non-spherical errors by way of a Cholesky decomposition argument. To bridge matrix theory and practical biostatistics, we apply this framework to baseline clinical data from the WHELD dementia study, demonstrating how the pseudoinverse resolves exact multicollinearity and why robust estimators are necessary for non-spherical intra-class correlations. This approach serves as a formal reference for researchers and an illustrative pedagogic tool.

Fragmentation of data is common in the U.S. healthcare system, which leads to substandard patient safety, excess administration waste, and impediments to public health monitoring. This paper proposes a relational database design, the Centralized-Decentralized Health Management System (CDHMS) that achieves a balance between conflicting requirements of local autonomy and federated access to data. The system is based on 15 normalized relations that are organized in 6 functional layers: core clinical infrastructure, Master Patient Index, Interoperability Mapping layer, Audit and Break-Glass logging subsystem, Patient consent and access-control framework, and a Role Based plus Attribute Based Access Control (RBAC+ABAC) model with 6 different user roles. The schema is deployed in MySQL Workbench 8.0 CE, with some sample mock data, and tested using 12 test queries. Results show the architecture enables no duplicate patient identities, reconciliation of incompatible coding vocabularies, granular patient consent management, and a tamper evident audit trail of all patient data access, including emergency overrides.

Chronic and acute wounds are a significant and increasing burden on health care systems worldwide. A key step in ensuring timely and evidence-based treatment planning and intervention is accurate, automated wound-type classification. Previous deep learning-based methods for wound classification have been mainly based on alone-employed convolutional image encoders or simple late-fusion methods combining image and location features. The approaches are not well explained and are not robust over a range of morphologically equivalent wound types. In this work, we introduce a multimodal, spatially-aware Graph Neural Network-Concept Bottleneck Model (GNN-CBM) framework. This architecture offers a combined reasoning about the appearance of the wound images and the context of the anatomical location, and provides an exposure of a bottleneck of latent clinical prototypes before final classification. We systematically compared the seven backbone architectures used in the visual encoding stage of the framework, such as CNNs (ResNet-50, EfficientNet-B0, EfficientNetV2-S, ConvNeXt-Tiny, DenseNet-121) and Vision Transformers (ViT-Base, Swin-Tiny) on the AZH Wound and Vascular Center dataset (Milwaukee, WI), which includes N = 311 unique patients and an augmented working set of 830 image-location pairs. In order to avoid data leakage, patient identifiers were used as grouping variable in a 5-fold stratified group cross-validation scheme, ensuring that patient images are not contained in more than one fold. Overall, Swin-Tiny outperforms the other backbones with the highest accuracy (0.809 ± 0.014), macro-F1 (0.805 ± 0.016) and mean AUC (0.954). Adjusted McNemar's testing (Bonferroni corrected across 6 comparisons) returned that Swin-Tiny performed significantly better than five of the six other backbones, except ConvNeXt-Tiny, which is not significant. An ablation study of the proposed Domain-Constrained Spatial Penalty Loss revealed that when the anatomical penalty term was removed, macro-F1 scores dropped from 0.805 to 0.792 for the best model, indicating the impact of the spatial-plausibility constraint. We discuss the mathematical formulation of our framework, the anatomical graph construction and its density statistics, the implications of unsupervised concept bottlenecks on real interpretability, hardware efficiency and highlight the importance of external, multi-center validation prior to clinical deployment.

Digital technologies have revolutionized the business environment and given Small and Medium-sized Enterprises (SMEs) in Africa new opportunities and challenges. The study explores approaches for African SMEs to harness digital ecosystems for revenue generation, cost savings, and business resilience. The study synthesises findings from a systematic narrative review of 54 peer-reviewed academic sources and 12 institutional reports, to examine how digital ecosystems evolve, their global manifestations and their unique characteristics in Africa. Key themes covered include the mobile-technology as the gateway to digital space in Africa, the evolution of Fintech and Agritech ecosystems, the impact of government policy and regulation, and persisting structural challenges to SMEs' digital participation, such as lack of access to finance, digital literacy gaps, and infrastructure gaps. The case studies of M-Pesa, Farmcrowdy, Twiga Foods, and Flutterwave provide evidence of the potential for a digital ecosystem to have tangible business impacts. Finally, the study outlines a policy framework and roadmap for digital transformation in the SME sector and recommends actions to be taken by governments, development partners, financial institutions, and the private sector.

Obesity is a major problem worldwide, particularly in Latin countries such as Mexico, Peru and Colombia. This study aims to evaluate five models (XGBoost, Random Forest, Multinomial Logistic Regression, Linear Discriminant Analysis (LDA) and Support Vector Machines (SVM)) from the statistical and machine learning fields to determine which of them is the most effective in classification of obesity levels in a multivariate classification model. The research uses a dataset of 2111 records consisting of 17 physiological and behavioral attributes which had to be preprocessed, such as one-hot encoding and multicollinearity filtering, and a powerful 10-fold cross validation methodology. It is shown that the results of ensemble methods are much better than those of the traditional parametric ones. XGBoost achieved the highest classification accuracy (97.54%), and an F_1-score of 0.9747, while Random Forest achieved the 2nd highest classification accuracy (95.64%). For all comparisons, McNemar's pairwise significance tests validated the superiority of XGBoost (with p<0.0001). Feature importance analysis was used to determine the most significant features as follows: weight, height, age and number of vegetables consumed per week (FCVC). While the study showcases the power of gradient boosting for capturing non-linear interactions in health data, the exclusive use of synthetic data (77% from SMOTE) poses drawbacks as it may overestimate performance and limit generalizability. The proposed ensemble architectures should be further evaluated on larger, all-nonsynthetic datasets to be clinically generalizable in the future.