This paper proposes a first fully automatic custom DL based PCO detection method on original retroillumination images. The class balanced dataset consists of 181 images from 100 patients at three different times after cataract surgery. The method was validated on internal dataset. This proof of concept provides a step towards developing a potential clinical application tool. The proposed method achieves an accuracy of 0.81, precision of 0.88, sensitivity of 0.74 and specificity of 0.88. The AUC was 0.83 (95% CI 0.66–0.96).The semi-automatic version of the custom DL model based PCO detection method was also provided and investigated. It needs a preprocessing of cutting out the region of interest of the retroillumination images resulting in images with cutout. The method is validated by a benchmark comparison to the recently published code free DL method presented by Huemer et. al. (2022), which only works on images with cutout. The semi-automatic method achieves an accuracy of 0.92, precision of 0.91, sensitivity of 0.95 and specificity of 0.88. The AUC was 0.96 (95% CI 0.86–1).Additionally also a novel AQUA based PCO detection method is constructed. The comparison of the novel AQUA based PCO detection method to the semi-automatic custom DL model based PCO detection method, both working on images with cutout, shows the superiority of the proposed custom DL model based method in terms of all investigated performance metrics.
Introduction Achieving Universal Health Coverage in fragile settings remains difficult, and Somalia has one of the lowest coverage profiles in sub-Saharan Africa. Evidence on household insurance coverage and its socioeconomic predictors is limited, and conventional models may not capture the complex nonlinear patterns in survey data. The objective of this study was to apply machine learning algorithms to predict health insurance coverage and identify the most influential predictors associated with household health insurance coverage. Methods A cross-sectional analytical study was conducted in Somalia utilizing nationally representative demographic survey data. The analytic sample included 6132 unweighted households, representing a nationally weighted sample of 7933 households after applying survey sampling weights. Six supervised classification algorithms were compared: logistic regression, decision tree, random forest, gradient boosting, k-nearest neighbors, and support vector machine. The dataset was stratified and split into an 80% training set and a 20% hold-out test set. Hyperparameters and classification thresholds were optimized strictly using 5-fold stratified cross-validation on the training set. To manage extreme class imbalance, class weighting, sample weights, and Precision-Recall threshold tuning were applied. The model performance was evaluated on the independent hold-out test set using accuracy, precision, recall, F1-score, Precision-Recall Area Under the Curve (PR-AUC) and area under the receiver operating characteristic curve (ROC-AUC). Interpretability was achieved using permutation importance and SHapley Additive exPlanations (SHAP). Results The weighted prevalence of household health insurance coverage was 2.73%, with 217 insured and 7716 uninsured households. Random forest showed the best overall discrimination, with a ROC-AUC of 87.91% (95% CI: 80.12–94.75), a PR-AUC of 33.56%, and an F1-score of 41.82%. It also achieved the highest recall (67.65%) for identifying insured households, with a precision of 30.26%. Gradient boosting performed similarly, with ROC-AUC of 87.76% and a PR-AUC of 32.75%. Across the top-performing models, region, household wealth, and residence were the most influential predictors, with additional contributions from household size, sex of the household head, and media exposure. SHAP analysis further indicated that wealth, region, and residence made substantial contributions to model predictions of insurance coverage. Conclusion Household insurance coverage in Somalia is extremely low and varies substantially by geography and socioeconomic position. Intelligence-based predictive modeling, particularly random forests and gradient boosting, provides an exploratory decision-support approach for understanding disparities and identifying geographic and socioeconomic patterns that may inform future health financing and resource-planning assessments, while the modest positive predictive value necessitates human verification before any operational use, though modest positive predictive value necessitates human verification to avoid resource misallocation.
High-dimensional microbiome abundance profiles provide critical insights into microbial dysbiosis and host-associated biological variation; however, identifying robust and interpretable microbial biomarkers remains challenging due to limited sample sizes, sparsity, compositional constraints, and complex nonlinear feature interactions. In this study, we present an explainable machine learning framework for identifying class-specific microbial signatures from microbiome abundance data. Multiple supervised classifiers, including Random Forest, XGBoost, support vector machines, and logistic regression, were systematically evaluated.To reduce optimistic bias associated with small-sample evaluation, model performance was assessed using repeated stratified cross-validation (5-fold × 10 repeats) and reported as mean ± standard deviation. Among the evaluated models, ensemble-based approaches, particularly Random Forest and XGBoost, demonstrated the most stable performance in the multiclass setting.To enhance interpretability, Shapley Additive exPlanations (SHAP) were applied to quantify both global and class-specific feature contributions. SHAP analysis identified a concise subset of microbial taxa that consistently influenced classification outcomes, providing biologically interpretable insights into microbiome variation across classes.The proposed framework integrates predictive modelling with transparent feature attribution, offering a reproducible strategy for explainable biomarker discovery in microbiome-based biomedical informatics research. Furthermore, the interpretability of the identified microbial signatures highlights the potential of the proposed approach to support clinically relevant decision-making and advance precision medicine applications in microbiome-associated diseases.
Background Cerebrovascular diseases like stroke and cerebral microbleeds pose global health challenges. Susceptibility-weighted imaging (SWI) enhances visualization of cerebrovascular features, but manual analysis is time-intensive and expertise-dependent. Objective This systematic review evaluates deep learning (DL) techniques for cerebrovascular feature detection in SWI, synthesizing architectures, preprocessing methods, performance metrics, and clinical applicability, while identifying challenges and future directions. Methods Following PRISMA guidelines (PROSPERO: CRD420251247494), we searched PubMed, Scopus, Web of Science, IEEE Xplore, and Cochrane Library for peer-reviewed studies using DL for SWI-based detection. The PICO framework guided eligibility. Quality was assessed using QUADAS-2. Results From 1587 records, 13 studies were included, focusing on cerebral microbleeds (CMBs), deep medullary veins, superficial cerebral veins, and dural arteriovenous fistulas. Architectures included 3D U-Net, DeepLabV3+, and GANs, with sensitivities of 78.85–96.67% and Dice coefficients of 0.46–0.81 for CMBs. Sample sizes ranged from 20 to 2546. Multicenter studies showed better generalizability, but small samples, inconsistent annotations, and scanner variability limited robustness. Clinical integration was preliminary. Conclusion DL shows promising accuracy for cerebrovascular feature detection in SWI, especially with CNN-based models. However, small sample sizes, single-center designs, inconsistent annotations, and limited external validation mean that conclusions regarding clinical utility remain preliminary. Future research should prioritize standardized datasets, external validation, and real-time clinical integration to enhance diagnostic precision.
