Pneumonia is a deadly disease affecting millions worldwide, caused by microorganisms and environmental factors. It leads to lung fluid build-up, making breathing difficult, and is a leading cause of death. Early detection and treatment are crucial for preventing severe outcomes. Chest X-rays are commonly used for diagnoses due to their accessibility and low costs; however, detecting pneumonia through X-rays is challenging. Automated methods are needed, and machine learning can solve complex computer vision problems in medical imaging. This research develops a robust machine learning model for the early detection of pneumonia using chest X-rays, leveraging advanced image processing techniques and deep learning algorithms that accurately identify pneumonia patterns, enabling prompt diagnosis and treatment. The research develops a CNN model from the ground up and a ResNet-50 pretrained model This study uses the RSNA pneumonia detection challenge original dataset comprising 26,684 chest array images collected from unique patients (56% male, 44% females) to build a machine learning model for the early detection of pneumonia. The data are made up of pneumonia (31.6%) and non-pneumonia (68.8%), providing an effective foundation for the model training and evaluation. A reduced size of the dataset was used to examine the impact of data size and both versions were tested with and without the use of augmentation. The models were compared with existing works, the model’s effectiveness in detecting pneumonia was compared with one another, and the impact of augmentation and the dataset size on the performance of the models was examined. The overall best accuracy achieved was that of the CNN model from scratch, with no augmentation, an accuracy of 0.79, a precision of 0.76, a recall of 0.73, and an F1 score of 0.74. However, the pretrained model, with lower overall accuracy, was found to be more generalizable.

With the increasing availability of wearable devices for data collection, studies in human activity recognition have gained significant popularity. These studies report high accuracies on k-fold cross validation, which is not reflective of their generalization performance but is a result of the inappropriate split of testing and training datasets, causing these models to evaluate the same subjects that they were trained on, making them subject-dependent. This study comparatively discusses this validation approach with a universal approach, Leave-One-Subject-Out (LOSO) cross-validation which is not subject-dependent and ensures that an entirely new subject is used for evaluation in each fold, validated on four different machine learning models trained on windowed data and select hand-crafted features. The random forest model, with the highest accuracy of 76% when evaluated on LOSO, achieved an accuracy of 89% on k-fold cross-validation, demonstrating data leakage. Additionally, this experiment underscores the significance of hand-crafted features by contrasting their accuracy with that of raw sensor models. The feature models demonstrate a remarkable 30% higher accuracy, underscoring the importance of feature engineering in enhancing the robustness and precision of HAR systems.

This study investigates sentiment analysis of X data about the National Health Service (NHS) during a politically charged period, using lexicon-based, machine learning, and deep learning approaches, as well as topic modelling and aspect-based sentiment analysis (ABSA). This study is distinct in its comparative evaluation of sentiment analysis techniques on NHS-related tweets during a politically sensitive period, offering insights into public opinion shaped by political discourse. A dataset of 35,000 tweets collected and analysed using various techniques, including VADER, TextBlob, Naive Bayes, Support Vector Machines, Logistic Regression, Ensemble Learning, and BERT. Unlike previous studies that focus on structured feedback or general sentiment, this research uniquely explores unstructured public discourse during an election period, capturing real-time political sentiment towards NHS policies. The sentiment distribution from lexicon-based methods depicted that the presence of stop words could affect model performance. While all models achieved high accuracy on the validation dataset, challenges such as class imbalance and limited labelled data impacted performance, with signs of overfitting observed. Topic modelling identified nine topic clusters, with “waiting list,” “service,” and “immigration” carrying negative sentiments. At the same time, words like “thank,” “support,” “care,” and “team” had the most positive sentiments, reflecting public delight in these areas. ABSA identified positive sentiments towards aspects like “useful service”. This study contributes a comparative framework for evaluating sentiment analysis techniques in politically contextualised healthcare discourse, offering insights for policymakers and researchers. The study underscores the importance of data quality in sentiment analysis. Future research should consider incorporating multilingual datasets, extending data collection periods, optimising deep learning models, and employing hybrid approaches to enhance performance.

Accurate detection of musculoskeletal abnormalities from radiographs remains challenging because abnormal findings can be subtle and radiographic appearance may vary considerably across examinations. This study aimed to develop and systematically evaluate a deep learning framework incorporating Grad-CAM-based visual attribution for automated musculoskeletal abnormality detection using the publicly available Musculoskeletal Radiographs (MURA) dataset. Four transfer learning architectures, namely DenseNet201, EfficientNetV2-S, ConvNeXt Tiny, and Swin Transformer Tiny, were evaluated for musculoskeletal abnormality classification. The framework incorporated Grad-CAM-based attribution and activation-based weak localization to highlight image regions associated with model predictions without requiring bounding-box annotations during training and a sensitivity-oriented decision threshold to prioritize the detection of abnormal examinations. Model performance was assessed using recall, precision, F1-score, area under the receiver operating characteristic curve (AUC), and confusion matrix analysis. A targeted decision-threshold ablation analysis compared the sensitivity-focused threshold of 0.40 with the conventional 0.50 baseline using the same validation-set probabilities. Previously published results were reviewed to provide contextual information rather than establish comparative superiority. EfficientNetV2-S achieved the highest precision (0.925), F1-score (0.892), and AUC (0.956) on the internal validation set among the evaluated architectures. DenseNet201 achieved the highest recall of 0.876. Previously published performance metrics were reviewed to contextualize the results of the present study. However, differences in datasets, anatomical coverage, preprocessing procedures, and evaluation protocols preclude direct conclusions regarding comparative superiority. At the sensitivity-focused threshold of 0.40, EfficientNetV2-S obtained an accuracy of 0.896 and an F1-score of 0.892 on the internal validation set. Lowering the threshold to 0.40 improved sensitivity from 0.754 to 0.861 and decreased false-negative predictions from 1092 to 618 when compared to the traditional threshold of 0.50. Alongside this improvement, false-positive predictions increased from 150 to 307 and specificity decreased from 0.966 to 0.931. Grad-CAM visualizations and activation-derived bounding boxes provided qualitative information about spatial attribution. However, their correspondence with pathological findings was not independently validated using expert annotations or quantitative localization metrics. On the independent held-out test set, EfficientNetV2-S achieved a recall of 0.826, precision of 0.886, F1-score of 0.855, and AUC of 0.927. The independent evaluation demonstrated lower performance than the internal validation results, highlighting the importance of separating model optimization from final performance assessment. The findings demonstrate the feasibility of integrating transfer learning, Grad-CAM-based visual attribution, activation-based weak localization, and sensitivity-oriented decision-making within a unified framework for musculoskeletal abnormality classification. Although the framework demonstrated promising classification performance, the visual attribution results remain exploratory and do not establish pathological localization accuracy or clinical interpretability. Further evaluation using external clinical datasets, expert-annotated pathological regions, and radiologist assessment is required to establish its generalizability, explanation validity, and potential clinical utility.