Aleka Melese Ayalew

Nordic healthcare systems that implement universal electronic health records need interoperable clinical documentation to ensure seamless care coordination among hospital networks, primary care, and home healthcare services. Current discharge summaries lack structural standardization, contributing to 20% of patients experiencing preventable adverse events within 30 days of hospital discharge. While AI-assisted documentation offers automation potential for resource-constrained healthcare systems, clinical deployment remains limited by hallucination rates of 18–64%. We present a hybrid LLM framework combining parameter-efficient fine-tuning with retrieval-augmented generation to produce structured 11-section discharge summaries compatible with national EHR systems. Evaluation across six transformer architectures on MIMIC-IV clinical notes demonstrates that hybrid RAG augmentation reduces hallucination rates compared to fine-tuning alone, with improvements up to 12.4% points (Gemma: 36.3% to 23.9%). The best configuration (Llama-3.1-8B Hybrid) achieves 18.5% hallucination rate with ROUGE-1 of 0.466. The framework addresses critical requirements for Nordic digital health infrastructure: (1) structured output compatible with national EHR platforms; (2) reduced documentation burden supporting physician workforce sustainability; (3) improved patient comprehension for aging populations; and (4) potential for wireless health integration. Results demonstrate technical maturity for pilot testing in Nordic healthcare testbed environments, supporting care pathway digitalization and population health management initiatives.

Medical question abbreviation aims to condense long, consumer-facing medical questions into concise forms while preserving clinical intent and meaning. Despite recent progress with large language models, existing approaches frequently introduce unsupported medical concepts, omit critical qualifiers, or subtly alter the original information need, errors that are poorly captured by standard lexical and semantic evaluation metrics. This paper focuses on the MeQSum benchmark, where verbose consumer health questions are abbreviated into concise professional forms, a task-critical for digital health platforms enabling remote consultations. We propose MEG-QA (Medical Evidence-Grounded Question Abbreviation) , a framework that aligns generation to both retrieved evidence and structured medical knowledge. MEG-QA integrates retrieval-augmented generation with knowledge graph grounding to guide abbreviations toward conceptually faithful and evidence-supported outputs. Instead of relying solely on prompt design or fine-tuning, MEG-QA constrains generation through alignment with normalized medical concepts and their relationships. To evaluate faithfulness beyond surface similarity, we introduce the CQA-Align Score , a concept-qualifier-attribution alignment metric that measures whether the generated questions preserve medical concepts, maintain critical qualifiers, and avoid unsupported additions. Experiments on the MedQSum dataset show that MEG-QA consistently improves faithfulness and intent preservation across models and adaptation paradigms, with substantial gains in CQA-Align that are not reflected by conventional metrics. These results highlight the importance of evidence-aligned and knowledge-generation for reliable medical question abbreviations.

Recently, large language models have gained momentum in several areas, including medical text summarization and clinical decision support in healthcare. In this systematic review, we focus on how LLMs are used for summarization and explainability, and on how they improve the precision of clinical decisions. Initially, by screening eight distinct databases, we identified 1601 studies using the PRISMA protocol. Following strict qualifying criteria, 61 studies were selected for the final review. Our results show that decoder‐only architectures dominate the field (60.7% of studies), with zero‐shot prompting emerging as the predominant approach (55.7%), while 24.6% of studies employed fine‐tuning. General clinical decision support (26.2%) and diagnostic assistance (19.7%) constitute the two most widely used clinical applications across multiple clinical specialties. Identified in 26.2% of studies, the intersection of summarization and explainability becomes a key focus area, with fine‐tuned models obtaining 38.6% higher average ROUGE‐1 scores compared to zero‐shot approaches, and studies including robust explainability features report 27.3% higher clinician acceptance rates. Despite great potential, major obstacles still exist, including model hallucinations (63.9% of studies), minimal workflow integration (only 14.8% of studies), and inadequate attention to regulatory paths (8.2%). Although LLMs offer significant potential for transforming clinical decision support via improved information handling and clear reasoning, realizing these goals requires overcoming essential challenges in clinical validation, system integration, and ethical oversight. This review aims to provide an extensive framework for understanding the current capabilities, limitations, and prospects of LLM‐based explainable summarization in clinical settings, guiding future research and practical deployment. This article is categorized under: Technologies > Artificial Intelligence Fundamental Concepts of Data and Knowledge > Explainable AI Application Areas > Health Care

Tuberculosis (TB) is a highly contagious disease that affects millions of individuals globally. Early detection is crucial in preventing its spread and improving patient outcomes. Radiologists often utilize X-ray imaging as a diagnostic tool for tuberculosis; however, the accuracy of the results may differ depending on the radiologist’s interpretation. To increase the precision of TB identification from X-ray images, a hybrid strategy combining Convolutional Neural Network, Histogram of Oriented Gradients, and Quantum Support Vector Machine (QSVM) classifier has been developed. Quantum machine learning is a rapidly growing field at the intersection of quantum computing and machine learning. By combining the strengths of both techniques, this approach aims to capture relevant features and shape and texture information of an image. The study used state-of-the-art models such as VGG16 and AlexNet, along with a handcrafted method (HOG) and a deep learning (CNN) method to detect TB from chest X-ray images. Six distinct tests were carried out by the authors to successfully diagnose tuberculosis. All models showed promise when the data was analyzed, but the best techniques involved augmentation, image preprocessing, contrast improvement, and noise filtering. The block-matching and 3D filtering approach was used to improve image edge preservation and reduce noise. In this study, Grad-CAM (gradient-weighted class activation map) was applied to the convolutional neural network model to identify the model’s important features (explainability) for decision-making. This focus on Explainable AI (XAI) is crucial for clinical adoption, as it provides radiologists with visual evidence of the model’s reasoning, fostering trust and enabling more informed decisions. The hybrid model was found to have achieved an exceptional level of accuracy in detecting and classifying TB, with an accuracy rate of 100% during training, 99.07% during validation, and 98.14% during testing. These results highlight the effectiveness of the hybrid model in accurately identifying TB and its potential to be a valuable tool in the fight against this disease.

Airborne infectious diseases are a significant threat to human beings. Nowadays, one of the deadliest airborne diseases, coronavirus (COVID-19), is resulting in a massive health crisis due to its rapid transmission. The World Health Organization for protection against the spread of airborne diseases has set several guidelines. The most effective preventive measure against airborne diseases, according to the World Health Organization, is wearing masks in public places and crowded areas. It is challenging to monitor people manually in these areas. In this study, we collect data from public and local sources to develop an occlusion-aware face mask detection model. This study presents a deep learning-based occlusion-aware face mask detection model designed to identify both proper and improper mask usage, even under partial facial occlusions. A dataset of 4,820 images, including occlusions from hands, objects, and mask misuse, was used to train and evaluate three convolutional neural network models: InceptionV3, MobileNetV2, and DenseNet121. Among them, DenseNet121 achieved the highest accuracy of 96.3% on test data. Therefore, our proposed study is used to investigate occlusion aware face mask classification using deep learning.

Breast cancer is the most prevalent form of tumor in women, and this is the leading cause of mortality in women. Accurately detecting and classifying breast cancer is crucial for effective treatment and diagnosis preparation. Internet of Things (IoT) wearable devices (smart bra) are considered one of the best methods for early detection and, thereby, reducing breast cancer mortality. This study proposes a unique approach for breast cancer classification that combines involutional neural networks (InvNets) with wearable devices to reduce the parameter-intensive nature of convolutional neural networks. Specifically, the involution kernel differs from the convolution kernel because it is location-specific and channel-agnostic. This location-specific operation enhances the network's ability to acquire detailed elements in medical images by adapting to diverse visual patterns across spatial regions. In that regard, magnetic resonance imaging (MRI) has become an important modality for breast cancer detection. This work investigates the use of IoT devices such as wearables and sensors to collect patient health data and monitor changes in breast tissue. If symptoms are present, an MRI scan is performed for a reliable diagnosis of breast cancer utilizing deep learning (InvNets). Our findings show that InvNets achieves 100% training accuracy, 98.5% validation accuracy, and 98.6% testing accuracy after data augmentation, indicating its potential for breast cancer classification. InvNets are highly successful for medical image processing, particularly in circumstances with limited computer resources, as seen by enhanced accuracy and reduced parameter count. As a result, the findings show that InvNets provided consistent and reliable features for breast cancer detection and faster diagnosis.