Background: Stroke is a leading cause of disability and death worldwide, where acute ischemic strokes are the most prevalent. Real-time health monitoring is crucial for early detection, timely alerts, and improved clinical management. This study developed an innovative multimodal system for the first time that integrates clinical data and facial recognition technology to improve stroke risk prediction. By combining clinical and physiological data with facial image features such as muscle asymmetries and stroke-related expressions, our goal is to establish more accurate early stroke prediction and reduce false alarms. Methodology: The multimodal system employs Convolutional Neural Networks (CNNs) for facial image analysis and a Multi-Layer Perceptron (MLP) integrates outputs from both data types into a comprehensive predictive model. The dataset includes records from patients with attributes like age, blood pressure, glucose, and cholesterol. Data preprocessing involved encoding categorical variables and standardization. The distillation-based multimodal model further improved performance, achieving an average accuracy of 0.96, precision of 0.95, recall of 0.97, and an F1 score of 0.95 across five cross-validation folds. This enhancement reflects a more generalized and stable predictive capacity, reducing the risk of overfitting and enhancing robustness. Discussion and Conclusion: The proposed distillation-based multimodal system offers an advanced framework for the assessment of stroke risk, particularly beneficial for early stage diagnosis by detecting both clinical and morphological indicators. The integration of clinical and facial image data not only improves predictive accuracy but also provides a balanced approach by leveraging visual and non-visual stroke indicators. These findings underscore the potential of the system for clinical applications, especially in remote or underserved areas where rapid and accessible stroke screening is essential. Future work will explore the integration of transformer-based architectures to further improve the generalizability and performance of the model on larger and more diverse datasets.

This paper presents a novel framework for medical data assessment that integrates automated AI analysis, SHAP-based interpretability, and human feedback to generate comprehensive medical reports. Our approach employs a logistic regression model evaluated on a heart disease dataset, demonstrating robust performance across training, validation, and test splits. The framework uses SHAP values to provide transparent, quantitative insights into the influence of each clinical parameter on the prediction outcome. By incorporating human feedback as the definitive ground truth, the system refines its outputs, thereby bridging the gap between automated analysis and evolving clinical expertise. This integration addresses common challenges in clinical data including missing entries, coding discrepancies, and heterogeneity across healthcare providers to ensure that the generated reports are both consistent and reliable. The resulting automated report not only reduces the documentation burden on healthcare professionals but also standardizes reporting workflows, ultimately enhancing diagnostic decision-making. Future work will focus on extending the multi-agent framework to encompass additional clinical tasks and on integrating reinforcement learning techniques to enable continuous model improvement based on real-time feedback.

Cancer genomics identifies all genes playing critical roles in carcinogenesis. The state-of-the-art cancer genomics profiling characterized many clinically and biologically relevant patterns that are not resolvable by morphology nor distinguishable under the microscope for cancer diagnosis. With that genomic information, doctors can develop an individualized treatment plan for cancer patients and provide precision medicine. However, several technical challenges (such as low tumor purity, batch effects and formalin-fixed, paraffin-embedded (FFPE) tissue restoration) potentially led to ambiguous diagnoses that needed to be solved in the clinical setting. The purpose of this study is to develop a robust tumor classification framework to improve cancer diagnosis and provide Explainable Artificial Intelligence (XAI) based interpretable results with increased transparency model interpretability of the classification. We utilized a large set of over six thousand tumor samples (DNA methylation and gene expression) from The Cancer Genome Atlas (TCGA). We implemented realistic variable selection by separating the training and test datasets and removed artificial and technical sources of variabilities to overcome batch effect issues while identifying the biological variation and making the prediction meaningful and robust. The Random Forest classifier produced about 95 and 96% accuracy for mRNA and methylation-based models respectively with minimum features of 50 methylation probes and gene expression signatures. We further developed an XAI strategy and applied it to a large brain cancer patient group to make an explainable patient-specific decision while tailoring the provided recommendations based on each patient's characteristics. This strategy demonstrates more accurate and practical molecular subtype classification with explainable AI for model interpretation.

Background: The recent revolution in Large Language Models (LLMs) is transforming industries, enhancing communication, and reshaping research methodologies. LLMs have found significant applications across various sectors, notably in finance for stock market predictions, and in healthcare, where complex medical data is analyzed for diagnosis at an early stage, improving diagnostic procedures, and personalized treatment planning. In healthcare, where complex medical data is analyzed for diagnosis at an early stage. Despite the immense potential, challenges such as overwhelming Big Data, model hallucinations, and ethical concerns about patient privacy and bias persist. Method: We implemented novel strategies like CanPrompt to mitigate the accuracy and hallucination concerns to ensure responsible deployment. The CanPrompt strategy utilizes prompt engineering combined with few-shot and in-context learning to significantly enhance model accuracy by generating more relevant answers. The models were tested against a specialized dataset from MedQuAD, focusing on cancer, and evaluated using metrics like ROUGE and BERTScore to assess the semantic and syntactic accuracy of generated responses against validated "Gold Answers". Through this approach, the study seeks to outline the potential and limitations of LLMs in improving cancer care. Result: After applying CanPrompt with models Mistral 7x8b, Falcon 40b, and Llama 3-8b, BERTScore results showed Mistral leading with an accuracy around 84%, Falcon slightly lower, and Llama the least, with respective precision scores also reflecting a similar trend. Conclusion: The study demonstrates the promise of LLMs in cancer care through the introduction of CanPrompt.

Diffuse gliomas can be divided based on presence or absence of mutation in isocitrate dehydrogenase (IDH) genes. IDH-mutant diffuse gliomas represent a wide range of clinical outcome, which is not accounted for by current clinical and pathologic parameters. To address this, we aim to identify and characterize a predictive signature of outcome in diffuse gliomas to better understand this heterogeneity in outcome. A total of 310 IDH mutant glioma samples with methylation data were used for the analysis together with 419 samples from The Cancer Genome Atlas (TCGA), utilizing methylation, mRNA, copy number variation (CNV) and mutation data to identify unique molecular signatures that predict patient outcome. Methylation analysis from our test cohort identified signatures from Cox regression analysis that split the glioma cohort into two prognostic groups that strongly predicted survival (p-value<0.0001). The CpG-based signatures were reliably validated using two independent validation datasets from TCGA and DKFZ (German Cancer Research Center) cohorts (both p-values < 0.0001). The results show that the methylation signatures that predict poor outcome also correlated with G-CIMP low status, elevated CNV instability and hypermethylation of a set of HOX gene probes. These results demonstrate the importance of HOX genes in the outcome of diffuse gliomas to identify relevant molecular subtyping indistinguishable under the microscope within a histology.

Electrical impedance tomography (EIT) is a noninvasive method to image conductivity distributions within a body. One promising application of EIT is to monitor ventilation in patients as a real-time bedside tool. Thus, it is essential that an EIT system reliably provide meaningful information, or alert clinicians when this is impossible. Because the reconstructed images are very sensitive to system instabilities (primarily from electrode connection variability and movement), EIT systems should continuously monitor and, if possible, correct for such errors. Motivated by this requirement, we describe a novel approach to quantitatively measure EIT data quality. Our goals are to define the requirements of a data quality metric, develop a metric q which meets these requirements, and an efficient way to calculate it. The developed metric q was validated using data from saline tank experiments and a retrospective clinical study. Additionally, we show that q may be used to compare the performance of EIT systems using phantom measurements. Results suggest that the calculated metric reflects well the quality of reconstructed EIT images for both phantom and clinical data. The proposed measure can thus be used for real-time assessment of EIT data quality and, hence, to indicate the reliability of any derived physiological information.