Hypertension, a life-threatening global health challenge, requires early detection to prevent severe cardiovascular complications. Fundus imaging reveals microvascular alterations, yet conventional diagnosis often misses subtle early changes. This study introduces a multimodal deep learning framework that integrates clinical data, fundus images, and demographic features to improve hypertension prediction. Unlike single-modality approaches, our method captures complementary risk factors from both structured and unstructured data. We evaluate machine learning and deep learning models on clinical data, confirming DL's superior accuracy. For fundus images alone, a CNN achieves 74.44% accuracy, highlighting the limitations of unimodal image analysis. To overcome this, we propose a fused CNN-LSTM architecture that models both spatial biomarkers and temporal clinical trends. The framework achieves robust performance, with an overall accuracy of 98% and minimal variation across datasets. Implemented in TensorFlow/Keras, the system adopts a modular, software-oriented design, ensuring flexibility, ease of maintenance, and seamless integration into clinical workflows. This holistic approach enables timely intervention, improves patient outcomes, and reduces healthcare burdens.

Facial Emotion Recognition (FER) software is an important part of modern software applications. It is used for intelligent user interfaces, diagnostics in psychiatry or psychology, human-computer interaction, and even in surveillance. The recent advancements in the use of deep learning, and the advanced architectures based on them, including Convolutional Neural Networks (CNNs) and transformer models have made the development of FER software much efficient and scalable. This review paper contributes to the existing literature by providing a comprehensive synthesis of Facial Emotion Recognition (FER) systems from a software engineering perspective spanning the period from 2015 to the present. Unlike prior surveys, our analysis emphasizes the software engineering perspective, covering aspects such as software architectures, system integration, deployment on edge devices, and MLOps practices for continuous testing and monitoring. It specifically targets software construction, software architecture, software systems integration, training frameworks, datasets employment, and also deployment challenges. It explores common datasets like FER2013, RAF-DB and AffectNet and the prevailing model architectures, which include CNN, Long-Short-Term Memory (LSTM), hybrid and transformers. Importantly, over half of the surveyed studies continue to rely on demographically narrow datasets (e.g., FER2013, JAFFE, CK+), which limits generalizability and raises fairness concerns across ethnicity, age, and gender. This review contributes by providing a comprehensive and up-to-date synthesis of FER systems while providing directions to the development of fair, lightweight, and robust FER systems suitable for real-world deployment.

Brain tumor segmentation plays an important role in clinical diagnosis, treatment planning, and neuro-oncology assessment using multimodal Magnetic Resonance Imaging (MRI) data. However, conventional centralized deep learning systems often face limitations associated with patient data privacy, secure inter-institutional collaboration, and limited model interpretability. This study presents a decentralized and privacy-preserving brain tumor segmentation framework that integrates Federated Learning (FL), a blockchain-inspired audit and coordination mechanism, and Explainable Artificial Intelligence (XAI) within a collaborative medical imaging environment. A 3D U-Net architecture was trained on the BraTS 2020 dataset under a simulated multi-institutional federated setting using non-IID MRI data distributions. Federated Averaging (FedAvg) was employed for global model aggregation, while a blockchain-inspired permissioned ledger mechanism was used to record model update hashes and aggregation metadata for auditability across communication rounds. Grad CAM and LIME were incorporated to provide interpretable visualization of tumor related regions contributing to segmentation predictions. Experimental evaluation demonstrated stable convergence behavior with an overall voxel accuracy of 98.52%, a weighted F1 score of 98.3%, and Dice coefficient improvement from 0.752 to 0.830 during federated training. The generated segmentation outputs showed strong agreement with manually annotated tumor regions while preserving decentralized data privacy. The proposed framework contributes a secure, interpretable, and collaborative segmentation pipeline for trustworthy brain tumor analysis across distributed healthcare environments.