Diabetes is a chronic condition and global health challenge, while prediabetes represents a critical stage for preventing disease progression. As prediabetes affects millions worldwide, accurate and early risk prediction tools are needed beyond blood sugar levels or HbA1c. Identifying at-risk individuals can support timely lifestyle and clinical interventions using individualized, noninvasive clinical features. This study proposes a novel multiclass Siamese Neural Network (SNN) approach for prediabetes prediction. By learning complex feature relationships from multidimensional healthcare data, the SNN model achieved 78.2\% accuracy, 88.9\% specificity, and an AUC of 0.91 using the DiabHealth dataset. These results demonstrate its ability to improve diagnostic precision despite imbalanced data and limited labeled samples. The findings highlight the potential of SNN-based predictive models for early detection, including previously unrecognized prediabetes risk features. Variable importance identified autonomic nervous system testing, oxidative stress, and diet as leading predictors ahead of fasting blood glucose. Integrating AI-driven diagnostics can improve decision-making, reduce long-term complications, and strengthen diabetes prevention efforts.

This paper explores the use of Artificial Intelligence (AI)-driven voice biomarkers for early detection and continuous monitoring of schizophrenia. Traditional diagnostic approaches are often subjective, timeconsuming, and limited in scalability. In contrast, AI offers automated, non-invasive alternatives that can detect vocal indicators linked to core symptoms. A systematic literature review was conducted using the PRISMA framework. Thereafter, the selected studies were evaluated through a systems approach guided by eight Human Factors and Ergonomics (HF/E) principles. The review revealed promising but heterogeneous model performance across classification, prediction, and correlation-based tasks using features extracted from speech across multiple languages and contexts. However, limitations related to generalizability, explainability, workload, and ethical issues were noted. While AI-based voice biomarkers show high potential for enhancing psychiatric diagnosis and monitoring, most solutions remain at the proof-of-concept stage. A systems approach is essential to bridge the gap between technological innovation and real-world clinical application, ensuring usability, safety, and scalability.

Medical question answering systems must provide responses that are not only understandable, but also grounded in reliable sources. This paper presents the MedQuAD Healthcare Assistant, a retrieval-augmented chatbot for patient-facing educational medical question answering. The system uses the MedQuAD dataset as a medical knowledge source, retrieves relevant question-answer records using dense vector search, and generates cited responses with Qwen2.5-7B-Instruct. The evaluation compares a No-RAG baseline with MiniLM and BGE-small retrieval configurations across multiple top-k settings, using retrieval metrics, RAGAS answer-quality metrics, latency, and human evaluation. BGE-small with k = 10 achieved the strongest overall balance, reaching 1.000 Exact QA Hit@10 and 0.911 Exact QA MRR, while also providing the highest context recall among RAG configurations. These results show that dense-retrieval RAG can improve reliable information access, while still requiring expert validation before clinical use.

Most university counseling centers struggle to meet student demand—limited hours, long waitlists, and not enough staff. This project explores whether a chatbot can help fill that gap, not by replacing therapists, but by offering psychoeducation: reliable mental health information and guided self-reflection, available around the clock. The system is built on Qwen3-8B, adapted with QLoRA fine-tuning on 750 filtered CounselChat examples and grounded through retrieval-augmented generation over 17 curated documents from NIMH, WHO, and other public health sources. A four-tier safety layer handles everything from general questions to crisis situations, routing high-risk messages directly to UAE emergency resources. In testing, the safety classifier correctly routed all 15 scenario messages, and the fine-tuned model showed no signs of overfitting (train/val loss ratio of 1.00).