Youssef Elmir

Background and ObjectivesLarge language models (LLMs) applied to clinical oncology may generate hallucinated explanations—fabricated justifications for classification decisions—posing risks in decision support. Generic models lack the domain-specific knowledge needed for reliable breast cancer diagnostics. Existing hallucination mitigation strategies operate at the input or output stage rather than at the architecture level. This study aims to develop and evaluate a Mixture-of-Experts (MoE) framework that mitigates explanation hallucination through domain-constrained specialization and to quantify the contribution of expert routing to hallucination reduction.MethodsWe propose BioCancer-MoLA, a parameter-efficient MoE framework comprising six Low-Rank Adaptation (LoRA) modules on a frozen OpenBioLLM8B backbone, each trained on one clinical subdomain (screening, diagnosis, staging, treatment, prognosis, follow-up). A knowledge-transfer router is initialized from a domain-trained RoBERTa classifier by extracting LayerNorm weights fusedwith adaptive random initialization. The model was trained on 21,300 annotated question-answer pairs derived from three breast cancer textbooks and evaluated on 319 clinical cases from a fourth independent textbook spanning non-invasive (150 cases), invasive (69 cases), and advanced-stage (100 cases) breast cancer. Two oncologists annotated model-generated explanations using a 5-point scale assessing diagnostic accuracy, information completeness, and recommendation clarity. A single-LoRA baseline trained on identical data without routing isolates the contribution of expert routing. All fine-tuned models were evaluated across five independent runs with paired bootstrap testing.ResultsBioCancer-MoLA achieves 91.30% accuracy on invasive cases and 58.00% on advanced-stage cases with only 57.6M trainable parameters (99.86% reduction versus full fine-tuning). Oncologist evaluation confirms clinical acceptability across all categories (mean scores 3.66–4.75). Expert routing produces more clinically grounded explanations than the single-LoRA baseline, isolating the architectural contribution to hallucination reduction.ConclusionsDomain-decomposed MoE architectures with knowledge-informed routing offer a structural approach to hallucination mitigation complementary to retrieval-augmented and post-hoc detection methods. Expert routing reduces hallucinated clinical reasoning compared to single-adapter baselines, supporting utility for trustworthy clinical decision support in oncology.

The use of vision large language models (VLLMs) for the analysis and diagnosis of respiratory patterns is affected by limited specialist training data and the lack of integrated medical expertise. This paper introduces a method that uses a multimodal VLLM to analyze electrocardiogram (ECG) data transformed to Gramian Angluar Fields (GAF) 2D image to determine respiratory patterns. Using advanced Shimmer3 sensing. Five VLLMs are fine-tuned using four different ECG datasets for training, each containing 5000 samples, to improve the interpretation of ECG signals. VLLMs are then evaluated against an independent set of data to determine normal, rapid, and difficult breathing patterns. The low-rank adaptation method is used to optimize VLLMs for the analysis of ECG signals obtained from the shimmer sensing platform. Results show that Idefics outperforms other fine-tuned VLLMs with a BLEU score of 0.572 for 2D ECG signals and 0.549 for 1D ECG signals, which demonstrates that the 2D image representation of the GAF improves the visualization of ECG respiratory patterns.

Surveillance systems often struggle with managing vast amounts of footage, much of which is irrelevant, leading to inefficient storage and challenges in event retrieval. This paper addresses these issues by proposing an optimized video recording solution focused on activity detection. The proposed approach utilizes a hybrid method that combines motion detection via frame subtraction with object detection using YOLOv9. This strategy specifically targets the recording of scenes involving human or car activity, thereby reducing unnecessary footage and optimizing storage usage. The developed model demonstrates superior performance, achieving precision metrics of 0.855 for car detection and 0.884 for person detection, and reducing the storage requirements by two-thirds compared to traditional surveillance systems that rely solely on motion detection. This significant reduction in storage highlights the effectiveness of the proposed approach in enhancing surveillance system efficiency. Nonetheless, some limitations persist, particularly the occurrence of false positives and false negatives in adverse weather conditions, such as strong winds.

Artificial Intelligence and infectious diseases prediction have recently experienced a common development and advancement. Machine learning apparition, along with deep learning emergence, extended many approaches against diseases apparition and their spread. And despite their outstanding results in predicting infectious diseases, conflicts appeared regarding the types of data used and how they can be studied, analyzed, and exploited using various emerging methods. This has led to some ongoing discussions in the field. This research aims not only to provide an overview of what has been accomplished, but also to highlight the difficulties related to the types of data used, and the learning methods applied for each research objective. It categorizes these contributions into three areas: predictions using Public Health Data to prevent the spread of a transmissible disease within a region; predictions using Patients’ Medical Data to detect whether a person is infected by a transmissible disease; and predictions using both Public and patient medical data to estimate the extent of disease spread in a population. The paper also critically assesses the potential of Artificial Intelligence and outlines its limitations in infectious disease management.

Traditional surveillance systems rely on human attention, limiting their effectiveness. This study employs convolutional neural networks and transfer learning to develop a real-time computer vision system for automatic handgun detection. A comprehensive analysis of online handgun detection methods is conducted, emphasizing reducing false positives and learning time. Transfer learning is demonstrated as an effective approach. Despite technical challenges, the proposed system achieves a precision rate of 84.74%, demonstrating promising performance comparable to related works, enabling faster learning and accurate automatic handgun detection for enhanced security. This research advances security measures by reducing human monitoring dependence, showcasing the potential of transfer learning-based approaches for efficient and reliable handgun detection.

The development of a road traffic model has numerous benefits, including the reduction of congestion in cities and precise estimated travel times. This study proposes a travel time prediction model based on a deep neural multilayer perceptron trained on GPS data collected in the city of Bechar, Algeria. GPS data was collected using an Android application and a learning model was built using Artificial Intelligence and deep learning techniques. The model was integrated into a web application that allows users to choose a route and make traffic and time predictions. Tests conducted in comparison with Google Maps and real travel experiences in Bechar showed promising results with high accuracy values for most tested routes. However, the accuracy was low in some tests due to the short duration of the data collection process. To improve the project, a data collection phase should be renewed to obtain high accuracy for all destinations and time periods.

Artificial Intelligence (AI) and infectious diseases prediction have recently experienced a common development and advancement. Machine learning (ML) apparition, along with deep learning (DL) emergence, extended many approaches against diseases apparition and their spread. And despite their outstanding results in predicting infectious diseases, conflicts appeared regarding the types of data used and how they can be studied, analyzed, and exploited using various emerging methods. This has led to some ongoing discussions in the field. This research aims not only to provide an overview of what has been accomplished, but also to highlight the difficulties related to the types of data used, and the learning methods applied for each research objective. It categorizes these contributions into three areas: predictions using Public Health Data to prevent the spread of a transmissible disease within a region; predictions using Patients' Medical Data to detect whether a person is infected by a transmissible disease; and predictions using both Public and patient medical data to estimate the extent of disease spread in a population. The paper also critically assesses the potential of AI and outlines its limitations in infectious disease management.

Surveillance systems often struggle with managing vast amounts of footage, much of which is irrelevant, leading to inefficient storage and challenges in event retrieval. This paper addresses these issues by proposing an optimized video recording solution focused on activity detection. The proposed approach utilizes a hybrid method that combines motion detection via frame subtraction with object detection using YOLOv9. This strategy specifically targets the recording of scenes involving human or car activity, thereby reducing unnecessary footage and optimizing storage usage. The developed model demonstrates superior performance, achieving precision metrics of 0.855 for car detection and 0.884 for person detection, and reducing the storage requirements by two-thirds compared to traditional surveillance systems that rely solely on motion detection. This significant reduction in storage highlights the effectiveness of the proposed approach in enhancing surveillance system efficiency. Nonetheless, some limitations persist, particularly the occurrence of false positives and false negatives in adverse weather conditions, such as strong winds.