Laboratoire de Recherche en Informatique

The integration of artificial intelligence (AI) and deep learning into healthcare is transforming how unstructured clinical data is processed, offering new opportunities to improve decision-making and patient care. However, the manual analysis of clinical text records poses significant challenges in terms of time, accuracy, and consistency. This paper presents an automated system for multi-label classification of procedures from clinical summaries, using a newly collected private dataset of digitized medical records. Our approach leverages recent advances in deep learning and natural language processing, including transformer-based models, to extract and classify actionable insights from unstructured clinical documentation. We address key challenges related to variability in documentation, data imbalance, and domain-specific language. This study contributes to intelligent processing of clinical documents, illustrating the practical integration of AI in real-world healthcare contexts.

In higher education, reading proficiency is essential for academic success. Yet, many students begin postsecondary studies with underdeveloped reading skills, limiting both knowledge acquisition and disciplinary engagement. This paper presents the design, implementation, and preliminary evaluation of an AIbased tutoring system for active reading, adapted from a speech therapy technology, called TutorIAt. Our system leverages speech recognition to guide students through critical reading processes, fosters self-regulated learning, and complements existing academic support services. We also discuss copyright management, ethical safeguards, and equitable use guidelines developed with faculty collaboration to ensure responsible deployment. Finally, we illustrate how the system engages students in using effective reading strategies as it provides feedback and phonetic analysis in less than 3 seconds after reading. Preliminary results show that such a system can promote more equitable, inclusive, and autonomous access to success in higher education.

The proliferation of fake profiles on social media platforms presents a growing challenge for digital ecosystems, where the detection of such profiles is critical to maintaining the integrity of online environments. This paper introduces a hybrid approach that integrates the Grasshopper Optimization Algorithm with various Machine Learning classifiers, including Support Vector Machine, Naive Bayes, and Random Forest. The nature-inspired metaheurisitic used is employed to optimize key hyperparameters of these classifiers, thereby enhancing their performance in detecting fake profiles. The proposed method is evaluated on a well defined balanced dataset, demonstrating significant improvements in classification performance, particularly in terms of accuracy, precision, recall, and F1-score. The results suggest that the proposed hybrid approach can effectively address the challenges associated with balanced and imbalanced datasets in fake profile detection. Furthermore, the study discusses potential directions for improving scalability and applying the approach to larger and more dynamic datasets in the future.

Simultaneous Localization and Mapping (SLAM) has long been a fundamental and challenging task in robotics literature, where safety and reliability are the critical issues for successfully autonomous applications of robots. Classically, the SLAM problem is tackled via probabilistic or optimization methods (such as EKF-SLAM, Fast-SLAM, and Graph-SLAM). Despite their strong performance in real-world scenarios, these methods may exhibit inconsistency, which is caused by the inherent characteristic of model linearization or Gaussian noise assumption. In this paper, we propose an alternative monocular SLAM algorithm which theoretically relies on interval analysis (iMonoSLAM), to pursue guaranteed rather than probabilistically defined solutions. We consistently modeled and initialized the SLAM problem with a bounded-error parametric model. The state estimation process is then cast into an Interval Constraint Satisfaction Problem (ICSP) and resolved through interval constraint propagation techniques without any linearization or Gaussian noise assumption. Furthermore, we theoretically prove the obtained consistency and propose a versatile method for numerical validation. To the best of our knowledge, this is the first time such a proof has been proposed. A plethora of numerical experiments are carried to validate the consistency, and a preliminary comparison with classical EKF-SLAM in different noisy situations is also presented. Our proposed iMonoSLAM shows outstanding performance in obtaining reliable solutions, highlighting the potential application prospect in safety-critical scenarios of mobile robots.

Online Social Networks (OSNs) are becoming increasingly important in business, government, and all areas of life. For-profit companies use them as rich sources of information and dynamic platforms to drive strategies in product design, innovation, relationship management, and marketing. However, analyzing and retrieving information from these platforms presents distinct challenges due to their inherent characteristics and dynamic nature. To address this, researchers have proposed various approaches for social information retrieval, ranging from term-based analysis to semantic-based methods. To overcome the limitations of existing techniques, the present study proposes a multilayer model that integrates graph analysis, semantic content, and deep learning. The general proposed approach is also presented. By combining learning-to-rank techniques with linked data, a robust framework for social information retrieval is constructed. This method enables a more nuanced understanding by leveraging both the rich contextual information provided by linked data and the structural characteristics of social networks. The proposed model is a flexible framework that can be easily extended to add or remove features and can be applied to various tasks. The experimental results confirm the effectiveness and efficiency of the proposed approach.

This study posits that the PID controller, designed to uphold drone stability, encounters a timing issue that warrants further tuning and enhancement. The study conducts a performance evaluation of PID controller gains for drone angle control, with the objective of optimising them to bolster the drone's speed, accuracy, and stability. To achieve this objective, a PID flight controller is proposed to manage the altitude dynamics of a UAV. The study's methodology primarily involves a comparative analysis across three levels: initially utilising a single PID controller for all three angles, then employing two PID controllers for all three angles where one manages pitch and roll angles while the other handles yaw angle, and finally implementing three PID sub-controllers for each angle (pitch, roll, and yaw). The comparative analysis aims to pinpoint the most effective PID controller configuration that enhances stability, responsiveness, and accuracy during flight. In comparison to prior research, the suggested adaptive PID flight controller showcases innovation and efficacy in the field.