INRIA a CCSD electronic archive server
Not a member yet
122212 research outputs found
Sort by
Evaluating and leveraging large language models in clinical pharmacology and therapeutics assessment: From exam takers to exam shapers
No full textInternational audienceAims In medical education, the ability of large language models (LLMs) to match human performance raises questions about their potential as educational tools. This study evaluates LLMs' performance on Clinical Pharmacology and Therapeutics (CPT) exams, comparing their results to medical students and exploring their ability to identify poorly formulated multiple‐choice questions (MCQs). Methods ChatGPT‐4 Omni, Gemini Advanced, Le Chat and DeepSeek R1 were tested on local CPT exams (third year of bachelor's degree, first/second year of master's degree) and the European Prescribing Exam (EuroPE + ). The exams included MCQs and open‐ended questions assessing knowledge and prescribing skills. LLM results were analysed using the same scoring system as students. A confusion matrix was used to evaluate the ability of ChatGPT and Gemini to identify ambiguous/erroneous MCQs. Results LLMs achieved comparable or superior results to medical students across all levels. For local exams, LLMs outperformed M1 students and matched L3 and M2 students. In EuroPE + , LLMs significantly outperformed students in both the knowledge and prescribing skills sections. All LLM errors in EuroPE + were genuine (100%), whereas local exam errors were frequently due to ambiguities or correction flaws (24.3%). When both ChatGPT and Gemini provided the same incorrect answer to an MCQ, the specificity for detecting ambiguous questions was 92.9%, with a negative predictive value of 85.5%. Conclusion LLMs demonstrate capabilities comparable to or exceeding medical students in CPT exams. Their ability to flag potentially flawed MCQs highlights their value not only as educational tools but also as quality control instruments in exam preparation
Embolization of the middle meningeal artery for chronic subdural hematoma: The OTEMACS multicenter, randomized, clinical trial protocol
No full textInternational audienceBackground The role of middle meningeal artery (MMA) embolization as an adjunct to standard treatment in chronic subdural hematoma (CSDH) is debated. Further randomized trials are needed to establish MMA embolization as an essential therapeutic option for CSDH. The OTEMACS study aims to assess the adjunctive benefit of MMA embolization in patients undergoing either conservative or surgical treatment for CSDH. Methods OTEMACS is a multicenter, prospective, randomized controlled clinical trial with an open-label and blinded endpoint evaluation (PROBE) design. Patients with symptomatic CSDH treated either with conservative or surgical treatment are randomized 1:1 to receive MMA embolization within 72 h (experimental arm) or standard of care alone (control arm). The primary efficacy outcome is a composite of clinical and radiological events, including surgical rescue or revision surgery within 90 ± 14 days postrandomization or radiological remaining of the CSDH thickness >10 mm at 90 ± 14 days postrandomization. The primary safety outcome included all-cause mortality. Secondary outcomes included the modified Rankin scale, Barthel index, EuroQol-5, and Mini Mental State Examination. The number of patients to be included is 440. Results The trial debuted in October 2021 in six centers, in France. A preplanned interim analysis was performed after the enrollment and completion of the follow-up of 220 patients, and the Data Safety Monitoring Board decided to stop the trial for efficacy. The final results will be made available upon completion of the enrollment. Conclusions OTEMACS will provide additional evidence for the clinical and radiological efficacy and safety of MMA embolization in patients with CSDH
Yara: An Ocean Virtual Environment for Research and Development of Autonomous Sailing Robots and Other Unmanned Surface Vessels
No full textInternational audienceOverall, a big challenge in building a sailboat USV relies on the development of an autonomous system for guidance, navigation, and control (GNC) because both sail and rudder angle must be cooperatively adjusted to correct the navigation direction -traditional propelled boats can be more easily controlled with a straightforward control task to set the rudder angle. Moreover, sailing upwind requires special maneuvers to reach a given target in that unfeasible direction. Reinforcement learning emerges as a promising technique for building autonomous GNCs for sailing robots, but training the neural network with a real sailboat is impractical due to long periods of training and safety reasons. Even traditional control-based approaches are mainly tested in simulated environments due to the difficulties in building and operating a real sailboat. The issue that arises is the fidelity of these simulated environments. In this context, we propose Yara, an oceanic virtual environment with a reliable physics simulation for developing, training, and evaluating autonomous agents to operate digital twins of sailing robots in reinforcement learning and other paradigms. An autonomous sailing robot digital twin is available within the virtual environment, with the foil dynamics constructed based on a real sailing robot. We coupled these foil dynamics in Gazebo's physics engine to compute the lift and drag forces acting on the sail, rudder, and keel. The simulated world feeds sensors such as cameras, wind sensors, and GPS. The Robot Operating System communicates these sensors' data through topics, facilitating users' implementation and testing of new GNC solutions. Yara provides a reliable solution for foil dynamic simulated physics that achieves a simulation speedup of 300 times on an i7 laptop with 8 GB of RAM, powered by a Nvidia RTX 3060 and running Ubuntu 20.04. With this speedup, it is possible to complete a million time steps of deep reinforcement learning training in approximately eight hours. Evaluation scenarios were presented to highlight specific features of the simulator, like the maneuverability of the sailing robot digital twin and applications to train, evaluate, and compare reinforcement learning agents and other control solutions.</div
Généralisation du filtre de FRANGI pour l'extraction des réseaux de fissures au sein d'images d'un glissement de terrain. Comparaison avec une méthode d'apprentissage profond
No full textInternational audienceFRANGI filter is a classical algorithm in image processing used to enhance tubular structures (vessels, cracks, etc.). We build upon the ideas of this filter by generalizing them in order to obtain, not a pixelwise response, but for each pair of neighboring pixels. This generalization allows for more precise information about the network of tubular structures. Working with graphs, it provides access to a range of graphical algorithms (e.g. for clustering). We apply these ideas to the extraction of the linear crack network in images of clay soils. The comparison with a deep learning method in combination with transfer learning shows equivalent or slightly better results when the transfer is performed with similar images (same types of soil, same lighting conditions, etc.). Our method has the significant advantage of not requiring annotated training data. Moreover, it relies on a very small number of parameters that can be set in an ad hoc manner.Le filtre de FRANGI est un algorithme classique en traitement d'image pour détecter les structures tubulaires (vaisseaux, fissures, etc.). Nous en reprenons les idées, en les généralisant afin d'obtenir une réponse non pas pixélique, mais en chaque couple de pixels voisins. Cette généralisation permet d'obtenir des informations plus précises sur le réseau de structures tubulaires, et de disposer d'un panel d'algorithmes pour graphes (pour construire un clustering hiérarchique par exemple). Nous appliquons cette nouvelle méthode à l'extraction du réseau linéique de fissures sur des images de sols argileux. La comparaison avec une méthode d'apprentissage profond couplée avec du transfer learning montre des résultats équivalents, voire légèrement supérieurs lorsque le "transfert" est fait avec des images similaires. Notre méthode a le grand avantage de ne nécessiter aucunes données d'apprentissage annotées. De plus, elle ne repose que sur un tout petit nombre de paramètres qui peuvent être fixés de façon ad hoc
Representation learning with a transformer by contrastive learning for money laundering detection
No full textInternational audienceThe present work tackles the money laundering detection problem. A new procedure is introduced which exploits structured time series of both qualitative and quantitative data by means of a transformer neural network. The first step of this procedure aims at learning representations of time series through contrastive learning (without any labels). The second step leverages these representations to generate a money laundering scoring of all observations. A two-thresholds approach is then introduced, which ensures a controlled false-positive rate by means of the Benjamini-Hochberg (BH) procedure. Experiments confirm that the transformer is able to produce general representations that succeed in exploiting money laundering patterns with minimal supervision from domain experts. It also illustrates the higher ability of the new procedure for detecting nonfraudsters as well as fraudsters, while keeping the false positive rate under control. This greatly contrasts with rule-based procedures or the ones based on LSTM architectures
Generalized UGK scheme in the diffusive limit
No full textThe unified gas kinetic scheme (UGKS) was initially designed to address multiscale challenges in rarefied gas dynamics and then extended to radiative transfert theory, as described by BGK like relaxation models. In this work, we extend its application to linear kinetic models with non isotropic scattering collision operators, as well as Fokker-Planck models . These problems typically exhibit a fully diffusive nature in the optically thick limit (corresponding to a small Knudsen number). It still leads to an asymptotic preserving (AP) property not only in this diffusive regime but also in the free transport limit. A series of numerical experiments confirm the effectiveness of the approach
GANSec: Enhancing Supervised Wireless Anomaly Detection Robustness through Tailored Conditional GAN Augmentation
No full textInternational audienceData augmentation techniques show potential in various domains, yet their application to enhance robustness in wireless anomaly detection remains underexplored. Wireless datasets often suffer from anomaly scarcity and class imbalance, hindering the training of reliable detection models. This work introduces GANSec, a novel conditional Generative Adversarial Networks (GAN) framework specifically designed to augment wireless time-series data. We investigate different neural network architectures (MLP, LSTM, CNN) and two conditional training objectives (Embedded Conditional, Classification Oriented) within GANSec, evaluating the framework using real-world 5G measurements for jamming anomaly detection. For evaluation, we train the downstream anomaly detector exclusively on GANSec-generated data and test its performance in a cross-scenario setting. Our evaluation demonstrates that models trained this way significantly outperform those trained on original or baseline augmentation data when tested under unseen network conditions. Specifically, our approach achieved up to 92.13% accuracy on the unseen dataset (i.e., data collected from a different distribution reflecting network conditions distinct from the training set), compared to 78% for models trained on raw data and 83.33% for the best-performing baseline, exhibiting substantially enhanced robustness and generalization
Predicting clinical outcomes from patient care pathways represented with temporal knowledge graphs
No full textInternational audienceBackground: With the increasing availability of healthcare data, predictive modeling finds many applications in the biomedical domain, such as the evaluation of the level of risk for various conditions, which in turn can guide clinical decision making. However, it is unclear how knowledge graph data representations and their embedding, which are competitive in some settings, could be of interest in biomedical predictive modeling. Method: We simulated synthetic but realistic data of patients with intracranial aneurysm and experimented on the task of predicting their clinical outcome. We compared the performance of various classification approaches on tabular data versus a graph-based representation of the same data. Next, we investigated how the adopted schema for representing first individual data and second temporal data impacts predictive performances. Results: Our study illustrates that in our case, a graph representation and Graph Convolutional Network (GCN) embeddings reach the best performance for a predictive task from observational data. We emphasize the importance of the adopted schema and of the consideration of literal values in the representation of individual data. Our study also moderates the relative impact of various time encoding on GCN performance
Leveraging UE-Level Collaborative Intelligence for Scalable Jamming Detection in 5G Networks
No full textInternational audienceThe proliferation of 5G networks enhances connectivity but also increases vulnerability to threats like jamming. Traditional detection strategies often rely on external hardware capturing RF fingerprints, a method becoming increasingly costly and impractical with expanding 5G coverage and heterogeneous IoT device growth. Our prior research demonstrated the feasibility of on-device jamming detection using native User Equipment (UE) log messages, which provide valuable insights but suffer from limited spatial awareness when used locally. This Work-In-Progress (WIP) paper proposes a novel collaborative framework to overcome these limitations. We outline a system where distributed UE report binary jamming indicators derived from standard signal quality measurements available in device logs. A central or distributed fusion mechanism aggregates these simple inputs to achieve network-wide jamming detection, map the affected area, and estimate the jammer's location. We argue that this approach offers scalability and improved situational awareness compared to local methods, while minimizing UE complexity and communication overhead per device. This paper details the proposed architecture, discusses suitable fusion and localization algorithms adapted for binary data, outlines key challenges, and presents our plan for evaluation