HAL-Paris 13
Not a member yet
33893 research outputs found
Sort by
Identifying biomarker-driven subphenotypes of cardiogenic shock: analysis of prospective cohorts and randomized controlled trials
No full textInternational audienceBackground Cardiogenic shock (CS) is a heterogeneous clinical syndrome, making it challenging to predict patient trajectory and response to treatment. This study aims to identify biological/molecular CS subphenotypes, evaluate their association with outcome, and explore their impact on heterogeneity of treatment effect (ShockCO-OP, NCT06376318).Methods We used unsupervised clustering to integrate plasma biomarker data from two prospective cohorts of CS patients: CardShock (N = 205 [2010-2012, NCT01374867]) and the French and European Outcome reGistry in Intensive Care Units (FROG-ICU) (N = 228 [2011-2013, NCT01367093]) to determine the optimal number of classes. Thereafter, a simplified classifier (Euclidean distances) was used to assign the identified CS subphenotypes in three completed randomized controlled trials (RCTs) (OptimaCC, N = 57 [2011-2016, NCT01367743]; DOREMI, N = 192 [2017-2020, NCT03207165]; and CULPRIT-SHOCK, N = 434 [2013-2017, NCT01927549]) and explore heterogeneity of treatment effect with respect to 28-day mortality (primary outcome).Findings Four biomarker-driven CS subphenotypes ('adaptive', 'non-inflammatory', 'cardiopathic', and 'inflammatory') were identified separately in the two cohorts. Patients in the inflammatory and cardiopathic subphenotypes had the highest 28-day mortality (p (log-rank test) = 0.0099 and 0.0055 in the CardShock and FROG-ICU cohorts, respectively). Subphenotype membership significantly improved risk stratification when</div
A Greedy Algorithm for Low-Crossing Partitions for General Set Systems
No full textSimplicial partitions are a fundamental structure in computational geometry, as they form the basis of optimal data structures for range searching and several related problems. Current algorithms are built on very specific spatial partitioning tools tailored for certain geometric cases. This severely limits their applicability to general set systems. In this work, we propose a simple greedy heuristic for constructing simplicial partitions of any set system. We present a thorough empirical evaluation of its behavior on a variety of geometric and non-geometric set systems, showing that it performs well on most instances. Implementation of these algorithms is available on Github
Sémantique catégorielle de logique de descriptions et raisonnement
No full textInternational audienceWe present a new description logic (DL) allowing for negation and a reasoning procedure for it. This construction is based on a rewriting of the usual set semantics of the DL ALC using category theory where concepts and subsumptions are respectively represented as objects and arrows of a category. This rewriting offers a categorical representation of the semantics that is more modular than the usual one based on set theory. This modularity allows us to define an NP-complete logic without the interaction between logical constructors reponsible for EXPTIME complexity.Nous présentons une nouvelle logique de description (DL) permettant la négation ainsi qu'une procédure de raisonnement pour celle-ci. La construction de cette nouvelle logique est basée sur la réécriture de la sémantique usuelle ensembliste de la DL ALC en langage catégoriel où les concepts et subsomptions sont respectivement représentés par des objets et des flèches d'une catégorie. Cette réécriture nous donne une plus grande modularité sur la nouvelle sémantique des différents constructeurs comparée à la sémantique ensembliste. La modularité nous permet de définir une nouvelle DL NP-complète avec la négation en éliminant les interactions entre les constructeurs logiques responsables de la complexité EXPTIME de la DL ALC
Machine learning score to predict in-hospital outcomes in patients hospitalized in cardiac intensive care unit
No full textInternational audienceAbstract Aims Although some scores based on traditional statistical methods are available for risk stratification in patients hospitalized in cardiac intensive care units (CICUs), the interest of machine learning (ML) methods for risk stratification in this field is not well established. We aimed to build an ML model to predict in-hospital major adverse events (MAE) in patients hospitalized in CICU. Methods and results In April 2021, a French national prospective multicentre study involving 39 centres included all consecutive patients admitted to CICU. The primary outcome was in-hospital MAE, including death, resuscitated cardiac arrest, or cardiogenic shock. Using 31 randomly assigned centres as an index cohort (divided into training and testing sets), several ML models were evaluated to predict in-hospital MAE. The eight remaining centres were used as an external validation cohort. Among 1499 consecutive patients included (aged 64 ± 15 years, 70% male), 67 had in-hospital MAE (4.3%). Out of 28 clinical, biological, ECG, and echocardiographic variables, seven were selected to predict MAE in the training set (n = 844). Boosted cost-sensitive C5.0 technique showed the best performance compared with other ML methods [receiver operating characteristic area under the curve (AUROC) = 0.90, precision–recall AUC = 0.57, F1 score = 0.5]. Our ML score showed a better performance than existing scores (AUROC: ML score = 0.90 vs. Thrombolysis In Myocardial Infarction (TIMI) score: 0.56, Global Registry of Acute Coronary Events (GRACE) score: 0.52, Acute Heart Failure (ACUTE-HF) score: 0.65; all P < 0.05). Machine learning score also showed excellent performance in the external cohort (AUROC = 0.88). Conclusion This new ML score is the first to demonstrate improved performance in predicting in-hospital outcomes over existing scores in patients admitted to the intensive care unit based on seven simple and rapid clinical and echocardiographic variables. Trial Registration ClinicalTrials.gov Identifier: NCT05063097
Double-network polysaccharide hydrogel for guided tissue repair
No full textInternational audienceThe increasing need for biocompatible and sustainable materials has highlighted the potential of natural-based polymers in tissue engineering, particularly due to their bioactivity, degradability, and ability to mimic the extracellular matrix. Polysaccharide-based membranes are especially promising for guided tissue regeneration (GTR) applications, thanks to their biocompatibility, resorbability, and capacity to recreate biological environments. However, their limited mechanical properties present challenges for practical handling during implantation. In this study, double-network polysaccharide hydrogels were developed to enhance the mechanical robustness of polysaccharide membranes for tissue engineering purposes. By optimizing synthesis parameters, a biphasic membrane was achieved, comprising a non-porous side to serve as a physical barrier and a porous side to facilitate cellular infiltration during GTR. Sterilization via gamma irradiation did not compromise the structural integrity or implantability of the membranes. Furthermore, in vivo studies using a mouse subcutaneous model demonstrated a barrier effect, confirming the suitability of these membranes for guided tissue repair. These findings demonstrate the potential of engineered polysaccharide membranes as versatile and effective materials in regenerative medicine
Correlates of food insecurity among university students in a socioeconomically disadvantaged area of the Paris suburbs: A cross-sectional study
No full textInternational audienceDespite increasing political and scientific interest in food insecurity (FI) among higher education students, data for Europe remain scant. We assessed the prevalence of FI at a French university located in a disadvantaged area in the outskirts of Paris, and its associations with perceived academic dropout and socioeconomic, demographic, and lifestyle factors. Methods: We conducted a cross-sectional online survey among 5068 students (22% of overall population, 66% women). FI status was defined by a three-level class variable (quantitative FI: “not having access to enough food”, qualitative FI: “not having access to the desired food”, food security: “having access to enough desired food”). Multivariate multinomial logistic regression models found associations between (i) FI and academic dropout and (ii) socioeconomic, demographic, academic-related data, cooking and eating conditions, and FI. Results: Students reporting quantitative FI (11%) or qualitative FI (35%) more often experienced academic dropout ( p <0.0001). Men more often reported quantitative FI ( p < 0.0001). Living in a collective residence, lacking sufficient household cooking facilities, experiencing financial difficulties, using food assistance, being an undergraduate, having obtained a high-school diploma abroad, not receiving food from family, regularly eating alone, and infrequent cooking were positively associated with both quantitative and qualitative FI ( p < 0.0001). Conclusion: We found elevated rates of FI among university students and an association between FI and academic dropout. Structural and behavioural factors were found to be associated with FI. These findings provide insight into the characteristics of those students most likely to experience FI and suggest testable preventive actions
Families' Experience of Family Therapy by Videoconference During the First Lockdown: A Qualitative Study
No full textInternational audienceTo ensure continuity of care during the Covid‐19 pandemic, family therapy sessions were maintained via videoconferencing. Few studies evaluate the effectiveness of this modality. This exploratory qualitative study gathered the experiences of families who underwent videoconferencing family therapy during the first French lockdown and in‐person therapy before and after. This study was conducted in France, within a department of adolescent medicine and psychiatry. Ten semi‐structured interviews were conducted with eight families and analyzed using Interpretative Phenomenological Analysis. Results show that the absence of travel constraints in videoconferencing limited engagement during sessions. At home, daily life disrupted immersion, close proximity censored conversations, and therapists' virtual presence was perceived as intrusive or alliance‐reinforcing. Patients described the sessions as a hiatus in therapy, citing therapists' exclusion from the family system, the paradox of being in therapy without full participation, and weak integration into the family narrative. However, videoconferencing provided crucial support during the crisis, encouraged initiative‐taking, enabled participation of typically absent members, and allowed for experimentation with new formats. Some patients proposed alternating between in‐person and remote sessions in the future. These challenges highlight the disruption of the therapeutic system during the abrupt transition. Therapists, shifting from active participants to observers, struggled to maintain a cohesive system. Yet, videoconferencing can redistribute control, fostering family initiative and balancing group and individual identities. This tool could complement traditional family therapy. Further research is needed to identify conditions, populations, and stages of therapy where videoconferencing is most effective
The contribution of a quantitative biographical approach in studying the social and health trajectories of immigrant and French-born women receiving breast cancer care in the greater Paris area
No full textInternational audienc
Apprentissage adaptatif appliqué à la détection de la fraude
No full textPayment fraud, which affects banking institutions, can occur through various channels (e.g., checks, credit cards, bank transfers) and may result in significant financial losses or customer inconvenience, particularly in cases of false alerts. Combating fraud is therefore essential for banks, leading to a continuous cycle of conflict between security teams, who implement detection and blocking algorithms, and fraudsters, who evolve their strategies to bypass these measures. Despite efforts to employ machine learning approaches, the losses caused by fraud remain substantial, amounting to billions of dollars annually. This raises important questions about the suitability of traditional machine learning (ML) models, which are inherently more static and, therefore, potentially less effective in a dynamic environment. In this thesis, we explore the field of adaptive machine learning, which specifically addresses problems involving dynamic cause-and-effect relationships. In this context, we studied and proposed solutions to evaluate the efficiency of batch incremental models compared to instance incremental models, taking into account real-world challenges in fraud detection, such as class imbalance and label delay. Moreover, as ML solutions become increasingly sophisticated, it has become common for data scientists to use interpretability tools (such as SHAP and LIME), even though the reliability of this approach remains questionable. We have thus investigated to what extent inherently interpretable (II) models can serve as a viable alternative to enhance trust in ML systems applied to fraud detection. To this end, we proposed an II model based on the attention mechanism, which offers the advantage of producing stable explanations, in addition to being both interpretable and predictively efficient. Finally, our study on real-world bank transfer data allowed us to test the proposed and existing methods against the constraints of the industrial environment, resulting in concrete recommendations for data scientists and/or researchers working on this topic.La fraude aux moyens de paiement, dont sont victimes les institutions bancaires, peut emprunter plusieurs canaux (par exemple : chèque, carte bancaire, virement) et entraîner d'importantes pertes financières ou des désagréments pour les clients, notamment en cas de fausses alertes. La lutte contre la fraude est donc une nécessité pour les banques et se traduit par un cycle conflictuel entre les équipes dédiées à la sécurité, qui mettent en place des algorithmes de détection et de blocage, et les fraudeurs, qui adaptent leurs stratégies pour les contourner. Malgré les efforts visant à utiliser des approches de machine learning, les pertes liées à la fraude restent significatives, se chiffrant à des milliards de dollars chaque année. Il devient alors crucial de questionner l'efficacité des modèles de machine learning (ML) classiques, qui sont par nature plus statiques et, par conséquent, peut-être moins adaptés à un environnement en constante évolution. Dans cette thèse, nous explorons le domaine du machine learning adaptatif, qui vise précisément à résoudre des problèmes comportant des relations de cause à effet dynamiques. À cet égard, nous avons étudié et proposé des solutions pour évaluer l'efficacité des modèles incrémentaux par batch ou par lot, comparés à ceux incrémentaux par instance, en tenant compte des défis réels de la détection de fraude tels que le déséquilibre des classes et le retard des étiquettes. Par ailleurs, avec la sophistication croissante des solutions ML, il devient courant pour les data scientists d'utiliser des outils d'interprétabilité (comme SHAP ou LIME), bien que la fiabilité de cette approche reste discutable. Nous avons donc exploré dans quelle mesure l'utilisation de modèles intrinsèquement interprétables (II) peut constituer une alternative pertinente pour renforcer la confiance dans les systèmes ML appliqués à la lutte contre la fraude. Dans cette optique, nous avons proposé un modèle II basé sur le mécanisme d'attention, qui offre l'avantage de produire des explications stables, en plus d'être à la fois interprétable et performant sur le plan prédictif. Enfin, notre étude sur des données réelles de virements bancaires a permis de confronter les méthodes proposées et existantes aux contraintes du monde industriel, aboutissant à la formulation de recommandations concrètes pour les data scientists et/ou chercheurs travaillant sur ce sujet