Chronic kidney disease (CKD) is often underdetected in hospitalized older adults due to incomplete coding and laboratory data. Using eHOP, the clinical data warehouse of Rennes University Hospital, we developed and validated a rule-based interpretable multimodal algorithm combining laboratory and textual information for CKD phenotyping. Among 15,112 hospital visits (7,556 with CKD and 7,556 without), the algorithm achieved an accuracy of 92.1 % in agreement with ICD-10, recall 96.5 %, and precision 86.6 %. Validation on a clinician-reviewed subset confirmed the reliability of classifications. Most CKD cases were detected from clinical text. This approach demonstrates the feasibility of transparent multimodal phenotyping within hospital data and provides a foundation for future scalable decision-support systems in chronic disease surveillance.

This study aims to compare healthcare utilization between patients with stage 3–5 chronic kidney disease (CKD) and patients with normal estimated glomerular filtration rate (n-eGFR) to identify differences in healthcare needs. This retrospective observational study included patients hospitalized in Western France between 2015 and 2018. CKD was defined by an estimated glomerular filtration rate (eGFR) < 60 mL/min/1.73 m2 measured at least 90 days apart or a recorded CKD diagnosis without eGFR > 90 mL/min/1.73 m2, corresponding to KDIGO stages 3–5. Patients with n-eGFR had eGFR consistently ≥ 90 mL/min/1.73 m2 without any acute kidney injury episode. Clinical hospital data were then linked to the French National Health Data System (SNDS) to determine 2015 healthcare utilization, categorized into inpatient (nephrology-related and -unrelated hospitalizations) and outpatient care (consultations with general practitioners, nephrologists, cardiologists, endocrinologists). Overall, 26,125 stage 3–5 CKD and 292,559 patients with n-eGFR were included. Hospitalization duration increased with age and CKD severity. Nephrology-related hospitalizations declined with age, while unrelated ones rose, especially among patients with stage 5 CKD aged > 85 years (~ 1,800 per 1,000 person-years). General practitioners’ consultation rates increased with age and CKD stage, peaking at 9,000 per 1,000 person-years in > 85-year-olds with stage 5 CKD. Nephrologist consultations were most frequent in younger stage 5 CKD patients and decreased with age. Stage 3–5 CKD and age strongly affect healthcare utilization, underscoring the need for tailored care and better coordination, especially in older patients and advanced CKD.

Cerebral vasospasm is a serious complication after aneurysmal subarachnoid haemorrhage (aSAH). We trained machine learning models on 168 patients (225 CTA timepoints) using standardized clinical data (OMOP CDM). XGBoost achieved the best performance (AUROC 0.79; 95% CI 0.65-0.91).

Clinical data warehouses (CDWs) provide a robust foundation for generating real-world epidemiological indicators from electronic health records. By exploiting routinely collected hospital data, these infrastructures enable near real-time monitoring of infectious diseases. Their value increases when complemented by external data streams, including web search queries or meteorological variables. Evidence from recent studies shows that combining these heterogeneous signals with statistical models, natural language processing (NLP), and machine learning (ML) enhances both the timeliness and accuracy of outbreak detection and forecasting. The project aims to combine novel methods and apply them to our local CDW by considering the changes in open data.

This study evaluates the Avatar method for generating synthetic health data while preserving privacy. Using a cancer prediction dataset of 1,500 patients, we analyzed the balance between data utility and privacy protection across different parameter settings. Results show that avatars can approximate the statistical structure of the original data (utility metric: 97.96%, Hellinger distance: 0.13) while reducing re-identification risks (privacy rate: 91.7%, hidden rate: 92.4%). However, the study highlights that the choice of parameters, particularly the neighborhood size k, is critical to achieving a suitable utility-privacy trade-off. Careful tuning is therefore required before applying Avatarization in real clinical contexts.

Within the INTERREG Caribbean VI SAVOIR project, French EDSAN digital health training modules were adapted into a multilingual programme aligned with European standards for the Latin America and Caribbean region. Two hybrid workshops in 2025 combined expert-led sessions and gamified learning on health data, cybersecurity, and regulation. Involving 102 participants from eight countries, the initiative achieved high engagement and satisfaction, supporting a scalable model for disseminating European digital health standards in the LAC region.

Clinical reports contain valuable patient information but are difficult to use due to their unstructured format and privacy constraints. We present MINE, a secure and interactive platform that enables experts to collaboratively develop, evaluate, and deploy clinical information extraction models. MINE integrates rule-based, BERT-based, and LLM-based methods within a unified environment. Early results on smoking status, diabetes, and hypertension achieved F-scores above 95%, demonstrating the feasibility of expert-guided extraction.

Building federated infrastructure for European ICU data requires establishing standards for semantic interoperability across heterogeneous clinical settings. Even with standardized terminologies like LOINC and SNOMED CT, a single clinical concept may correspond to hundreds of alternative codes, precluding interoperability without explicit harmonization guidance. We developed the INDICATE Minimal Data Dictionary through an iterative, consensus-based process involving clinical experts and data scientists. The dictionary defines standardized OHDSI concept sets across nine clinical domains and supports six concrete use cases. Implementation is supported by a web-based application that provides expert-curated guidance for terminology selection and mapping. This approach reduces terminology ambiguity while remaining adaptable to local ICU practices, enabling scalable data harmonization across 15 European data providers from 12 countries.

BACKGROUND:Patient recruitment for clinical trials remains a major challenge, with 86% of trials failing to meet enrollment targets on time. In over 77% of cases, recruitment difficulties stem from matching problems between trials and patients. Case-Based Reasoning (CBR) offers a distinct patient-to-patient approach by determining eligibility through comparison with previously enrolled patients, yet this methodology remains underexplored in contemporary oncology trial matching despite its potential advantages. OBJECTIVE:To compare the performance of two CBR approaches-random forest (RF) and target patient similarity (TPS)-in predicting patient eligibility for recent oncology clinical trials using real-world electronic health record data. METHODS:We selected three breast cancer clinical trials (2019-2022) from our institutional registry. Patient data were extracted from our clinical data warehouse, including structured data (laboratory results, diagnosis codes, procedures, treatments) and unstructured clinical narratives processed using natural language processing. For each trial, we trained RF classifiers and TPS models using repeated hold-out validation (25 splits, 70/30 train-test). Performance was evaluated using discriminative metrics (AUC, positive precision, recall, F1-score) and ranking metrics (P@5, P@10, MAP, MRR, NDCG@5, NDCG@10). We analyzed model performance across varying numbers of eligible patients in training datasets (2 to 70% of the total number of eligible patients). RESULTS:Both approaches demonstrated strong discriminative performance across three trials, with average AUCs of 84.1 % for RF and 76.4 % for TPS, driven primarily by high recall (82.3 % and 77.7 %, respectively). However, positive precision remained low (13.3 % and 9.9 %), reflecting high false-positive rates due to class imbalance. RF showed superior ranking performance, particularly for the trial with the largest eligible cohort (n = 542; P@5 = 78.6 %, MRR = 88.0 %), compared to TPS (P@5 = 47.9 %, MRR = 69.2 %). Both approaches reached performance plateaus with only around 10 eligible patients in training datasets. Variable importance analysis revealed that treatment-related features, diagnostic codes, and procedures were consistently the most important predictors, with relevant patterns identified even with minimal training data. CONCLUSIONS:CBR approaches can effectively support patient pre-screening for oncology clinical trials, with RF demonstrating moderately superior performance over TPS. Both methods show robust discriminative performance with small training datasets, though ranking performance varies substantially across trials. Our findings suggest that CBR approaches may benefit from integration with query-based or prompt-based methods during early recruitment phases when training data is scarce.

Surgical Site Infections (SSIs) are a key target for Healthcare-Associated Infections surveillance. In France, the SPICMI program monitors SSI using hospital medico-administrative and microbiological data. At Martinique University Hospital (MUH) surveillance remains time-consuming manual due to limited data integration. This study implemented a semi-automated SSI detection method at MUH using the SPICMI protocol to identify SSIs suspected cases. The new algorithm detected 85 high and 36 moderate suspicion cases. Manual review confirmed 27 high and 3 moderate cases, improving SSI monitoring and aligning local practices with national surveillance standards.

Epstein-Barr virus infects over 90% of the adults and timely detection of severe infection or reinfection is essential to avoid complications. Since January 2024, the new serology kit to detect EBV was used at Rennes University Hospital, and virologists observed discrepancies between serologies and PCR results. We developed an algorithm using the clinical data warehouse to identify patients with negative EBV serology, but who could be positive at PCR. The algorithm is based on searching keywords associated with the symptoms/characteristics of the EBV infections in clinical notes of patients with a negative EBV serology. By crossing the criteria, we detected potential FN patients. The virologists are alerted by an email sent automatically and could provide better care.

The Observational Medical Outcomes Partnership Common Data Model (OMOP CDM) requires drug data to be encoded in RxNorm for interoperable and dose-aware analyses. In France, medications are identified using UCD and CIP codes, but no direct mapping exists between UCD and RxNorm. We developed an automated three-stage pipeline to map French UCD codes to RxNorm Clinical Drugs while preserving dosage information. The process combined reference-based mapping using the French Drug Nomenclature (RUIM) and OHDSI Standardized Vocabularies, rule-based similarity matching, and agentic AI reasoning, with expert validation at each stage. Applied to the 38,345 UCD codes from the French national nomenclature, the pipeline achieved a 97.4% mapping rate with 86.0% precision when evaluated on the 500 most frequently used codes. This approach provides a reproducible and scalable framework for drug terminology harmonization in OMOP ETL processes.

Dengue fever, also known as “tropical flu”, is an infectious disease caused by the dengue virus, which most often occurs in tropical and subtropical regions. To combat this virus and prevent dengue epidemics in France, Santé publique France (SpFrance) has set up both individual and collective actions. In Martinique, the Programme de Surveillance, d'Alerte et de Gestion des Epidémies de dengue (PSAGE dengue), set up by SpFrance in 2006, coordinates these collective actions. It brings together a few local players involved in both clinical surveillance (general practitioners, Martinique University Hospital, laboratories, etc.) and vector surveillance (Centre de Démoustication et de Recherches Entomologiques - Lutte Antivectorielle; CEDRE-LAV). At present, there is a delay of several weeks between the increase in the number of dengue cases detected in the field and the declaration of the epidemic phase by the PSAGE. The main objective of this study is to use Machine Learning algorithms to predict dengue epidemics in Martinique. Different Machine Learning algorithms using heterogeneous real-life data (administrative and clinical data, laboratory data, Google Trends data and entomological data) were evaluated and compared on their ability to predict the 2019-2021 dengue epidemic. The best-performing model was the random forest model (correlation at 0.933 [0.915 ;0.947]). This study highlighted the value of integrating real-life data into the surveillance and prediction of dengue epidemics in Martinique. The results show that the use of multiple data sources, such as clinical, entomological and Google Trends data, improves disease surveillance. Thus, the implementation of these different sources can contribute to the prediction of future epidemics.

Background: Clinical data warehouses store large volumes of unstructured text containing valuable information about patients' medical status. Traditional extraction systems based on named entity recognition (NER) identify medical terms but often fail to capture the contextual cues needed for accurate interpretation. Existing approaches to context-aware extraction differ in their reliance on expert annotation, computational power, and lexical resources, leading to uneven feasibility across institutions. Combined with heterogeneity in documentation practices and data-sharing restrictions, these limitations hinder the scalability and reuse of trained models. There is thus a need for practical frameworks that can be deployed and adapted locally within medical institutions. Objective: This study aimed to introduce the Medical Status Extraction Pipeline (MSEP), a methodological framework that extracts patients' medical status from clinical narratives through sentence classification and supports the local deployment of hybrid extractors, illustrated through an institutional case study. Methods: MSEP extracts medical status by classifying sentences into predefined categories (presence, absence, or unknown) for each targeted condition. The pipeline combines modules for data selection, expert annotation, and model development, with parameters customizable to different settings. It was applied within our institutional environment on 6 conditions: smoking, hypertension, diabetes, heart failure, chronic obstructive pulmonary disease, and family history of cancer, using 12,119 manually annotated sentences from the eHOP Clinical Data Warehouse (Rennes University Hospital). Three types of extractors were compared: fine-tuned CamemBERT, large language model (LLM) prompt, and a rule-based baseline, evaluated through stratified 3-fold cross-validation, measuring precision, recall, specificity, macro F-score, balanced accuracy, as well as manual annotation time and model inference speed. Results: Among the tested approaches, the CamemBERT-based extractor achieved the best overall performance, with macro F-scores above 0.94 for 5 of the 6 medical conditions. The study also highlights that when a medical status is very sparsely represented in the training data, rule-based extractors can outperform learned models (average macro F-score 0.94 vs 0.73 for family history of cancer). This shows the pragmatic value of choosing the extraction method according to data availability. Manual annotation time per sentence ranged from 1.2 to 2.9 seconds within the pipeline (2.23 to 4.25 seconds for informative sentences), compared with 7.8 to 16.5 seconds for named entity recognition-based systems. In our institutional experiments, the minimum time to complete all pipeline modules, from dataset construction to final extractor refinement, was 8 hours. Conclusions: In our institutional case study, MSEP enabled rapid construction of datasets and extractors across multiple clinical conditions while reducing the effort required for local development. Its modular and configurable design allowed the adoption of hybrid extraction approaches and adaptation to different resource settings. These features highlight MSEP's value as a research tool and upstream component that facilitates local deployment of clinical information extraction workflows.

Federated Learning (FL) allows institutions to train shared models without exchanging raw data, making it a promising approach for healthcare applications that involve sensitive electronic health records (EHRs). However, despite this distributed design, the gradients exchanged during training can still reveal private information. In this study, we analyze how vulnerable transformer-based language models are to gradient inversion attacks, focusing on the Decepticons method, which can reconstruct original training text from shared gradients. We simulate a cross-silo FL setup with three types of French clinical reports (genetic, anesthesia, and birth records) to evaluate how batch size and sequence length affect reconstruction quality. Our experiments show that a malicious server can recover clinical text with high accuracy: token-level recovery exceeded 95% when training with batch size 1 and remained above 60% for sequences of up to 512 tokens. Reconstructed examples contained identifying elements (names, dates, genetic markers), revealing serious privacy risks for real-world use. These results emphasize that FL alone is insufficient for sensitive clinical text and that privacy-preserving defenses must be integrated before real-world deployment.

Medication data are fragmented across national and hospital databases, limiting consistent longitudinal assessment of drug exposure. Using data from the multicentre HUGOSHARE infrastructure linking university hospitals in Western France to the French national claims (SNDS), we quantified longitudinal medication exposure in oncology patients (2015-2019). Drug identifiers were harmonised via Thériaque (CIP/UCD to ATC level 5) to enable cross-source comparability. Among 57,960 patients, the prevalence of polypharmacy (≥5 drugs/day) was 34% pre-admission, 43% during hospitalisation, and 16% post-discharge; hyper-polypharmacy (≥10 drugs/day) was 8%, 35%, and 4%, respectively. This study demonstrates the feasibility of scalable, interoperable medication analytics across care transitions and opens perspectives for monitoring drug-related risks and cumulative exposure in real-world data.

Background:Preventing adverse drug reactions requires accurate monitoring of drug exposure throughout patient care. Conventional metrics, measured at admission or discharge, fail to capture the dynamic and cumulative nature of drug burden during hospitalization. Improving exposure assessment is essential to support clinical decision-making and medication safety. Clinical data warehouses (CDWs), which integrate detailed drug administration records, enable the retrospective reuse of hospital data to develop more granular and dynamic measures of in-hospital drug exposure. Objective:This exploratory proof-of-concept study aimed to introduce 2 cumulative drug exposure metrics computed from CDW: cumulative drug exposure (CDE) and cumulative drug exposure density (CDED). The study also aimed to compare these metrics with conventional metrics, primarily as a methodological development for characterizing prescribing determinants in hospitalized patients. Methods:We conducted a retrospective study using the eHOP CDW at Rennes University Hospital. Adults hospitalized for hematological malignancies were included. Four prescribing determinants were analyzed: polypharmacy (PP), hyperpolypharmacy (HPP), drug-drug interactions (DDIs), and potentially inappropriate medications (PIMs). CDE quantified the number of days each determinant was present, while CDED normalized this value to hospital length of stay. Analyses combined descriptive statistics, Spearman correlations, and factorial analysis of mixed data (FAMD). Results:Mean CDE values were 10.5 days for PP, 5.7 for HPP, 64.7 for DDIs, and 19.0 for PIMs (≥65 years). CDED values ranged from 0.3 to 3.2. Conventional metrics at admission were weakly correlated with cumulative exposure measures (eg, DDIs: rs=0.04, P=.752; PP: rs=-0.04, P=.757; HPP: rs=0.11, P=.364). Stronger, significant correlations emerged at discharge (CDE DDIs: rs=0.44, CDED: rs=0.46; both P<.001). PIMs showed strong significant correlations at both time points. FAMD highlighted that cumulative metrics contributed independently to the principal components, capturing dynamics of drug exposure not reflected by conventional indicators. Conclusions:CDE and CDED, derived from real-time CDW data, offer reproducible and scalable alternatives to conventional metrics for characterizing drug exposure in patients hospitalized for severe conditions. They provide a more accurate characterization of drug burden and hold promise for pharmacoepidemiological research and clinical decision support.

Clinical data warehouses (CDWs) often contain missing data due to non-systematic collection. Characterizing this missingness through descriptors of rate, pattern, and mechanism helps assess dataset quality and understand real-world data. We first identified relevant qualification methods for clinical data through a scoping review. We then analyzed the behavior of the descriptors they provide, by implementing them in an automated descriptor extraction pipeline. To evaluate our approach, four missing data scenarios were simulated from the same complete dataset, reflecting clinically plausible situations. Descriptor variability across scenarios was tested using paired non-parametric tests (Friedman for numeric, Cochran's Q for binary descriptors). In total, 230 descriptors were extracted for each of 40 incomplete datasets, and 37.4% showed significant adjusted p-values after Benjamini-Hochberg correction. The most discriminant descriptors accurately reflected the scenario structures, demonstrating that descriptor-based qualification can provide interpretable insights into data quality in clinical warehouses.

We present a novel LLM-based approach for medical concept extraction that combines multiple anti-hallucination strategies. Our Streamlit web application enables flexible pipeline configuration through ensemble methods, Chain-of-Verification, contextual grounding, and LLM-as-a-Judge mechanisms. This modular framework addresses the critical challenge of hallucinations in clinical text processing, particularly for smaller models required in privacy-constrained hospital environments.

Chronic kidney disease (CKD) affects 8%–10% of the French population, posing considerable problems to healthcare systems due to its high morbidity and resource demands. However, less is known about healthcare utilization patterns among CKD patients compared to non-CKD patients. This study aims to compare healthcare utilization between these groups to identify differences in healthcare needs. This retrospective observational study included data from all hospitalized patients in several teaching hospitals of the west region of France between January 1, 2015, and December 31, 2018, our study period. Patients’ clinical data were extracted from hospitals’ clinical data warehouses (CDWs) and linked to the National Health Data System (SNDS). The SNDS is a comprehensive database maintained by the French Social Security system that contains all reimbursed outpatient care and inpatient care in France. We identified CKD patients using CDWs’ data. We defined CKD as having an estimated glomerular filtration rate (eGFR) <60 mL/min/1.73 m² measured at least 90 days apart or a documented CKD diagnosis. Patients were stratified by CKD stages (3, 4, or 5). Non-CKD patients were identified as having eGFR values consistently ≥90 mL/min/1.73 m² within study period. Using the SNDS data, healthcare utilization in 2015 was assessed in terms of inpatient care (nephrology-related and unrelated hospitalizations) and outpatient care (consultations with general practitioners (GP), nephrologists, cardiologists, and endocrinologists). Care utilization rates were calculated by age groups and CKD stages. We identified 292,559 non-CKD patients (mean age 53.3 years, 54.7% female) and 26,125 CKD patients (mean age 75.2 years; 36% female). Among those, 53.5% were stage 3, 27% stage 4 and 19.5% stage 5. Hospitalization duration increased with both age and CKD severity, with the longest stays observed in advanced CKD stages (mean duration 7–10 days) among older patients (≥75 years). Nephrology-related hospitalization rates decreased with age beyond 50, while unrelated-nephrology hospitalization rate increased with both age and CKD stages. Patients aged >85 years with stage 5 CKD showed the highest rates of unrelated-nephrology hospitalizations (2000 per 1000 person-years) compared to stage 3 CKD (800) and non-CKD patients (790). Minimal differences in hospitalization rates were observed between non-CKD and stage 3 CKD patients as age increased. GP consultation rates increased with age and CKD severity, peaking in stage 5 patients aged >85 years (over 9000 per 1000 person-years). Nephrologist consultations were mainly observed in stage 5 patients, with higher rates in younger age groups (over 3000 per 1000 person-years for age ≤50). Consultation rates with cardiologists increased with age, peaking in stage 4 CKD among patients aged 75–85 years, at approximately 600 per 1000 person-years. Patients over 85 years old had the lowest consultation rates with endocrinologists, regardless of CKD status. This is the first study in France to compare healthcare utilization between CKD and non-CKD patients. As age increased, CKD patients' rates of nephrologist consultations and related-nephrology hospitalizations decreased, suggesting a shift toward hospitalizations and consultations for comorbidities associated with CKD. Additionally, the results showed that hospitalization rates were similar at older ages, while GP consultation rates were comparable at younger ages between CKD and non-CKD patients. Further analysis will explore gender differences and adherence to French national care guidelines to identify areas of improvement in CKD care.