Search

In France, structured data on emergency room visits are aggregated at the national level to build a syndromic surveillance system for different health events. For visits motivated by a traumatic event, information on the circumstances is stored in free text clinical notes. Automating the processing of these notes should allow the enrichment of surveillance tools. In development at Inserm and the Emergency Department of the Bordeaux University Hospital, The TARPON (for Automatic Processing of Emergency Room Notes for a National Observatory, in French) project aims to meet this objective by using the latest deep learning tools applied to automatic language analysis. To exploit these data, an automatic de-identification system, guaranteeing the protection of personal data, is necessary. We present here a comparison study of models allowing the de-identification of clinical texts in French., En France, les données structurées concernant les visites aux urgences sont agrégées au niveau national pour construire un système de surveillance syndromique de différents événements de santé. Pour les visites motivées par un événement traumatique, les informations sur les circonstances sont stockées dans des notes cliniques en texte libre. Automatiser le traitement de ces notes devrait permettre l'enrichissement des outils de surveillance. En développement à l'Inserm et au Service des urgences du CHU de Bordeaux, le projet TARPON (Traitement Automatique des Résumés de Passages aux urgences pour un Observatoire National) vise à répondre à cet objectif par le biais des derniers outils d'apprentissage profond appliqués à l'analyse automatique du langage. Pour exploiter ces données, un système de désidentification automatique, garantissant la protection des données personnelles est nécessaire. Nous présentons ici une étude de comparaison de modèles permettant la désidentification des textes cliniques en français.

In France, structured data on emergency room visits are aggregated at the national level to build a syndromic surveillance system for different health events. For visits motivated by a traumatic event, information on the circumstances is stored in free text clinical notes. Automating the processing of these notes should allow the enrichment of surveillance tools. In development at Inserm and the Emergency Department of the Bordeaux University Hospital, The TARPON (for Automatic Processing of Emergency Room Notes for a National Observatory, in French) project aims to meet this objective by using the latest deep learning tools applied to automatic language analysis. To exploit these data, an automatic de-identification system, guaranteeing the protection of personal data, is necessary. We present here a comparison study of models allowing the de-identification of clinical texts in French., En France, les données structurées concernant les visites aux urgences sont agrégées au niveau national pour construire un système de surveillance syndromique de différents événements de santé. Pour les visites motivées par un événement traumatique, les informations sur les circonstances sont stockées dans des notes cliniques en texte libre. Automatiser le traitement de ces notes devrait permettre l'enrichissement des outils de surveillance. En développement à l'Inserm et au Service des urgences du CHU de Bordeaux, le projet TARPON (Traitement Automatique des Résumés de Passages aux urgences pour un Observatoire National) vise à répondre à cet objectif par le biais des derniers outils d'apprentissage profond appliqués à l'analyse automatique du langage. Pour exploiter ces données, un système de désidentification automatique, garantissant la protection des données personnelles est nécessaire. Nous présentons ici une étude de comparaison de modèles permettant la désidentification des textes cliniques en français.

To build a French national electronic injury surveillance system based on emergency room visits, we aim to develop a coding system to classify their causes from clinical notes in free-text. Supervised learning techniques have shown good results in this area but require a large amount of expert annotated dataset which is time consuming and costly to obtain. We hypothesize that the Natural Language Processing Transformer model incorporating a generative self-supervised pre-training step can significantly reduce the required number of annotated samples for supervised fine-tuning. In this preliminary study, we test our hypothesis in the simplified problem of predicting whether a visit is the consequence of a traumatic event or not from free-text clinical notes. Using fully re-trained GPT-2 models (without OpenAI pre-trained weights), we assess the gain of applying a self-supervised pre-training phase with unlabeled notes prior to the supervised learning task. Results show that the number of data required to achieve a ginve level of performance (AUC>0.95) was reduced by a factor of 10 when applying pre-training. Namely, for 16 times more data, the fully-supervised model achieved an improvement, Version of the published manuscript

National audience; Des modèles basés sur l'architecture Transformer qui intègrent une étape de pré-entrainement non supervisé à objectif prédictif, tels que le GPT-2 (Generative Pretrained Transformer 2) ont atteint récemment des succès remarquables. Nous avons adapté et mis en oeuvre un modèle de traitement automatique du langage naturel (NLP pour Natural Language Processing) permettant de déterminer si un texte libre clinique est de nature traumatique ou non. Nous avons comparé cette approche, nécessitant un nombre d'échantillons annotés réduit, à une approche entièrement supervisée. Nos résultats (basés sur l'AUC et le F1-score) montrent qu'il est possible d'adapter un modèle polyvalent tel que le GPT-2 pour créer un outil puissant de classification de notes de texte libre en français avec seulement un très faible nombre d'échantillons labélisés.

In order to build a national injury surveillance system based on emergency room (ER) visits we are developing a coding system to classify their causes from clinical notes in free-text. Supervised learning techniques have shown good results in this area but require large number of annotated dataset. New levels of performance have been recently achieved in neural language models (NLM) with models based on the Transformer architecture incorporating an unsupervised generative pre-training step. Our hypothesis is that methods involving a generative self-supervised pre-training step can significantly reduce the required number of annotated samples for supervised fine-tuning. In this case study, we assessed whether we could predict from free-text clinical notes whether a visit was the consequence of a traumatic or non-traumatic event. Using fully re-trained GPT-2 models (without OpenAI pre-trained weightings), we compared two scenarios: Scenario A (26 study cases of different training data sizes) consisted in training the GPT-2 on the trauma/non-trauma labeled (up to 161 930) clinical notes. In Scenario B (19 study cases), a first step of self-supervised pre-training phase with unlabeled (up to 151 930) notes and the second step of supervised fine-tuning with labeled (up to 10 000) notes. Results showed that, Scenario A needed to process >6 000 notes to achieve good performance (AUC>0.95), Scenario B needed only 600 notes, gain of a factor 10. At the end case of both scenarios, for 16 times more data (161 930 vs. 10 000), the gain from Scenario A compared to Scenario B is only an improvement of 0.89% in AUC and 2.12% in F1 score. To conclude, it is possible to adapt a multi-purpose NLM model such as the GPT-2 to create a powerful tool for classification of free-text notes with only very small number of labeled samples.

International audience; Dans cet article, nous présentons le schéma d'annotation utilisé pour étudier les interactions aliments-médicaments (Food-drug interaction-FDI). Le corpus se compose de 639 résumés d'articles scientifiques issus de Medline. Nous avons défini un schéma d'annotation constitué de 21 catégories d'entités et de 21 types de relations appliquées sur 9 catégories d'entités. Ces schémas ont été appliqués sur des documents rédigés en anglais ou en français, ouvrant la voie à un corpus multilingue annoté au moyen des mêmes catégories. Nous présentons également quelques expériences d'identification automatique des types de relations. L'adaptation de domaine à partir des interactions médicament-médicament (DDI) permet d'avoir un schéma d'annotation des relations selon 4 types. L'extraction automatique de ces relations conduit à une F1-mesure de 0.79 obtenue avec un modèle SVM précédé d'un processus de sélection de descripteurs SFM.

Non-compliance situations happen when patients do not follow their prescriptions and take actions that lead to potentially harmful situations. Although such situations are dangerous, patients usually do not report them to their physicians. Hence, it is necessary to study other sources of information. We propose to study online health fora. The purpose of our work is to explore online health fora with supervised classification and information retrieval methods in order to identify messages that contain drug non-compliance. The supervised classification method permits detection of non-compliance with up to 0.824 F-measure, while the information retrieval method permits detection non-compliance with up to 0.529 F-measure. For some fine-grained categories and new data, it shows up to 0.65-0.70 Precision.

International audience; Clinical information in electronic health records (EHRs) is mostly unstructured. With the ever-increasing amount of information in patients' EHRs, manual extraction of clinical information for data reuse can be tedious and time-consuming without dedicated tools. In this paper, we present SmartCRF, a prototype to visualize, search and ease the extraction and structuration of information from EHRs stored in an i2b2 data warehouse.

International audience; In the health-related field, drug non-compliance situations happen when patients do not follow their prescriptions and do actions which lead to potentially harmful situations. Although such situations are dangerous, patients usually do not report them to their physicians. Hence, it is necessary to study other sources of information. We propose to study online health fora with information retrieval methods in order to identify messages that contain drug non-compliance information. Exploitation of information retrieval methods permits to detect non-compliance messages with up to 0.529 F-measure, compared to 0.824 F-measure reached with supervized machine learning methods. For some fine-grained categories and on new data, it shows up to 0.70 Precision.

International audience; Introduction: Drug interactions could account for 1% of hospitalizations in the general population and 2-5% of hospital admissions in the elderly. However, few data are available on the drugs concerned and the potential severity of the interactions encountered. We thus first aimed to estimate the prevalence of dispensings including drugs Contraindicated or Discommended because of Interactions (CDI codispensings) and to identify the most frequently involved drug pairs. Second, we aimed to investigate whether the frequency of CDI codispensings appeared higher or lower than the expected for the drugs involved.Methods: We carried out a study using a random sample of all drugs dispensings registered in a database of the French Health Insurance System between 2010 and 2015. The distribution of the drugs involved was described considering active principles, detailing the 20 most frequent ones for both contraindicated or discommended codispensings (DCs). To investigate whether the frequency of CDI codispensings appeared higher or lower than the expected for the drugs involved, we developed a specific indicator, the Drug-drug interaction prevalence study-score (DIPS-score), that compares for each drug pair the observed frequency of codispensing to its expected probability. The latter is determined considering the frequencies of dispensings of the individual drugs constituting a pair of interest.Results: We analyzed 6,908,910 dispensings: 13,196 (0.2%) involved contraindicated codispensings (CCs), and 95,410 (1.4%) DCs. For CCS, the most frequently involved drug pair was "bisoprolol+flecainide" = 5,036); four out of five of the most represented pairs involved cardiovascular drugs. For DCS, the most frequently involved drug pair was "ramipril+spironolactone" = 4,741); all of the five most represented pairs involved cardiovascular drugs. The drug pair involved in the CC with the highest score value was "citalopram+hydroxyzine" (DIPS-score: 3.7; 2.9-4.6); that with the lowest score was "clarithromycin+simvastatin" (DIPS-score: 0.2; 0.2-0.3). DIPS-score median value was 0.4 for CCs and 0.6 for DCs.Conclusion: This high prevalence of CDI codispensings enforces the need for further risk-prevention actions regarding drug-drug interactions (DDIs), especially for arrhythmogenic or anti-arrhythmic drugs. In this perspective, the DIPS-score we develop could ease identifying the interactions that are poorly considered by clinicians/pharmacists and targeting interventions.

Background Pharmacovigilance consists in monitoring and preventing the occurrence of adverse drug reactions (ADR). This activity requires the collection and analysis of data from the patient record or any other sources to find clues of a causality link between the drug and the ADR. This can be time-consuming because often patient data are heterogeneous and scattered in several files. To facilitate this task, we developed a timeline prototype to gather and classify patient data according to their chronology. Here, we evaluated its usability and quantified its contribution to routine pharmacovigilance using real ADR cases. Methods The timeline prototype was assessed using the biomedical data warehouse eHOP (from entrepôt de données biomédicales de l’HOPital) of the Rennes University Hospital Centre. First, the prototype usability was tested by six experts of the Regional Pharmacovigilance Centre of Rennes. Their experience was assessed with the MORAE software and a System and Usability Scale (SUS) questionnaire. Then, to quantify the timeline contribution to pharmacovigilance routine practice, three of them were asked to investigate possible ADR cases with the “Usual method” (analysis of electronic health record data with the DxCare software) or the “Timeline method”. The time to complete the task and the data quality in their reports (using the vigiGrade Completeness score) were recorded and compared between methods. Results All participants completed their tasks. The usability could be considered almost excellent with an average SUS score of 82.5/100. The time to complete the assessment was comparable between methods (P = 0.38) as well as the average vigiGrade Completeness of the data collected with the two methods (P = 0.49). Conclusions The results showed a good general level of usability for the timeline prototype. Conversely, no difference in terms of the time spent on each ADR case and data quality was found compared with the usual method. However, this absence of difference between the timeline and the usual tools that have been in use for several years suggests a potential use in pharmacovigilance especially because the testers asked to continue using the timeline after the evaluation. Electronic supplementary material The online version of this article (10.1186/s12911-018-0667-x) contains supplementary material, which is available to authorized users.

International audience

International audience; Misuses occur when patients do not respect prescriptions and commit actions which can lead to harmful results. Even if such situations are dangerous, patients do not inform medical doctors about such events. To obtain some information on misuses, it becomes necessary to study other sources of information. We propose to concentrate on discussion fora. The purpose of our work is to explore health fora with machine learning methods and to identify messages where users describe or mention drug misuses. Our approach detects the mesuses with up to 0.773 F-measure. This approach can help in routine detection of misuses and to provide material exploitable by health professionals.

International audience; Pharmacovigilance consists in monitoring and preventing the occurrence of adverse drug reactions. This activity can be time-consuming because it requires the collection of both patient and medication information. In this paper, we present two visualization and data mining applications to make this task easier for the practitioner. These tools have been developed and tested using the biomedical data warehouse eHOP (Hospital Biomedical Data Warehouse) of the Rennes University Hospital Centre. The first application is a tool to visualize the patient electronic health record in the form of a timeline. All patient data is collected and displayed chronologically. The usability test of the timeline has been very positive (SUS score 82.5) and the tool is now available for practitioners in their daily practice. The second application is a tool to visualize and search the sequences of a patient cohort. The visual interface allow user to quickly visualize sequences. A query builder allows user to search for sequences in relation with a reference sequence, such as a prescription sequence followed by an abnormal biological value. The sequences are then visually aligned with this reference sequence and ranked by similarity. The GSP (Generalized Sequential Pattern) and Apriori algorithms allow us to display a summary of the sequences list by searching for common sequences and associations. The tool was tested on a use case which consisted in detection of inappropriate drug administration. Compared to a random order, we showed this ranking system saved the practitioner time in this task (to analyze one sequence, 3.49 +/- 3.54 vs. 2.26 +/- 2.86 s, p = 0.0003). These two visualization and data mining applications will help the daily practice of pharmacovigilance.

OBJECTIVES: To introduce and summarize current research in the field of Public Health and Epidemiology Informatics. METHODS: The 2017 literature concerning public health and epidemiology informatics was searched in PubMed and Web of Science, and the returned references were reviewed by the two section editors to select 14 candidate best papers. These papers were then peer-reviewed by external reviewers to provide the editorial team with an enlightened vision to select the best papers. RESULTS: Among the 843 references retrieved from PubMed and Web of Science, two were finally selected as best papers. The first one analyzes the relationship between the disease, social/mass media, and public emotions to understand public overreaction (leading to a noticeable reduction of social and economic activities) in the context of a nation-wide outbreak of Middle East Respiratory Syndrome (MERS) in Korea in 2015. The second paper concerns a new methodology to de-identify patient notes in electronic health records based on artificial neural networks that outperformed existing methods. CONCLUSIONS: Surveillance is still a productive topic in public health informatics but other very important topics in Public Health are appearing. For example, the use of artificial intelligence approaches is increasing.

Background Improvement of coordination of all health and social care actors in the patient pathways is an important issue in many countries. Health Information (HI) technology has been considered as a potentially effective answer to this issue. The French Health Ministry first funded the development of five TSN (“Territoire de Soins Numérique”/Digital health territories) projects, aiming at improving healthcare coordination and access to information for healthcare providers, patients and the population, and at improving healthcare professionals work organization. The French Health Ministry then launched a call for grant to fund one research project consisting in evaluating the TSN projects implementation and impact and in developing a model for HI technology evaluation. Methods EvaTSN is mainly based on a controlled before-after study design. Data collection covers three periods: before TSN program implementation, during early TSN program implementation and at late TSN program implementation, in the five TSN projects’ territories and in five comparison territories. Three populations will be considered: “TSN-targeted people” (healthcare system users and people having characteristics targeted by the TSN projects), “TSN patient users” (people included in TSN experimentations or using particular services) and “TSN professional users” (healthcare professionals involved in TSN projects). Several samples will be made in each population depending on the objective, axis and stage of the study. Four types of data sources are considered: 1) extractions from the French National Heath Insurance Database (SNIIRAM) and the French Autonomy Personalized Allowance database, 2) Ad hoc surveys collecting information on knowledge of TSN projects, TSN program use, ease of use, satisfaction and understanding, TSN pathway experience and appropriateness of hospital admissions, 3) qualitative analyses using semi-directive interviews and focus groups and document analyses and 4) extractions of TSN implementation indicators from TSN program database. Discussion EvaTSN is a challenging French national project for the production of evidenced-based information on HI technologies impact and on the context and conditions of their effectiveness and efficiency. We will be able to support health care management in order to implement HI technologies. We will also be able to produce an evaluation toolkit for HI technology evaluation. Trial registration ClinicalTrials.gov ID: NCT02837406, 08/18/2016.

International audience