Francophone SIG Workshop at MIE 2022

De-Identification of Electronic Health Records (EHRs) has become a necessity for data privacy protection and government regulatory compliance. To solve this problem, several natural language processing techniques have been applied to English medical text and achieved high performance. However, the limited availability of labeled data sets in other languages and the expensive manual annotation costs have severely limited the applicability of these supervised techniques in other languages, such as French. To address this issue, we created an automatically labeled dataset from the French hospitalization reports, annotated for the extraction of Personal Identifiable Information (PII) for instance, (full names of doctors, full names of patients, dates, cities, postal addresses, emails and phone numbers) and we trained a BERT+Bi-LSTM+CRF Named Entity Recognition (NER) Model using this automatically labeled data. To evaluate the model performance, we manually annotated a subset of hospitalization reports and we used Averaged F-score as a metric to evaluate our NER system. The experimental results show that the model can achieve high performance (about 94\% in Averaged F1-score) using a large training dataset. This procedure can reduce the cost and time of manual annotation of personal information in medical reports. In the future, we would like to investigate active learning to improve our model in the loop during annotation and federated learning for a collaborative training approach.

Sharing clinical data is fundamental for clinical practice and for biomedical research. Nevertheless, the required deidentification of data is so challenging that few robust tools are existing neither shared data. We present a hybrid approach and an available functional tool for French, based on 3 parallel techniques: knowledge-based, rule-based and machine-learning-based. These three approaches as well as the hybrid one are evaluated against three corpuses.

The implementation of a reliable identification process is the basis for any secure patient information sharing system. Each individual is indeed unique and could be identified with a unique number (identifier). To address these challenges, we developed a method of unique patient identification that is suited to the Burkina Faso context. The suggested method takes its roots in the work of the Group for the Modernization of the Hospital Information System (GMSIH) [1]. The developed model assigns a "Unique Identifier" (IDpatient) to each patient based upon their profile of identification attributes (name, birth date, gender, etc.). The IDpatient is a sequence of twenty characters plus a two-character security "key". A reliability test was performed to take into account identity anomalies (duplicate, collision).

Emergency phone triage aims at identifying quickly patients with critical emergencies. Patient triage is not an easy task, especially in situations involving children, mostly due to the lack of training and the lack of clinical guidelines for children. To overcome these issues, we aim at designing and assessing an interactive interface for displaying recommendations on emergency phone triage in pediatrics. Four medical students formalized local guidelines written by the SAMU of Paris, into a decision tree and designed an interface according to usability principles. The navigation within the interface was designed to allow the identification of critical emergencies at the beginning of the decision process, and thus ensuring a quick response in case of critical emergencies. The interface was assessed by 10 medical doctors: they appreciated the ergonomics (e.g., intuitive colors), and found easy to navigate through the interface. Nine of them would like to use this interface during phone call triage. In the future, this interface will be improved and implemented in emergency call centers.

Increasing efficiency and completeness of identifying potentially eligible patients for entry into pneumo-oncology trials is possible by developing an automatic language processing tool to check inclusion criteria. Algorithms based on regular expressions were implemented and evaluated for eight features to automate extraction among studies protocols. The detection performance of the features reached 79% of F-measure on macro average among protocols. Based on previous researches, the tool identified at least one eligible subject from the Tours hospital data warehouse for the majority of protocols and at least one protocol was proposed for 26% of those subjects. Testing the tool on different centers would be necessary to assess its external validity and allow interoperability of data warehouses in the perspective of multicentre recruitment.

This paper presents the acute Heart Failure (aHF) corpus of French clinical texts produced during clinical care. The aHF corpus is annotated for two tasks: text classification and information extraction. The aHF corpus currently consists of 1,639 hospital stays reports annotated for the text classification task, and 100 hospital stays reports annotated for the information extraction task. We detail the methodology set up with clinicians and the annotation of the corpus. This corpus is a fundamental step for the GAVROCHE study, whose main objective is assessing the prognosis value of blood sugar variability in patients hospitalized for aHF.