Search

The use of international laboratory terminologies inside hospital information systems is required to conduct data reuse analyses through inter-hospital databases. While most terminology matching techniques performing semantic interoperability are language-based, another strategy is to use distribution matching that performs terms matching based on the statistical similarity. In this work, our objective is to design and assess a structured framework to perform distribution matching on concepts described by continuous variables. We propose a framework that combines distribution matching and machine learning techniques. Using a training sample consisting of correct and incorrect correspondences between different terminologies, a match probability score is built. For each term, best candidates are returned and sorted in decreasing order using the probability given by the model. Searching 101 terms from Lille University Hospital among the same list of concepts in MIMIC-III, the model returned the correct match in the top 5 candidates for 96 of them (95%). Using this open-source framework with a top-k suggestions system could make the expert validation of terminologies alignment easier.

In 722 cities of Minas Gerais (Brazil), primary care patients can have their ECGs remotely interpreted by cardiologists of the Telehealth Network of Minas Gerais (TNMG), a public telehealth service. As of December 2014, more than 1.9 million ECGs were interpreted. This study analyzed the database of all ECGs performed by the TNMG on primary care patients from 2009 to 2013 (n=1,101,993). Structured patient data and the results of automated ECG interpretation by the Glasgow Program are described. Mean patient age is 51 years old, 59% of them are women. The average body mass index is 25.9 kg/m2, with an average increase of 0.15 kg/m2 per civil year. Those patients notably have hypertension (33.2%), family history of coronary artery disease (14.5%), smoking (6.9%), diabetes (5.8%), obesity (5.8%) or Chagas Disease (3.0%). Seventy percent of ECGs are normal. This percentage is higher in women (72.3%) and decreases in average by 7.4 every 10 years of life. There are notably 12% of possible myocardial infarction, 10% of possible left ventricular hypertrophy and 8% of possible supraventricular extra systole.

Administrative data can be used for the surveillance of the outcomes of implantable medical devices (IMDs). The objective of this work is to build a web-based tool allowing for an exploratory analysis of time-dependent events that may occur after the implementation of an IMD. This tool should enable a pharmacoepidemiologist to explore on the fly the relationship between a given IMD and a potential outcome. This tool mine the French nationwide database of inpatient stays from 2008 to 2013. The data are preprocessed in order to optimize the queries. A web tool is developed in PHP, MySQL and Javascript. The user selects one or a group of IMD from a tree, and can filter the results using years and hospital names. Four result pages describe the selected inpatient stays: (1) temporal and demographic description, (2) a description of the geographical location of the hospital, (3) a description of the geographical place of residence of the patient and (4) a table showing the rehospitalization reasons by decreasing order of frequency. Then, the user can select one readmission reason and display dynamically the probability of readmission by mean of a Kaplan-Meier curve with confidence intervals. This tool enables to dynamically monitor the occurrence of time-dependent complications of IMD.

Diagnoses and medical procedures collected under the French system of information are recorded in a nationwide database, the "PMSI national database", which is accessible for exploitation. Quality of the data in this database is directly related to the quality of coding, which can be of poor quality. Among the proposed methods for the exploitation of health databases, data mining techniques are particularly interesting. Our objective is to build sequential rules for missing diagnoses prediction by data mining of the PMSI national database.Our working sample was constructed from the national database for years 2007 to 2010. The information retained for rules construction were medical diagnoses and medical procedures. The rules were selected using a statistical filter, and selected rules were validated by case review based on medical letters, which enabled to estimate the improvement of diagnoses recoding.The work sample was made of 59,170 inpatient stays. The predicted ICD codes were E11 (non-insulin-dependent diabetes mellitus), I48 (atrial fibrillation and flutter) and I50 (heart failure).We validated three sequential rules with a substantial improvement of positive predictive value: {E11,I10,DZQM006}={E11} {E11,I10,I48}={E11} {I48,I69}={I48}We were able to extract by data mining three simple, reliable and effective sequential rules, with a substantial improvement in diagnoses recoding. The results of our study indicate the opportunity to improve the data quality of the national database by data mining methods.

The objective of the work is to extract healthcare process quality indicators from the literature, and to evaluate which of them could be automatically computed using routinely collected data from electronic health records (EHRs). A minimal set of data commonly available in EHRs is first defined. The initial bibliographic query enables to identify 8,744 papers, among which 126 papers describe 440 process indicators. 22.3% of indicators can be automatically computed. The computation of the indicators mostly require diagnoses (99%), drug prescriptions (59%), medical procedures (48%), administrative data (30%), laboratory results (20%), free-text reports with basic keyword research (19%), linkage with the patient's previous stays (11%) and dependence assessment (3%). 77.7% of indicators cannot be automatically computed, mostly because they require a linkage with outpatient data (61%), structured data that are usually not available (43%), unstructured data (26%) or the trace of an information that was given to the patient (8%).

The objective of this study is to analyse the length of patient stay in Pediatric emergency department according to diagnosis and the number of patients over a 3 year-period. A survival tree was used, to explore the underlying construct of overcrowding depending of the length of patient stay. The tree was used to cluster 55.183 patients with respect to length of stay where partitioning is based on covariates such as the number of patients, the diagnosis and existence of complementary exams. The hazard ratio test was used to determine optimal partition. The approach is illustrated using Electronic Medical Record Software database available at the Pediatric Emergency Department of Lille University Hospital.

Hospitals have at their disposal large databases that may be considered for reuse. The objective of this work is to evaluate the impact of a drug on a specific laboratory result by analyzing these data. This analysis first involves building a record of temporal patterns, including medical context, of drug prescriptions. Changes in outcome due to these patterns of drug prescription are assessed using short phases of the inpatient stay compared to monotonous changes in the laboratory result. To illustrate this technique, we investigated potassium chloride supplementation and its impact on kalemia. This method enables us to assess the impact of a drug (in its frequent context of prescription) on a laboratory result. This kind of analysis could play a role in post-marketing studies.

Management of vitamin K antagonists (VKA) is difficult, and overdoses can have dramatic hemorrhagic consequences. These works form part of a European computerized medical data processing project, which aims to develop IT tools for describing adverse drug events (ADEs). Materials and methods A tool enabling retrospective research of potential ADE cases was developed, using complex ADE detection rules taking into account chronological parameters: the ADE scorecards. The rules were applied on 14,748 medical records from a community hospital. We evaluated the predictive positive value of the rules related to INR elevation by an expert review of the detected cases. The severity of the clinical consequences was evaluated. Results 49 cases were detected, among which 11 cases were ADEs. The predictive positive value of the rules is 22.44%, mostly related to antibiotics and amiodarone introduction. The four cases of clinical damage related to a drug were properly designated by the rules. Discussion - Conclusion Our study shows the great potential of developing complex rules for the identification of adverse drug events in large medical databases.

Several papers propose to take contexts into account for adverse drug events (ADE) detection and prevention, notably to decrease over-alerting of clinical decision support systems (CDSS). However, no statistical argument has been published till now. This works demonstrates, based on statistical analysis, that contextualization is necessary for ADE detection and prevention by 3 steps. A database of 115,447 inpatients stays from 6 hospitals, and 236 ADE detection rules are used. Step 1: the patients differ significantly between and within hospitals, regarding their medical background, their medication and several outcomes. Step 2: The estimated ADE rates vary between and within hospitals. Step 3: even when comparable conditions are present, the probability of ADE occurrence may differ between and within hospitals. Those 3 steps demonstrate that contextualization is necessary, and pave the way for a statistics-based method to contextualize ADE prevention (CDSS) and ADE detection tools.

Purpose Medical free-text records enable to get rich information about the patients, but often need to be de-identified by removing the Protected Health Information (PHI), each time the identification of the patient is not mandatory. Pattern matching techniques require pre-defined dictionaries, and machine learning techniques require an extensive training set. Methods exist in French, but either bring weak results or are not freely available. The objective is to define and evaluate FASDIM, a Fast And Simple De-Identification Method for French medical free-text records. Methods FASDIM consists in removing all the words that are not present in the authorized word list, and in removing all the numbers except those that match a list of protection patterns. The corresponding lists are incremented in the course of the iterations of the method. For the evaluation, the workload is estimated in the course of records de-identification. The efficiency of the de-identification is assessed by independent medical experts on 508 discharge letters that are randomly selected and de-identified by FASDIM. Finally, the letters are encoded after and before de-identification according to 3 terminologies (ATC, ICD10, CCAM) and the codes are compared. Results The construction of the list of authorized words is progressive: 12h for the first 7000 letters, 16 additional hours for 20,000 additional letters. The Recall (proportion of removed Protected Health Information, PHI) is 98.1%, the Precision (proportion of PHI within the removed token) is 79.6% and the F -measure (harmonic mean) is 87.9%. In average 30.6 terminology codes are encoded per letter, and 99.02% of those codes are preserved despite the de-identification. Conclusion FASDIM gets good results in French and is freely available. It is easy to implement and does not require any predefined dictionary.

Adverse drug events (ADEs) are a public health is sue. Their detection usually relies on voluntary reporting or medical chart reviews. The objective of this paper is to automatically detect cases of ADEs by data mining. 115 447 complete past hospital stays are extracted from six French, Danish, and Bulgarian hospitals using a common data model including diagnoses, drug administrations, laboratory results, and free-text records. Different kinds of outcomes are traced, and supervised rule induction methods (decision trees and association rules) are used to discover ADE detection rules, with respect to time constraints. The rules are then filtered, validated, and reorganized by a committee of experts. The rules are described in a rule repository, and several statistics are automatically computed in every medical department, such as the confidence, relative risk, and median delay of outcome appearance. 236 validated ADE-detection rules are discovered; they enable to detect 27 different kinds of outcomes. The rules use a various number of conditions related to laboratory results, diseases, drug administration, and demographics. Some rules involve innovative conditions, such as drug discontinuations.

Although several methods exist for Adverse Drug events (ADE) detection due to past hospitalizations, a tool that could display those ADEs to the physicians does not exist yet. This article presents the ADE Scorecards, a Web tool that enables to screen past hospitalizations extracted from Electronic Health Records (EHR), using a set of ADE detection rules, presently rules discovered by data mining. The tool enables the physicians to (1) get contextualized statistics about the ADEs that happen in their medical department, (2) see the rules that are useful in their department, i.e. the rules that could have enabled to prevent those ADEs and (3) review in detail the ADE cases, through a comprehensive interface displaying the diagnoses, procedures, lab results, administered drugs and anonymized records. The article shows a demonstration of the tool through a use case.

The automated detection of Adverse Drug Events (ADE) is an important issue in medical informatics. The objective of this work is to automatically detect ADEs and to present the results to physicians. 90,000 stays are extracted from the EHR of 5 French and Danish hospitals and loaded into a common repository, using a common data model. Then data mining procedures such as decision trees are used in order to get ADE detection rules that are filtered and validated by an expert committee. The procedure enables to produce 236 ADE detection rules that are able to detect 27 different kinds of outcomes. Contextualized statistics are computed for every rule in every medical department separately. The physicians of the medical departments are provided with that information by means of a web-based tool named “ADE Scorecards”. The tool is presented in the article through a use case and several screenshots. Based on a list of rules and a repository of stays, it allows for displaying the important rules, the related statistics, and the complete information about the suspicious cases. The knowledge is contextualized, i.e. it depends on the medical department. The tool is deployed in a French hospital and is currently being evaluated through a prospective impact assessment.

Adverse drug events (ADEs) are a public health issue. The objective of this work is to data-mine electronic health records in order to automatically identify ADEs and generate alert rules to prevent those ADEs. The first step of data-mining is to transform native and complex data into a set of binary variables that can be used as causes and effects. The second step is to identify cause-to-effect relationships using statistical methods. After mining 10,500 hospitalizations from Denmark and France, we automatically obtain 250 rules, 75 have been validated till now. The article details the data aggregation and an example of decision tree that allows finding several rules in the field of vitamin K antagonists.

Our main objective is to detect adverse drug events (ADEs) in former hospital stays. As ADEs are rare, that supposes to screen thousands of electronic health records (EHRs). For that purpose, we need to define a data model that has two main objectives: (1) being able to describe hospital stays from various hospitals (2) being tuned so as to prepare the data mining process: as ADEs are not flagged in the datasets, the data model must be optimized for ADE detection. The article presents the phases of the design and the data model that results from this work. It is compatible with many hospitals. It deals with diagnoses, drug prescriptions, lab results and administrative information. It allows for data mining and ADE detection in EHRs.

Adverse drug events are a public health issue (98,000 deaths in the USA every year). Some computerized physician order entry (CPOEs) coupled with clinical decision support systems (CDSS) allow to prevent ADEs thanks to decision rules. Those rules can come from many sources: academic knowledge, record reviews, and data mining. Whatever their origin, the rules may induce too numerous alerts of poor accuracy when identically applied in different places. In this work we formalized rules from various sources in XML and enforced their execution on several medical departments to evaluate their local confidence. The article details the process and shows examples of evaluated rules from various sources. Several needs are enlightened to improve confidences: segmentation, contextualization, and evaluation of the rules over time.