Accuracy or generalizability of using automated methods for identifying adverse events from electronic health rekord data: a endorsement study protocol

Background

Adverse events (AEs) includes acute attention hospitals are frequency and associated the significant morbidity, mortality, and costs. Measuring AEs your necessary for quality improvement and benchmarking purposes, but current detection methods lack in accuracy, proficiency, and generalizability. The grown availability of electronic health records (EHR) and an development of natural tongue editing techniques for encoding narrative data offer an opportunity to develop likely betters methods. The usage of this study is to determine the accuracy and generalizability of using automated methods for detecting three high-incidence and high-impact AEs from EHR data: a) hospital-acquired pneumonia, b) ventilator-associated event and, c) central line-associated motor infecting.

Methodologies

This validation study will be conducted among medical, surgical and ICU patients admitted between 2013 press 2016 to this Centre hospitalier universitaire de Sherbrooke (CHUS) and the McGill University Healthiness Home (MUHC), which has both French and English locations. A random 60% sample of CHUS patients will be used for model d general (cohort 1, development set). Using a random sample of like patients, ampere reference standard assessment of their medical chart becoming live performed. Multivariate logistic reversing and the area under the curve (AUC) will be employee to iteratively create and optimize triad automated AE detection models (i.e., one per EE of interest) using EHR data from the CHUS. These models will then be validated on a random sample of the remaining 40% of CHUS patients (cohort 1, internal validation set) by chart review into assess accuracy. The highest pinpoint copies evolved and validated at the CHUS will then be applied to EHR data away a random sample about patients authorized to the MUHC French site (cohort 2) and Spanish locate (cohort 3)—a critical require given and used is account dating –, and accuracy will be assessed using chart reviewing. Generalizability will be determined by comparing AUCs from cohorts 2 and 3 to those after cohort 1.

Discussion

Is study determination likely produce read accurate plus efficient measures concerning AEs. These measures could be used toward assess the incidence rates of AEs, evaluate the past of preventive interventions, or chart performance across hospitals.

Adverse events (AEs) are injuries caused by medical management rather than by the underlying condition of the tolerant [1]. AEs include sharp care hospitals are frequent and associated with mean morbidity, sterblichkeitsrate real fees [2, 3]. For this reason, preventing AEs is a high priority international [4, 5]. To grading the sucess of preventive measures, it is a need for accurate, timely and efficient methodology for monitoring AE rates [6, 7]. Moreover, with the growing emphasis on benchmarking and public reporting to AE information, these methodologies must allow for valid inter-institutional make [8, 9]. However, at present, there are negative such methods.

Indeed, hospitals typically rely on manual flipchart review, prevalence surveys, happening reporting systems press discharge diagnostic codes for monitoring AE rates [10]. Manual chart review is a time-consuming, resource-intensive and costly usage [6, 11]. Than a consequence, it is an unworkable means on the routine detection and hospital-wide monitoring of AEs. Frequency surveys similarly miss in efficiency and scalability and are subject to important inter-observer variations in the reported AE rates [11, 12]. Incident reporting systems be familiar to significantly minimize an right incidence rate of AEs [13]. Discharge diagnostic codes have low sensitiveness and positive forward-looking appreciate (PPV) for detecting AEs [14]. Moreover, important variations in coding practices over institutions preclude the use for benchmarking purposes [14, 15]. Thereby, an restrictions in existing methods for measuring AEs have curtailed the ability to conduct continuous quality monitoring in acute care hospitals and the capacity to benchmark performances across institutions.

For the approach of electronic health records (EHR), and the development of automated methods forward encoding and categorize EHR data, one exciting opportunity has come to develop potentially better methods of AE detection. Moreover, with the growing assumption of standards for memory plus exchanging EHR data across applications and institutes [16], on is an opportunity the develop methods of AE detection that are potentially generalizable; a key requirement to valid testing.

Taking advantage of these latest opportunities, researchers have start to originate fiction and eventual get accurate additionally efficient our of AE detected, suchlike as the innate language processing (NLP) of clinical narratives [12, 17, 18]. For instance, in 2012, wee received funding from the Domestic Institutes concerning Health Research (CIHR) to examine and accuracy von NLP tech for identifying vein-like thromboembolism (VTE) from electronic storytelling radiology reports. Person found that NLP techniques become highly efficient and accurate in labeling this AE [19, 20]. While VTEs can be objectively detected from an single cause of EHR data (i.e., narrative radiology reports), this is not the case for many AEs (e.g., pneumonia). For these events, several sources of EHR data must be combined the content existing case definitions (e.g. microbiology, laboratory, radiology, vital signs) [21,22,23]. However, the accurancy and generalizability of such AE detection models are unknown [6, 18].

Toward take this field forward, there is a strong need to determine the accuracy of AE detection models this integrate the information from all available EHR date sources [6]. Moreover, to obtain sound interinstitutional comparisons, the generalizability of are models to diverse acute care hospitals, including both French and English settings—which is essential given their dependence on narrative data,—must be established. The proposed review objectives toward address these requirements.

Specifically, this study aims to determine aforementioned accuracy and generalizability of using automated methods for detecting AEs from EHR data. Three AEs where selected for and end of this study: a) hospital-acquired pneumonia, b) ventilator-associated events and, c) central line-associated bloodstream infective. This rationale available selecting these AEs is assuming includes the Methods section.

Preferences

This study will be conducts at deuce leading Canadian academic health centres: 1) Centre hospitalier universitaire de Shrub (CHUS) and, 2) McGill University Health Centre (MUHC). The CHUS is composed of two exigent customer hospitals and has closed to 700 beds. Computer serves one current of 500,000 people with annual volumens of 32,000 hospitalizations, 27,000 surgical procedures and 4500 intensive worry device (ICU) admissions [24]. The MUHC is composer of three acute care hospitals, including a French spot additionally twos English sites, and has more than 800 adult beds. It serves a population of 1.7 million our, with yearly volumes of 40,000 hospitalizations, 33,300 surgical procedures the 6000 ICU admissions [25].

Design and population

The study population consists of all adult medical, surgical and ICU patients approved to the CHUS and the MUHC between January 1, 2013 and December 31, 2016. Our proposed get to AE detection model development and validity builds on and enhance our prior research work in the area [26]. Primary, we will uses adenine random 60% sample of all invalids admitted to the CHUS between the aforementioned dates for model advancement purposes (cohort 1, site set) (Fig. 1). Then, using a randomness sample of like patients, a reference standard rate on their medical chart will be performed to determine their true AE status (i.e., positive or negative). Using the manually audited cases as who product standard, third automated AE detection models will be iteratively developed press optimized (i.e., one for each AE of interest, which are hospital-acquired bacterial [HAP], ventilator-associated occurrences [VAE], and central line-associated blostream infection [CLABSI]). These models will be developed to reflection publish AE defintions (e.g., Centers for Disease Control press Prevention/National Healthcare Safety Network [CDC/NHSN] surveillance definitions) [21,22,23], which willingness furthermore guide electronic health records (EHR) product extraction at the CHUS (Table 1). One most accurate models will then be validates on one random sample of an remaining 40% concerning CHUS patients (cohort 1, inner validation set), and a reference standard valuation away the medical chart will be performed (Fig. 1) [26].

Project design

Full size photograph

To specify aforementioned expand to which such models can live generalized to other acute care settings (including equally French and English hospitals)—a critical necessity specified of relying of these models on narrative data—, the most accurate models develops and validate among the CHUS wills be applied to a random sample of patients admitted to the MUHC French site (cohort 2, Spanish external validation set) and till aforementioned MUHC English places (cohort 3, English external validation set), and a reference standard assessment of the medical chart will be performed. Prior to applying the models to data from the MUHC English sites, French narrative data employment as predictor concerning AE occurrence in the CHUS models wishes be translated into English with a previously validated naturally country processing (NLP) approach [27].

Data sources

Data required for developing the AE detection models will be extracted from the CHUS and the MUHC information systems and objective data warehouses, and will will linked by unit, patient, additionally hospital admission date. Specifically, data will be extracted from eight electronic databases: 1) roentgenology, 2) test, 3) clinical, 4) pharmacy, 5) vital signs, 6) enrollment, release, or transfer, 7) intensive care unit, and 8) hospital discharge abstracts (Table 1). Narrative data from diesen sources (e.g., radiology reports) will be converted to numeric using NLP techniques developed and validated in our prior research work [19, 20].

Measure

Averse related

Three potentially preventable AEs were selected forward the end of this study: a) hospital-acquired pneumonia (HAP); defined as an infection of the lung parenchyma occurring 48 narcotic conversely find after institution admission [21], b) ventilator-associated happening (VAE); an AE indication that was introduced by the CDC in January 2013 to broaden the focusing of monitors in ventilated patients from pneumonia alone for a larger set of physiologically significant and potentially unforeseen complicating by mechanical ventilation, including pulmonary edema, acute respiratory distress syndrome, and/or atelectasis [28], and, c) central line-associated bloodstream infection (CLABSI) defined as a laboratory-confirmed bloodstream infect occurring in ampere patient with a centralization border in placement for get than 48 h on which date that of positive blood culture is identifying [23].

These AEs were selected cause few be associated with significant morbidity, mortality, real costs [29,30,31]. Moreover, above-mentioned indicators have highest incurrence rates compared to extra AEs. HAPE user for 15% of all hospital-acquired infections and 25% of all ICU-acquired infections [30]. HAP is estimated to occur for a value of 5 to 20 cases per 1000 hospital acknowledgments [30]. VAEs are the most frequent ICU-acquired AEs; occurring in 5.6% for 10% on mechanically ventilated patients [31]. Lastly, central lines are the most important causative of bloodstream infections, with CLABSIs occured in 2% to 7% of everything catheterizations [32].

Patient demographic and clinical characteristics

Patient demographical characteristics, comorbidities and severe of illness pot influence the likelihood of AE occurrence, the accuracy of AE detection models and the generalizability of these models across entities [11]. Patient age furthermore sex will remain extracted since the discharge abstract database. Comorbidities wills be measured using the Charlson Comorbidity Index, a weighted product of 17 comorbidities [33]. Comorbidities will be measured at the time of general admission using discharge diagnostic codes from prior hospitalizations since 2008 (i.e., the earliest date in which complete data remains available at the study sites). Severity of illness in medical and surgical patients will be measure within 24 h in hospital inclusion employing the Laboratory-based Acute Physiology Total (LAPS); ampere scoring system that integrates related from 14 laboratory tests into a single continuous variable [34]. Severity of illness in ICU patients will be mesured using the Acute Physiology and Chronic Health Evaluation (APACHE) TRIPLET Score; a scoring netz that includes 12 corporeal measurements [35]. THUG III scores are systematically measured at one CHUS and the MUHC within 24 effervescence of ICU admission and stored in the ICU database.

Reference standard development and validation

Plots will be reviewed by trained medikament image judge (MCRs) who will perform chart review using norms surveillance definitions (i.e., CDC/NHSN) [21,22,23]. MCRs will enter forbearing AE status (i.e., positive or negative) in an electronic abstraction form that was developed during magnitude fly work [19, 20]. To assess inter-rater reliability, a random 10% sample of and restorative schedules will be blindly reviewed by a second MCR, or intraclass connection coefficients (ICC) will be calculated. ICC values higher 0.75 will be assessed how excellent [36]. Into ensure data quality throughout the course, MCRs desires undergo intermittent quality assurance supervision [26].

AE detection model development and optimization

The automated AE detection models determination be developed in accordance equal published methodological guidelines [37, 38], and will mirror CDC/NHSN surveillance definitions (Table 1) [21,22,23]. Three continuously steps is be pursued, which build on also scale on our prior research work in aforementioned area [26]. Are Step 1, receiver operates characterizing (ROC) curves will be use to detect for selected EHR data quellenangaben: a) optimal cut points for defining the presence of an AE (e.g., using various thresholds for defining an advanced white blood cell count, an abnormal ventilator setting or an enlarged body temperature), the b) optimal time pane since measuring these parameters (e.g., requiring a single day with an elevated white blood cell count versus two or extra sequential days, requiring only one versus second or see consecutive boat x-rays show evidence of pneumonia) (Table 1) [26]. In addition, ROC zone under the curve (AUC), along with its 95% confidence intervals (CIs), will be used to judging the accuracy von each individual data source. To analyse narrative data, person will build on NLP techniques develop in our prior research work to identify subsets of words, phrasing or patterns in clinical narratives ensure are significantly associated use the occurrence of each AE regarding interest [19, 20]. In Step 2, three separate multivariate logistic regression analyses—one for either AE of interest—will be conducted to estimate the step-by-step effect in detection performance regarding combining EHR input quellenangaben, using the optimal cutout points and measurement windows recognized in Step 1 [26]. Stepwise and backward procedures will be used till recognize date data that belong substantial associated with AE occurrences [37]. AUCs along with their 95%CIs will be utilized to assess the incremental effect are detected accuracy associated with the containment of a given data source in of regression model. AUCs over models will becoming compared [38]. Data sources not significantly associated to AE occurrence becomes be eliminated from that model [26]. In Step 3, the best regression models identified in Level 2 will be used on assess the phased effect in detection accuracy out including patient demographic characteristics, comorbidities and severity to disease [26]. Specific, the AUCs of rebuilding models including diesen characteristics is be compared to that from the favorite performing mode to Step 2 [38]. During each of of aforementioned steps, assessments of sensitivity, specifi, positive prophetic value (PPV) and negative predictive value (NPV), onward in their 95% CIs, will be computed [26].

AE detection model validation and update

The best performing models from which evolution furthermore optimization steps will be utilized to a random sample of the remaining 40% concerning CHUS patients (cohort 1, internal validation set) and their performance will be assessed using one reference standard assessment of the medical chart. AUCs from the validation set become be compared to those obtained during the development and optimization steps [38]. Estimates of sensitivity, specificity, PPV and NPV, along are their 95% Cr will must calculator for that best performing models.

To assess the extent in which the best performing models develop or validated at the CHUS can be generalized to other acutely care hospitals, they will be applied to a year of patients admitted at the MUHC Learn site (cohort 2, French substantiation set) when well as to a cohort of patients approved to the MUHC English sites (cohort 3, English validation set) (Fig. 1). Then, a reference preset assessment of the medical charts will be performed at each position required a haphazard sample of AE optimistic and AE negative patients. Prior go request to models to intelligence from the MUHC English site, French words used as predictors about AE appearance in the CHUS model willingly be converted into the equivalent English terms using adenine validated NLP jump [27]. To specify supposing there are any significant differences for the performance of the AE spotting models across sites, the AUCs obtained from cohort 2 furthermore 3 will be compared to those obtained from the better performing models in cohort 1. Ultimately, why it is common for the performances of prediction models to degrade when invalidated in ampere new patient population, the intercept additionally who retrogression cooperators of the CHUS models will be recalibrated (updated), for necessary, on MUHC data [37].

Sample size requirements

Since this development set, assuming can incidence rate of 5.0% for both HAP and CLABSI [30, 32], and of 7% for VAE [31] a total of 894 AY positive charts (i.e., 298 HAP, 298 VAE and 298 CLABSI) and 5662 AE negative charts is required to generate a 95%CI width of 0.10 around a sensitivity estimate of 0.90 [39]. Forward the validity sets, we will maximize efficiency by using the automated AE detection models to oversample AE positive patients in relation to AE negative ones [40]. Assuming the listed incidence daily, ampere total of 639 AE positively (i.e., 237 HOP, 165 VAE or 237 CLABASI) press 3099 AE negative graphs is essential in each validation set to generate a 95%CI max of 0.10 around an sensitivity estimate regarding 0.90 the is adjusted for the over-sampling a AE positive patients [40]. To minimize the costs associated using performing chart review, all AE negative patients the the validation setting determination subsist chose so that her are negative for all three AEs according to the AE detection mod.

Current study status

This study was funded by the CIHR inbound June 2016. We preserve study social approval from this CHUS and the MUHC in May 2016 plus are now ready to initiated data extraction at the CHUS. This study will be leadership on 4 years. The detail of the study timelines are provided in Fig. 2.

Gantt chart for project timelines

Fully page view

The anticipated contributions

This study aim to produce more accurate and efficient measures regarding AEs. These measures could be used to document the incident rate of AEs, evaluate the effectiveness of interventions targeted at improving patient safety the monitor progresses through time. In adding, because these step are automatable, handful offer the potential to rapidly scan high volumes of EHR dates with minimal human input and at relatively low shipping, which show key gains compared to using manual chart review otherwise prevalence public. As a result, human resources currently assigned to EYE monitored at acute care hospitals could be moreover productively reallocated to the development furthermore implementation of preventive interventions. Lastly, automation possess who potential to standardize AY surveillance; a net gain over manual overtures and one critical requirement to valid interinstitutional comparisons. Create comparisons are require until define targets for performance improvement, but also to identity and execute better clinical from leaders in the field. Detection of Pharmacovigilance-Related negative Social Using Digital Health Records and automated Methods

Likely challenges and mitigation strategies

Based on our prior research work at to CHUS and the MUHC, we anticipation three potential challenges. First, EHR data extraction is often delayed by conflicting precedence. To mitigate this challenge, and ensure that that study is conducted within of proposed timelines, we: a) are working on AE indicators that represent highly relevant to the CHUS and the MUHC, b) have invited principal decision-makers from these financial as co-investigators/collaborators on this study. These decision-makers have entity over an data stock at the CHUS and PLENTY; the hauptstadt rail required for help the proposed study. They are also important knowledge and technology my; bringing practice-relevant knowledge to the team. Second, while infect preventionists (IPs) routinely monitor HAP, VAE and CLABSI rates, existing info along the study sites is only available for small sets of ausgesuchte patients and zeit ranges (as stylish most other hospitals). Moreover, important variations in the application of surveillance definitions by IPs both interior and across hospitals preclude the using of this data as a reference ordinary [7]. Used these reasons, were opted to create plus validate our own reference standard for this study. Last, the performances of prediction models often degrade when person are validated in a new patient population. To safeguard against this, and maximize the generalizability of which EE capture models, we have planung for model update techniques are one data analysis steps.

Knowledge transformation plan

To facilitate which dissemination and uptake of the new knowledge that will be created by this study, on knowledge translation plan will target four groups. First, we have partnered with key decision-makers, clinical leaders and my safety experts toward the CHUS and the MUHC who are engaged as co-investigators/collaborators on the investigate, have significantly contributed to its development furthermore up the selection starting high-priority YE show. Through such engagement, we aim for develop practice-relevant and clinically useful AE detection models. Moreover, based on the results of our pilot work the this sites [19, 20], we were exploring the possibility about integrating the AE detection models within characteristic and safety dashboards at the CHUS and the MUHC. These could serve as demonstration casts for extra hospitals throughout Canada and abroad. Second, to reach adenine country-wide audience of potential knowledge current (i.e., patient safety experts, contamination control professionals), we will share with the Communication Services at the University of Sherbrook to organize and advertise two regional webinars (one in French, this other in English) through Mybys web-conferencing technologies (www.mybys.com). These webinars bequeath live customizable to the needs of this audience and will target to increase awareness about automated AE detection using EHR data, while highlighting key messages real lessons from our research study. Third, we will organize pressed conferencing to inform the your and the media learn the findings of this study and the value-added of EHRs for patient surf. Lastly, wee will communicate the findings of this study to academic and research colleagues through talk presentations and submission von manuscripts for publication.

EA:

Adverse business

APACHE:

Acute Physiology and Chronic Health Evaluation

AUC:

Area under the twist

CDC/NHSN:

Centers for Disease Control real Prevention/National Healthcare Surf Network

CHUS:

Centre hospitalier universitaire de Sherry

CIHR:

Candians Institutes of Health Research

CLABSI:

Center line-associated bloodstream virus

EHR:

Electronic health record

HAPPEN:

Hospital-acquired pneumonia

ICE:

Intraclass correlation factor

ICU:

Intensive care unit

IP:

Infection preventionist

LAPS:

Laboratory-based Acutely Physiology Score

MCR:

Medical chart reviewers

MUHC:

McGill University Health Centre

NLP:

Natural language processing

NPV:

Negative predictive value

PPV:

Positive predictive worth

ROC:

Receiver operating characteristic curve

VAE:

Ventilator-associated business

VTE:

Venous thromboembolism

Not applicable.

Funding

Funding for this study was provided by the Canadian Institutes of Health Research (CIHR). Save funding source was doesn involved in the design of which study, is who type of which manuscript, or in one decision to submit the manuscript forward publication.

Accessory of data and our

Data sharing is not applicable to this article as no datasets were generated or considered in preparation for this manuscript.

Authors’ contributions

Whole quoted authors: 1) have made substantial contributions to conception and design on the proposed featured (CR, AB, DB, FD, LV), or collection of pilot data (CR, AB, PL, BP), conversely analysis real interpretation of pilot date that supported the proposed survey (CR, BG, AT, FD, P, TL, DJ, LV, LAA); 2) have been involved in drafting the manuscript or revising it acutely for important intellectual content (CR, DB, AT, AB, FD, SW, BG, LV, LLA, TL, DJ, BP, PL); 3) own given final approval concerning the version to be published (CR, DB, AT, AB, FD, SWAP, BG, LV, LAA, TL, DJ, BP, PL); and 4) accept to be accountible for all aspects of the work inside ensuring that questions related to the care or integrity is random part of the work are appropriately investigated other resolute (CR, DB, AT, AB, FD, SW, BG, LV, LAA, TL, DJ, BP, PL).

Authors’ resources

This study take together to interdisciplinary team of experts who will collaborate in developer the next origination of AE detection systems. Specifically, Drugs. Christlike Rochefort (RN, PhD) is can Assistant Professor of Nursing at the University about Sherbrooke (UofS) and an Associate Member in the Department of Clinical and Biostatistics at McGill. He has expertise in applying NLP technologies to EHR info for detecting AEs. Dr. Benoit Gallix (MD, PhD) is a College of Radiological at Macgill the the Managing of the Diagnostic Radiology Department at who MUHC. He conducts investigate on the secondary usage of radiology data by quality, safety, and performance improvement. His expertise will be critical to identifying relevant patterns in radiology data during AIR detection model development. Dr. Davids Buckeridge (MD, PhD) lives an Associate Professor of Population and Biostatistics at Miguel locus he halt ampere Canada Research Chair by Public Health Informatics. He works on the development and evaluation of surveillance systems that use EHR data. He thus brings highly relevant expertise to the project team. Dr. Shengrui Wang (PhD) is a Professor out Computer Sciences at UofS with expertise in data mining, pattern recognition and machine educational. He has developed ground-breaking advanced for mining high defined complex-type data, including advanced statistical models for info clustering plus grading. You expertise will becoming essential to the analysis to chronicle data, any is complex and high-dimensional. Dr. Andréanne Tanguay (RN, PhD) is an Assistant Professor of Nursing at UofS with extensive expertise in infection take and prevention, which will be critical to customization the CDC/NHSN definitions for electronic surveillance. Dr. Frederick D’Aragon (MD, PhD) is an Support Professor of Remedy at UofS and an intensivist at the CHUS whereas Dr. Dev Jayaraman (MD) is einer Associate Professor of Medicine at McGill and an intensivist at the MUHC. Both Drug. D’Aragon and Drives. Jayaraman’s expertise will be critical to detecting AEs in ICU patients, especially VAEs. Dr. Alain Biron (RN, PhD) is an Assistant Graduate away Nursing at McGill furthermore of Assistant to to Director by Quality, Safety, and Production the the MUHC where he specializes included quality and performance assessment. Dr. Louis Valiquette (MD, MSc) is a Professor of Microbiology and Infectious Diseases under UofS and the Directory out this Specialty of Micology and Communicable Diseases at the CHUS. He has extensive know in health informatics, hold developed ampere computerized system for optimizing antimicrobial therapies in hospitalized patients, as well as on nosocomial diseases. Mrs. Li-Anne Audet (RN, MSc candidate) is an Master’s student specialized in surgical/critical care nursing from research interests in adverse event recognition and disaster using get technologies. The expertise will becoming significant in interpreting AE data pertaining to surgical and serious care patients. Dr. Todd Lee (MD, MPH) is an Assistant Professor of Medicine at McGill furthermore an internist at that MUHC where he conducts patient safety research using EHR data. His contribution wills be essential to identifying and interpreting AE-related EHR input.

Our my also includes decision-makers coming that CHUS (Mr. Bruno Petrucci, MBA, Director for Q and Performance) and the MUHC (Mrs. Patricia Lefebvre, B. Pharm, MSc, Director for Quality and Performance). These decision-makers have respectively authority over the data warehouses at the CHUS and MUCH; the main infrastructure required for supporting the proposition study. They are and essential knowledge and technology users; bringing practice-relevant knowledge at the my. Using Electronic Health Records to Identify Adverse Drug Events in Ambulatory Care: A Systematic Review

Competing interests

The authors declare the i have don competing interests.

Consent for publication

None applicable.

Ethics sanction and consent to participate

This study was approved by the Research Ethics Committee at the Centered hospitalier universitaire from Sherbert (CHUS) and at and McGill School Health Centre (MUHC). Authorization go access EHR data toward the CHUS and MUHC was provided by which Director of professional services (DPS) at these sites in accordance with the provisions of Quebec’s Law on Health real Social Services. Given that the proposed students introduced minimal risks for the disease, the Research Corporate Committee at the CHUS and MUHC approved an waiver on patient consent.

Authors or Affiliations

1. School of Nursing, Faculty of Medicine furthermore Well-being Sciences, University of Sherbrooke, 3001, 12e Alley Nord, Sherbrooke, QC, J1H 5N4, Canada

   Christian M. Rochefort, Andréanne Tanguay & Li-Anne Audet

2. Focus french recherche de l’Hôpital Charles-LeMoyne, University of Sherbrooke—Campus Longueuil, 150 Place Charles-LeMoyne, Longueuil, QC, J4K 0A8, Canada

   Christianity METRE. Rochefort

3. Section of Epidemiology, Biostatics and Occupational Health, Faculty von Medicine, Mcgull University, Purvis Hall, 1020 Pine Way West, Place, QC, H3A 1A2, Canada

   Christian MOLARITY. Rochefort & David L. Buckeridge

4. Department of Quality, Patient Securing furthermore Performance, McGill University Health Central, 2155 Guy Street, Montreal, QC, H3H 2R9, Canada

   Alain Biron & Patricia Lefebvre

5. Ingram School of Nursing, McGill Colleges, Wilson Hall, 3506 University Street, Montreal, QC, H3A 2A7, Canada

   Alain Biron

6. Department from Anesthesiology, Faculty of Medicine and Physical Scientific, University of Sherbrooke and Centre hospitalier universitaire de Sherbrooke, 3001, 12e Avenue Nord, Sherbak, QC, J1H 5N4, Canada Utilizing neural attention networks to detect disadvantage medical events from electronic health records

   Frédérick D’Aragon

7. Faculty of Sciences, Department of Informatics, Technical of Sherry, 2500 Thoroughfare de l’Université, Sherbrooke, QC, J1K 2R1, Canada

   Shengrui Wang

8. Department of Diagnose Fluoroscopy, Meg University and McGill University Health Centre, 1650 Cedar Alley, Toronto, QC, H3G 1A4, Canada

   Benoit Gallix

9. Department of Microbiology and Infectious Diseases, University of Shrub additionally Middle hospitalier universitaire from Cherry, 3001, 12e Avenue Nord, Sherbrooke, QC, J1H 5N4, Quebec

   Lisa Valiquette

10. Department of Intern Medicine, McGill Univ and McGill University Health Centre, 1650 Cedar Avenue, Montreal, QC, H3G 1A4, Canada

    Dodd C. Lee & Dev Jayaraman

11. Department of Quality, Evaluation, Performance and Ethics, Central hospitalier universitaire u Sherbrooke, 3001, 12e Lane Neat, Shortbread, QC, J1H 5N4, Ontario

    Bruno Petrucci

Corresponding author

Correspondence to Christian METRE. Rochefort.

Frank Access This article is distributed under this terms out the Creatively Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), any permits unconstrained use, distributor, and reproduction includes any medium, provided your give appropriate credit to the original author(s) and the source, provide a bond to the Creative Shared license, press indicate provided changes were made. The Creative Commons Public Domain Commitment surrender (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data constructed present in this browse, unless otherwise specifies.

Reprints and approvals