Care Electronic Medical Record Database Research Articles

Validating pertussis data measures using electronic medical record data in Ontario, Canada 1986–2016

BackgroundPertussis is a reportable disease in many countries, but ascertainment bias has limited data accuracy. This study aims to validate pertussis data measures using a reference standard that incorporates different suspected case severities, allowing for the impact of case severity on accuracy and detection to be explored. MethodsWe evaluated 25 pertussis detection algorithms in a primary care electronic medical record database between January 1, 1986 and December 30, 2016. We estimated sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). We used sensitivity analyses to explore areas of uncertainty and evaluated reasons for lack of detection. ResultsThe algorithm including all data measures achieved the highest sensitivity at 20.6%. Sensitivity increased to 100% after reclassifying symptom-only cases as non-cases, but the PPV remained low. Age at first episode was significantly associated with detection in half of the tested scenarios, and false negatives often had some history of immunization. ConclusionsSensitivity improved by reclassifying symptom-only cases but remained low unless multiple data sources were used. Results demonstrate a trade-off between PPV and sensitivity. EMRs can enhance detection through patient history and clinical note data. It is essential to improve case identification of older individuals with vaccination history to reduce ascertainment bias.

Assessing Primary Care Blood Pressure Documentation for Hypertension Management During the COVID-19 Pandemic by Patient and Provider Groups

BackgroundPrimary care electronic medical record (EMR) data can be used to identify, manage, and screen hypertension cases. However, this approach relies on completeness and accessibility of documented blood pressure (BP) values. With the large switch to virtual care due to the COVID-19 pandemic, we assessed BP documentation in primary care EMRs during the pandemic, across patient and physician groups. MethodsHypertension-related visits were identified during the pre-pandemic (January 2017 to February 2020) and pandemic (March 2020 to December 2021) periods from a primary care EMR database in Ontario, Canada. Clustered logistic regression models were used to analyze the relationship of physician and patient characteristics with an outcome variable of documented BP. A chart review of 3200 hypertension visits without a BP recorded in structured data fields was conducted to determine if BP was recorded in progress notes. ResultsPre-pandemic, 75.7% of hypertension-related visits (113,966 of 150,511) had a BP recorded in structured documentation, but this significantly decreased to 36.4% (26,660 of 73,239) during the pandemic (odds ratio [OR] = 0.18, 95% confidence interval [CI]: 0.18-0.19). For virtual visits, 14.3% (6357 of 44,572) had a documented BP, vs 74.0% (20,056 of 27,089) for in-person visits. Chart review found that 55.9% of hypertension visits had no associated BP in structured documentation, but did have a BP recorded in the progress note. Male providers, compared to female providers, were less likely to record BPs pre-pandemic (OR = 0.45, 95% CI: 0.32-0.63) and during the pandemic, for both virtual visits (OR = 0.48, 95% CI: 0.32-0.71) and in-person visits (OR = 0.46, 95% CI: 0.33-0.64). ConclusionsBP documented in primary care EMRs declined during the pandemic, most likely due to high rates of virtual visits impacting hypertension detection and management.

Effects of the COVID-19 Pandemic on Primary Health Care for Chronic Conditions in Canada: Protocol for a Retrospective Pre-Post Study Using National Practice-Based Research Network Data.

Since the COVID-19 pandemic began, there have been concerns that interruptions to the health care system may have led to changes in primary care, especially for care of chronic conditions such as diabetes and heart failure. Such changes may have longer term implications for population health. This study aims to describe the impacts of the COVID-19 pandemic on indicators of primary care access, comprehensiveness, and appropriateness among adult patients, as well as on specific indictors of chronic conditions. Additionally, this study aims to determine whether any identified changes were associated with patient sociodemographic characteristics and multimorbidity. This is a retrospective, single-arm, pre-post study using Canadian Primary Care Sentinel Surveillance Network (CPCSSN) data. CPCSSN is a research network supported by a primary care electronic medical record database, comprising over 1500 physicians and nearly 2 million patients. We are examining changes in care (eg, frequency of contacts, laboratory tests and investigations, referrals, medications prescribed, etc) among adults. We will also examine indicators specific to evidence-based recommendations for care in patients with diabetes and those with heart failure. We will compare rates of outcomes during key periods of the pandemic between March 13, 2020, and December 31, 2022, with equal time periods before the pandemic. Differences will be examined among specific subgroups of adults, including by decade of age, number of comorbidities, and socioeconomic status. Regression models appropriate to outcome distributions will be used to estimate changes, adjusting for potential confounders. This analysis is part of a mixed-methods study with a qualitative component investigating how patients with diabetes with or without concurrent heart failure perceived the impact of the pandemic on access to primary care and health care-related decisions. This study was approved by the Hamilton Integrated Research Ethics Board (14782-C). The start date of this study was October 5, 2022, and the prospective end date is January 31, 2024. As of May 2023, the study cohort (n=875,934) is defined, data cleaning is complete, and exploratory analyses have begun. Extended analyses using 2022 data are planned once the new data becomes available. We will disseminate results through peer-reviewed publications and academic conference, as well as creating evidence briefs, infographics, and a video for policy maker and patient audiences. This study will investigate whether the COVID-19 pandemic has resulted in changes in the provision of primary care in Canada and whether these potential changes have led to gaps in care. This study will also identify patient-level characteristics associated with changes in care patterns across the COVID-19 pandemic. Indicators specific to chronic conditions, namely diabetes and heart failure, will also be explored to determine whether there were changes in care of these conditions. ClinicalTrials.gov NCT05813652; https://clinicaltrials.gov/ct2/show/NCT05813652. RR1-10.2196/49131.

An analysis of sex differences and socioeconomic deprivation among Canadian children with high blood pressure: a retrospective, cross-sectional study.

Understanding which children are at highest risk for high blood pressure (HBP) can inform surveillance and treatment. This study evaluated sex differences in childhood HBP and its associations with socioeconomic status. This retrospective cross-sectional study assessed 74,233 children with data from a national primary care electronic medical record database. Differences between sex and material and social deprivation scores for children with and without HBP were examined. Covariates included age, BMI z-score, diabetes, hyperlipidemia, and depression. HBP was defined as > 90th percentile for < 13-year-olds, and ≥ 120/80 for age ≥ 13years on 2 separate occasions between 2010 and 2017. The prevalence of HBP was 10.2% in males and 7.6% in females (p < 0.0001). Children with HBP had higher BMI z-scores (0.66 vs. 0.18, p < 0.0001), and higher rates of diabetes (1.31 vs. 0.54%, p < 0.0001), depression (9.89 vs. 7.11%, p < 0.0001), and hyperlipidemia (2.82 vs. 0.86%, p < 0.0001). In univariate regression analyses, boys in the most materially deprived quintile had increased odds of HBP (OR 1.24 (95% CI 1.08-1.43)), whereas females did not (OR 1.11 (95% CI 0.95-1.29)). In multivariate regression, male sex was associated with HBP with adjusted OR of 1.39 (95% CI 1.24-1.55). After statistical adjustment, material deprivation was no longer significant (aOR 1.05, 95% CI 0.94-1.17). Male sex is associated with HBP in Canadian children. This study also suggests a possible association between material deprivation and HBP, particularly in boys. Further study is required to better understand this relationship. A higher resolution version of the Graphical abstract is available as Supplementary information.

Transfer or tailor? Implementing a technology‐supported intervention for noncommunicable diseases across contexts

Transfer or tailor? Implementing a technology‐supported intervention for noncommunicable diseases across contexts

Impact of government-funded insulin pump programs on insulin pump use in Canada: a cross-sectional study using the National Diabetes Repository

IntroductionInsulin pump access in type 1 diabetes may be inequitable. We studied the association between government funding programs for insulin pumps and rates of insulin pump use and disparities between...

Identifying cases of spinal cord injury or disease in a primary care electronic medical record database

Objective To identify cases of spinal cord injury or disease (SCI/D) in an Ontario database of primary care electronic medical records (EMR). Design A reference standard of cases of chronic SCI/D was established via manual review of EMRs; this reference standard was used to evaluate potential case identification algorithms for use in the same database. Setting Electronic Medical Records Primary Care (EMRPC) Database, Ontario, Canada. Participants A sample of 48,000 adult patients was randomly selected from 213,887 eligible patients in the EMRPC database. Interventions N/A. Main Outcome Measure(s) Candidate algorithms were evaluated using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F-score. Results 126 cases of chronic SCI/D were identified, forming the reference standard. Of these, 57 were cases of traumatic spinal cord injury (TSCI), and 67 were cases of non-traumatic spinal cord injury (NTSCI). The optimal case identification algorithm used free-text keyword searches and a physician billing code, and had 70.6% sensitivity (61.9–78.4), 98.5% specificity (97.3–99.3), 89.9% PPV (82.2–95.0), 94.7% NPV (92.8–96.3), and an F-score of 79.1. Conclusions Identifying cases of chronic SCI/D from a database of primary care EMRs using free-text entries is feasible, relying on a comprehensive case definition. Identifying a cohort of patients with SCI/D will allow for future study of the epidemiology and health service utilization of these patients.

1062-P: Attainment of Glycemic Targets among Adults with Diabetes in Canada: A Cross-Sectional National Diabetes Repository Study

Objective: International societies and clinical practice guidelines recommend a Hemoglobin A1c target of ≤7.0% for the majority of adults with type 1 (T1D) and type 2 diabetes (T2D). However, the proportion meeting this target and its predictors have not been comprehensively described outside of the USA or in population-based data sources. Methods: Adults with diabetes were identified in the National Diabetes Repository, a primary care electronic medical record (EMR) database from 5 Canadian provinces, from January 2015 to June 2019. Individuals were categorized as T1D (using a previously validated algorithm), T2D not prescribed insulin, or T2D prescribed insulin. Using the most recent HbA1c, proportions of individuals with HbA1c ≤7.0% were determined and compared using Chi-square tests. Multivariable logistic regression models were used to assess the association between predictors and meeting the HbA1c target. Results: Of 83,273 eligible adults in the National Diabetes Repository, 1484 had T1D, 10,473 had T2D prescribed insulin and 71,316 had T2D not prescribed insulin. Proportions meeting the HbA1c target of ≤7.0% were 23.9%, 24.9%, and 62.4%, respectively (p=0.4 for T1D vs. T2D prescribed insulin, p&amp;lt;0.0001 for other comparisons). In multivariable logistic regression models, depression was associated with reduced likelihood of meeting the HbA1c target for T1D (OR 0.74, 95% CI 0.55-0.98). Male sex, lower income quintile, and insulin use were associated with reduced likelihood of meeting the HbA1c target for T2D (0.85(0.83-0.88); 0.83(0.79-0.88); 0.23(0.22-0.24)). Conclusions: In a Canadian population-based primary-care EMR database, less than one-fourth of adults with T1D and T2D prescribed insulin, whereas two-thirds of adults with T2D not prescribed insulin, attain the recommended HbA1c target of ≤7.0%. Further research and implementation approaches are required to optimize outcomes in diabetes, particularly for those requiring insulin. Disclosure A. Weisman: None. B. A. Perkins: Advisory Panel; Self; Abbott Diabetes, Boehringer Ingelheim International GmbH, Insulet Corporation, Novo Nordisk Canada Inc., Other Relationship; Self; Abbott Diabetes, Medtronic, Novo Nordisk Canada Inc. Funding Diabetes Action Canada

Association Between Bariatric Surgery and Major Adverse Diabetes Outcomes in Patients With Diabetes and Obesity

There are high-quality randomized clinical trial data demonstrating the effect of bariatric surgery on type 2 diabetes remission, but these studies are not powered to study mortality in this patient group. Large observational studies are warranted to study the association of bariatric surgery with mortality in patients with type 2 diabetes. To determine the association between bariatric surgery and all-cause mortality among patients with type 2 diabetes and severe obesity. This retrospective, population-based matched cohort study included patients with type 2 diabetes and body mass index (BMI; calculated as weight in kilograms divided by height in meters squared) 35 or greater who underwent bariatric surgery from January 2010 to December 2016 in Ontario, Canada. Multiple linked administrative databases were used to define confounders, including age, baseline BMI, sex, comorbidities, duration of diabetes diagnosis, health care utilization, socioeconomic status, smoking status, substance abuse, cancer screening, and psychiatric history. Potential controls were identified from a primary care electronic medical record database. Data were analyzed in 2020. Bariatric surgery (gastric bypass and sleeve gastrectomy) and nonsurgical management of obesity provided by the primary care physician. The primary outcome was all-cause mortality. Secondary outcomes were cause-specific mortality and nonfatal morbidities. Groups were compared through a multivariable Cox proportional Hazards model. A total of 6910 patients (mean [SD] age at baseline, 52.04 [9.45] years; 4950 [71.6%] women) were included, with 3455 patients who underwent bariatric surgery and 3455 match controls and a median (interquartile range) follow-up time of 4.6 (3.22-6.35) years. In the surgery group, 83 patients (2.4%) died, compared with 178 individuals (5.2%) in the control group (hazard ratio [HR] 0.53 [95% CI, 0.41-0.69]; P < .001). Bariatric surgery was associated with a 68% lower cardiovascular mortality (HR, 0.32 [95% CI, 0.15-0.66]; P = .002) and a 34% lower rate of composite cardiac events (HR, 0.68 [95% CI, 0.55-0.85]; P < .001). Risk of nonfatal renal events was also 42% lower in the surgical group compared with the control group (HR, 0.58 [95% CI, 0.35-0.95], P = .03). Of the groups that had the highest absolute benefit associated with bariatric surgery, men had an absolute risk reduction (ARR) of 3.7% (95% CI, 1.7%-5.7%), individuals with more than 15 years of diabetes had an ARR of 4.3% (95% CI, 0.8%-7.8%), and individuals aged 55 years or older had an ARR of 4.7% (95% CI, 3.0%-6.4%). These findings suggest that bariatric surgery was associated with reduced all-cause mortality and diabetes-specific cardiac and renal outcomes in patients with type 2 diabetes and severe obesity.

Blood Eosinophil Counts and Their Variability and Risk of Exacerbations in COPD: A Population-Based Study

BackgroundThere is controversy regarding the role of blood eosinophil levels as a biomarker of exacerbation risk in chronic obstructive pulmonary disease (COPD). Our aim was to quantify blood eosinophil levels and determine the risk of exacerbations associated with these levels and their variability. MethodsObservational, retrospective, population-based study with longitudinal follow-up in patients with COPD identified in a primary care electronic medical record database in Catalonia, Spain, covering 80% of the general population. Patients were classified into 4 groups using the following cut-offs: (a) <150cells/μl; (b) ≥150 and <300cells/μl; (c) ≥300 and <500cells/μl; (d) ≥500cells/μl. ResultsA total of 57,209 patients were identified with a mean age of 70.2 years, a mean FEV1(% predicted) of 64.1% and 51.6% had at least one exacerbation the previous year. The number of exacerbations in the previous year was higher in patients with the lowest and the highest eosinophil levels compared with the intermediate groups. During follow-up the number of exacerbations was slightly higher in the group with the lowest blood eosinophil levels and in those with higher variability in eosinophil counts, but ROC curves did not identify a reliable threshold of blood eosinophilia to discriminate an increased risk of exacerbations. ConclusionsOur results do not support the use of blood eosinophil count as a reliable biomarker of the risk of exacerbation in COPD in a predominantly non-exacerbating population. Of note was that the small group of patients with the highest variability in blood eosinophils more frequently presented exacerbations.

Developing and Validating a Primary Care EMR-based Frailty Definition using Machine Learning.

IntroductionIndividuals who have been identified as frail have an increased state of vulnerability, often leading to adverse health events, increased health spending, and potentially detrimental outcomes.ObjectiveThe objective of this work is to develop and validate a case definition for frailty that can be used in a primary care electronic medical record database.MethodsThis is a cross-sectional validation study using data from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN) in Southern Alberta. 52 CPCSSN sentinels assessed a random sample of their own patients using the Rockwood Clinical Frailty scale, resulting in a total of 875 patients to be used as reference standard. Patients must be over the age of 65 and have had a clinic visit within the last 24 months. The case definition for frailty was developed using machine learning methods using CPCSSN records for the 875 patients.ResultsOf the 875 patients, 155 (17.7%) were frail and 720 (84.2%) were not frail. Validation metrics of the case definition were: sensitivity and specificity of 0.28, 95% CI (0.21 to 0.36) and 0.94, 95% CI (0.93 to 0.96), respectively; PPV and NPV of 0.53, 95% CI (0.42 to 0.64) and 0.86, 95% CI (0.83 to 0.88), respectively.ConclusionsThe low sensitivity and specificity results could be because frailty as a construct remains under-developed and relatively poorly understood due to its complex nature. These results contribute to the literature by demonstrating that case definitions for frailty require expert consensus and potentially more sophisticated algorithms to be successful.

Methods to improve the quality of smoking records in a primary care EMR database: exploring multiple imputation and pattern-matching algorithms

BackgroundPrimary care electronic medical record (EMR) data are emerging as a useful source for secondary uses, such as disease surveillance, health outcomes research, and practice improvement. These data capture clinical details about patients’ health status, as well as behavioural risk factors, such as smoking. While the importance of documenting smoking status in a healthcare setting is recognized, the quality of smoking data captured in EMRs is variable. This study was designed to test methods aimed at improving the quality of patient smoking information in a primary care EMR database.MethodsEMR data from community primary care settings extracted by two regional practice-based research networks in Alberta, Canada were used. Patients with at least one encounter in the previous 2 years (2016–2018) and having hypertension according to a validated definition were included (n = 48,377). Multiple imputation was tested under two different assumptions for missing data (smoking status is missing at random and missing not-at-random). A third method tested a novel pattern matching algorithm developed to augment smoking information in the primary care EMR database. External validity was examined by comparing the proportions of smoking categories generated in each method with a general population survey.ResultsAmong those with hypertension, 40.8% (n = 19,743) had either no smoking information recorded or it was not interpretable and considered missing. Those with missing smoking data differed statistically by demographics, clinical features, and type of EMR system used in the clinic. Both multiple imputation methods produced fully complete smoking status information, with the proportion of current smokers estimated at 25.3% (data missing at random) and 12.5% (data missing not-at-random). The pattern-matching algorithm classified 18.2% of patients as current smokers, similar to the population-based survey (18.9%), but still resulted in missing smoking information for 23.6% of patients. The algorithm was estimated to be 93.8% accurate overall, but varied by smoking status category.ConclusionMultiple imputation and algorithmic pattern-matching can be used to improve EMR data post-extraction but the recommended method depends on the purpose of secondary use (e.g. practice improvement or epidemiological analyses).

Blood Eosinophil Counts and Their Variability and Risk of Exacerbations in COPD: A Population-Based Study

BackgroundThere is controversy regarding the role of blood eosinophil levels as a biomarker of exacerbation risk in chronic obstructive pulmonary disease (COPD). Our aim was to quantify blood eosinophil levels and determine the risk of exacerbations associated with these levels and their variability. MethodsObservational, retrospective, population-based study with longitudinal follow-up in patients with COPD identified in a primary care electronic medical record database in Catalonia, Spain, covering 80% of the general population. Patients were classified into 4 groups using the following cut-offs: (a) <150cells/μl; (b) ≥150 and <300cells/μl; (c) ≥300 and <500cells/μl; (d) ≥500cells/μl. ResultsA total of 57,209 patients were identified with a mean age of 70.2 years, a mean FEV1(% predicted) of 64.1% and 51.6% had at least one exacerbation the previous year. The number of exacerbations in the previous year was higher in patients with the lowest and the highest eosinophil levels compared with the intermediate groups. During follow-up the number of exacerbations was slightly higher in the group with the lowest blood eosinophil levels and in those with higher variability in eosinophil counts, but ROC curves did not identify a reliable threshold of blood eosinophilia to discriminate an increased risk of exacerbations. ConclusionsOur results do not support the use of blood eosinophil count as a reliable biomarker of the risk of exacerbation in COPD in a predominantly non-exacerbating population. Of note was that the small group of patients with the highest variability in blood eosinophils more frequently presented exacerbations.

Accuracy of US Administrative Claims Codes for the Diagnosis of Autoinflammatory Syndromes.

To determine the accuracy of case definitions for autoinflammatory syndromes (AISs) based on administrative claims codes compared with rheumatology records in the electronic medical record (EMR). An AIS screening filter of administrative codes was applied to a large tertiary care EMR database to extract all possible AIS cases. We manually chart reviewed all patients who were evaluated by a rheumatologist to determine their reference standard diagnosis of adult onset Still's disease (AOSD), Behçet's disease (BD), and familial Mediterranean fever (FMF). We calculated sensitivity, specificity, positive predictive values, negative predictive values, and area under the receiver operating characteristic curve of specific codes for diagnosing AIS subtypes. We identified 273 individuals with possible AIS, of which 72 (26.4%) had a true AIS diagnosis, including 24 with AOSD, 32 with BD, and 9 with FMF. For all 3 AIS subtypes, the estimates of specificities and negative predictive values for specific administrative codes were excellent (>95%). Sensitivity estimates were excellent (>89%) for BD and FMF codes and lower for AOSD (46%-50%). Positive predictive values were excellent for BD (>99%) and AOSD (>86%) and lower for FMF (>53%). Area under the receiver operating characteristic curve estimates were excellent for BD (97%-98%) and FMF (93%) and very good for AOSD (75%). This is the first study to characterize the accuracy of specific administrative codes for the diagnosis of AOSD, BD, and FMF in a large tertiary care EMR. Validation in external EMRs and linked EMR-administrative databases is needed to enable future clinical outcomes research of AIS.

Developing a case definition for type 1 diabetes mellitus in a primary care electronic medical record database: an exploratory study.

Identifying cases of disease in primary care electronic medical records (EMRs) is important for surveillance, research, quality improvement and clinical care. We aimed to develop and validate a case definition for type 1 diabetes mellitus using EMRs. For this exploratory study, we used EMR data from the Southern Alberta Primary Care Network within the Canadian Primary Care Sentinel Surveillance Network (CPCSSN), for the period 2008 to 2016. For patients identified as having diabetes mellitus according to the existing CPCSSN case definition, we asked family physicians to confirm the diabetes subtype, to create the reference standard. We used 3 decision-tree classification algorithms and least absolute shrinkage and selection operator logistic regression to identify variables that correctly distinguished between type 1 and type 2 diabetes cases. We identified a total of 1309 people with type 1 or type 2 diabetes, 110 of whom were confirmed by their physicians as having type 1 diabetes. Two machine learning algorithms were useful in identifying these cases in the EMRs. The first algorithm used "type 1" text words or age less than 22 years at time of initial diabetes diagnosis; this algorithm had sensitivity 42.7% (95% confidence interval [CI] 33.5%-52.5%), specificity 99.3% (95% CI 98.6%-99.7%), positive predictive value 85.5% (95% CI 72.8%-93.1%) and negative predictive value 94.9% (95% CI 93.5%-96.1%). The second algorithm used a combination of free-text terms, insulin prescriptions and age; it had sensitivity 87.3% (95% CI 79.2%-92.6%), specificity 85.4% (95% CI 83.2%-87.3%), positive predictive value 35.6% (95% CI 29.9%-41.6%) and negative predictive value 98.6% (95% CI 97.7%-99.2%). We used machine learning to develop and validate 2 case definitions that achieve different goals in distinguishing between type 1 and type 2 diabetes in CPCSSN data. Further validation and testing with a larger and more diverse sample are recommended.

Developing a Primary Care EMR-based Frailty Definition using Machine Learning

IntroductionFrailty is a geriatric syndrome that is predictive of heightened vulnerability for disability, hospitalization, and mortality. Annually an estimated 250,000 frail Canadians die, and this estimate is expected to double in the next 40 years, as Canadians grow older. Currently there is no single accepted clinical definition of frailty.&#x0D; Objectives and ApproachThe objective of this study was to develop an operational definition of frailty using machine learning that can be applied to a primary care electronic medical record (EMR) database.&#x0D; The Canadian Primary Care Sentinel Surveillance Network (CPCSSN) is a pan-Canadian network of primary care practices that collect de-identified patient information (such as encounter diagnoses, health conditions, and laboratory data) from EMRs.&#x0D; 780 patients from CPCSSN have were randomly selected and assessed by physicians using the Rockwood Clinical Frailty Scale (as frail or not frail), and their clinical characteristics from CPCSSN used to develop the definition using machine-learning.&#x0D; ResultsA total of 8,044 clinical features were extracted from these tables: billing, problem list, encounter diagnosis, labs, medications and referrals. A chi-squared automatic interaction detector (CHAID) approach was selected as the best approach. The bootstrapping process used a cost matrix that prioritized high sensitivity and positive predictive value. 10-fold cross validation was used for validity measures. Key features factored into the algorithm included: diagnosis of dementia (ICD-9 code 290), medications furosemide and vitamins, and use of key word “obstruction” within the billing table. The validation measures with 95% confidence intervals are as follows: sensitivity of 28% (95% CI: 21% to 36%), specificity of 94% (95% CI: 93% to 96%), positive predictive value of 53% (95% CI: 42% to 64%), negative predictive value of 86% (95% CI: 83% to 88%).&#x0D; Conclusion/ImplicationsNo other primary care specific frailty screening tools have sufficient validity. These results suggest heterogeneous diseases require clearly defined features and potentially more sophisticated algorithms to account for heterogeneity. Further research utilizing continuous features and continuous frailty scores may be more suitable in the creation of a case detection algorithm.

Learning Unsupervised Representations from Biomedical Text

IntroductionHealthcare settings are becoming increasingly technological. Interactions/events involving healthcare providers and the patients they service are captured as digital text. Healthcare organizations are amassing increasingly large/complex collections of biomedical text data. Researchers and policy makers are beginning to explore these text data holdings for structure, patterns, and meaning.&#x0D; Objectives and ApproachEMRALD is a primary care electronic medical record (EMR) database, comprised of over 40 family medicine clinics, nearly 400 primary care physicians and over 500,000 patients. EMRALD includes full-chart extractions, including all clinical narrative information/data in a variety of fields.&#x0D; The input data (raw text strings) are discrete, sparse and high dimensional. We assessed scalable statistical models for high dimensional discrete data, including fitting, assessing and exploring models from three broad statistical areas: i) matrix factorization/decomposition models ii) probabilistic topic models and iii) word-vector embedding models.&#x0D; ResultsEMRALD is comprised of 12 text data streams. EMRALD text data is structured into 84 million clinical notes (3.5 billion word/language tokens) and is approximately 18Gb in storage size. We employ a “text as data” pipeline, i) mapping raw strings to sequences of word/language tokens, ii) mapping token sequences to numeric arrays, and finally iii) using numeric arrays as inputs to statistical models.&#x0D; Fitted topic models yield useful thematic summaries of the EMRALD corpora. Topics discovered reflect core responsibilities of primary care physicians (e.g. women’s health, pain management, nutrition/diet, etc.).&#x0D; Fitted vector embedding models capture structure of discourse/syntax. Related words are mapped to similar locations of vector spaces. Analogical reasoning is possible in the embedding space.&#x0D; Conclusion/Implications“Text as data” requires an understanding of statistical models for discrete, sparse, high dimensional data. We fit a variety of unsupervised statistical models to biomedical text data. Preliminary results suggest that the learned low dimensional representations of the biomedical text data are effective at uncovering meaningful patterns/structure.

Methods to Describe Referral Patterns in a Canadian Primary Care Electronic Medical Record Database: Modelling Multilevel Count Data.

A referral from a family physician (FP) to a specialist is an inflection point in the patient journey, with potential implications for clinical outcomes and health policy. Primary care electronic medical record (EMR) databases offer opportunities to examine referral patterns. Until recently, software techniques were not available to model these kinds of multi-level count data.OBJECTIVE: To establish methodology for determining referral rates from FPs to medical specialists using the Canadian Primary Care Sentinel Surveillance Network (CPCSSN) EMR database.METHOD: Retrospective cohort study, mixed effects and multi-level negative binomial regression modelling with 87,258 eligible patients between 2007 and 2012. Mean referrals compared by patient sex, age, chronic conditions, FP visits, and urban/rural practice location. Proportion of variance in referral rates attributable to the patient and practice levels.RESULTS: On average, males had 0.26, and females 0.31 referrals in a 12-month period. Referrals were significantly higher for females, increased with age, FP visits, and number of chronic conditions (p<.0001). Overall, 14% of the variance in referrals could be attributed to the practice level, and 86% to patient level characteristics.CONCLUSIONS: Both patient and practice characteristics influenced referral patterns. The methodologic insights gained from this study have relevance to future studies on many research questions that utilize count data, both within primary care and broader health services research. The utility of the CPCSSN database will continue to increase in tandem with data quality improvements, providing a valuable resource to study Canadian referral patterns over time.

From patient care to research: a validation study examining the factors contributing to data quality in a primary care electronic medical record database.

BackgroundElectronic Medical Records (EMRs) are increasingly used in the provision of primary care and have been compiled into databases which can be utilized for surveillance, research and informing practice. The primary purpose of these records is for the provision of individual patient care; validation and examination of underlying limitations is crucial for use for research and data quality improvement. This study examines and describes the validity of chronic disease case definition algorithms and factors affecting data quality in a primary care EMR database.MethodsA retrospective chart audit of an age stratified random sample was used to validate and examine diagnostic algorithms applied to EMR data from the Manitoba Primary Care Research Network (MaPCReN), part of the Canadian Primary Care Sentinel Surveillance Network (CPCSSN). The presence of diabetes, hypertension, depression, osteoarthritis and chronic obstructive pulmonary disease (COPD) was determined by review of the medical record and compared to algorithm identified cases to identify discrepancies and describe the underlying contributing factors.ResultsThe algorithm for diabetes had high sensitivity, specificity and positive predictive value (PPV) with all scores being over 90%. Specificities of the algorithms were greater than 90% for all conditions except for hypertension at 79.2%. The largest deficits in algorithm performance included poor PPV for COPD at 36.7% and limited sensitivity for COPD, depression and osteoarthritis at 72.0%, 73.3% and 63.2% respectively. Main sources of discrepancy included missing coding, alternative coding, inappropriate diagnosis detection based on medications used for alternate indications, inappropriate exclusion due to comorbidity and loss of data.ConclusionsComparison to medical chart review shows that at MaPCReN the CPCSSN case finding algorithms are valid with a few limitations. This study provides the basis for the validated data to be utilized for research and informs users of its limitations. Analysis of underlying discrepancies provides the ability to improve algorithm performance and facilitate improved data quality.

PRM45 - Assessment of a Canadian Primary Care Electronic Medical Record Database for Use in Observational Studies

PRM45 - Assessment of a Canadian Primary Care Electronic Medical Record Database for Use in Observational Studies