Large-scale Electronic Health Record Research Articles

Patient Specific Comorbidities Impact Overall Survival in Myelofibrosis

Background:Treatment decisions in primary myelofibrosis (PMF) are guided by several prognostic systems based on disease-specific risk factors, including complete blood counts and cytogenetics. Patient specific comorbidities, e.g. non-hematopoietic organ dysfunction, are not incorporated into current prognostic models. Likewise, PMF risk stratification has not yet integrated large scale electronic health record (EHR) clinical data to refine these scoring systems. We have identified a PMF cohort within the Synthetic Derivative (SD), a de-identified, research-dedicated mirror of the EHR at Vanderbilt University Medical Center that contains 2.9 million individual records with 148 million ICD codes, and 125 million clinical notes. As a proof of concept, we leveraged the SD to develop a PMF cohort. We then aimed to identify novel patient specific comorbidities that may be associated with reduced overall survival (OS) in PMF via a phenome-wide association (PheWAS) study. Methods:We interrogated the SD for PMF via an algorithm that relied on ICD codes, natural language processing of physician notes, and medication history to identify high probability cases. Confirmation of PMF was based on strict hematologist review using 2016 WHO criteria. To this end, only patients with accessible hematopathology reports and cytogenetics, and more than 1 visit to the institution were enrolled. Patients with transformation to AML at presentation (e.g. within 30 days) were excluded. To evaluate each patient's overall comorbidity burden, we interrogated patient phecodes, which are grouped ICD9 codes shown to better mimic clinical phenotypes (PMID 20335276). Specifically, we extracted all ICD9 codes within 360 days of PMF diagnosis or referral and converted them to phecodes using the map available at https://phewascatalog.org/phecodes. PMF disease-related phecodes or codes that corresponded to DIPSS dependent variables were excluded. We identified 375 phecodes at PMF diagnosis, and conducted a PheWAS study to test the association of each phecode with survival. Survival was calculated as the interval between PMF diagnosis or referral and death or last follow-up (censor); patients who underwent hematopoietic stem cell transplant (HSCT) or transformed to AML were censored at that respective date. Survival from PMF diagnosis was estimated using the Kaplan-Meier method. We utilized the Cox proportional hazards model adjusted for the DIPSS predictors and evaluated the association of each comorbidity with the overall survival and reported those that are statistically significant after multiple testing adjustment (Bonferroni corrected P<0.00013). Results:We identified 193 cases of PMF from 1995-2016 that met strict pathology inclusion criteria. PMF median age of diagnosis was 59 (range 24-87), 42% were female, and 35 patients were referred greater than one year after diagnosis. Median OS was 39 months (range 1-265), with 23 patients developing AML and 40 patients treated with HSCT. Comorbidity analysis adjusted for DIPSS factors revealed five phecodes associated with reduced OS with Bonferroni correction (Figure 1); intracranial hemorrhage (HR 28.7; 95% CI 7-116; P=2.83E-06) invasive fungal infection (HR 41.1; 95% CI 7-235; P=2.90E-05), cerebral degeneration or hydrocephalus (HR 15.1; 95% CI 3-60;P=8.56E-05), encephalopathy or coma (HR 15; 95% CI 3-59; P=0.0001) and renal failure (HR 4; 95% CI 2-8; P=0.0001). Within the renal failure cohort, uric acid levels within 12 months of PMF diagnosis were elevated (N=18, mean 9.3 mg/dl) compared to the remaining PMF cases (N=132; mean 7.6 mg/dl) P=0.016. An additional 21 patient specific comorbidity patterns noted near Bonferroni cutoff (P<0.01), with highlights including pulmonary congestion (HR 6; 95% CI 1.7-21), pulmonary heart disease (HR 5.9; 95% CI 0.6-58), cardiomyopathy (HR 5; 95% CI 1.3-19), congestive heart failure (HR 4.3; 95% CI 1.3-14), and pneumonia (HR 3.3; 95% CI 1.4-7). Summary:We successfully leveraged our PMF cohort to identify potential high risk patient-specific comorbidities not included in current risk models. In addition, we illustrated the capacity to use these techniques to identify potential clinical intervention (e.g. chronic hyperuricemia and its impact on renal function). This study demonstrates the potential to refine prognostication and treatment decisions via EHR data to study large populations of rare myeloid disease. Disclosures Savona: Celgene Corporation: Membership on an entity's Board of Directors or advisory committees; Incyte Corporation: Membership on an entity's Board of Directors or advisory committees, Research Funding; Karyopharm Therapeutics: Consultancy, Equity Ownership, Membership on an entity's Board of Directors or advisory committees; Selvita: Membership on an entity's Board of Directors or advisory committees; Takeda: Membership on an entity's Board of Directors or advisory committees, Research Funding; Boehringer Ingelheim: Patents & Royalties; AbbVie: Membership on an entity's Board of Directors or advisory committees; Sunesis: Research Funding; TG Therapeutics: Membership on an entity's Board of Directors or advisory committees, Research Funding.

48.1 LEVERAGING LARGE-SCALE ELECTRONIC HEALTH RECORD (EHR) DATA TO IMPROVE QUALITY MEASUREMENT IN CHILD MENTAL HEALTH CARE

The goal of this session is to establish a framework for the use of large-scale electronic health record (EHR) data for patient-oriented outcomes research and quality improvement in child and adolescent psychiatry.

MOBILIZING CLINICAL INFORMATICS TO IMPROVE CHILD MENTAL HEALTH CARE

This session will stimulate interest in electronic health record (EHR) data science in child and adolescent psychiatry, explore integration of web-based and mobile applications into the EHR, identify barriers, and probe policy and financial implications of this new translational field. Over the last decade, child and adolescent psychiatry has shifted from documentation in paper charts to widespread adoption of EHRs. Clinical informatics has empowered clinicians to extract large-scale EHR data and to adapt and modify EHRs to meet subspecialty needs, especially to advance measurement-based care. Infrastructure and financial and clinical barriers remain to mobilizing EHRs in support of quality mental health care for children and adolescents.

Fine-Tuning Bidirectional Encoder Representations From Transformers (BERT)-Based Models on Large-Scale Electronic Health Record Notes: An Empirical Study.

BackgroundThe bidirectional encoder representations from transformers (BERT) model has achieved great success in many natural language processing (NLP) tasks, such as named entity recognition and question answering. However, little prior work has explored this model to be used for an important task in the biomedical and clinical domains, namely entity normalization.ObjectiveWe aim to investigate the effectiveness of BERT-based models for biomedical or clinical entity normalization. In addition, our second objective is to investigate whether the domains of training data influence the performances of BERT-based models as well as the degree of influence.MethodsOur data was comprised of 1.5 million unlabeled electronic health record (EHR) notes. We first fine-tuned BioBERT on this large collection of unlabeled EHR notes. This generated our BERT-based model trained using 1.5 million electronic health record notes (EhrBERT). We then further fine-tuned EhrBERT, BioBERT, and BERT on three annotated corpora for biomedical and clinical entity normalization: the Medication, Indication, and Adverse Drug Events (MADE) 1.0 corpus, the National Center for Biotechnology Information (NCBI) disease corpus, and the Chemical-Disease Relations (CDR) corpus. We compared our models with two state-of-the-art normalization systems, namely MetaMap and disease name normalization (DNorm).ResultsEhrBERT achieved 40.95% F1 in the MADE 1.0 corpus for mapping named entities to the Medical Dictionary for Regulatory Activities and the Systematized Nomenclature of Medicine—Clinical Terms (SNOMED-CT), which have about 380,000 terms. In this corpus, EhrBERT outperformed MetaMap by 2.36% in F1. For the NCBI disease corpus and CDR corpus, EhrBERT also outperformed DNorm by improving the F1 scores from 88.37% and 89.92% to 90.35% and 93.82%, respectively. Compared with BioBERT and BERT, EhrBERT outperformed them on the MADE 1.0 corpus and the CDR corpus.ConclusionsOur work shows that BERT-based models have achieved state-of-the-art performance for biomedical and clinical entity normalization. BERT-based models can be readily fine-tuned to normalize any kind of named entities.

A regression framework to uncover pleiotropy in large-scale electronic health record data.

Pleiotropy, where 1 genetic locus affects multiple phenotypes, can offer significant insights in understanding the complex genotype-phenotype relationship. Although individual genotype-phenotype associations have been thoroughly explored, seemingly unrelated phenotypes can be connected genetically through common pleiotropic loci or genes. However, current analyses of pleiotropy have been challenged by both methodologic limitations and a lack of available suitable data sources. In this study, we propose to utilize a new regression framework, reduced rank regression, to simultaneously analyze multiple phenotypes and genotypes to detect pleiotropic effects. We used a large-scale biobank linked electronic health record data from the Penn Medicine BioBank to select 5 cardiovascular diseases (hypertension, cardiac dysrhythmias, ischemic heart disease, congestive heart failure, and heart valve disorders) and 5 mental disorders (mood disorders; anxiety, phobic and dissociative disorders; alcohol-related disorders; neurological disorders; and delirium dementia) to validate our framework. Compared with existing methods, reduced rank regression showed a higher power to distinguish known associated single-nucleotide polymorphisms from random single-nucleotide polymorphisms. In addition, genome-wide gene-based investigation of pleiotropy showed that reduced rank regression was able to identify candidate genetic variants with novel pleiotropic effects compared to existing methods. The proposed regression framework offers a new approach to account for the phenotype and genotype correlations when identifying pleiotropic effects. By jointly modeling multiple phenotypes and genotypes together, the method has the potential to distinguish confounding from causal genotype and phenotype associations.

Outpatient Engagement and Predicted Risk of Suicide Attempts in Fibromyalgia.

Patients with fibromyalgia (FM) are 10 times more likely to die by suicide than the general population. The purpose of this study was to externally validate published models predicting suicidal ideation and suicide attempts in patients with FM and to identify interpretable risk and protective factors for suicidality unique to FM. This was a case-control study of large-scale electronic health record data collected from 1998 to 2017, identifying FM cases with validated Phenotype KnowledgeBase criteria. Model performance was measured through discrimination, including the receiver operating area under the curve (AUC), sensitivity, and specificity, and through calibration, including calibration plots. Risk factors were selected by L1 penalized regression with bootstrapping for both outcomes. Secondary utilization analyses converted time-based billing codes to equivalent minutes to estimate face-to-face provider contact. We identified 8,879 patients with FM, with 34 known suicide attempts and 96 documented cases of suicidal ideation. External validity was good for both suicidal ideation (AUC 0.80) and attempts (AUC 0.82) with excellent calibration. Risk factors specific to suicidal ideation included polysomatic symptoms such as fatigue (odds ratio [OR] 1.29 [95% confidence interval (95% CI) 1.25-1.32]), dizziness (OR 1.25 [95% CI 1.22-1.28]), and weakness (OR 1.17 [95% CI 1.15-1.19]). Risk factors specific to suicide attempt included obesity (OR 1.18 [95% CI 1.10-1.27]) and drug dependence (OR 1.15 [95% CI 1.12-1.18]). Per utilization analyses, those patients with FM and no suicidal ideation spent 3.5 times more time in follow-up annually, and those without documented suicide attempts spent more than 40 times more time face-to-face with providers annually. This is the first study to successfully apply machine learning to reliably detect suicidality in patients with FM, identifying novel risk factors for suicidality and highlighting outpatient engagement as a protective factor against suicide.

Performance of a Natural Language Processing Method to Extract Stone Composition From the Electronic Health Record

To demonstrate the utility of a natural language processing (NLP) algorithm for mining kidney stone composition in a large-scale electronic health records (EHR) repository. We developed StoneX, a pattern-matching method for extracting kidney stone composition information from clinical notes. We trained the extraction algorithm on manually annotated text mentions of calcium oxalate monohydrate, calcium oxalate dihydrate, hydroxyapatite, brushite, uric acid, and struvite stones. We employed StoneX to identify patients with kidney stone composition data and mine >125 million notes from our institutional EHR. Analyses performed on the extracted patients included stone type conversions over time, survival analysis from a second stone surgery, and disease associations by stone composition to validate the phenotyping method against known associations. The NLP algorithm identified 45,235 text mentions corresponding to 11,585 patients. Overall, the system achieved positive predictive value >90% for calcium oxalate monohydrate, calcium oxalate dihydrate, hydroxyapatite, brushite, and struvite; except for uric acid (positive predictive value = 87.5%). Survival analysis from a second stone surgery showed statistically significant differences among stone types (P = .03). Several phenotype associations were found: uric acid-type 2 diabetes (odds ratio, OR = 2.69, 95% confidence intervals, CI = 1.91-3.79), struvite-neurogenic bladder (OR = 12.27, 95% CI = 4.33-34.79), struvite-urinary tract infection (OR = 7.36, 95% CI = 3.01-17.99), hydroxyapatite-pulmonary collapse (OR = 3.67, 95% CI = 2.10-6.42), hydroxyapatite-neurogenic bladder (OR = 5.23, 95% CI = 2.05-13.36), brushite-calcium metabolism disorder (OR = 4.59, 95% CI = 2.14-9.81), and brushite-hypercalcemia (OR = 4.09, 95% CI = 1.90-8.80). NLP extraction of kidney stone composition from large-scale EHRs is feasible with high precision, enabling high-throughput epidemiological studies of kidney stone disease. These tools will enable high fidelity kidney stone research from the EHR.

Long-Term Exposure to Elevated Systolic Blood Pressure in Predicting Incident Cardiovascular Disease: Evidence From Large-Scale Routine Electronic Health Records.

BackgroundHow measures of long‐term exposure to elevated blood pressure might add to the performance of “current” blood pressure in predicting future cardiovascular disease is unclear. We compared incident cardiovascular disease risk prediction using past, current, and usual systolic blood pressure alone or in combination.Methods and ResultsUsing data from UK primary care linked electronic health records, we applied a landmark cohort study design and identified 80 964 people, aged 50 years (derivation cohort=64 772; validation cohort=16 192), who, at study entry, had recorded blood pressure, no prior cardiovascular disease, and no previous antihypertensive or lipid‐lowering prescriptions. We used systolic blood pressure recorded up to 10 years before baseline to estimate past systolic blood pressure (mean, time‐weighted mean, and variability) and usual systolic blood pressure (correcting current values for past time‐dependent blood pressure fluctuations) and examined their prospective relation with incident cardiovascular disease (first hospitalization for or death from coronary heart disease or stroke/transient ischemic attack). We used Cox regression to estimate hazard ratios and applied Bayesian analysis within a machine learning framework in model development and validation. Predictive performance of models was assessed using discrimination (area under the receiver operating characteristic curve) and calibration metrics. We found that elevated past, current, and usual systolic blood pressure values were separately and independently associated with increased incident cardiovascular disease risk. When used alone, the hazard ratio (95% credible interval) per 20–mm Hg increase in current systolic blood pressure was 1.22 (1.18–1.30), but associations were stronger for past systolic blood pressure (mean and time‐weighted mean) and usual systolic blood pressure (hazard ratio ranging from 1.39–1.45). The area under the receiver operating characteristic curve for a model that included current systolic blood pressure, sex, smoking, deprivation, diabetes mellitus, and lipid profile was 0.747 (95% credible interval, 0.722–0.811). The addition of past systolic blood pressure mean, time‐weighted mean, or variability to this model increased the area under the receiver operating characteristic curve (95% credible interval) to 0.750 (0.727–0.811), 0.750 (0.726–0.811), and 0.748 (0.723–0.811), respectively, with all models showing good calibration. Similar small improvements in area under the receiver operating characteristic curve were observed when testing models on the validation cohort, in sex‐stratified analyses, or by using different landmark ages (40 or 60 years).ConclusionsUsing multiple blood pressure recordings from patients’ electronic health records showed stronger associations with incident cardiovascular disease than a single blood pressure measurement, but their addition to multivariate risk prediction models had negligible effects on model performance.

Targeted learning with daily EHR data.

Electronic health records (EHR) data provide a cost- and time-effective opportunity to conduct cohort studies of the effects of multiple time-point interventions in the diverse patient population found in real-world clinical settings. Because the computational cost of analyzing EHR data at daily (or more granular) scale can be quite high, a pragmatic approach has been to partition the follow-up into coarser intervals of pre-specified length (eg, quarterly or monthly intervals). The feasibility and practical impact of analyzing EHR data at a granular scale has not been previously evaluated. We start filling these gaps by leveraging large-scale EHR data from a diabetes study to develop a scalable targeted learning approach that allows analyses with small intervals. We then study the practical effects of selecting different coarsening intervals on inferences by reanalyzing data from the same large-scale pool of patients. Specifically, we map daily EHR data into four analytic datasets using 90-, 30-, 15-, and 5-day intervals. We apply a semiparametric and doubly robust estimation approach, the longitudinal Targeted Minimum Loss-Based Estimation (TMLE), to estimate the causal effects of four dynamic treatment rules with each dataset, and compare the resulting inferences. To overcome the computational challenges presented by the size of these data, we propose a novel TMLE implementation, the "long-format TMLE," and rely on the latest advances in scalable data-adaptive machine-learning software, xgboost and h2o, for estimation of the TMLE nuisance parameters.

The dynamics of institutional pressures and stakeholder behavior in national electronic health record implementations: A tale of two countries

Through electronic health record implementation, national healthcare systems are aiming for care integration and enhancement. However, the path to large-scale electronic health record implementation is seldom smooth, involving multiple stakeholders with diverse interests and influences. This study proposes a framework that draws on both stakeholder and institutional theories to understand the complex dynamics of stakeholder interactions and institutional pressures over time during electronic health record systems implementation. This framework is utilized to analyze the national electronic health record programs of Singapore and England, which provide contrasting perspectives on how two top-down system implementations took place with different outcomes. Our results suggest that in the Singapore case, the presence of boundary spanners, supporting implementation agency that included IT staff from healthcare organizations, and greater engagement with medical professionals were associated with more positive dynamics of stakeholder interactions (e.g. limited pushback from professionals or the press) during electronic health record implementation than in England. Differences in the healthcare structures and systems, electronic health record project organization, and the combined influences of institutional pressures shed light on the varying implementation paths and outcomes in the two cases. This study adds to the health information technology literature through a comparative examination of the organizational and social processes during complex national healthcare integration projects. It also contributes to the institutional and stakeholder literatures in several ways, in particular by depicting the processes and outcomes of the dynamics of isomorphic pressures played out under different institutional conditions. Finally, our proposed framework provides a useful conceptual tool for analyzing such complex IT implementations across multiple stakeholders.

Ascertainment of Aspirin Exposure Using Structured and Unstructured Large-scale Electronic Health Record Data.

Aspirin impacts risk for important outcomes such as cancer, cardiovascular disease, and gastrointestinal bleeding. However, ascertaining exposure to medications available both by prescription and over-the-counter such as aspirin for research and quality improvement purposes is a challenge. Develop and validate a strategy for ascertaining aspirin exposure, utilizing a combination of structured and unstructured data. This is a retrospective cohort study. In total, 1,869,439 Veterans who underwent usual care colonoscopy 1999-2014 within the Department of Veterans Affairs. Aspirin exposure and dose were obtained from an ascertainment strategy combining query of structured medication records available in electronic health record databases and unstructured data extracted from free-text progress notes. Prevalence of any aspirin exposure and dose-specific exposure were estimated. Positive predictive value and negative predictive value were used to assess strategy performance, using manual chart review as the reference standard. Our combined strategy for ascertaining aspirin exposure using structured and unstructured data reached a positive predictive value and negative predictive value of 99.2% and 97.5% for any exposure, and 92.6% and 98.3% for dose-specific exposure. Estimated prevalence of any aspirin exposure was 36.3% (95% confidence interval: 36.2%-36.4%) and dose-specific exposure was 35.4% (95% confidence interval: 35.3%-35.5%). A readily accessible approach utilizing a combination of structured medication records and query of unstructured data can be used to ascertain aspirin exposure when manual chart review is impractical.

Structured Approach for Evaluating Strategies for Cancer Ascertainment Using Large-Scale Electronic Health Record Data.

Cancer ascertainment using large-scale electronic health records is a challenge. Our aim was to propose and apply a structured approach for evaluating multiple candidate approaches for cancer ascertainment using colorectal cancer (CRC) ascertainment within the US Department of Veterans Affairs (VA) as a use case. The proposed approach for evaluating cancer ascertainment strategies includes assessment of individual strategy performance, comparison of agreement across strategies, and review of discordant diagnoses. We applied this approach to compare three strategies for CRC ascertainment within the VA: administrative claims data consisting of International Classification of Diseases, Ninth Revision (ICD9) diagnosis codes; the VA Central Cancer Registry (VACCR); and the newly accessible Oncology Domain, consisting of cases abstracted by local cancer registrars. The study sample consisted of 1,839,043 veterans with index colonoscopy performed from 1999 to 2014. Strategy-specific performance was estimated based on manual record review of 100 candidate CRC cases and 100 colonoscopy controls. Strategies were further compared using Cohen's κ and focused review of discordant CRC diagnoses. A total of 92,197 individuals met at least one CRC definition. All three strategies had high sensitivity and specificity for incident CRC. However, the ICD9-based strategy demonstrated poor positive predictive value (58%). VACCR and Oncology Domain had almost perfect agreement with each other (κ, 0.87) but only moderate agreement with ICD9-based diagnoses (κ, 0.51 and 0.57, respectively). Among discordant cases reviewed, 15% of ICD9-positive but VACCR- or Oncology Domain-negative cases had incident CRC. Evaluating novel strategies for identifying cancer requires a structured approach, including validation against manual record review, agreement among candidate strategies, and focused review of discordant findings. Without careful assessment of ascertainment methods, analyses may be subject to bias and limited in clinical impact.

Next Generation Myelofibrosis Risk Analysis in the Electronic Health Record

Background: Myelofibrosis (MF) is a devastating myeloproliferative neoplasm that is hallmarked by marrow fibrosis, symptomatic extramedullary hematopoiesis, and risk of leukemic transformation, most commonly driven by janus kinase 2 (JAK2) pathway mutations. MF risk classification systems guide prognosis, decisions regarding allogeneic stem cell transplantation, and disease modifying agents. Key systems include the Dynamic International Prognostic Scoring System (DIPSS) 2009, DIPSS plus 2010, Genetics-Based Prognostic Scoring System (GPSS) 2014, and Mutation-Enhanced International Prognostic Scoring System (MIPSS) 2014. System contributions include dynamic scoring (DIPSS), cytogenetics (DIPSS Plus), and high risk molecular mutations (GPSS and MIPSS). To power the next generation of MF risk prognostication, and ascertain new prognostic factors, large scale electronic health record (EHR) and genomic data will need integration. As a proof of concept, we leveraged our de-identified research EHR (2.9 million records) and linked genomic biobank (288,000 patients) to develop an all-inclusive phenotype-genotype-prognostic system for MF and recapitulate DIPSS, DIPSS Plus, GPSS and MIPSS.Methods: Our previously described methods (Bejan et al. AACR 2018) utilized natural language processing to algorithmically identify 306 MF patients. A subset (N=125) had available DNA for genotyping. We automatically extracted: age greater than 65, leukocyte count (WBC) greater than 25x109/L, hemoglobin (Hgb) less than 10g/dL, platelets (PLT) less than 100 x 109/L, circulating myeloid blasts ≥ 1%, and 10% weight loss compared to baseline as a proxy for constitutional symptoms. Transfusion data was not included. Karyotype data was manually reviewed. Next generation sequencing (NGS) was performed on biobanked peripheral blood DNA with the Trusight Myeloid Panel (Illumina®). Genotyped samples were restricted to dates after MF diagnosis. Multivariate Cox proportional hazard analysis was performed on all clinical and genomic variables. DIPSS plus was calculated without adjustment but lacked transfusion data. DIPSS, GPSS and MIPSS scores were calculated by published methods.Results: Multivariate Cox proportional hazard regression identified Hgb (HR=6.4; P=0.006), myeloid blasts (HR=3.8; P=0.03), and ASXL1 (HR=5.2; P=0.02) as significant in our cohort with regard to overall survival (OS). We noted a strong trend for high risk karyotype (HR=5.6; P=0.07). Our DIPSS model median survival (N=120) for each subgroup; low risk (median survival not met), intermediate-1 (108 months), intermediate-2 (47 months) and high risk (6 months) P=0.0002 (Figure 1a). DIPSS Plus (N=122) integrated karyotype data and PLT count with similar survival with the exception of high risk (4 months) P=0.00003 (Figure 1b). The percentage of patients with driver mutations in JAK2V617F (57%), CALR (3%) and MPLW515 (7.2%); JAK2WT, CALRWT and MPLWT triple negative (34%); high molecular risk ASXL1 (15%), EZH2 (6%), IDH1/2 (7%), SRFS2 (17%); other variants of interest TET2 (9.6%), TP53 (29%) and DNMT3A (16.8%). MIPSS (N=125; 48 months follow up) noted low risk, intermediate-1, and intermediate-2 (median survival not met) and high risk (32 months) P=0.0001 (Figure 1c). GPSS (N=125; 48 months follow up) did not demonstrate statistical separation among groups (Figure 1d).Discussion: This proof of concept transformed raw EHR records into clinical risk scores for MF. The addition of retrospective DNA analysis via NGS opens the possibility of multi-institutional EHR-biobank studies to most accurately create a system to define MF risk. Our sample size limited the significance of age, PLTs, poor risk mutations and other variables previously shown to impact OS. Likewise, we lacked the capacity to track transfusion dependence, previously shown to have prognostic relevance. Still, prognostication via the EHR mimics common scoring systems in MF and supports correct MF case selection, accurate laboratory extraction and reproducible genotyping of biobanked samples. Similar to the original GPSS report, our low risk cohort was small (N=2) and will benefit from expansion of genotyping underway. Finally, this phenotype-genotype-prognostic paradigm represents a technical advance and a unique opportunity to deploy patient specific comorbidities from lifetime EHR records to further refine risk across all myeloid disease. [Display omitted] DisclosuresSavona:Boehringer Ingelheim: Consultancy; Celgene: Consultancy, Membership on an entity's Board of Directors or advisory committees; Incyte: Membership on an entity's Board of Directors or advisory committees, Research Funding.

Modeling EHR with the openEHR approach: an exploratory study in China

BackgroundThe openEHR approach can improve the interoperability of electronic health record (EHR) through two-level modeling. Developing archetypes for the complete EHR dataset is essential for implementing a large-scale interoperable EHR system with the openEHR approach. Although the openEHR approach has been applied in different domains, the feasibility of archetyping a complete EHR dataset in a hospital has not been reported in academic literature, especially in a country where using openEHR is still in its infancy stage, like China. This paper presents a case study of modeling an EHR in China aiming to investigate the feasibility and challenges of archetyping a complete EHR dataset with the openEHR approach.MethodsWe proposed an archetype modeling method including an iterative process of collecting requirements, normalizing data elements, organizing concepts, searching corresponding archetypes, editing archetypes and reviewing archetypes. Two representative EHR systems from Chinese vendors and the existing Chinese EHR standards have been used as resources to identify the requirements of EHR in China, and a case study of modeling EHR in China has been conducted. Based on the models developed in this case study, we have implemented a clinical data repository (CDR) to verify the feasibility of modeling EHR with archetypes.ResultsSixty four archetypes were developed to represent all requirements of a complete EHR dataset. 59 (91%) archetypes could be found in Clinical Knowledge Manager (CKM), of which 35 could be reused directly without change, and 23 required further development including two revisions, two new versions, 18 extensions and one specialization. Meanwhile, 6 (9%) archetypes were newly developed. The legacy data of the EHR system in hospitals could be integrated into the CDR developed with these archetypes successfully.ConclusionsThe existing archetypes in CKM can faithfully represent most of the EHR requirements in China except customizations for local hospital management. This case study verified the feasibility of modeling EHR with the openEHR approach and identified the fact that the challenges such as localization, tool support, and an agile publishing process still exist for a broader application of the openEHR approach.

Kluster: An Efficient Scalable Procedure for Approximating the Number of Clusters in Unsupervised Learning

Abstract The majority of the clinical observation data stored in large-scale Electronic Health Record (EHR) research data networks are unlabeled. Unsupervised clustering can provide invaluable tools for studying patient sub-groups in these data. Many of the popular unsupervised clustering algorithms are dependent on identifying the number of clusters. Multiple statistical methods are available to approximate the number of clusters in a dataset. However, available methods are computationally inefficient when applied to large amounts of data. Scalable analytical procedures are needed to extract knowledge from large clinical datasets. Using both simulated, clinical, and public data, we developed and tested the kluster procedure for approximating the number of clusters in a large clinical dataset. The kluster procedure iteratively applies four statistical cluster number approximation methods to small subsets of data that were drawn randomly with replacements and recommends the most frequent and mean number of clusters resulted from the iterations as the potential optimum number of clusters. Our results showed that the kluster's most frequent product that iteratively applies a model-based clustering strategy using Bayesian Information Criterion (BIC) to samples of 200–500 data points, through 100 iterations, offers a reliable and scalable solution for approximating the number of clusters in unsupervised clustering. We provide the kluster procedure as an R package.

Comprehensive methodology to monitor longitudinal change patterns during EHR implementations: a case study at a large health care delivery network

ObjectiveTo test a systematic methodology to monitor longitudinal change patterns on quality, productivity, and safety outcomes during a large-scale commercial Electronic Health Record (EHR) implementation. Materials and MethodsOur method combines an interrupted time-series design with control sites and 41 consensus outcomes including quality (11 measures), productivity (20 measures), and safety (10 measures). The intervention consisted of a phased commercial EHR implementation at a large health care delivery network. Four medium-size hospitals and 39 clinics from 5 geographic regions implementing the new EHR were compared against a parallel control consisting of one medium-size and one large hospital and 10 clinics that had not implemented the new EHR at the time of this study. We collected monthly data from February 2013 to July 2017. ResultsThe proposed methodology was successfully implemented and significant changes were observed in most measured variables. A significant change attributable to the intervention was observed in 12 (29%) measures in three or more regions; in 32 (78%) measures in two or more regions; and in 40 (98%) measures in at least one region. A similar pattern (i.e., same impact in three or more regions) was detected for nine (22%) measures, a mixed pattern (i.e., same impact in two regions, and different impact in other regions) was detected for nine (22%) measures, and an inconsistent pattern (i.e., did not detect the same impact across regions) was detected for 23 (56%) measures. DiscussionUsing a formal methodology to assess changes in a set of consensus measures, we detected various patterns of impact and mixed time-sensitive effects. With an increasing adoption of EHR systems, it is critical for health care organizations to systematically monitor their EHR implementations. The proposed method provides a robust and consistent approach to monitor EHR implementations longitudinally allowing for continuous monitoring after the system becomes stable in order to avoid unexpected effects. ConclusionOur results and methodology can guide the broader medical and informatics communities by informing what and how to continuously monitor EHR impact on quality, productivity, and safety.

Documentation of penicillin adverse drug reactions in electronic health records: inconsistent use of allergy and intolerance labels.

The majority of patients with penicillin allergy labels tolerate penicillins. Inappropriate avoidance of penicillin is associated with increased hospitalisation, infections and healthcare costs. To examine the documentation of penicillin adverse drug reactions (ADR) in a large-scale hospital-based electronic health record. Penicillin ADR were extracted from 96 708 patient records in the Enterprise Patient Administration System in South Australia. Expert criteria were used to determine consistency of ADR entry and suitability for further evaluation. Of 43 011 unique ADR reports, there were 5023 ADR to penicillins with most being entered as allergy (n = 4773, 95.0%) rather than intolerance (n = 250, 5.0%). A significant proportion did not include a reaction description (n = 1052, 20.9%). Using pre-set criteria, 10.1% of reports entered as allergy had a reaction description that was consistent with intolerance and 31.0% of the entered intolerances had descriptions consistent with allergy. Virtually all ADR (n = 4979, 99.1%) were appropriate for further evaluation by history taking or immunological testing and half (50.7%, n = 2549) had documented reactions suggesting low-risk of penicillin allergy. The frequency of penicillin allergy label in this data set is consistent with the known overdiagnosis of penicillin allergy in the hospital population. ADR documentation was poor with incomplete entries and inconsistent categorisation. The concepts of allergy and intolerance for ADR classification, whilst mechanistically valid, may not be useful at the point of ADR entry by generalist clinicians. Systematic evaluation of reported ADR is needed to improve the quality of information for future prescribers.

Spatiotemporal patterns of childhood asthma hospitalization and utilization in Memphis Metropolitan Area from 2005 to 2015

ABSTRACTObjective: To identify the key risk factors and explain the spatiotemporal patterns of childhood asthma in the Memphis metropolitan area (MMA) over an 11-year period (2005–2015). We hypothesize that in the MMA region this burden is more prevalent among urban children living south, downtown, and north of Memphis than in other areas. Methods: We used a large-scale longitudinal electronic health record database from an integrated healthcare system, Geographic information systems (GIS), and statistical and space-time models to study the spatiotemporal distributions of childhood asthma at census tract level. Results: We found statistically significant spatiotemporal clusters of childhood asthma in the south, west, and north of Memphis city after adjusting for key covariates. The results further show a significant increase in temporal gradient in frequency of emergency department (ED) visits and inpatient hospitalizations from 2009 to 2013, and an upward trajectory from 4 per 1,000 children in 2005 to 16 per 1,000 children in 2015. The multivariate logistic regression identified age, race, insurance, admit source, encounter type, and frequency of visits as significant risk factors for childhood asthma (p < 0.05). We observed a greater asthma burden and healthcare utilization for African American (AA) patients living in a high-risk area than those living in a low-risk area in comparison to the white patients: AA vs. white [odds ratio (OR) = 3.03, 95% confidence interval (CI): 2.75–3.34]; and Hispanic vs. white (OR = 1.62, 95% CI: 1.21–2.17). Conclusions: These findings provide a strong basis for developing geographically tailored population health strategies at the neighborhood level for young children with chronic respiratory conditions.

Patient-Specific Predictive Modeling Using Random Forests: An Observational Study for the Critically Ill.

BackgroundWith a large-scale electronic health record repository, it is feasible to build a customized patient outcome prediction model specifically for a given patient. This approach involves identifying past patients who are similar to the present patient and using their data to train a personalized predictive model. Our previous work investigated a cosine-similarity patient similarity metric (PSM) for such patient-specific predictive modeling.ObjectiveThe objective of the study is to investigate the random forest (RF) proximity measure as a PSM in the context of personalized mortality prediction for intensive care unit (ICU) patients.MethodsA total of 17,152 ICU admissions were extracted from the Multiparameter Intelligent Monitoring in Intensive Care II database. A number of predictor variables were extracted from the first 24 hours in the ICU. Outcome to be predicted was 30-day mortality. A patient-specific predictive model was trained for each ICU admission using an RF PSM inspired by the RF proximity measure. Death counting, logistic regression, decision tree, and RF models were studied with a hard threshold applied to RF PSM values to only include the M most similar patients in model training, where M was varied. In addition, case-specific random forests (CSRFs), which uses RF proximity for weighted bootstrapping, were trained.ResultsCompared to our previous study that investigated a cosine similarity PSM, the RF PSM resulted in superior or comparable predictive performance. RF and CSRF exhibited the best performances (in terms of mean area under the receiver operating characteristic curve [95% confidence interval], RF: 0.839 [0.835-0.844]; CSRF: 0.832 [0.821-0.843]). RF and CSRF did not benefit from personalization via the use of the RF PSM, while the other models did.ConclusionsThe RF PSM led to good mortality prediction performance for several predictive models, although it failed to induce improved performance in RF and CSRF. The distinction between predictor and similarity variables is an important issue arising from the present study. RFs present a promising method for patient-specific outcome prediction.

PhenoTree: Interactive Visual Analytics for Hierarchical Phenotyping From Large-Scale Electronic Health Records

Electronic health records (EHRs) capture comprehensive patient information in digital form from a variety of sources. Increasing availability of EHRs has facilitated development of data and visual analytic tools for healthcare analytics, such as clinical decision support and patient care management systems. Many healthcare analytic tools are used to investigate fundamental problems, such as study of patient population, exploring complicated interactions among patients and their medical histories, and extracting structured phenotypes characterizing the patient population. In this paper, we propose PhenoTree , a novel data-driven, hierarchical, and interactive phenotyping tool, that enables physicians and medical researchers to participate in the phenotyping process of large-scale EHR cohorts. The proposed visual analytic tool allows users to interactively explore EHR cohorts, and generate, interpret, evaluate, and refine phenotypes by building and navigating a phenotype hierarchy. Specifically, given a cohort or subcohort, PhenoTree employs sparse principal component analysis (SPCA) to identify key clinical features that characterize the population. The clinical features provide a natural way to generate deeper phenotypes at finer granularities by expanding the phenotype hierarchy. To facilitate the intensive computation required for interactive analytics, we design an efficient SPCA solver based on a variance reduced stochastic gradient technique. The benefits of our method are demonstrated by analyzing two different EHR patient cohorts, a public and a private dataset containing EHRs of $101\,767$ and $223\,076$ patients, respectively. Our evaluations show that PhenoTree can detect clinically meaningful hierarchical phenotypes.