---
title: "EHR Phenotype Choice Alters Outcomes: Three Invasive Breast Cancer Definitions Compared in TriNetX"
id: "plos-one-6-electronic-health-record-phenotype-choice-and-real-world-outcome-estimates-a"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-6-electronic-health-record-phenotype-choice-and-real-world-outcome-estimates-a"
content_type: "clinical_feed_article"
specialty: "Oncology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361883"
published_at: "2026-10-30T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# EHR Phenotype Choice Alters Outcomes: Three Invasive Breast Cancer Definitions Compared in TriNetX
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-6-electronic-health-record-phenotype-choice-and-real-world-outcome-estimates-a
- **Specialty:** [Oncology](https://medichelpline.com/clinical-feed/oncology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361883)
- **Published At:** 2026-10-30T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- This retrospective benchmarking study used the TriNetX Global Collaborative Network to compare three electronic health record operational definitions of incident **invasive breast cancer**: a single diagnosis-code rule (Dx1), a repeated-diagnosis rule within 90 days (Dx2), and a diagnosis-plus-treatment rule (DxTx). - Female patients aged 20 years or older were eligible after a 12-month washout period excluding carcinoma in situ; data were queried on May 3, 2026, and all analyses used the same snapshot. - Cohort yields differed substantially: Dx1 identified 1,450,233 patients, Dx2 identified 678,739, and DxTx identified 578,576 patients, reflecting increasing stringency across definitions. - Pairwise one-to-one propensity-score matching was performed for DxTx versus Dx1, Dx2 versus Dx1, and DxTx versus Dx2 using 13 baseline covariates (age, five TriNetX race categories, hypertension, diabetes, stroke, dyslipidemia, chronic kidney disease, BMI, and serum albumin). - After matching, standardized mean differences for all covariates were below 0.10 in each comparison; matched pair counts were 565,417 (DxTx vs Dx1), 678,739 (Dx2 vs Dx1), and 555,617 (DxTx vs Dx2). - Primary outcomes were time-to-all-cause mortality and time-to-first emergency department visit, assessed from day 180 to 1, 3, and 5 years after index; follow-up began at day 180 to accommodate DxTx treatment ascertainment. - More restrictive phenotype definitions were associated with higher hazards of death and higher hazards of emergency department visits after matching; for mortality, hazard ratios across windows ranged 1.26–1.32 (DxTx vs Dx1), 1.13–1.14 (Dx2 vs Dx1), and 1.13–1.17 (DxTx vs Dx2). - Emergency department visit hazard patterns paralleled mortality but were attenuated. - Negative-control outcomes (appendicitis, femoral neck fracture) were generally close to the null; small appendicitis associations involving Dx1 lost statistical significance after multiple-testing correction. - The authors emphasize these differences reflect selection and ascertainment mechanisms inherent to phenotype specification rather than causal treatment effects and recommend treating phenotype choice as a core analytic decision with transparent reporting and sensitivity analyses.
## Clinical Analysis & Structured Key Points
Electronic health record phenotype choice and real-world outcome estimates: A propensity score–matched comparison of three invasive breast cancer definitions in TriNetX | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Electronic health record-based oncology studies depend on computable case definitions, but different operational definitions may identify different patient populations and change downstream estimates. This retrospective benchmarking cohort study evaluated three definitions of incident invasive breast cancer in the TriNetX Global Collaborative Network: at least one malignant breast cancer diagnosis code (Dx1), a repeated diagnosis-code definition within 90 days (Dx2), and a diagnosis code plus breast cancer-directed treatment evidence (DxTx). Female patients aged 20 years or older were identified after a 12-month washout period and exclusion of patients with carcinoma in situ only. Pairwise one-to-one propensity-score matching was performed using 13 baseline covariates. Outcomes were all-cause mortality and time to the first emergency department visits, assessed from day 180 after the index date to 1, 3, and 5 years after the original index date. Appendicitis and femoral neck fracture were included as negative-control outcomes. Dx1, Dx2, and DxTx identified 1,450,233, 678,739, and 578,576 patients, respectively. After matching, more restrictive definitions were consistently associated with higher hazards of mortality and emergency department visits. For mortality, hazard ratios ranged from 1.26 to 1.32 for DxTx versus Dx1, 1.13 to 1.14 for Dx2 versus Dx1, and 1.13 to 1.17 for DxTx versus Dx2 across assessment windows. Emergency department visit estimates showed a similar but attenuating pattern. These associations should be interpreted as differences between phenotype-defined populations arising from selection and ascertainment mechanisms rather than as treatment effects. Negative-control outcomes were generally close to the null; small appendicitis associations observed in comparisons involving Dx1 did not remain statistically significant after correction for multiple testing. These findings indicate that plausible electronic health record-based breast cancer definitions are not analytically interchangeable. Phenotype specification should therefore be treated as a core analytic decision and evaluated through transparent reporting and sensitivity analyses. Citation: Chiang J-K, Chiang P-C, Koo M (2026) Electronic health record phenotype choice and real-world outcome estimates: A propensity score–matched comparison of three invasive breast cancer definitions in TriNetX. PLoS One 21(10): e0361883. https://doi.org/10.1371/journal.pone.0361883 Editor: Srinivasa Rao Karumuri, KL University: KL Deemed to be University, INDIA Received: June 18, 2026; Accepted: October 16, 2026; Published: October 30, 2026 Copyright: © 2026 Chiang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: The data used in this study were obtained from the TriNetX Global Collaborative Network, a third-party federated electronic health record platform. The authors are not permitted to publicly share patient-level TriNetX data. Access to the data was provided through the authors’ institutional access to TriNetX under the applicable data-use agreement. The authors did not receive special privileges in accessing the data that would not be available to other researchers with appropriate TriNetX access. Researchers interested in accessing the same or comparable data may apply for access directly through TriNetX, LLC, subject to institutional agreement, applicable data-use permissions, and legal and ethical requirements. Information regarding access is available from TriNetX, LLC, 100 Cambridgepark Drive, Suite 501, Cambridge, MA 02140, USA; email: info@trinetx.com . The aggregate results underlying the analyses reported in this study, together with the operational phenotype definitions, coding criteria, and outcome specifications that can legally be shared, are provided in the manuscript and Supporting Information . Funding: The author(s) received no specific funding for this work. Competing interests: The authors have declared that no competing interests exist. Introduction Electronic health records (EHRs) have become an important source of real-world data for oncology research, supporting studies of disease epidemiology, treatment patterns, comparative effectiveness, and health care utilization [ 1 , 2 ]. However, EHRs are generated primarily for clinical care, administration, and billing rather than research. Identifying a cancer cohort from structured EHR data therefore requires a computable phenotype: a rule that translates the clinical construct of interest into specific data elements, codes, time windows, and logical criteria [ 3 ]. The choice of phenotype determines who is included, how the index date is assigned, which patients are eligible for matching, and how downstream outcome estimates are generated. Invasive breast cancer is a useful condition in which to examine these issues. It is one of the most frequently studied cancers in EHR-based research and remains a major global health burden [ 4 , 5 ]. At the same time, EHR-based identification of breast cancer must address several clinically meaningful distinctions, including invasive cancer versus carcinoma in situ, incident versus prevalent disease, and confirmed diagnoses versus suspected or rule-out diagnoses. Small differences in coding logic may therefore produce meaningfully different cohorts. Operational definitions of breast cancer in EHR-based studies range from minimal diagnosis-code rules to definitions that combine diagnosis codes with treatment evidence [ 6 , 7 ]. A broad single-code definition is expected to improve sensitivity but may retain rule-out diagnoses or miscoded entries. A repeated-code definition, requiring a second confirmatory diagnosis code within a specified interval, is expected to improve specificity but may exclude patients with limited follow-up, fragmented care, or incomplete capture of subsequent encounters. A diagnosis-plus-treatment definition may improve clinical plausibility by requiring evidence of cancer-directed treatment, but it may also select patients with more complete treatment capture, more intensive care pathways, or different disease severity. In a Japanese claims-based validation study, the positive predictive value for breast cancer differed substantially between prevalent-case and incident-case algorithms, from 86.4% to 59.9%, illustrating how design choices in claims-based cancer identification can materially change algorithm performance [ 8 ]. Federated EHR research platforms such as the TriNetX Global Collaborative Network make these choices particularly consequential. Large multisite EHR networks allow rapid construction of very large cohorts, but large sample sizes also make small definitional differences statistically detectable. In addition, heterogeneity in coding practices, encounter capture, and treatment recording across health care organizations may affect how an operational definition performs across settings. Despite this, EHR-based breast cancer studies often apply a single case definition without examining how alternative specifications change cohort composition or downstream effect estimates. Negative-control outcomes, which are not expected to be causally affected by the exposure or classification strategy under study, provide one approach for assessing whether observed differences may reflect residual imbalance, selection, or other sources of bias not captured by measured covariates [ 9 ]. Rather than establishing diagnostic validity against a clinical gold standard, this study benchmarked the analytic consequences of three plausible EHR-based operational definitions of incident invasive breast cancer spanning increasing levels of diagnostic and treatment confirmation: a single diagnosis-code definition (Dx1), a repeated diagnosis-code definition (Dx2), and a diagnosis code combined with treatment evidence (DxTx). Using the TriNetX Global Collaborative Network [ 10 ] with one-to-one propensity score matching on 13 baseline covariates, we compared cohort yield, baseline characteristics, all-cause mortality, and emergency department visits across definitions. Appendicitis and femoral neck fracture were included as negative-control outcomes. Methods Study design and data source This retrospective benchmarking cohort study evaluated whether alternative diagnosis-based operational definitions of invasive breast cancer yielded different downstream outcome estimates in real-world electronic health record data. De-identified data were queried and aggregate analysis results were extracted from the TriNetX Global Collaborative Network (TriNetX, LLC) on May 3, 2026. All analyses were performed on the same date. Accordingly, all cohort counts, matching results, and outcome estimates reflect the data available in the network on that date. TriNetX is a federated research platform that provides access to aggregated electronic health record data from participating health care organizations across inpatient and outpatient care settings. Available data included demographics; diagnoses coded using the International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM); procedures coded using the International Classification of Diseases, 10th Revision, Procedure Coding System (ICD-10-PCS) and/or Current Procedural Terminology (CPT), medications; laboratory measurements mapped to Logical Observation Identifiers Names and Codes (LOINC); and health care utilization records. Study population The source population comprised female patients aged 20 years or older in the TriNetX Global Collaborative Network. Male patients and patients younger than 20 years were excluded. We used an incident invasive breast cancer-oriented design. Invasive breast cancer was identified using ICD-10-CM codes for malignant neoplasm of the breast (C50.*), and codes for carcinoma in situ of the breast (D05.*) were excluded to avoid inclusion of non-invasive disease. Eligible patients were required to have at least 12 months of observable history before cohort entry; this requirement was applied to all patients without exception. Washout period and index date A 12-month washout period before the index date was applied for all case definitions. During this period, patients were required to have no recorded C50.* diagnosis code. Patients with D05.* codes only were not considered to meet the invasive breast cancer definition. The index date was defined separately for each operational definition, as described below. Case definitions Three alternative operational definitions of invasive breast cancer were selected a priori to evaluate how different requirements for diagnosis and treatment evidence influence cohort composition and downstream outcome estimates. The single-diagnosis-code definition (Dx1) required at least one ICD-10-CM C50.* diagnosis code. The repeated-diagnosis-code definition (Dx2) required at least two C50.* diagnosis records, with a repeat diagnosis recorded within 90 days of the initial qualifying diagnosis. The diagnosis-plus-treatment definition (DxTx) required at least one C50.* diagnosis code together with at least one qualifying breast cancer-directed treatment: surgery, chemotherapy, or breast cancer-related medication within 90 days after the index date, or radiotherapy within 180 days after the index date. Dx2 and DxTx represented distinct ascertainment strategies based on repeated diagnosis coding and treatment capture, respectively. These additional requirements were not assumed to confer greater diagnostic validity. For all three definitions, invasive breast cancer was identified using ICD-10-CM C50.* codes, and D05.* codes were excluded to avoid inclusion of carcinoma in situ only. Pathology findings, tumor stage, tumor grade, and receptor-status criteria were not used to determine cohort membership. The complete diagnosis, procedure, and medication concepts used in the operational definitions are provided in S4 Table . Covariates and propensity score matching Propensity score matching was performed separately for three pairwise comparisons: DxTx versus Dx1, Dx2 versus Dx1, and DxTx versus Dx2. The same 13 prespecified baseline covariate terms were included in each propensity score model: age at index; five race categories recorded in TriNetX (White, Black or African American, Asian, other race, and unknown race); hypertension; diabetes mellitus; stroke; dyslipidemia; chronic kidney disease; body mass index; and serum albumin. The race categories were predefined by TriNetX and were not reclassified by the investigators. For each comparison, patients from the two phenotype-defined cohorts were matched one-to-one on their propensity scores using the TriNetX built-in matching function, which employed a greedy nearest-neighbor algorithm with a caliper width of 0.1 pooled standard deviations. Each comparison therefore generated its own matched sample, which was used for the corresponding outcome analyses. Covariate balance before and after matching was assessed using absolute standardized mean differences, with values below 0.10 indicating adequate balance on the measured covariates. The underlying platform implementation is proprietary, and individual patient-level matching assignments could not be independently verified by the authors. Outcomes and follow-up The primary clinical outcomes were all-cause mortality and emergency department use. Mortality was evaluated as time to recorded death from any cause; cause-specific mortality, cancer progression, cancer recurrence, progression-free survival, and hospitalization were not evaluated. Emergency department use was evaluated as time to the first recorded emergency department encounter within each assessment window and therefore did not measure the total number of recurrent emergency department visits. Outcomes were assessed across three prespecified windows: day 180–1 year, day 180–3 years, and day 180–5 years after the original index date. Follow-up began 180 days after the index date to account for the treatment-ascertainment period used in the DxTx definition. The exact TriNetX concepts and coding criteria used to define all clinical and negative-control outcomes are provided in S5 Table . Negative-control outcomes Appendicitis and femoral neck fracture were selected as negative-control outcomes using the diagnosis concepts specified in S5 Table . Each was analyzed as time to the first recorded occurrence within the corresponding assessment window. These outcomes were used to assess whether the alternative case definitions were associated with outcomes that were not expected to be directly determined by the breast cancer phenotype specification after propensity score matching. Ethical considerations The study protocol was reviewed and approved by the Institutional Review Board of the Research Ethics Committee of Buddhist Dalin Tzu Chi Hospital, Taiwan (No. B11402057). Statistical analysis Patient characteristics were summarized descriptively for each operational definition using counts and percentages for categorical variables and summary statistics for continuous variables, where available. Baseline differences before and after matching were evaluated using standardized mean differences. No formal sample size calculation was performed because this was a retrospective analysis of all eligible patients meeting the predefined case definitions in TriNetX. Within each matched pairwise comparison, all-cause mortality, emergency department visits, appendicitis, and femoral neck fracture were analyzed as time-to-event outcomes. Because the negative-control analyses comprised 18 tests across two outcomes, three pairwise phenotype comparisons, and three assessment windows, p values for these analyses were additionally adjusted using the Holm procedure to control the family-wise error rate. Both nominal and Holm-adjusted p values were considered when interpreting the negative-control findings. Kaplan-Meier methods were used to estimate event-free probabilities for each follow-up window, and hazard ratios (HRs) with 95% confidence intervals (CIs) were estimated to compare outcomes between definitions. Log-rank tests were used to compare event-free curves between groups. In each pairwise comparison, the first-named definition was treated as the comparison group and the second-named definition as the reference group. E-values were calculated for the point estimate and for the confidence-limit bound closest to the null to assess the minimum strength of association that an unmeasured confounder would need to have with both the case-definition group and the outcome, conditional on the measured covariates, to fully explain the observed association [ 11 ]. A two-sided p value < 0.05 was considered statistically significant. Event-free probabilities were exported from TriNetX and plotted using R (version 4.5.3). Results Cohort assembly As of the May 3, 2026 query date, the TriNetX Global Collaborative Network included 82,364,945 female patients aged 20 years or older, who constituted the source population for the three operational definitions. After application of the 12-month washout period and exclusion of patients with carcinoma in situ only, the Dx1, Dx2, and DxTx definitions identified 1,450,233, 678,739, and 578,576 patients, respectively. As expected from the increasing confirmation requirements of the three definitions, the DxTx cohort was the smallest and the Dx1 cohort was the largest. The selection of patients meeting each operational definition is shown in Fig 1 . Download: PNG larger image TIFF original image Fig 1. Cohort assembly for three operational definitions of incident invasive breast cancer in TriNetX. Qualifying treatment evidence comprised at least one of the following: surgery, chemotherapy, or breast cancer-related medication within 90 days after the index date, or radiotherapy within 180 days after the index date. https://doi.org/10.1371/journal.pone.0361883.g001 Baseline characteristics and propensity score matching Before matching, the three case-definition groups differed on several baseline demographic and clinical variables. After one-to-one propensity score matching on the 13 prespecified baseline covariates, standardized mean differences for all covariates were below the prespecified 0.10 threshold in every pairwise comparison, indicating adequate balance. Matching produced 565,417 pairs for the DxTx versus Dx1 comparison, 678,739 pairs for the Dx2 versus Dx1 comparison, and 555,617 pairs for the DxTx versus Dx2 comparison. Detailed baseline characteristics before and after matching are provided in S1 – S3 Tables , and covariate balance before and after matching is summarized visually using Love plots in Fig 2 . Download: PNG larger image TIFF original image Fig 2. Love plots of covariate balance before and after propensity-score matching across three pairwise comparisons. Absolute standardized mean differences are shown for (a) DxTx versus Dx1, (b) Dx2 versus Dx1, and (c) DxTx versus Dx2. The vertical dashed line indicates the 0.10 threshold for acceptable covariate balance. Dx1, single-diagnosis-code definition; Dx2, repeated-diagnosis-code definition; DxTx, diagnosis-plus-treatment definition. https://doi.org/10.1371/journal.pone.0361883.g002 All-cause mortality Higher mortality was observed in the more stringent definition group in ev
## Related Clinical Research

- [Potassium-Enriched Salt Could Prevent Millions of Cardiovascular and Kidney Disease Deaths Globally](https://medichelpline.com/clinical-feed/medical-news-today-0-potassium-enriched-salt-could-help-prevent-millions-of-deaths-globally-per-year.md)
- [Spatial Clustering of CKM Risk Burden Across US Census Tracts: Aging, Environment, and Socioeconom](https://medichelpline.com/clinical-feed/medrxiv-10-aging-structural-and-neighborhood-environmental-correlates-of-cardiovascular.md)
- [Medicinal Plants and Diabetic Nephropathy: Phytochemicals, Mechanisms, and Therapeutic Potential](https://medichelpline.com/clinical-feed/frontiers-in-immunology-2-medicinal-plants-in-diabetic-nephropathy-bioactive-compounds-mechanisms-and.md)
- [ZEUS trial: IL-6 blocker ziltivekimab lowered inflammation but did not cut major CV events](https://medichelpline.com/clinical-feed/aha-news-1-researchers-say-inflammation-as-a-treatment-target-still-important-despite.md)
- [Cheryl A.M. Anderson to receive AHA 2026 Chairman’s Award for public health and nutrition leadersh](https://medichelpline.com/clinical-feed/aha-news-4-public-health-and-nutrition-research-champion-to-receive-2026-chairman-s-award.md)

## Navigation
- [← Back to Oncology Feed](https://medichelpline.com/clinical-feed/oncology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.