Publications

BACKGROUND: Administrative billing codes commonly misclassify complex diseases like heart failure (HF) in the electronic health record (EHR). Artificial intelligence (AI) approaches, including natural language processing (NLP), may improve disease adjudication.

OBJECTIVES: We validated an existing NLP model and specifically examined accuracy across HF subtypes (preserved vs reduced ejection fraction: HFpEF, EF ≥50%; HFrEF, EF <50%) in an EHR-based sample.

METHODS: We identified individuals who sought primary care between 2004-2014 at Beth Israel Deaconess Medical Center. Incident HF hospitalizations were identified using International Classification of Diseases (ICD) codes, and gold-standard HF adjudication was performed in 500 random cases. We applied a preexisting NLP model to classify each case and then further fine-tuned the model. We calculated diagnostic metrics for overall HF and by subtype.

RESULTS: Among 156,409 individuals, 5014 incident HF cases were identified. The fine-tuned NLP model outperformed primary ICD codes, yielding a superior kappa (0.73 vs 0.49), negative predictive value (93% vs 73%), and sensitivity (89% vs 52%), with similar positive predictive value (79% vs 87%) and specificity (86% vs 94%). Notably, NLP performance was robust across HF subtypes, whereas primary ICD codes had worse performance for identification of HFpEF. When applied to the full cohort, NLP identified 84% more HFrEF and 99% more HFpEF candidate cases than primary ICD codes.

CONCLUSION: NLP accurately classifies HF cases across subtypes, with improved accuracy after fine-tuning. AI-based approaches represent a critical opportunity to improve disease ascertainment at scale, nearly doubling the number of HF cases identified in the EHR.

BACKGROUND: Heart failure is classically divided into 2 subtypes based on the ejection fraction (EF): HF with reduced ejection fraction (HFrEF) and HF with preserved ejection fraction (HFpEF), each with varying risk-mitigation strategies. Electrocardiogram-based artificial intelligence (AI) may facilitate HF subtype-specific risk prediction, enabling tailored prevention.

METHODS AND RESULTS: Based on a previously validated HF-prediction architecture, we developed an ECG-based AI model to predict 10-year risk of incident HFrEF (LVEF <50%) vs HFpEF (LVEF ≥50%) (Electrocardiogram to Heart Failure Subtype, ECG2HF-Subtype). ECG2HF-Subtype was developed in 98,190 individuals at Massachusetts General Hospital (MGH) and validated in 3 test sets: MGH, Brigham and Women's Hospital (BWH), and Beth Israel Deaconess Medical Center (BIDMC) among individuals without prior HF. HF was identified by using a validated natural language processing model, and subtypes were classified by LVEF proximate to incident HF. Discrimination was quantified using time-dependent area under the receiver operating characteristic curve (AUROC) and average precision (AP), with comparisons to age and sex, and a multinomial logistic regression classifier, including variables in the Predicting Risk of cardiovascular disease EVENTs (PREVENT-HF) equations. In aggregate, models were evaluated in MGH (n = 16,739, ages 56 ± 17, 54% women), BWH (n = 75,195, 56 ± 16, 59%), and BIDMC (n = 30,133, 56 ± 17, 54%). ECG2HF-Subtype consistently discriminated incident HFrEF (MGH 10-year AUROC 0.89 [0.86-0.91], 10-year AP: 0.11 [95%CI 0.09-0.15], cumulative incidence 1.73%; BWH: 0.88 [0.87-0.89], 0.13 [0.11-0.14], 2.23%; BIDMC: 0.88 [0.86-0.89], 0.14 [0.12-0.17], 2.19%), and HFpEF events (MGH 10-year AUROC 0.83 [0.81-0.85], 10-year AP: 0.10 [95% CI 0.09-0.13], cumulative incidence 3.17%; BWH: 0.79 [0.77-0.80], 0.06 [0.06-0.07], 2.75%; BIDMC: 0.81 [0.80-0.82], 0.10 [0.09-0.12], 3.49%), with consistently greater AUROC vs age and sex. When compared to PREVENT-HF variables, ECG2HF-Subtypes had favorable discrimination of HFrEF (eg, 10-year AUROC in MGH+BWH 0.86 [0.84-0.88] vs 0.78 [0.76-0.80]), and modestly lower for HFpEF (10-year AUROC in MGH+BWH 0.78 [0.76-0.80] vs 0.81 [0.79-0.82]).

CONCLUSIONS: By predicting the risk of HFpEF and HFrEF, AI-enabled ECG analysis may enable targeted prevention to reduce HF-related morbidity.

BACKGROUND: Heart failure is classically divided into two subtypes based on the ejection fraction (EF): HF with reduced ejection fraction (HFrEF) and HF with preserved ejection fraction (HFpEF), each with varying risk mitigation strategies. Electrocardiogram-based artificial intelligence (AI) may facilitate HF subtype-specific risk prediction, to enable tailored prevention.

METHODS AND RESULTS: Based on a previously validated HF prediction architecture, we developed an ECG-based AI model to predict 10-year risk of incident HFrEF (LVEF <50%) versus HFpEF (LVEF ≥50%) (Electrocardiogram to Heart Failure Subtype, "ECG2HF-Subtype"). ECG2HF-Subtype was developed in 98,190 individuals in Massachusetts General Hospital (MGH), and validated in three test sets: MGH, Brigham and Women's Hospital (BWH), and Beth Israel Deaconess Medical Center (BIDMC), among individuals without prior HF. HF was identified using a validated natural language processing model, and subtypes were classified by LVEF proximate to incident HF. Discrimination was quantified using time-dependent area under the receiver operating characteristic curve (AUROC) and average precision (AP), with comparisons to age and sex, and a multinomial logistic regression classifier including PREVENT-HF variables. In aggregate, models were evaluated in MGH (n=16,739, age 56±17, 54% women), BWH (n=75,195, 56±16, 59%) and BIDMC (n=30,133, 56±17, 54%). ECG2HF-Subtype consistently discriminated incident HFrEF (MGH 10-year AUROC 0.89 [0.86-0.91], 10-year AP: 0.11 [95%CI 0.09-0.15], cumulative incidence 1.73%; BWH: 0.88 [0.87-0.89], 0.13 [0.11-0.14], 2.23% BIDMC: 0.88 [0.86-0.89], 0.14 [0.12-0.17], 2.19%) and HFpEF events (MGH 10-year AUROC 0.83 [0.81-0.85], 10-year AP: 0.10 [95% CI 0.09-0.13], cumulative incidence 3.17%; BWH: 0.79 [0.77-0.80], 0.06 [0.06-0.07], 2.75%), BIDMC: 0.81 [0.80-0.82], 0.10 [0.09-0.12], 3.49%), with consistently greater AUROC versus age and sex. When compared to PREVENT-HF variables, ECG2HF-Subtypes had favorable discrimination of HFrEF (e.g., 10-year AUROC in MGH+BWH 0.86 [0.84-0.88] vs. 0.78 [0.76-0.80]), and modestly lower for HFpEF (10-year AUROC in MGH+BWH 0.78 [0.76-0.80] vs. 0.81 [0.79-0.82]).

CONCLUSIONS: By predicting risk of HFpEF and HFrEF, AI-enabled ECG analysis may enable targeted prevention to reduce HF-related morbidity.

BACKGROUND: The PREVENT-HF equations estimate 10-year risk of overall heart failure (HF), inclusive of both HF with preserved and reduced ejection fraction (HFpEF; HFrEF). Given differential risk factor profiles, pathophysiology, and therapeutic strategies for each HF subtype, we sought to examine the performance of PREVENT-HF for subtype-specific risk assessment.

METHODS: We included data from five pooled longitudinal cohort studies (baseline exam 1989 to 2013) and two longitudinal primary-care electronic health record-based samples (Beth Israel Deaconess Medical Center [BIDMC], baseline 2004 to 2014; Mass General Brigham [MGB], baseline 2001 to 2018). We assessed the performance of the PREVENT-HF equations in predicting 10-year incident HFrEF and HFpEF, including model discrimination (area under the receiver operating characteristic curve [AUROC]; average precision [AP]/event rate) and calibration (integrated calibration index [ICI]), in adults aged 30-79 years without prevalent cardiovascular disease or HF.

RESULTS: The pooled cohort studies included 23,652 (55% women, age 60 ± 13 years), of whom 1,232 developed HF (432 [35%] HFpEF, 522 [42%] HFrEF). The BIDMC sample included 29,815 (55% women, 50 ± 13 years), of whom 259 (1%) developed HF (133 [51%] HFpEF, 101 [39%] HFrEF). MGB included 138,999 (62% women, 52 ± 13 years), of whom 1050 developed HF (586 [56%] HFpEF, 381 [36%] HFrEF). PREVENT-HF demonstrated better discrimination for HFpEF than HFrEF across all samples: in the pooled cohorts, the AUROC for HFpEF was 0.824 (95% CI 0.807, 0.839) versus 0.789 (95% CI 0.770, 0.807) for HFrEF; in BIDMC, the AUROC for HFpEF was 0.897 (95% CI 0.874, 0.918) versus 0.833 (95% CI 0.791, 0.869) for HFrEF; in MGB, the AUROC for HFpEF was 0.864 (95% CI 0.848, 0.880) versus 0.797 (95% CI 0.771, 0.820) for HFrEF. Similar results were reflected in the AP/event rate ratios: 2.92 (95% CI 2.47, 3.54) for HFpEF compared to 2.66 (95% CI 2.33, 3.14) for HFrEF in the pooled cohorts; 8.08 (95% CI 6.12, 12.30) for HFpEF compared to 5.12 (95% CI 3.84, 8.17) for HFrEF in BIDMC; and 6.84 (95% CI 5.66, 8.28) for HFpEF compared to 4.12 (95% CI 3.37, 5.16) for HFrEF in MGB. Calibration was similar for HFpEF and HFrEF across samples. For example, in the pooled cohorts, the ICI for HFpEF was 0.041 (95% CI 0.039, 0.043) and for HFrEF was 0.038 (95% CI 0.036, 0.040).

CONCLUSIONS: PREVENT-HF demonstrated excellent discrimination for both HF subtypes across community-based cohorts and real-world samples, with consistently better discrimination for HFpEF as compared to HFrEF. Given PREVENT-HF was developed to estimate overall HF risk rather than subtype-specific risk, future studies are needed to establish subtype-specific risk thresholds.

BACKGROUND: Despite advances in diagnostic capabilities and therapeutic options for aortic stenosis (AS), evidence suggests substantial underdiagnosis and care delays in real-world practice. This study examines patterns and disparities in AS diagnosis following echocardiographic identification, quantifying delays in diagnosis and subsequent care.

METHODS: Using the Optum® Market Clarity Dataset (2016-2023), we identified patients with first echocardiographic evidence of moderate or severe AS. Time-to-event analyses examined delays between echocardiographic findings and formal diagnosis (defined by ICD code assignment), subsequent imaging, and interventions. Analyses were stratified by severity, demographics, and care setting.

RESULTS: Among 2366 moderate and 1327 severe AS patients, 22.0% and 38.3% received formal diagnoses within one year of echocardiogram, respectively. Cumulative incidence of AS diagnosis at one year was lower in women vs men (moderate: 17.3% vs 27.8%; severe: 33.8% vs 43.0%) and non-White vs White patients (moderate: 11.4% vs 24.6%; severe: 25.8% vs 40.4%; all P ≤ 0.0001). Formal diagnosis was associated with increased likelihood of subsequent care (composite of follow-up imaging or aortic valve replacement) (moderate AS: HR 1.68, P < 0.0001; severe AS: HR 2.25, P < 0.0001).

CONCLUSIONS: This analysis reveals substantial delays in formal ICD-coded AS diagnosis following its echocardiographic detection, with concerning disparities across demographic groups. These findings suggest gaps between clinical recognition and formal documentation of AS, with implications for care continuum.

CLINICAL PERSPECTIVES: What Is New?•Using a large, real-world integrated claims and EMR database, this study demonstrates that fewer than one-quarter of patients with echocardiographically confirmed moderate AS and fewer than 40% with severe AS receive a formal ICD-coded diagnosis within one year, with substantially lower rates among women and non-White patients.What Are the Clinical Implications?•These findings reveal a critical gap between echocardiographic detection of AS and formal clinical documentation, suggesting that administrative claims data alone may substantially underestimate the true burden of AS in the population.•Formal AS diagnosis was independently associated with a significantly higher likelihood of receiving subsequent care - including surveillance imaging and aortic valve replacement - underscoring that closing the documentation gap may be a meaningful lever for improving outcomes and reducing disparities in AS management.

BACKGROUND: Ambulatory cardiac monitoring (ACM) is fundamental to arrhythmia evaluation, yet robust comparative evidence among alert-based monitoring strategies is lacking.

OBJECTIVE: To compare diagnostic yield and downstream clinical and economic outcomes among arrhythmia diagnosis-naïve Medicare Fee-For-Service (FFS) beneficiaries receiving mobile cardiac telemetry (MCT) or ambulatory event monitoring (AEM).

METHODS: Using 100% Medicare FFS claims from 2020 to 2022, this retrospective cohort study included beneficiaries aged 65 years or older, excluding those with prior arrhythmia diagnoses, antiarrhythmic medication use, or rhythm-management procedures. Cohorts were iRhythm MCT, Other MCT, and AEM; inverse probability of treatment weighting (IPTW) balanced baseline characteristics. Outcomes included diagnostic yield at 90 days, retesting at 180 days, and clinical events and costs at one year.

RESULTS: Among 441,792 eligible beneficiaries, 17,773 received iRhythm MCT (4%), 368,830 Other MCT (83%), and 55,189 AEM (13%). After IPTW, 90-day diagnostic yield was higher with iRhythm MCT (49.5%) than Other MCT (40%; HR 0.79 [95% CI, 0.77-0.81]) and AEM (39.1%; HR 0.76 [95% CI, 0.74-0.78]). Retesting at 180 days was more frequent with iRhythm MCT (8.8%) than Other MCT (6.3%; HR 0.70 [95% CI, 0.66-0.74]) and AEM (7.2%; HR 0.81 [95% CI, 0.76-0.87]). Pacemaker implantation was more common with iRhythm MCT; no differences were identified in hospitalizations, emergency visits, or heart failure.

CONCLUSIONS: Diagnostic yield and retesting varied across ACM platforms, though residual confounding cannot be excluded. No differences were identified in outcomes, utilization, or costs. Randomized trials are warranted to determine true differences in platform performance.

Randomized controlled trials (RCTs) remain the cornerstone of causal inference on the safety and efficacy of medicinal products. But their limited follow-up, controlled settings, and narrowly defined data collection to balance the burden to patients and maintain study feasibility often results in unaddressed important questions for health authorities and other clinical decision makers. Linkage of RCT data to routinely collected health data [real-world data (RWD)] offers a mechanism for addressing these gaps by extending observations and outcomes assessment into routine practice. This paper synthesizes methodological and operational considerations for RCTs with RWD linkage, drawing on deterministic, probabilistic, referential, and privacy-preserving record linkage methodologies and on four case studies that span different indications and regional settings: a long-term follow-up of a human papillomavirus vaccine trial using deterministic linkage via universal Personal Identity Numbers; a U.S. linkage of a respiratory syncytial virus vaccine trial to administrative health claims using privacy-preserving tokenization; and two transcatheter aortic valve replacement trials linked to Medicare claims to reproduce randomized treatment-effect estimates and to assess transportability to the broader Medicare population. We offer actionable methodological considerations and recommendations, drawing on the case studies and the broader linkage literature. Progress beyond the current state will require global health authority guidance specific to linking clinical trials with routinely collected data, shared benchmarks for evaluating linkage approaches, routine reporting and assessments of impact of use cases on regulatory and reimbursement decisions. Linkage of clinical trial data with routinely collected data sources can meaningfully accelerate clinical development and inform health authority and other clinical decision making.

BACKGROUND: The utility of routine troponin testing to identify recurrent myocardial infarction (MI) after an incident MI is unclear. We assessed the incidence and prognosis of recurrent MIs identified from centralized troponin review in patients from the Myocardial Ischemia and Transfusion (MINT) trial.

METHODS: The MINT trial randomized patients with acute MI and anemia to a liberal versus restrictive red blood cell transfusion strategy. Suspected recurrent MIs were identified through both site-report and centralized review of troponin levels collected for 3 days following randomization. Differences in cardiac, non-cardiac, and all-cause death at 30 and 180 days were compared across patients with any site-reported MI, only centrally identified MI, and no recurrent MI.

RESULTS: Among 3,504 patients, 275 (7.8%) had a recurrent MI within 30 days; 119 (43.3%) by site-report, and 156 (56.7%) by central troponin review only. Rates of cardiac and all-cause death at 30 and 180 days were highest for patients with site-reported MI, intermediate for centrally identified MI, and lowest for no recurrent MI; rates of non-cardiac death did not vary. Patients with only centrally identified recurrent MI had an increased risk of cardiac death at 30 days (RR 1.9, 95% CI 1.0-3.4) and 180 days (RR 1.7, 95% CI 1.1-2.7) compared to those without recurrent MI.

CONCLUSIONS: In patients with acute MI and anemia, centralized troponin review identified more than half of all recurrent MI events. Patients with centrally identified MI had a higher risk of cardiac death than those with no recurrent MI.

TRIAL REGISTRATION: ClinicalTrials.gov NCT02981407 https://clinicaltrials.gov/study/NCT02619136.

BACKGROUND: Reported adverse events (AEs) to the Food and Drug Administration Adverse Event Reporting System (FAERS) have suggested an increased rate of serious AEs (SAEs) during the COVID-19 pandemic, but the extent to which this may be related to overall changes in reporting during this time period is uncertain. Accordingly, we aimed to evaluate trends in SAE reporting across commercially available ultrasound enhancing agent (UEA) brands as a function of overall trends in AE reporting.

METHODS: We retrospectively analyzed the FAERS public database, 2014-2024, to evaluate risks of UEAs overall and by brand, compared to similar contrast media.

RESULTS: Between 2014 and 2024, 21,960,760 AEs were reported to FAERS, of which 11,450,891 (52.1%) were categorized as SAEs. Overall SAE reports to FAERS increased from 678,953 in 2014 to 1,368,393 in 2021 before subsequently declining to 1,065,845 in 2024 (-7.9% change from 2021 to 2024). During the same period, overall death reports to FAERS increased from 124,055 in 2014 to 195,207 in 2018 before declining to 147,046 in 2024. During this period of decline, there was a 23.9% relative increase in SAEs to UEAs which peaked in 2023 at 350 before declining to 326 in 2024. Deaths attributed to SAEs increased from 1 in 2014 to 19 in 2023 before declining to 9 in 2024. Overall, these data suggest that 11.2% of the observed increase in SAEs to UEAs can be attributed to reporting changes. Despite changes in relative risks for SAEs, absolute SAE rates remained small and lower than other types of contrast media.

CONCLUSIONS: In this analysis of the FAERS dataset, 2014-2024, 11.2% of SAEs to UEAs were attributable to temporal changes in AE reporting. Absolute risks are small and declining, suggesting broad safety of UEAs as a class. Collectively, these results support continued use of UEAs, but motivate improved safety screening and preparedness to mitigate small but existing risks.