The digitization of the medical record has represented a paradigm shift in patient care. The benefits of this transition, which spanned a decade and received significant financial investment from our health care system and unmeasurable clinician hours spent in adapting workflows, is now poised to transform the landscape of medical research. The work by Ambrosy and colleagues1 represents a key advance in this domain.
The study of the epidemiological mechanisms underlying cardiovascular disease requires the ability to accurately capture risk factors and the development of clinical disease and outcomes. These processes have often been achieved through large surveys, longitudinal cohort studies, and clinical data sources compiled from structured elements of the medical record.2
The last category is most commonly represented by claims-based data sets,3 wherein clinical entities are summarized as standardized codes, or through manually abstracted clinical registries. These clinical data sources enhanced the generalizability of assessments of disease trajectories, which were initially drawn from nonrepresentative cohort studies, and allowed the study of therapeutic approaches and the effects of health policies on patient outcomes.4 These data sources are simplified to ensure that data can be stored, transmitted, and processed for clinical studies that rely on their use. However, this has resulted in the reduction of multifaceted clinical scenarios to simpler abstractable elements, thereby lacking context about complexity of disease.5 The study of heart failure represents a distinctly challenging domain in which a heterogeneous disease is studied based on limited diagnostic categorization in medical claims.6 This challenge has endured because of a lack of alternative, richer data sources constructed using a clinical focus.
The study by Ambrosy et al1 represents an important proof of concept that a digital strategy of studying cardiovascular diseases can finally be deployed at scale. The study uses a decade of data from the Kaiser Permanente Northern California health system, representing 21 hospitals and associated ambulatory practices on the same electronic health record (EHR) platform to study epidemiologic and health care patterns in heart failure. The investigators use multidimensional data from the EHR, representing both structured elements and unstructured data from clinical notes. Their approach has a unique strength—it leveraged the details of clinical presentation of patients with a diagnosis of heart failure to specifically identify acutely decompensated heart failure. They achieved this through the creation of a broad, clinically meaningful definition of worsening heart failure. This definition included symptoms and signs of hypervolemia noted in specific sections of clinical documentation, including presenting patient concerns and physical examination findings, all drawn using natural language processing of clinical notes. This approach was complemented by the study of objective changes in vital signs and laboratory measures, markers of congestion and cardiomegaly on chest radiography, and the use of treatments that are frequently used in worsening heart failure. The definitions are comprehensive, are rigorously assessed with manual abstraction, and allow the study of hospitalization for heart failure using a definition that is clinically meaningful and actionable.
The key findings of the study1 are reassuring about the use of principal diagnosis of heart failure to study worsening heart failure because nearly 90% of such hospitalizations were indeed for decompensated heart failure. This finding was borne out in trends in heart failure hospitalizations, which were consistent with those based on the more nuanced definition of worsening heart failure in the current study. Despite the high specificity of a principal diagnosis of heart failure, the sensitivity of the approach is called into question. The authors found that twice as many hospitalizations were for worsening heart failure with the use of the natural language processing algorithms than those based on a principal diagnosis of heart failure. The additional group of patients represents those with a secondary diagnosis of heart failure and a different primary diagnosis but an algorithm-defined decompensation of heart failure. A greater dissection of these cases is challenging because the principal diagnosis may still represent the reason for admission, but the authors’ approach represents an advance in being able to identify hospitalizations for which acutely decompensated heart failure in the record has implications for quality improvement and disease monitoring in heart failure.
There are important implications of automating the accurate identification of a cohort with heart failure. First, the obvious benefit lies in the ability to conduct large-scale studies of the patterns of disease and their risk factors. Second, such digital transformation of health care can cause targets of quality improvement programs for hospitals to be well aligned with national policies designed to improve the care and outcomes of patients. For heart failure, this approach may signify the opportunity to enhance the use of evidence-based medical therapy that continues to be underused. Third, the accurate digital phenotyping of disease entities also allows investigating pathophysiological mechanisms, biology, and therapies to be more accurately assessed within clinically generated data. Fourth, the study engenders confidence about the ability to study rapidly and in real time the effect of health interventions on patient outcomes and to infer associations of strategies to curb diseases.
Although the study1 is comprehensive, an appraisal of its limitations offers avenues to ensure sustainability and scalability of the approach. First, like many studies that focus on computational phenotypes of disease, an underrecognized barrier is that model development and testing rely on an extracted cohort of patients rather than the broad population of patients in a large health care system. In the current study,1 this barrier is reflected in the use of a diagnosis of heart failure to define the denominator. Therefore, the prospective clinical utility of the model in identifying patients with worsening heart failure is unknown because patients without an encounter diagnosis of heart failure but admitted with worsening heart failure were not included in the development of their algorithms. A health analytics platform that allows accessibility to the broader population is essential to ensuring that the diagnostic algorithms will work in real time, when such diagnostic codes are not available.7
Second, it is conceivable that diagnostic coding differences across sites are manifested in the differences in their unstructured documentation strategies. Whether the algorithm generalizes to other sites is immensely important to assess the scalability of the approach, and whether it would usurp the reliance on claims and registries is also important. Third, the use of a proprietary natural language processing algorithm creates another challenge. The authors have provided a detailed description of their approach, including the terms that were incorporated in the model.1 However, the role of the algorithm itself in achieving good performance is hard to disassociate from the medical terms that define heart failure. Therefore, it is essential to replicate the natural language processing components of the algorithm using open-source tools because many such tools with broad functionality are available in the public domain.
Fourth, the study uses an operational definition of systolic heart failure that relies on the most recent echocardiogram.1 However, patients with echocardiograms outside the health care system or those with recovery of systolic function can be missed using this approach. Therefore, the results describing the heart failure subtypes should be interpreted with caution. Fifth, despite an advance in incorporating different data elements in the EHR, multidimensional longitudinal information in the EHR that spans different encounters is underexplored as a strategy to enhance the performance of these models.
In summary, the study ushers in an era where the automated digital ascertainment and phenotyping of cardiovascular disease will result in rapid-cycle, adaptable scientific discovery in clinical sciences. This will be key for achieving a better understanding of the disease and its treatment and the early detection of emerging health care challenges.
Published: November 22, 2021. doi:10.1001/jamanetworkopen.2021.35561
Open Access: This is an open access article distributed under the terms of the CC-BY License. © 2021 Khera R. JAMA Network Open.
Corresponding Author: Rohan Khera, MD, MS, Section of Cardiovascular Medicine, Department of Internal Medicine, Yale School of Medicine, 195 Church St, Fifth Floor, New Haven, CT 06510 (email@example.com).
Conflict of Interest Disclosures: Dr Khera reported receiving support from the National Heart, Lung, and Blood Institute of the National Institutes of Health under award 1K23HL153775 and is a founder of Evidence2Health, a precision digital health analytics platform.
et al. A natural language processing–based approach for identifying hospitalizations for worsening heart failure within an integrated health care delivery system. JAMA Netw Open
. 2021;4(11):e2135152. doi:10.1001/jamanetworkopen.2021.35152Google Scholar
et al. Association of the Hospital Readmissions Reduction Program With mortality during and after hospitalization for acute myocardial infarction, heart failure, and pneumonia. JAMA Netw Open
. 2018;1(5):e182777-e182777. doi:10.1001/jamanetworkopen.2018.2777
HM. Agile health care analytics: enabling real-time disease surveillance with a computational health platform. J Med Internet Res
. 2020;22(5):e18707. doi:10.2196/18707