Skip to main content

The use of Natural Language Processing-based algorithms and outpatient clinical encounters for worsening heart failure: Insights from the Utilize-WHF Study

Parikh R et al

Parikh R, Ambrosy A, HeeSung S, Narayanan A, Masson R, Lam P-Q, Kheder K, Iwahashi A, Hardwick A, Fitzpatrick J, Avula H, Shen X, Sanghera N, Cristino J, Go A



Background: The focus on reducing 30-day readmissions for heart failure (HF) has unintentionally shifted clinical care from the inpatient to the outpatient setting. We used a natural language processing (NLP)-based approach to systematically identify episodes of outpatient worsening HF (WHF).


We reviewed all emergency department (ED) visits and hospitalizations for observation status in 2019 at an integrated health care delivery system in Northern California. All qualifying clinical encounters for patients ≥18 years with a discharge diagnosis code of HF were included. Patients with end-stage renal disease were excluded. WHF was defined as ≥1 symptom, ≥2 objective findings including ≥1 sign, AND ≥1 change in HF-related therapy (new administration of an intravenous diuretic and/or augmentation of an oral diuretic). Vital signs, laboratory values, and pharmacy prescriptions ±7 days were accessed. Clinical documentation within 72 hours were analyzed using I2E software (Linguamatics, v5.4.1). Manual chart review by two physicians with final adjudication by a third was performed for a random sample of 200 cases to validate the approach. The observed positive predictive value = 100%, negative predictive value = 94%, and accuracy = 95%.


There were 38,652 qualifying encounters in 2019 including 26,599 ED visits (69%) and 12,053 observation stays (31%). The age was 75±14, 50% were women, 45% were non-white, and the left ventricular ejection fraction (%) was 51±15. Among 5,861 and 32,791 clinical encounters, respectively, for a primary and secondary diagnosis of HF, there were 4,424 (75%) and 3,983 (12%) confirmed episodes of WHF. The proportion of cases meeting the definition for outpatient WHF was higher for observation stays, compared to ED visits, for both encounters with HF listed as a primary (91% vs. 64%) and as a secondary diagnosis (23% vs. 8%).


The burden of ED visits and observation stays for WHF was found to be twofold higher using a standardized multidimensional definition compared to primary discharge diagnosis codes. Future research is necessary to clarify the clinical utility of an NLP-based approach for ascertaining ambulatory clinic appointments for WHF.

Ready to get started?

Request a Demo

Questions? Ask our experts