JMIR Publications

ABSTRACT

Background:

Assessment of New York Heart Association (NYHA) functional class in heart failure (HF) has relied on electronic medical record (EMR) data, omitting psychosocial determinants now collectible through patient-facing applications.

Objective:

This study aimed to assess the feasibility of collecting measures of anxiety, perceived social support, and health literacy through a patient-facing digital health application and to determine whether these measures improve identification of advanced NYHA functional limitation beyond EMR data in adults with HF.

Methods:

This single-center cross-sectional feasibility pilot study enrolled adults with HF in inpatient and outpatient settings. Clinical data came from the EMR and patient-reported measures from the application. The originally specified 30-day readmission outcome yielded a single event, so the primary outcome was self-reported NYHA Class III-IV versus Class I-II, assessed concurrently with the predictors. Stage 1 used prespecified EMR variables (age, sex, HF duration, comorbidity count, guideline-directed medical therapy, and diuretic use); Stage 2 added the patient-reported measures. Logistic regression, elastic net, random forest, and gradient-boosted tree models were evaluated for discrimination, calibration, Brier score, and reclassification using repeated cross-validation, nested cross-validation, bootstrap optimism correction, and sensitivity analyses.

Results:

Of 100 consented participants, 93 were analyzed, and 42 (45%) reported NYHA Class III-IV. Core application instruments were completed by 92%-93% of consented participants. The Stage 2 elastic net (primary model) achieved a cross-validated area under the receiver operating characteristic curve (AUC) of 0.66 (95% confidence interval [CI] 0.55-0.77) versus 0.57 (95% CI 0.45-0.70) for Stage 1, and the primary test was nonsignificant (AUC difference +0.09; DeLong P=.09). The continuous net reclassification improvement was 0.45 (95% CI 0.08-0.83), but its event component was zero for both linear models. Two tree-based comparators reached nominal significance (P=.003 and P=.007) but were not considered confirmatory: the random forest EMR-only AUC was at chance (0.49), and gradient-boosted tree calibration and Brier score worsened with the added predictors. Under nested cross-validation, the Stage 2 elastic net AUC was 0.62, with a mean per-repeat increment of +0.10, and the bootstrap optimism-corrected Stage 2 AUC was 0.69. In multiply imputed multivariable logistic regression, the Generalized Anxiety Disorder 7-item scale score was the only significant patient-reported predictor (odds ratio 1.15 per point, 95% CI 1.02-1.30; P=.02).

Conclusions:

In this cross-sectional feasibility pilot study, completion of the application-collected instruments was high, but adding the patient-reported measures to EMR-derived clinical variables produced a numerical, nonsignificant improvement in discrimination of advanced self-reported NYHA functional limitation. Anxiety was the only patient-reported measure with an individual association across several analytic approaches, although this association was not uniformly robust and interpretation is limited by concurrent collection of predictor and outcome through the same application. Larger multisite studies with independently assessed functional outcomes and external validation are needed.