Heart failure

Optimizing the Machine Learning of an Electronic Health Record Audit for Heart Failure in Primary Care

This article was originally published here

ESC Heart Fail. November 23, 2021. doi: 10.1002 / ehf2.13724. Online ahead of print.


OBJECTIVES: The diagnosis of heart failure (HF) is an important problem in primary care. We previously demonstrated a 74% increase in CI diagnoses recorded in primary care electronic health records (EHRs) following an extended audit process. What remains unclear is the accuracy of the recorded HF pre-audit and which DSE variables are most important in the extended audit strategy. This study aims to describe the diagnostic classification sequence of HF at different stages, to assess misclassification of HF by general practitioners (GPs) and to test the predictive performance of an optimized audit.

METHODS AND RESULTS: This is a secondary analysis of the OSCAR-HF study, a prospective observational trial involving 51 participating general practitioners. OSCAR used an extensive audit based on risk factors, signs, symptoms and medications typical of HF in the EHR of GPs. This resulted in a list of possible CI patients, which participating GPs had to classify as CI or not CI. We compared the diagnoses of HF recorded before and after the evaluation of general practitioners. For our analysis of audit performance, we used general practitioner assessment of CI as the primary outcome and audit queries as dichotomous predictor variables for a machine gradient decision tree algorithm (GBM ) and a logistic regression model. Of the 18,011 patients eligible for the audit intervention, 4,678 (26.0%) were identified as patients with possible HF and submitted for evaluation by general practitioners during the stage of audit. There were 310 patients with HF registered before the GP assessment, of which 146 (47.1%) were judged not to have HF by their GP (over-registration). There were 538 patients with HF recorded after GP assessment, of which 374 (69.5%) had not recorded HF before GP assessment (under-recording). The GBM and the logistic regression model had comparable predictive performance (area under the curve 0.70 [95% confidence interval 0.65-0.77] and 0.69 [95% confidence interval 0.64-0.75], respectively). This was not significantly affected by the reduction in the set of predictor variables to the 10 most important variables identified in the GBM model (free text and coded cardiomyopathy, ischemic heart disease and atrial fibrillation, digoxin, mineralocorticoid receptor antagonists and combinations of renin-angiotensin systemic inhibitors and beta-blockers with diuretics). This optimized set of queries was sufficient to identify 86% (n = 461/538) of the HF self-assessed population by general practitioners with a 33% (n = 1537/4678) reduction in the number of cases. screening.

CONCLUSIONS: The diagnostic coding of CI in primary health care records is inaccurate with a high degree of under-recording and over-recording. An optimized set of queries made it possible to identify more than 80% of the HF population self-assessed by general practitioners.

IDPM: 34816632 | DOI: 10.1002 / ehf2.13724

Source link