Diagnostic accuracy of case-identification algorithms for heart failure in the general population using routinely...

Heart failure (HF), affecting 1-4% of adults in industrialized countries, is a major public health priority. Several algorithms based on administrative health data (HAD) have been developed to detect patients with HF in a timely and inexpensive manner, in order to perform real-world studies at the population level. However, their reported diagnostic accuracy is highly variable.

To assess the diagnostic accuracy of validated HAD-based algorithms for detecting HF, compared to clinical diagnosis, and to investigate causes of heterogeneity.

We included all diagnostic accuracy studies that utilized HAD for the diagnosis of congestive HF in the general adult population, using clinical examination or chart review as the reference standard. A systematic search of MEDLINE (1946-2023) and Embase (1947-2023) was conducted, without restrictions. The QUADAS-2 tool was employed to assess the risk of bias and concerns regarding applicability. Due to low-quality issues of the primary studies, associated with both the index test and the reference standard definition and conduct, and to the high level of clinical heterogeneity, a quantitative synthesis was not performed. Measures of diagnostic accuracy of the included algorithms were summarized narratively and presented graphically, by population subgroups.

We included 24 studies (161,524 patients) and extracted 36 algorithms. Algorithm selection was based on type of administrative data and DOR. Six studies (103,018 patients, 14 algorithms) were performed in the general outpatient population, with sensitivities ranging from 24.8 to 97.3% and specificities ranging from 35.6 to 99.5%. Eight studies (14,957 patients, 10 algorithms) included hospitalized patients with sensitivities ranging from 29.0 to 96.0% and specificities ranging from 65.8 to 99.2%. The remaining studies included subgroups of the general population or hospitalized patients with cardiologic conditions and were analyzed separately. Fourteen studies had one or more domains at high risk of bias, and there were concerns regarding applicability in 9 studies.

The considerable percentage of studies with a high risk of bias, together with the high clinical heterogeneity among different studies, did not allow to generate a pooled estimate of diagnostic accuracy for HAD-based algorithms to be used in an unselected general adult population.

PROSPERO CRD42023487565.