How Well Do ICD‐9‐CM Codes Predict True Congenital Heart Defects? A Centers for Disease Control and Prevention‐Based Multisite Validation Project

Abstract

Background The Centers for Disease Control and Prevention's Surveillance of Congenital Heart Defects Across the Lifespan project uses large clinical and administrative databases at sites throughout the United States to understand population‐based congenital heart defect (CHD) epidemiology and outcomes. These individual databases are also relied upon for accurate coding of CHD to estimate population prevalence. Methods and Results This validation project assessed a sample of 774 cases from 4 surveillance sites to determine the positive predictive value (PPV) for identifying a true CHD case and classifying CHD anatomic group accurately based on 57 International Classification of Diseases, Ninth Revision, Clinical Modification (ICD‐9‐CM) codes. Chi‐square tests assessed differences in PPV by CHD severity and age. Overall, PPV was 76.36% (591/774 [95% CI, 73.20–79.31]) for all sites and all CHD‐related ICD‐9‐CM codes. Of patients with a code for complex CHD, 89.85% (177/197 [95% CI, 84.76–93.69]) had CHD; corresponding PPV estimates were 86.73% (170/196 [95% CI, 81.17–91.15]) for shunt, 82.99% (161/194 [95% CI, 76.95–87.99]) for valve, and 44.39% (83/187 [95% CI, 84.76–93.69]) for “Other” CHD anatomic group ( X 2 =142.16, P <0.0001). ICD‐9‐CM codes had higher PPVs for having CHD in the 3 younger age groups compared with those >64 years of age, ( X 2 =4.23, P <0.0001). Conclusions While CHD ICD‐9‐CM codes had acceptable PPV (86.54%) (508/587 [95% CI, 83.51–89.20]) for identifying whether a patient has CHD when excluding patients with ICD‐9‐CM codes for “Other” CHD and code 745.5, further evaluation and algorithm development may help inform and improve accurate identification of CHD in data sets across the CHD ICD‐9‐CM code groups.

A machine learning model for predicting congenital heart defects from administrative data

“…Though excluded from our data, non-specific CHD codes, codes for extra-cardiac vascular anomalies, and the code used to indicate a normal patent foramen ovale variant (745.5) are often included in ICD-9-CM code group (745.XXÀ747.XX) used to identify CHD in administrative and clinical data despite evidence that the PPV of code group 745.XXÀ747.XX for a CHD is as low as 48.7% (Khan et al, 2018). After excluding the non-CHD codes in ICD-9-CM code group 745.XXÀ747.XX, and excluding isolated 745.5 from our data, the PPV was still less than ideal at 76.5%, although this was further improved to 86.5% by excluding the entire group of 'other' CHD codes in Rodriguez et al (2022). While PPV can be increased by restricting the data to certain codes, or to specific encounter types and numbers of encounters with CHD codes, entire subsets of TP CHD cases (e.g., all individuals with atrial septal defects) would be excluded and the resulting data would be skewed toward more severe cases.…”

“…For GA, NC, and NY, the age groups were 1−10‐year‐olds, 11−19‐year‐olds, 20−64‐year‐olds, and >64‐year‐olds, while ages for UT were 11−19‐year‐olds and 20−64‐year‐olds. Further methodological details can be found in Rodriguez et al (2022). During the study period, all contributing healthcare systems utilized ICD‐9‐CM codes.…”

“…Race and ethnicity data were collected to understand applicability of developed models to a population. The final data set available for analysis had five additional cases not included in Rodriguez et al (2022).…”

“…Features with higher mean absolute SHAP values like # CHD-coded outpatient healthcare encounters, 'other' CHD anatomic group diagnosis, and # CHD-coded healthcare encounters are more influential. Other features that influence prediction include several CCS categories: having a musculoskeletal system and connective tissue diagnosis; a circulatory disease diagnosis; an electrocardiography diagnosis; "factors influencing healthcare" which includes a subcategory "medication management"; and a respiratory system diagnosis (Rodriguez et al, 2022). CHD, congenital heart defects; SHAP, SHapley Additive exPlanation values.…”

Extracorporeal Life Support for Cardiogenic Shock in Adult Congenital Heart Disease—An ELSO Registry Analysis

“…ICD-9 and 10 codes were used to identify ACHD patients and their primary congenital diagnosis. Recent studies have demonstrated a reasonable degree of accuracy for these codes 27 ; however, there is a risk of inappropriately labeling non-ACHD patients and excluding ACHD patients from our cohort because of inaccurate coding. We have attempted to minimize this risk through sensitivity analysis with the exclusion of diagnoses most likely to inappropriately classify ACHD patients (congenital aortic stenosis and low-risk anatomy).…”