319 - Tiny Babies, Big Data: The ability of ICD billing codes to identify genetic diagnoses

Publication Number: 1303.319

Elizabeth Brokamp, Boston Children's Hospital, Boston, MA, United States; Retna Arun, Boston Children's Hospital, Shrewsbury, MA, United States; Monica Wojcik, Boston Children's Hospital, Boston, MA, United States; Gwendolyn Strickland, Boston Children's Hospital, Boston, MA, United States; Anna-Thérèse Mehra, Boston Children's Hospital, Boston, MA, United States

Background: Rare genetic diseases often present in the neonatal intensive care unit (NICU). Accurate identification of neonates with genetic diseases (GD) in electronic health records (EHR) would enable a more complete understanding of their phenotypic spectrum, advancing care and personalized medicine. However, accurate and comprehensive identification of neonates with GD in the EHR still eludes the bioinformatic field. Prior studies have used International Classification of Diseases (ICD) billing codes as proxies, though their accuracy for detecting confirmed GDs is uncertain.

Objective: To evaluate the recall rate and accuracy of ICD codes for identifying neonates with confirmed GDs, we compared ICD code patterns between neonates with and without GD.

Design/Methods: Retrospective analysis was conducted on NICU patients admitted to a level IV NICU (2011-2025) who underwent genetic evaluations. Data from 1,344 neonates were collected, including confirmed GDs and ICD codes, which were mapped to phecodes (aggregated ICD codes). Analyses were restricted to phecodes with ≥2 unique instances, and GDs captured by phecodes were compared with confirmed GDs. A phenome-wide association study (PheWAS) compared phecode patterns between neonates with GDs (n=239) and without (n=863) across 191 non-genetic phecodes present in ≥10 individuals, adjusting for sex, age at admission, gestational age, and NICU length of stay.

Results: During their time in the NICU, 239 neonates received a GD and phecodes were able to correctly identify 82 (34.3%) (Figure 1). Phecodes from NICU billing alone predicted 34 (15.9%) of 214 GDs diagnosed within one year after NICU discharge.

Among 152 individuals with genetic phecodes, 25 (16.4%) were misclassified- 8 (5.3%) without a GD and 17 (11.1%) with an incorrect GD (Figure 2).

Although no phecodes reached statistical significance (p < 2.6×10⁻⁴), two showed strong associations with GD: atrioventricular septal defects (ASDs; p = 3.4×10⁻⁴) and cleft palate/lip (CP/L; p = 6.5×10⁻³). Several showed strong negative associations: tracheoesophageal fistula/esophageal atresia and stenosis (TEF/EA; p = 7.7×10⁻³), apnea of newborn (p = 9.0×10⁻³), encephalopathy (p = 1.4×10⁻²), TE fistula (p = 1.9×10⁻²), and bronchopulmonary dysplasia (BPD; p = 1.9×10⁻²) (Figure 3).

Conclusion(s): The use of ICD codes to identify NICU infants with GDs is neither sensitive nor accurate, though phecode analysis demonstrated stronger accuracy than sensitivity. Our data highlight clinical features of NICU infants more commonly seen in those with a GD (ASD, cleft palate) and those that are not (BPD, TEF/EA, encephalopathy).

Figure 1. Confirmed genetic disease and genetic phecode identification
Out of 239 neonates that received a genetic diagnosis in the NICU, phecodes correctly identified 82 (34.3%). While technically correct, 6 (2.5%) genetic phecodes could be more specific.

Genetic phecode accuracy
Out of 152 assigned genetic phecodes, 117 are correct, giving a positive predictive value of 77.0%.

Genetic disease vs no genetic disease phenotype-wide association study
A pheWAS plot for neonates with GDs. Phecode chapters are categorized by color along the x-axis: BI= BloodImmune, CA= Cancer, CM= CongenitalMalformation, CV= Cardiovascular, DE= Dermatological, EM= EndocrinMetabolic, GI= Gastrointestinal, GU= Genitourinary, ID= Infections, MB= Mental, MS= Musculoskeletal, NB= Neonatal, NS= Neurological, RE= Respiratory, SO= SenseOrgans, SS= Symptoms. The blue dotted line represents the significance threshold of p<= 2.6x10-4 .