Hierarchical Clustering Analysis to Inform Classification of Congenital Malformations for Surveillance of Medication Safety in Pregnancy

Background: Medication-safety studies in pregnancy typically focus on a single or selected outcomes. However, treatment decision making in pregnancy requires a sophisticated trade-off between benefits and all possible risks. Large healthcare utilization databases allow for simultaneous screening across a broad range of potential adverse pregnancy outcomes using scan statistic approaches such as TreeScanTM. When using such a safety surveillance approach, hierarchical classification of the outcomes based on shared pathogenetic mechanisms increases the power to detect signals for clinically related outcomes.

Objectives: We explored the use of hierarchical clustering analysis (HCA) to help inform the classification of congenital malformations, an outcome of most immediate concern.

Methods: Using a pregnancy cohort of >1.8 million mother-child dyads nested within the Medicaid Analytic eXtract 2000-2014, we identified 251,878 children with International Classification of Diseases, Ninth Revision (ICD-9) malformation codes (740.x-759.x) documented within 3 months after birth. After assessing the Pearson correlation coefficient for each unique pair of malformation codes, we applied agglomerative HCA, and created a cluster dendrogram. The dendrogram was manually reviewed to screen out clusters of codes used interchangeably (e.g., "unspecified" and "other") and clusters based on extremely sparse data.

Results: We identified the following clusters: (1) Malformations belonging to the same organ, organ system or body part (e.g., urinary tract anomalies, cleft lip and cleft palate, components within cardiac anomalies) which are grouped in existing clinical coding systems such as the ICD; (2) known combinations and sequences across different organ systems (e.g., splenic and cardiac anomalies in the context of situs inversus) not grouped in existing coding systems; (3) not previously recognized combinations of malformations occurring together more often than expected by chance (e.g., Meckel's diverticulum and urachus anomaly); and (4) malformations that seemed to stand on their own (e.g., pyloric stenosis, abdominal wall defects, transverse limb defects).

Conclusions: This human-assisted HCA based approach for hierarchical outcome classification suggested splitting and lumping certain malformations that are not considered as such in classification systems organized by body system. Augmenting existing clinical classifications with clusters identified through large data exploration may be a promising approach when defining phenotypes for scanning-based safety surveillance activities and causal inference studies of congenital malformations and, by extension, other outcomes.