Performance of the FIND-FH machine learning algorithm for the identification of individuals with suspected familial hypercholesterolemia

BACKGROUND

Familial hypercholesterolemia (FH) is an inherited cholesterol disorder that is markedly underdiagnosed.

OBJECTIVE

This study evaluated the real-world performance of the Find, Identify, Network, Deliver-FH (FIND-FH) score, a novel machine learning algorithm, in identifying individuals with high likelihood of FH.

METHODS

The FIND-FH model was applied to electronic health record (EHR) data from UT Southwestern Medical Center. Manual chart review was performed on those deemed high probability of FH (score >0.35) to assess accuracy of FH diagnosis by modified Simon-Broome and Dutch Lipid Clinic Network (DLCN) criteria. Individual characteristics were compared across quintiles of the FIND-FH score. Individuals deemed suitable for FH outreach were identified using predetermined clinical criteria denoting adequate clinical probability of FH.

RESULTS

Of the 93,418 individuals in the EHR dataset, the FIND-FH algorithm identified 340 with high probability of FH. These 340 individuals had a mean age of 49.8 years, were 59% male, and had a highest low-density lipoprotein cholesterol (LDL-C) of 168.4 mg/dL (±51.9). A total of 20-32% met modified Simon-Broome or DLCN criteria for at least possible FH based on available EHR data. Several variables differed significantly by FIND-FH score quintile, including Simone-Broome and DLCN probability. In the reviewed cohort, 191 (56%) had enough clinical suspicion for FH to warrant outreach. Among these, 101 (53%) had highest LDL-C <190 mg/dL and would be missed by LDL-C-based FH screening strategies.

CONCLUSION

In a large healthcare system EHR cohort, most individuals identified as higher risk for FH by the FIND-FH algorithm were deemed appropriate for further evaluation, despite the majority not meeting FH diagnostic criteria using available EHR data.