Improving Familial Hypercholesterolemia Diagnosis Using an EMR-based Hybrid Diagnostic Model.

CONTEXT: Familial hypercholesterolemia (FH) confers a greatly increased risk for premature cardiovascular disease, but remains very underdiagnosed and undertreated in primary care populations.
OBJECTIVE: We assessed whether using a hybrid model consisting of 2 existing FH diagnostic criteria coupled with electronic medical record (EMR) data would accurately identify patients with FH in a Midwest US metropolitan healthcare system.
METHODS: We conducted a retrospective, records-based, cross-sectional study using datasets from unique EMRs of living patients. Using Structured Query Language to identify components of 2 currently approved FH diagnostic criteria, we created a hybrid model to identify individuals with FH.
RESULTS: Of 264 264 records analyzed, between 794 and 1571 patients were identified as having FH based on the hybrid diagnostic model, with a prevalence of 1:300 to 1:160. These patients had a higher prevalence of premature coronary artery disease (CAD) (38-58%) than the general population (1.8%) and higher than those having a high CAD risk but no FH (10%). Although most patients were receiving lipid-lowering therapies (LLTs), only 50% were receiving guideline-recommended high-intensity LLT.
CONCLUSION: Using the hybrid model, we identified FH with a higher clinical and genetic detection rate than using standard diagnostic criteria individually. Statin and other LLT use were suboptimal and below guideline recommendations. Because FH underdiagnosis and undertreatment are due partially to the challenges of implementing existing diagnostic criteria in a primary care setting, this hybrid model potentially can improve FH diagnosis and subsequent early access to appropriate treatment.

Familial hypercholesterolemia (FH), the most common genetic lipid disorder, is an autosomal dominant condition characterized by severe, lifelong elevation of low-density lipoprotein cholesterol (LDL-C). It is considered a significant public health problem for the United States and has been designated a tier 1 genomic application by the US Centers for Disease Control and Prevention (1). Although FH can remain asymptomatic for decades (2), it is estimated to affect more than 25 million individuals worldwide. The prevalence of FH varies across ethnic groups (1) and less than 10% of individuals with FH in the United States are diagnosed (3-5). Heterozygous FH has the highest prevalence of premature mortality among genetic conditions (4, 6, 7), and affected individuals have a 20-fold increased risk for premature atherosclerotic cardiovascular disease (ASCVD) (6). Yet, the heterozygous phenotype is widely variable, which contributes to its diagnostic difficulty (8, 9). Because FH morbidity and mortality result mainly from long-term exposure to elevated LDL-C (cumulative LDL-C burden) and other risk factors, rather than to specific genetic defects (7, 10, 11), it is important to assess other ASCVD risk factors in persons with FH (7, 10, 12). With early diagnosis and lipid-lowering therapy (LLT) treatment, total cholesterol-years is reduced, and the risk for atherosclerosis and coronary heart disease (CHD) decreases to rates comparable with the general population (6). Of the 3 currently accepted diagnostic criteria for FH (the UK’s Simon Broome system, MEDPED [Make Early Diagnosis to Prevent Early Death], and the Dutch Lipid Clinic Network Score [DLCNS]), the DLCNS is the most easily adaptable for population surveys and for unselected general populations (6, 13-15). The American Heart Association (AHA) provides simpler clinical functional diagnostic criteria for FH (7), but these currently are not widely implemented (16).

The prevalence of FH in an unselected sample from the local population of our Midwest regional healthcare system has not been determined directly. The goal of this study is to test the utility of a hybrid diagnostic model to determine FH prevalence and treatment characteristics in this population. The hybrid model comprised 2 approved diagnostic criteria combined with electronic medical record (EMR) data. The study was approved by the St. Elizabeth Institutional Review Board and a waiver for informed consent approved, given the retrospective data abstraction.

Materials and Methods

Study Population

We conducted a retrospective, records-based, cross-sectional study using datasets from unique EMRs of living patients presenting at a US metropolitan healthcare system (n = 2 020 239) (17). Using a dynamic EMR-based clinical decision support tool, and inpatient and outpatient records showing any lipid profile from a clinical encounter in the St. Elizabeth Healthcare system between January 1, 2009, and April 30, 2020 (n = 289 299) were enrolled in a clinical query using Structured Query Language (17).

The query identified all patient records showing components of the DLCNS or AHA diagnostic criteria for FH (7, 10, 18, 19) (Fig. 1). Data included a patient history of premature ischemic CHD or cerebral or peripheral vascular disease (Fig. 1, Category A) and family history of a first-degree relative with known premature CHD or ischemic vascular disease (Fig. 1, Category B). To standardize collection of family history data, graduate students from local colleges of pharmacy and other graduate students were trained to collect the family history of premature CVD or cerebrovascular accident (CVA) from at-risk patients (those with DLCNS = 5 or with LDL-C ≥190 mg/dL) (5). We calculated an estimated LDL-C for all individuals with an active statin prescription, adjusting for treatment effect, using their last LDL-C multiplied by 1.43 (17, 18, 20-23). This automated LDL-C correction increased the number of records meeting the phenotypic diagnostic criteria for FH (19). Maximum LDL-C (whether EMR-documented or last estimated pretreatment) (Fig. 1, Category C) was used to fulfill the biochemical criteria for both diagnostic systems (5). To determine the likelihood of FH, we used the highest DLCNS in each category (A, B, and C; blue circles in Fig. 1) to calculate the total DLCNS for each record. We also checked each record for AHA criteria: LDL ≥190 mg/dL (once or twice) and a personal or family history of premature CVD. Although we had genotypes for some patients, we did not use these to calculate their DLCNS in this phase of the study. Accurate data for physical exams also were not fully available and were not included in the study.

Figure 1.

Clinical query showing EMR data (red boxes) used to identify records for inclusion in the analysis (orange box). Abbreviations: CAD, coronary artery disease; CVD, cardiovascular disease; CVA, ischemic cerebrovascular accident; LDL-C, low-density lipoprotein C; PAD, peripheral arterial disease; TG, total triglyceride.

Records were excluded for patients with DLCNS ≥3 (n = 2018), and uncontrolled secondary causes of dyslipidemia (including significant proteinuria and significantly uncontrolled hypothyroidism) (n = 1062), and those with total triglyceride (TG) >400 mg/dL more than once during the study timeframe (n = 956), since this might indicate familial combined hyperlipidemia (Fig. 1) (Table 1 (24)).

Of the records submitted for analysis (n = 264 264), we used an algorithm to stratify patients into study groups according to EMR data and hybrid combinations of existing diagnostic criteria (Fig. 2): modified DLCNS (10) (excluding genotype); AHA (using either 1 [AHA-1] or 2 [AHA-2] verified levels of max LDL-C ≥190 mg/dL) (25); and genotype, when available (26, 27). When indicated by the hybrid criteria or by clinical evaluation, we sent patients’ results to either primary care or specialty providers for optional referral to a lipid clinic or to a precision medicine clinic for further consultation or genetic testing (5). Using this algorithm, we formed 3 categories of 5 study groups (Fig. 3).

Figure 2.

Using EMR data, patients were stratified into 1 of 5 groups* based on various hybrids of existing FH diagnostic criteria (modified DLCNS, AHA, and/or genotype). Abbreviations: CAD, coronary artery disease; CVD, cardiovascular disease; CVS, ischemic cerebrovascular stroke; LDL-C, low-density lipoprotein C; PAD, peripheral arterial disease; TG, total triglyceride. *Study groups: Confirmed FH: Group 1—FH genotype confirmed and if not then DLCNS ≥6 and if not then LDL-C ≥190 mg/dL more than once in addition to a personal or family history of premature CVD (AHA-2) and if not then LDL-C ≥190 mg/dL at least once in addition to a personal or family history of premature CVD (AHA-1). Group 2—DLCNS ≥6 and if not then LDL-C ≥190 mg/dL more than once in addition to a personal or family history of premature CVD (AHA-2). Group 3—DLCNS ≥6 and if not then LDL ≥190 mg/dL at least once in addition to a personal or family history of premature CVD (AHA-1). High CVD risk, but FH not confirmed: Group 4—including nonadopted patients with DLCNS 3,4, or (5 and not adopted) (correlating with possible FH) and not meeting any of the above criteria. General population: Group 5—DLCNS <3 and not meeting any of the above criteria. †AHA-1; LDL-C ≥190 mg/dL at least once and a personal or family history of premature CVD. ‡ AHA-2; LDL-C ≥190 mg/dL more than once and a personal or family history of premature CVD.

Figure 3.

(A) Venn diagram showing the total number and interactions among different features; (B, C, and D) stacked Venn diagrams showing the incremental detection rate for FH using the various hybrids of existing FH diagnostic criteria (modified DLCNS, AHA) and proportional contribution of each feature to the final composition of each study group; and (E) a table showing the total count for each variable (yellow highlight). The DLCNS 5 (adopted) feature was not used to contribute to the total count of group 1, because it did not add much beyond AHA-1 and AHA-2 have contributed. Count per variable: Genetically positive for FH = 4 + 8 + 1 + 2 + 5 + 2 + 3 = 25; DLCNS ≥6 = 221 + 188 + 142 + 2 + 5 + 3 = 561; DLCNS 5 and adopted = 10 + 5 = 15; AHA-2‡ = 226 + 5 + 142 + 5 + 2 = 380; AHA-1† = 755 + 10 + 226 + 5 + 142 + 5 + 2 + 188 + 2 + 1 = 1336; DLCNS 3,4, or (5 and nonadopted) = 19 589 + 8 + 1 + 2 = 19 600; DLCNS < 3 = 244 095 + 4 = 244 099. Abbreviations: AHA, American Heart Association; DLCNS, Dutch Lipid Clinic Network Score. †AHA-1; LDL-C ≥190 mg/dL at least once and a personal or family history of premature CVD. ‡AHA-2; LDL-C ≥190 mg/dL more than once and a personal or family history of premature CVD.

To assess the incremental clinical and genetic detection rate of our proposed diagnostic model (hybrid combinations of existing diagnostic criteria with EMR data), we compared patients identified with FH using the hybrid model against patients identified by the modified DLCNS ≥6 criterion only (reference group) (Fig. 3).

We identified comorbidities in the study population (Table 2 (24)), including coronary artery disease (CAD), diabetes mellitus (type 1 or type 2), essential hypertension, congestive heart failure, and obesity. These comorbidities are present in our EMR problems list, which is continuously updated and rigorously reviewed by providers and professional coders to ensure it always reflects our local population.

We also assessed tobacco use and exposure, as well as use of different LLTs: primarily statins, ezetimibe, and proprotein convertase subtilisin/kexin type 9 inhibitors. Statin intensity was classified as high intensity (atorvastatin 40 mg or 80 mg, rosuvastatin 20 mg or 40 mg, simvastatin 80 mg) or moderate to low intensity (any lower dose) according to the American College of Cardiology and AHA cholesterol guidelines (28). When available, test results for lipoprotein(a) (Lp(a)) and other significant cardiovascular risk factors also were examined for enrolled patients (10, 12, 29-32).

Data Analysis

Data were analyzed using Minitab®18 Statistical Software. Based on the inclusion–exclusion criteria (Fig. 1), 95% CIs for proportions were created for each group to estimate the prevalence of statin usage and other characteristics. Additional factors (eg, CAD/CVA, comorbidities) were incorporated into these estimates for subsequent subgroup analyses. Groups 1 to 3 are not independent, since inclusion criteria allow multiple group classifications to facilitate comparisons with national characteristics. Nonoverlapping CIs for groups 1, 2, or 3 with either groups 4 or 5 indicate significant differences in prevalence rates for all the study parameters.

Results

Of 289 299 total records screened, 25 035 were excluded (23 017 deceased; 1062 uncontrolled secondary dyslipidemia; 956 familial combined hyperlipidemia phenotype). Of 264 264 remaining records analyzed, 20 161 patients had a DLCNS ≥3 and 244 103 had a DLCNS <3 (Fig. 1).

Phenotype Features

Figure 4 and Table 3 (24) show FH phenotype features in the different study groups using different diagnostic criteria. The majority of individuals in all groups were women, and the mean age in groups 1 to 4 was older than in group 5 (Fig. 4A). Total cholesterol, LDL-C, and nonhigh-density lipoprotein cholesterol (HDL-C) were comparable in groups 1 and 3, but lower in these 2 groups than in group 2; higher than in group 4; and higher in groups 1 to 4 than in group 5 (Fig. 4B and 4C). Serum TG was lower, and HDL-C was higher in group 5 than in groups 1 to 4 (Fig. 4B and 4D). Lp(a) values and frequency of testing were comparable in groups 1 to 3 and in those with DLCNS ≥6 (reference group) (CI 4.7-9.0%; CI for Lp(a) levels 68-80 mg/dL), but higher and more frequently tested than in group 4, and higher and more frequently tested in groups 1 to 4 than in group 5 (Fig. 4E). The incidence of CVD and premature CAD was similar, and highest in groups 1 and 3, followed by group 2, then group 4, with the lowest incidence in group 5 (Fig. 4G). There was no evidence of any difference in the prevalence of nonpremature CAD among groups 1 to 4, but a slightly higher prevalence in group 5 than in groups 1 and 3. Obesity, diabetes (type 1 and type 2), congestive heart failure, hypertension, and smoking exposure were comparable in groups 1, 2, and 3, but higher than in group 4, and higher in groups 1 to 4 than in group 5 (Fig. 4H and 4I). All case groups (1-4) had comparable, but higher, average hierarchical condition category scores (Fig. 4F) and a higher blood pressure (last mean, mean systolic, and diastolic arterial) compared with group 5 (Fig. 4J and 4K).

Figure 4.

Features of the FH phenotype and treatment characteristics according to various diagnostic criteria (using data from Table 3 (24)). In each panel, the X axis represents the study group and the Y axis represents the studied variables. Abbreviations: BP, blood pressure; DM, diabetes mellitus; HCC, hierarchical condition category; HDL, high-density lipoprotein; HTN, hypertension; LDL, low-density lipoprotein; MAP, mean arterial pressure. *Atorvastatin 40 mg or 80 mg, or rosuvastatin 20 mg or 40 mg, or simvastatin 80 mg.

Genotype–Phenotype Interactions

Figure 5 shows the correlations among genotype (for patients who had genetic testing) (Invitae Corporation) (n = 213) and phenotypic features detected by various diagnostic criteria. Twenty-five patients had the pathogenic FH mutation. The most common mutations were in the LDL receptor gene, followed by the apolipoprotein B gene (33). Two patients had double mutations (1 pathogenic and 1 variant of unknown significance [VUS]). Nineteen patients (76%) had a unique mutation (ie, not reported in the Exome Aggregation Consortium dataset) in the 4 genes associated with FH. In these patients, 9 (47%) had mutations pathogenic for FH and 11 (58%) had VUS. Of those with VUS, 6 (55%) met other clinical criteria for FH. Figure 3A shows the interactions among phenotypic features (for study groups 1-5) in patients who had a positive genotype, including those who had cascade screening. Cascade screening identified 4 individuals in 3 families with FH (Figs. 5 and 3A). All had LDL-C <190 mg/dL (and were therefore categorized in the DLCNS-unlikely group), and therefore would have been missed without cascade screening.

Figure 5.

Interactions among the FH genotype and phenotypic features. *Cascade case: individuals who initially did not meet clinical indications for genetic testing, but were tested based on a family member positive for the FH mutation (positive or VUS). Study groups: Confirmed FH: Group 1—FH genotype confirmed and if not then DLCNS ≥6 and if not then LDL-C ≥190 mg/dL more than once in addition to a personal or family history of premature CVD (AHA-2) and if not then LDL-C ≥190 mg/dL at least once in addition to a personal or family history of premature CVD (AHA-1). Group 2—DLCNS ≥6 and if not then LDL-C ≥190 mg/dL more than once in addition to a personal or family history of premature CVD (AHA-2). Group 3—DLCNS ≥6 and if not then LDL ≥190 mg/dL at least once in addition to a personal or family history of premature CVD (AHA-1). High CVD risk, but FH not confirmed: Group 4—including nonadopted patients with DLCNS 3, 4 or (5 and not adopted) (correlating with possible FH) and not meeting any of the above criteria. General population: Group 5—DLCNS < 3 and not meeting any of the above criteria. Abbreviations: AHA, American Heart Association; APOB, apolipoprotein B; FH, familial hypercholesterolemia; LDL-C, low-density lipoprotein C; LDLR, low-density lipoprotein receptor; LDLRAP1, low-density lipoprotein receptor adaptor protein 1; Lp(a), lipoprotein a; PCSK9, proprotein convertase subtilisin/kexin type 9); VUS, variant of unknown significance. †AHA-1: FH diagnosed using the AHA criterion LDL-C ≥190 mg/dL at least once. ‡AHA-2: FH diagnosed using the AHA criterion LDL-C ≥190 mg/dL at least twice. §LDL-C ≥190 mg/dL and not meeting criteria for AHA-2, AHA-1, or DCLNS. ||Unique mutation: unique mutation not reported in the Exome Aggregation Consortium dataset (https://gnomad.broadinstitute.org/).

Mutation Rates

The green highlighted cells in Fig. 5 represent the pathogenic mutation rates for FH among those in the different study groups who received genetic testing. Group 1 had the highest mutation rate (25%) (applying the algorithm using the hybrid criteria), followed by those who would had only a DLCNS ≥6 (22%). In patients with an LDL-C ≥190 mg/dL, 12% to 14% had positive genetic mutation rates (with or without other AHA criteria), aside from the hybrid model.

Comparison of Study Groups With Reference Group

Comparing the incremental clinical detection rate of the hybrid diagnostic model with that of the modified DLCNS ≥6 (excluding genotype, n = 561) (34, 35) showed a 42% increased detection rate in group 2 (n = 794) and a 178% increased detection rate in group 3 (n = 1 560) (Figs. 3C and 3D, respectively). Comparing the incremental positive genetic detection rate of the hybrid model with that of the modified DLCNS ≥6 (excluding genotype, 10 patients) (Fig. 3A) showed a 20% to 30% improved genetic detection rate in groups 2 and 3 (n = 12 and n = 13, respectively), excluding cascade screening, and a 110% improved detection rate in group 1 (n = 21), excluding cascade screening (Fig. 3A). We do not believe these increased detection rates are associated with the increased Lp(a) levels or testing frequency in groups 1 to 3 (Fig. 4E), since these did not differ from that of those with DLCNS ≥6.

Treatment Characteristics

Figure 6 and Table 4 (24) show treatment characteristics in the absence of comorbidities. Use of any general statin, high-intensity statin, and ezetimibe was comparable in groups 1, 2, and 3, with slightly higher use in these groups than in group 4, and much higher usage in groups 1 to 4 than in group 5 (Fig. 6B). For all groups, statins were prescribed most frequently for individuals aged 40-75 years, but much less frequently for individuals aged <40 years or >75 years (Fig. 6C).

Figure 6.

Treatment characteristics among groups in the absence of comorbidities (using data from Table 4 (24)). In each panel, the X axis represents the study group and the Y axis represents the studied variables. Study groups: Confirmed FH: Group 1—FH genotype confirmed and if not then DLCNS ≥6 and if not then LDL-C ≥190 mg/dL more than once in addition to a personal or family history of premature CVD (AHA-2) and if not then LDL-C ≥190 mg/dL at least once in addition to a personal or family history of premature CVD (AHA-1). Group 2—DLCNS ≥6 and if not then LDL-C ≥190 mg/dL more than once in addition to a personal or family history of premature CVD (AHA-2). Group 3—DLCNS ≥6 and if not then LDL ≥190 mg/dL at least once in addition to a personal or family history of premature CVD (AHA-1). High CVD risk, but FH not confirmed: Group 4—including nonadopted patients with DLCNS 3,4 or (5 and not adopted) (correlating with possible FH) and not meeting any of the above criteria. General population Group 5—DLCNS < 3 and not meeting any of the above criteria. Abbreviations: BP, blood pressure; DM, diabetes mellitus; HCC, hierarchical condition category; HDL, high-density lipoprotein; HTN, hypertension; LDL, low-density lipoprotein; MAP, mean arterial pressure.

There was no evidence of a difference across the groups in having established care with primary care providers (PCPs), endocrinologists, or cardiologists. Generally, less than 50% of patients had established care with PCPs, less than 10% had established care with cardiology or endocrinology specialists, and 3% to 13% had future PCP visits scheduled (Fig. 6A). There was no evidence across the groups of any difference in the use of the MyChart™ personal health record.

We observed the same pattern for prescriptions of general statins, high-intensity statins, and ezetimibe in the presence of studied comorbidities (Fig. 4L, 4M, and 4N) and in studied comorbidities other than CAD/CVA (Figure 1 and Table 5 (24)). In general, the use of moderate-intensity statin was highest in group 4 compared with other groups (Fig. 4M).

Discussion

FH is a recognized national and international public health concern (36). Its clinical management varies widely among nations, and is encumbered by underdiagnosis and undertreatment (10, 36-38). Suboptimal identification in the United States may be due to the lack of systematic screening (6, 39), but there are few studies regarding the use of FH diagnostic criteria in current US practice (6, 40-42). In this study, when 2 approved FH diagnostic criteria were used individually, neither captured the FH prevalence expected for our population with adequate detection rates (both clinically and genetically). However, when combined in a unique hybrid model that included consistent application of an algorithm using an EMR-based clinical decision support tool along with genetic testing, we adequately detected the estimated FH prevalence for our population. We thereby improved FH identification for individuals and for genetically related family members (6).

National and international disease registries are considered essential for driving improvements in FH care and policy (6, 7, 36, 37, 43-46). Although some registries rely primarily on genetic testing, others, such as ours, utilize primarily clinical indicators and genotype, when available (3, 6, 37, 42, 43, 47). We included the DLCNS in our hybrid model, because it is the primary diagnostic tool for FH in most countries and has been shown in some studies to be more accurate than other published criteria (6, 15, 20, 37, 38, 46, 48). Yet, despite its high precision, the DLCNS is limited by suboptimal sensitivity and is difficult to implement in clinical practice (25). Conversely, the AHA functional criteria are easier to apply in practice, but require further studies examining their correlation with the FH genotype (16). Using a hybrid diagnostic system, we identified 794 to 1560 patients with FH (Figs. 2 and 3), a prevalence of 0.6% to 0.3%, which is similar to the FH prevalence reported in the NHANES data (20), the Mayo Employee and Community Health (SEARCH) study (42), and other studies (1:200 = 0.5% to 1:311 = 0.3%) (6, 10, 20, 26, 36, 40, 42, 46, 49). Data from these studies, as in ours, show a higher FH prevalence (3%) among individuals with ASCVD (46).

In addition to our study, the US SEARCH study (42) and other studies in United Kingdom (14) and Australia (15) used algorithms on extracted EMR data to identify patients with FH in a primary care setting. The SEARCH study included only patients with LDL-C ≥190 mg/dL (42); however, because we used our hybrid model to calculate the DLCNS on the entire study population, we identified 1422 patients with a DLCNS ≥3 (despite having an LDL-C <190 mg/dL) who might be candidates for genetic testing, if clinically indicated (41). In contrast to the SEARCH study (that included only measured LDL-C), we and Brett et al. (15) adjusted for pretreatment LDL-C levels, in alignment with other national studies (17, 20, 23, 41). However, we included both the DLCNS and the AHA criteria in our hybrid model (rather than the DLCNS, alone). Consequently, our detection rates for FH were higher than those reported in the SEARCH study. Three patients in the SEARCH study received genetic testing; however, because our approach provided patients a referral option for genetic testing, 213 patients received genetic testing (Fig. 5). This referral approach is similar to that reported by Myers et al., in which the treatment exception to the Health Insurance Portability and Accountability Act privacy (HIPAA) rule was utilized (5).

Although machine-learning algorithms (5, 39, 50) that use EMR data to screen for FH phenotypes provide innovative precision-screening approaches, they do not replace clinical evaluation or existing diagnostic criteria (5). These algorithms are structured primarily to identify the FH phenotype via numerous clinical classifiers (procedures and diagnostic codes, prescriptions, and laboratory findings) (5, 51) and some require a minimum set of features to achieve high diagnostic accuracy (eg, 20-75 features in the model by Myers and Banda) (5, 51). To improve diagnostic detection, our study used phenotype features combined with widely approved FH diagnostic criteria (including a modified DLCNS) to directly identify patients. EMR-based screening tools, including ours, leverage longitudinal clinical data that can span 3 to 21 years (5, 42). The longer the timeframe, the more valuable the clinical data become for identifying patients with FH (5). This is especially important given our and others’ results (5, 42, 51) showing that laboratory data are the most frequent identifying factors, followed by clinical features gleaned from healthcare encounters (ie, age, gender, diagnosis, and procedures) (5, 42).

Comparing clinical and treatment characteristics among different groups, the median age of individuals with FH in our study (groups 1-3) was 55-57 years, similar to that reported in the Cascade Screening for Awareness and Detection of FH Registry (CASCADE-FH) (45), but older than that reported for the SEARCH study (46 years) (42) and younger than that reported from other machine-learning projects (61 years) (5). Older age at diagnosis may reflect late referral in the course of treatment or preventive care, and the low prevalence of lipid data in patients younger than 40 years old (5).

Genetic Testing

This study provides an evidence-based platform for FH genetic testing. Approximately 6% to 10% of individuals with an FH phenotype (groups 1-3) received genetic testing (Fig. 5). Results confirm that our EMR decision support tool provides a systematic process to collect clinical data and thereby determine if genetic testing is needed (26). Consistent with other studies, our results show that, in the absence of comorbidities, most patients receive primary care rather than specialty care (7, 36), regardless of the diagnostic criteria used. Although primary care is considered an optimal setting for FH screening (36, 44, 52), knowledge and awareness of FH among PCPs are limited (42, 53, 54). Accurate manual screening for all DLCN or AHA criteria in the entire primary care population is not feasible and guidelines recommend FH genotype screening only for certain populations (26, 27). The scientific expert panel of the Journal of the American College of Cardiology and other authorities suggest that FH clinical and genetic diagnostic criteria are complementary, with genetic testing dependent on the pretest probability of FH indicated by clinical criteria (similar to those collected in our study) and other clinical factors (26, 41, 52, 55, 56). Although an FH diagnosis is not excluded if genetic testing does not detect a pathogenic FH variant (5, 26, 42), some studies suggest that relying only on clinical criteria may lead to a false-positive FH diagnosis (40, 55), while other studies show a positive correlation between DLCNS and a positive FH genotype (56-59). Using the DLCNS ≥6 diagnostic criterion (Fig. 5), our results showed a 22% FH mutation rate, similar to that shown in studies from the United States (5) and from The Netherlands (18). The FH mutation rate was 25% in group 1 and 15% in groups 2 or 3, respectively. Using only the clinical criterion, LDL-C ≥190 mg/dL, in the absence of other criteria, yielded a 14% FH mutation rate. This rate was much higher than the <2% mutation rate reported by Khera et al. in 12 studies of CAD-free control subjects and in population-based cohort studies (41). This might reflect referral bias in our genetically tested population, because testing was provided primarily to those with severe hypercholesterolemia referred to the lipid clinic for treatment due to their young age or due a positive personal or family history of premature CAD (in the absence of other risk factors), and most of these had monogenic hypercholesterolemia (9, 41, 60). Combining clinical DLCNS, functional AHA criteria (groups 1, 2, and 3), and an optional referral to a specialty lipid or precision medicine clinic for patients with primary severe hypercholesterolemia, improved the clinical detection rate (by up to 178%) and genetic detection rate (by up to 110%) when compared with using only the modified DLCNS for diagnosis (Fig. 3A and 3D).

In our study and in others (26, 29), some patients with DLCNS ≥6 were negative for the FH genotype. This might be due to other genetic determinants, such as polygenic hypercholesterolemia, high Lp(a), apolipoprotein E , undiscovered FH genes, sitosterolemia, or lysosomal acid lipase deficiency (Fig. 5) (26, 29, 59).

Although there are numerous reported difficulties inherent in collecting a family history of premature CAD (6, 26, 40-42, 61), we found a family history extant in 45% of groups 1 and 3, 64% of group 2, and in 56% of the DLCNS definite/probable group, all of which are comparable or slightly higher than that reported in the CASCADE-FH registry (41%) or in the SEARCH study (38%) (6, 42, 45). Our slightly higher rates are attributed mainly to the active collection of family history of premature CVD from high-risk patients. On the other hand, available machine-learning studies for FH screening did not capture some data crucial to conventional diagnostic criteria (such as family history) (5).

Associated Comorbidities

The prevalence of modifiable ASCVD-related comorbidities in our study was higher than in the general population (group 5), an outcome similar to the NHANES data and the CASCADE-FH registry (20, 45). The diagnostic criterion used for groups 1 and 3 (primarily a single value for LDL-C ≥190 mg/dL) was more sensitive for premature CAD (53-58%) than that used for group 2 (38-45%) or group 4 (10-10.8%), respectively. Groups 1 to 4 had a higher prevalence of premature CAD than group 5 (1.8%). The prevalence of premature CAD reported by the CASCADE-FH registry (38%) is lower than our results for groups 1 and 3, but comparable with our results for group 2. Similarly, the prevalence reported for the Danish population (28% using the DLCNS (6, 10, 18, 45)) is lower than the prevalence found in groups 1 to 3. This might be due to the hybrid criteria used in our study and in the CASCADE-FH compared with using the DLCNS alone.

Lipoprotein(a) is a well-established, independent, continuous causal risk factor for CVD (62-64), particularly in individuals with FH (65-68). High Lp(a) (>50 mg/dL) can adversely affect CAD rates, as well as the natural history of FH, and consequent treatment plans (7, 10, 12, 26, 29, 32, 62, 63, 65, 69-71). Although Lp(a) was tested more frequently in our study, with high averages (>50 mg/dL) for individuals with FH (groups 1-3) compared with groups 4 and 5, the percentage of individuals screened was low (5-7%). This might reflect a general lack of knowledge among providers about the importance of high Lp(a) levels in individuals with a high risk for CVD, including those with FH (69, 72).

Regardless of the FH diagnostic criteria used and the presence or absence of comorbidities, 76% to 90% of individuals with FH in our study were receiving LLT. This figure is similar to that reported in the NHANES (20) and SEARCH study data (82%) (42), and is similar to that reported for the CASCADE-FH registry (45) (75%). Individuals with FH (groups 1-3) were treated with high-intensity LLT more frequently than those at high risk for CVD, but not yet meeting FH diagnostic criteria (group 4) (73). When compared with the CASCADE-FH registry (45) and the SEARCH study (42), use of high-intensity statins in our study was more frequent (47-54% vs 42%), but use of moderate-intensity statins was similar (~33%). This might reflect the pattern of increased statin use between 1999 and 2014, and/or the difference in inclusion criteria between studies (20). In individuals with no comorbidities, high-intensity statins were used less frequently (28-44%) (45). This highlights previously reported suboptimal treatment in individuals who do not have identifiable comorbidities and may reflect a knowledge gap about FH diagnosis (22, 54), in addition to evidence of an inverted U-shaped curve showing younger and older individuals were not treated as aggressively as recommended by guidelines (28).

Study Limitations

Geneotype testing was limited in this study and without full genetic analysis of the entire study cohort, we could not accurately differentiate between monogenic and polygenic FH (20, 74). Other studies report difficulties (15) in contacting patients with FH, partly due to a lack of awareness among clinicians and patients of the need for comprehensive assessment (6, 7, 26, 37, 54). We currently are expanding access to genetic testing for individuals who meet FH clinical criteria using the hybrid diagnostic criteria present in our algorithm (ie, in groups 1, 2, or 3). Some misclassification bias (42) may be present, since we collected data during routine medical care and not primarily for research purposes, but we expect this to be very minimal, given the rigorous selection criteria. Excluding patients with secondary dyslipidemia also could potentially exclude patients with FH and concomitant secondary dyslipidemia (19). We did not assess statin intolerance from collected EMR data (75) or the effect(s) of treatment intensification after identifying at-risk individuals, an effort that has been successful in other studies (14, 15). Lastly, some studies report limited accuracy concerning the calculation of pretreatment LDL-C (19, 20, 76).

Study Strengths

The approach employed in this study fulfills most of the current policy recommendations for FH screening: a systemic, time- and cost-effective, accurate, targeted, universal screening strategy for FH index cases in primary care settings within a large healthcare system and optional referral to a lipid clinic or to a precision medicine clinic. We also addressed common problems associated with FH underdiagnosis by accessing longitudinal data regarding patients’ LDL-C, adjusting for treatment effect on LDL-C, and family history of premature CAD (51). Our EMR-based registry offers a platform for genetic testing at a cost that continues to decrease over time (26), and uses a comprehensive and longitudinal EMR with high interoperability since the Structured Query Language code can be shared easily with other healthcare systems using same EMR (6, 7, 36, 44, 46, 52, 77).

Conclusions

Using a hybrid of validated FH diagnostic criteria combined with data routinely collected using our EMR provides an innovative, comprehensive, and systematic screening model for FH and can improve differential diagnosis of FH in the general population. Using the hybrid model in a structured fashion, we identified the estimated prevalence of FH in our population and documented LLT treatment far below guideline-recommended FH management (73). Leveraging EMR data for population-based screening can help optimize the pretest clinical probability for FH genetic testing (40) and can identify those at high risk for CVD, thus achieving earlier diagnosis and treatment for this high-risk population.