Kullo, Iftikhar J.

Abstract Current studies regarding the secondary use of electronic health records (EHR) predominantly rely on domain expertise and existing medical knowledge. Though significant efforts have been devoted to investigating the application of machine learning algorithms in the EHR, efficient and powerful representation of patients is needed to unleash the potential of discovering new medical patterns underlying the EHR. Here, we present an unsupervised method for embedding high-dimensional EHR data at the patient level, aimed at characterizing patient heterogeneity in complex diseases and identifying new disease patterns associated with clinical outcome disparities. Inspired by the architecture of modern language models—specifically transformers with attention mechanisms, we use patient diagnosis and procedure codes as vocabularies and treat each patient as a sentence to perform the patient embedding. We applied this approach to 34,851 unique medical codes across 1,046,649 longitudinal patient events, including 102,739 patients from the electronic Medical Records and GEnomics (eMERGE) Network. The resulting patient vectors demonstrated excellent performance in predicting future disease events (median AUROC = 0.87 within one year) and bulk phenotyping (median AUROC = 0.84). We then illustrated the utility of these patient vectors in revealing heterogeneous comorbidity patterns, exemplified by disease subtypes in colorectal cancer and systemic lupus erythematosus, and capturing distinct longitudinal disease trajectories. External validation using EHR data from the University of Washington confirmed robust model performance, with median AUROCs of 0.83 and 0.84 for bulk phenotyping tasks and disease onset prediction, respectively. Importantly, the model reproduced the clustering results of disease subtypes identified in the eMERGE cohort and uncovered variations in overall mortality among these subtypes. Together, these results underscore the potential of representation learning in EHRs to enhance patient characterization and associated clinical outcomes, thereby advancing disease forecasting and facilitating personalized medicine.

BACKGROUND: Coronary atherosclerotic burden and adverse coronary heart disease events are related phenotypes with likely shared genetic cause. METHODS: We analyzed 6021 patients with available coronary angiography, genotyping, and exome sequencing data. We tested for associations of polygenic risk scores for coronary heart disease (PRS CHD ) with multiple measures of coronary artery disease (CAD) severity. We assessed the joint associations of PRS CHD and pathogenic/likely pathogenic variants in 3 familial hypercholesterolemia genes, with CAD severity. We performed mediation analyses to explore whether CAD severity mediated the association of PRS CHD with prevalent coronary heart disease and incident myocardial infarction. RESULTS: A 1-SD increase in PRS CHD was associated with multiple measures of CAD severity, including the log Gensini score (β, 0.31 [95% CI, 0.28–0.33]). Carrying a pathogenic/likely pathogenic familial hypercholesterolemia variant was associated with a higher log Gensini score after adjustment for PRS CHD (β, 0.21 [95% CI, 0.03–0.38]). A 1-SD increase in PRS CHD was associated with incident myocardial infarction over a mean follow-up of 9.2 years (hazard ratio, 1.20 [95% CI, 1.13–1.27]; P =5×10 − 10 ), and the Gensini score mediated 90% of this association. CONCLUSIONS: PRS CHD was associated with multiple measures of CAD severity. The association of PRS CHD with incident myocardial infarction was almost fully mediated by CAD severity, indicating a considerable genetic overlap between the 2 phenotypes.

Introduction: We validated the “Familial Hypercholesterolemia (FH) Decision Aid” to promote shared decision making (SDM) about lipid lowering therapy, based on patient-clinician encounters. Methods: Validation of the DA included an assessment of whether the content and design of the tool facilitated SDM and patient engagement and to determine how well the tool integrated with clinical workflow. Validation was conducted during patient-clinician encounters in the Mayo preventive cardiology clinic. As inclusion criteria we considered an upcoming appointment in the clinic, an LDL-C level ≥ ~160 mg/dL, and no prior diagnosis of FH. Prior to the encounters, each clinician viewed a demonstration of the online, interactive DA. During the encounters, the DA was shared with patients using the clinic room computer screen, either in-person or virtually. Patient-clinician encounters were video recorded for further analysis by two design researchers and experts in SDM. Results: Ten patients were enrolled in the study (7 male; average age 51 years). The average patient LDL-C level was 207 mg/dL, range 157 - 248 mg/dL. Review and qualitative analysis of the encounter recordings yielded 4 key findings: i) clinicians primarily used the cardiovascular risk estimate in the DA to guide conversation with patients; ii) highly trained specialists needed to change their usual approach to care and conversation to integrate the DA with their clinical workflows; iii) as clinicians became more familiar with the DA, they moved through the tool at a more efficient pace; and iv) each clinician developed his/her own style of navigating the DA, either moving through it one screen at a time or focusing on one specific section for the duration of the conversation. The DA effectively guided conversation and facilitated active patient engagement in 6 of 10 encounters. Social distancing and use of telemedicine encounters, due to the COVID-19 pandemic, limited patient engagement in 2 of 10 encounters. Overall, the DA promoted SDM regarding FH management in 9 of 10 encounters. Conclusion: Using patient-clinician encounters, we found the FH DA to facilitate SDM regarding lipid-lowering therapy. The tool has potential to improve care of FH patients and increase patient engagement and knowledge.

Introduction: Severe hypertriglyceridemia (SHTG), defined as plasma triglyceride (TG) level ≥ 500 mg/dl, is a risk factor for cardiovascular disease. In the NHANES dataset (2001 - 2006), the prevalence of SHTG was 1.7%. However, the prevalence of primary SHTG in the community is unknown. We conducted a community-based study to estimate the prevalence of primary SHTG, in Olmsted County, Minnesota. Methods: We used the Rochester Epidemiology Project (REP) records linkage system to identify individuals with TG ≥ 500 mg/dl between 1998-2015. Index date was defined as the first documentation of TG ≥500 mg/dl during the study period. We excluded individuals with mixed hyperlipidemia (TG ≥500 mg/dl and non-high density lipoprotein (HDL) cholesterol ≥190 mg/dl). HTG was defined as secondary in the presence of hypothyroidism, biliary obstruction, nephrotic syndrome, uncontrolled diabetes, pregnancy, liver disease, renal failure, BMI ≥35 and medication use (isotretinoin, tetracyclines, anti-retrovirals). Primary HTG was defined as absence of an identifiable secondary etiology. Characteristics of the primary SHTG individuals were ascertained from the electronic health record. Results: SHTG was identified in 3802 individuals of whom 2511 had mixed hyperlipidemia, 743 had a secondary cause for SHTG and the remaining 548 had primary SHTG. Using the 2015 population census of 151,436 for Olmsted County, the prevalence of SHTG was 2.5%, of secondary SHTG was 0.49% and of primary SHTG was 0.36%. Only half of the primary SHTG group had a diagnosis of dyslipidemia in the electronic health record and 44.3% were on lipid-lowering therapy on index date. Characteristics of individuals with primary SHTG are summarized in Table 1. Conclusions: These data highlight that primary SHTG is relatively common (≈1 in 300, similar to the prevalence of familial hypercholesterolemia), underdiagnosed and undertreated.

Introduction: Prior reports of the genetic etiology of hypercholesterolemia are affected by referral bias, inclusion of individuals with secondary hypercholesterolemia and variability in identifying pathogenic variants. We sought to assess the prevalence of monogenic and polygenic etiologies in a community based cohort of adults with primary hypercholesterolemia (LDL-C ≥155 mg/dL after excluding secondary causes). Methods: We identified 1719 individuals aged 18-70 years with primary hypercholesterolemia from SE Minnesota. The Dutch Lipid Clinic Network (DLCN) score was ascertained by an electronic health record (EHR) based algorithm followed by manual review. Familial hypercholesterolemia (FH) was considered present with a DLCN score ≥6 whereas a DLCN score of 3-5 was considered possible FH. Participants underwent sequencing of FH genes ( LDLR, APOB, PCSK9 ) and genotyping of 12 SNPs known to be associated with LDL-C. Rare (frequency <0.1%) and putatively functional variants were identified and pathogenicity was assigned based on the ACMG guidelines using relevant database, literature and EHR review. A polygenic score (PGS) for LDL-C (derived from 12 SNPs) >90 th percentile was considered an indicator of polygenic etiology. Results: A pathogenic/likely pathogenic variant (16 in LDLR , 1in APOB and 2 in PCSK9 ) was present in 24 (1.3%) individuals, a PGS >90 th percentile was present in 265 (15.4%) and 4 (0.2%) had both a monogenic and polygenic etiology. The overall prevalence of an identifiable genetic etiology was 17%. Based on clinical criteria the overall prevalence of FH (DLCN ≥6) was 7.0% (6.1% probable and 0.9% definite) and that of possible FH was 30.0%. In the possible, probable and definite FH categories, a monogenic etiology was present in 2.4%, 9.5% and 26.7% and a polygenic etiology in 19.1%, 21.0% and 13.3% respectively. 52.0% of those with a monogenic etiology and 9.7% of those with a polygenic etiology met the DLCN criteria for FH. Conclusions: This study provides for the first time estimates of monogenic and polygenic etiology of hypercholesterolemia in adults from the community, excluding cases of secondary hypercholesterolemia and using standardized methods of variant annotation.

Introduction: Familial Hypercholesterolemia (FH) is a common inherited disorder of lipid metabolism. FH patients are at an increased risk of atherosclerotic cardiovascular disease (ASCVD) due to elevated low-density lipoprotein cholesterol (LDL-C) levels. Patients with FH may not reach target LDL-C levels despite maximally tolerated statin (±ezetimibe) therapy (MTT). The AHA scientific statement on FH highlighted the need for developing decision aids (DAs) to facilitate shared decision making (SDM) regarding additional therapy including PCSK9 inhibitors. Methods: A user-centered design process was initiated to understand FH treatment decisions and to develop prototype interventions. Patients with FH had their diagnosis confirmed by genetic testing. Encounters between FH patients and health care providers were video recorded and analyzed to assess how decisions are made regarding lipid lowering therapy. Nine encounters were recorded: 3 with a genetic counselor and 6 with an FH specialist. Informed by clinical encounters and literature review, we developed content for a DA related to available therapeutic options for FH patients on MTT. Results: A prototype online DA was created to support in-encounter discussion of a patient’s risk of adverse cardiovascular events and which therapies may be best suited to the patient (Figure). 5 and 10-year ASCVD risk is estimated using the SAFEHEART Risk Score (SRS). Risk is communicated in numbers using a 100-person icon array and the language “Out of 100 persons like you...”. Therapeutic options include watchful waiting, bile acid sequestrants, and PCSK9 inhibitors. For each option, the degree of risk reduction, costs, and side effects are presented. The tool is currently being refined by analyzing additional encounters with FH patients. Conclusion: We developed an interactive prototype DA to facilitate SDM for FH patients on MTT. After further iterative testing the tool will be deployed in the electronic health record.

Background: Atherosclerotic cardiovascular disease (ASCVD) often affects > 1 arterial location (polyvascular disease, PVD). The risk of adverse outcomes increases with the number of arterial territories involved. The genetic susceptibility to PVD is unknown. Hypothesis: We hypothesized that a genetic risk score representing cumulative effects of susceptibility variants to ASCVD is associated with presence of PVD. Methods: We searched genome-wide association study (GWAS) catalog to identify genetic variants associated with coronary heart disease (CHD), ischemic stroke, peripheral arterial disease (PAD) or abdominal aortic aneurysm (AAA). No variants were reported to be associated with PVD in ≥ 3 arterial territories at GWAS significance. 10 loci were associated with PVD in 2 arterial territories. We calculated a multi-locus GRS using 10 lead single nuclear polymorphisms (SNP) from these distinct loci (1 per each loci) in 7486 participants in Mayo Vascular Disease Biorepository (age 68±10 years, 61% men, 31% with CHD; 27% carotid artery stenosis; 39% PAD; 14% AAA; 38% without any of these vascular diseases). Results: Of all participants 34% had PVD (2%, 10% and 22% with exact 4, 3 and 2 arterial territories involved respectively). A higher GRS was associated with greater number of arterial territories involved (P for trend of ANOVA <0.001). Patients with PVD in ≥ 3 arterial locations had higher GRS compared to those with 2 arterial locations involved after adjustment for age, sex and cardiovascular risk factors (P =0.008). One unit increase in GRS was associated with 9% higher likelihood for each additional arterial territory involved after adjustment for all covariates (odds ratio from ordinal regression model= 1.09, P <0.001). The result remained significant after further adjustment for combinations of AAA, CHD, carotid arterial stenosis and PAD (all P ≤ 0.01). 5 Of the 10 SNPs were each associated with presence of PVD after adjustment for all covariates (all P <0.01). None of them remained significant after further adjustment for combinations of PVD in different arterial territories (all P ≥ 0.4). Conclusions: A GRS derived from 10 SNPs of ASCVD is associated with PVD,suggesting genetic contribution to the dimension of PVD that may lead to adverse outcome.

Introduction: Atherosclerotic cardiovascular disease (ASCVD) is the leading cause of death worldwide. The genetic contribution to mortality in patients with ASCVD is unknown. Hypothesis: Abdominal aortic aneurysm (AAA) often co-exists with ASCVD and two with shared genetic susceptibility underlying disease development. We hypothesiz that a genetic risk score (GRS) for susceptibility variants to AAA and ASCVD increases the risk for all-cause mortality in patients with ASCVD. Methods: We analyzed high-density genotyping data in 3053 participants (age 71 ±10 years, 72% men) with ASCVD in Mayo Vascular Disease Biorepository. We used 4 single nucleotide polymorphisms (SNPs) (rs2383207, rs7025486, rs599839, rs1466535) previously shown to be associated with AAA and ASCVD at GWAS significance (P ≤10^-8) in the NHGRI GWAS catalog to build the GRS. We defined ASCVD as a history of myocardial infarction / coronary revascularization / ischemic stroke / peripheral arterial disease with revascularization. Results: Patients with GRS > median level had similar age (70±10 vs. 71±10), prevalence of men (72% vs. 73%), hypertension (83% vs. 82%), diabetes (32% vs. 31%) and smokers (74% vs. 72%) compared to those with GRS ≤ median (all P>0.2), except for dyslipidemia (91% vs. 88%, P =0.02). After a median follow-up of 4.56 years (interquartile: 3.07-6.26), 620 patients died. Patients with GRS > median had poorer survival compared to those with GRS ≤ median (Figure: Kaplan-Meier curve); and with 23% higher risk for death after adjustment for age, sex and cardiovascular risk factors (hazard ratio HR, 95% CI: 1.23, 1.05-1.45). The association of GRS with all-cause mortality remained significant after excluding patients with known AAA (n=672) (HR, 95%CI: 1.22, 1.01-1.48). Conclusions: A GRS for susceptibility variants of AAA and ASCVD is a prognostic marker in patients with ASCVD.

Objectives: We investigated the association of 47 candidate proteomic markers in pathways of inflammation, thrombosis, lipoprotein metabolism, vascular remodeling, and vasoregulation, with the ankle-brachial index (ABI) and peripheral arterial disease (PAD) in a bi-ethnic cohort of African Americans (AA) and non-Hispanic Whites (NHW). Methods: Multiple assay platforms were used to measure 47 candidate proteomic markers in 1324 AA and 1237 NHW participants belonging to hypertensive sibships. ABI was measured using a standard protocol, with values >1.4 excluded and PAD defined as ABI <0.9. Multivariable regression analyses (with generalized estimating equations) were employed to identify markers independently associated with ABI and PAD after adjusting for age, BMI, smoking, hypertension, diabetes, history of heart attack and stroke, medication use, total and HDL cholesterol, and estimated GFR as well as physical activity, alcohol consumption and education level. Results: Higher circulating levels of CRP, myeloperoxidase, NT-proBNP, osteoprotegerin, D-Dimer and fibrinogen were each associated with lower ABI in both AA and NHW. Additional markers significantly associated with lower ABI were; in AA - serum amyloid A, ICAM, IL-6, TNFR-I, TNFR-II, MCP-1, Lp(a), pro-atrial natriuretic peptide, pro-arginine vasopressin, pro-adrenomedullin, pro-endothelin, osteonectin, Factor VIII; and in NHW- Hsp27, ox-LDL and resistin. Markers associated with presence of PAD were: in AA - CRP, NT-proBNP, Lp(a), pro-arginine vasopressin and osteoprotegerin; in NHW -CRP, NT-proBNP, Hsp27, myeloperoxidase, oxidized-LDL, Factor VIII, and D-Dimer. When the significant markers were included in a final model, the R 2 for ABI increased from 15.9 to 19.4 for AA and from 18.1 to 22.1 for NHW; the c-statistic for PAD increased from 0.791 to 0.833 inAA and from 0.841 to 0.866 in NHW. Conclusions: Novel protein markers significantly add to prediction of ABI and PAD in both AA and NHW and may have clinical utility in diagnosis and prognosis of PAD as well as in identifying new therapies for atherosclerotic vascular disease.

ImportanceClonal hematopoiesis of indeterminate potential (CHIP) may contribute to the risk of atrial fibrillation (AF) through its association with inflammation and cardiac remodeling.ObjectiveTo determine whether CHIP was associated with AF, inflammatory and cardiac biomarkers, and cardiac structural changes.Design, Setting, and ParticipantsThis was a population-based, prospective cohort study in participants of the Atherosclerosis Risk in Communities (ARIC) study and UK Biobank (UKB) cohort. Samples were collected and echocardiography was performed from 2011 to 2013 in the ARIC cohort, and samples were collected from 2006 to 2010 in the UKB cohort. Included in this study were adults without hematologic malignancies, mitral valve stenosis, or previous mitral valve procedure from both the ARIC and UKB cohorts; additionally, participants without hypertrophic cardiomyopathy and congenital heart disease from the UKB cohort were also included. Data analysis was completed in 2023.ExposuresCHIP (variant allele frequency [VAF] ≥2%), common gene-specific CHIP subtypes (DNMT3A, TET2, ASXL1), large CHIP (VAF ≥10%), inflammatory and cardiac biomarkers (high-sensitivity C-reactive protein, interleukin 6 [IL-6], IL-18, high-sensitivity troponin T [hs-TnT] and hs-TnI, N-terminal pro–B-type natriuretic peptide), and echocardiographic indices.Main Outcome MeasureIncident AF.ResultsA total of 199 982 adults were included in this study. In ARIC participants (4131 [2.1%]; mean [SD] age, 76 [5] years; 2449 female [59%]; 1682 male [41%]; 935 Black [23%] and 3196 White [77%]), 1019 had any CHIP (24.7%), and 478 had large CHIP (11.6%). In UKB participants (195 851 [97.9%]; mean [SD] age, 56 [8] years; 108 370 female [55%]; 87 481 male [45%]; 3154 Black [2%], 183 747 White [94%], and 7971 other race [4%]), 11 328 had any CHIP (5.8%), and 5189 had large CHIP (2.6%). ARIC participants were followed up for a median (IQR) period of 7.0 (5.3-7.7) years, and UKB participants were followed up for a median (IQR) period of 12.2 (11.3-13.0) years. Meta-analyzed hazard ratios for AF were 1.12 (95% CI, 1.01-1.25; P = .04) for participants with vs without large CHIP, 1.29 (95% CI, 1.05-1.59; P = .02) for those with vs without large TET2 CHIP (seen in 1340 of 197 209 [0.67%]), and 1.45 (95% CI, 1.02-2.07; P = .04) for those with vs without large ASXL1 CHIP (seen in 314 of 197 209 [0.16%]). Large TET2 CHIP was associated with higher IL-6 levels. Additionally, large ASXL1 was associated with higher hs-TnT level and increased left ventricular mass index.Conclusions and RelevanceLarge TET2 and ASXL1, but not DNMT3A, CHIP was associated with higher IL-6 level, indices of cardiac remodeling, and increased risk for AF. Future research is needed to elaborate on the mechanisms driving the associations and to investigate potential interventions to reduce the risk.