Abstract
Background
Lipoprotein(a) [Lp(a)] is underutilized in short-term atherosclerotic cardiovascular disease (ASCVD) risk prediction.
Objectives
This study investigates Lp(a) contribution to short-term ASCVD event prediction using contemporary real-world data and machine learning (ML).
Methods
A cohort of 731,983 individuals from a claims database was used to investigate the association of Lp(a) with incident ASCVD and all-cause mortality using Cox proportional hazards models. Novel ML models were developed to predict incident ASCVD events at 1, 2, and 3 years after Lp(a) testing. The models were validated in an independent cohort of 53,930 patients.
Results
An increase of 50 nmol/L in Lp(a) was independently associated with incident ASCVD events (HR: 1.072; 95% CI: 1.059-1.084) and all-cause mortality (HR: 1.041; 95% CI: 1.015-1.068) after adjustment for age, sex, and race/ethnicity. Novel ML models featuring Lp(a) predicted incident ASCVD events at 1, 2, and 3 years with robust discrimination (C-statistic: 0.83-0.84) in both the derivation and validation cohorts. Modest underestimation of risk was observed in the validation cohort for the 1-year model (calibration slope 1.25). Lp(a) contributed more to 1-year ASCVD prediction than smoking, diabetes, and other lipid parameters. Inclusion of Lp(a) in the 1-year model led to an integrated discrimination improvement of 0.03 and an optimal net reclassification improvement of 10% at a risk threshold of 26%.
Conclusions
Lp(a) is a significant predictor of short-term ASCVD risk. Assessing Lp(a) and imminent ASCVD risk may assist in identifying patients who may benefit from escalation of preventative therapies.
Key words: atherosclerosis, lipoprotein (a), machine learning, real-world data, risk prediction
Central Illustration
Elevated lipoprotein (a) [Lp(a)] is a prevalent and potent genetically determined risk factor that increases atherosclerotic cardiovascular disease (ASCVD) risk independent of other traditional CV risk factors.1, 2, 3, 4, 5, 6 Lp(a) exhibits structural similarities to low-density lipoprotein (LDL);5 however, apolipoprotein(a) moieties and higher oxidized phospholipid content contribute to the six-fold higher atherogenicity observed with Lp(a) compared to LDL on a per particle basis.7 A recent analysis of the Lp(a) HERITAGE cohort demonstrated that one in four individuals with clinically established ASCVD have elevated Lp(a).8 Despite these compelling observations, elevated Lp(a) remains underdiagnosed and underutilized in ASCVD risk prediction.9, 10, 11 Prior studies demonstrate that incorporation of Lp(a) improves performance of ASCVD risk prediction models in primary and secondary prevention settings over 5- to 30-year time frames;12, 13, 14, 15, 16, 17 however, the impact of Lp(a) on ASCVD event risk over shorter risk horizons spanning one to 3 years has not been systematically examined.
The assessment of short-term ASCVD risk can identify individuals who will likely benefit from earlier and more aggressive therapies. Beyond improvements in clinical outcomes, assessment and discussion of imminent risk with patients may also be more effective at promoting lifestyle changes.18 Although short-term acute coronary syndrome risk prediction models exist,19, 20, 21 none use Lp(a) as a clinical variable in risk prediction.
In this study, we leveraged a large claims database of 731,983 patients with Lp(a) measurements and longitudinal clinical follow-up to address gaps in Lp(a) utilization for short-term ASCVD event prediction. We employed regression and tree-based machine learning (ML) techniques to develop robust models that feature Lp(a) in the prediction of major ASCVD events over 1-, 2-, and 3-year time intervals after Lp(a) testing in both primary and secondary prevention settings. We evaluated the performance of these novel models and systematically examined the relative contribution of Lp(a) to short-term ASCVD risk across a broad range of subpopulations. We then validated the models using an independent electronic health record data set with 53,930 individuals with Lp(a) measurements to assess the performance of these novel short-term ASCVD risk prediction tools across diverse patient populations and real-world data environments (Central Illustration).
Central Illustration.
Lipoprotein(a) Increases Short-Term ASCVD Risk in Novel Machine Learning Models
Elevated Lp(a) was prevalent and independently associated with ASCVD events in a large derivation cohort of over 730,000 individuals from the HealthVerity claims database. This derivation cohort was used to develop novel, machine learning predictive models that predicted short-term ASCVD events over 1 to 3 years after Lp(a) testing with robust discrimination. The ML models were validated in an independent database with electronic health records and linked claims (Market Clarity). Lp(a) contributed more to 1-year ASCVD event risk than diabetes, smoking, and all other lipid parameters. EHR = electronic health records; NRI = net reclassification improvement; PCI = percutaneous coronary intervention; other abbreviations as in Figure 1, Figure 2, Figure 3, Figure 4.
Methods
Study populations and data sources
The study involved analysis of only deidentified databases with no identifiable elements and received a nonresearch determination after submission of a Human Subject Research Determination form to the Stanford University Institutional Review Board (IRB-81716). Data for the derivation cohort (Figure 1) consisted of 2 separate sources: deidentified closed insurance claims (medical and pharmacy claims) with deterministically linked cholesterol laboratory data and mortality data (HealthVerity) spanning October 1, 2017, to August 31, 2024. This data set includes >240 million individuals. The study consisted of individuals with more than 1 year of data (wash-in period) prior to the earliest outpatient Lp(a) assessment (index date) and claims data after the Lp(a) test. The medical insurance claims data included International Classification of Diseases (ICD) codes and Current Procedural Terminology (CPT)/Healthcare Common Procedure Coding System codes which were used to ascertain clinical variables; the pharmacy insurance claims information included National Drug Code data which were used to ascertain prescription fills; the laboratory data included cholesterol laboratory results from large independent lab facilities.
Figure 1.
Study Flow Diagrams
STROBE study flow diagram for the derivation (left) and validation (right) cohorts. Lp(a) = lipoprotein(a); STROBE = Strengthening the reporting of observational studies in epidemiology.
The validation data set (Figure 1) was derived from Optum’s Market Clarity Data (Optum Market Clarity) which consists of deidentified electronic health record information deterministically linked with medical and pharmacy data. The data span January 1, 2005, to March 30, 2024. The Market Clarity data set includes >114 million individuals. Individuals in the Market Clarity database who met all criteria required for inclusion in the derivation cohort were included in the validation cohort for the study. For each patient, the index date of the study was selected as the date of the earliest Lp(a) result recorded.
ASCVD Outcomes
The primary composite outcome consisted of incident fatal and nonfatal myocardial infarction (MI) and ischemic stroke. The secondary composite outcome included incident fatal and nonfatal MI and ischemic stroke, and revascularization procedures (percutaneous coronary interventions, coronary artery bypass, peripheral interventions, and carotid interventions). A list of ICD/CPT codes used to define these outcomes is provided in Supplemental Table 1. Incident MI and ischemic stroke definitions were restricted to ICD codes on inpatient claims.22 All-cause mortality was assessed in the derivation cohort but was not available in the validation data set.
Analysis methods
This was a retrospective cohort study performed on deidentified data. Consistent with American Heart Association and American College of Cardiology guidelines, elevated Lp(a) was defined as a value >125 nmol/L or 50 mg/dL.9 The primary study population included all patients aged 18 years or over with at least 1 year of data and an outpatient Lp(a) lab test measured in either nmol/L or mg/dL. Variables of interest included medical conditions, family history of ASCVD, medications, procedures, and laboratory values, relevant ICD-9/ICD-10 codes, CPT/Healthcare Common Procedure Coding System codes, National Drug Codes, Systematized Nomenclature of medicine codes, Logical Observation Identifiers Names and Codes, and structured and unstructured text fields. Additionally, clinical logic was embedded to identify conditions which may not be charted within the Optum Market Clarity data set (eg, a hemoglobin A1C value of 7.5% classified an individual as having diabetes even if no ICD9/10 code for diabetes was captured). Medication usage was ascertained via the presence of a prescription fill between 183 days before the index date.
Statistical analysis
To determine the association between Lp(a) and the outcomes of interest, we performed a stepwise analysis with a separate Cox proportional hazards regression model fitted to each outcome. In the first model, we performed a multivariable regression analysis adjusted for age, sex, and race/ethnicity. We then used a forward selection approach that adjusted for 32 additional variables relevant to ASCVD including baseline medications, relevant comorbidities, and laboratory studies (Supplemental Table 2). Analysis was performed with Python 3.11 and the “lifelines” package.23
Propensity score analysis
As a secondary analysis, individuals were separated into 2 groups (Lp(a) ≥125 nmol/L and Lp(a) <125 nmol/L). Propensity scores were used to control for confounding and isolate the effect of elevated Lp(a). Propensity scores were estimated using a logistic regression model with the group status (Lp(a) ≥125 nmol/L or Lp(a) <125 nmol/L) as the dependent variable.24 Details regarding the propensity score analysis are in the Supplemental Methods.
ML short-term ASCVD event prediction model development
Short-term ASCVD event risk models were developed using tree-based ML techniques. Three separate models were trained on the derivation cohort: 1-, 2-, and 3-year ASCVD event prediction tools. For each model, the data set was split into a training and test data set by time; to keep the training and test data sets appropriately sized, a different cutoff date was used for each of the 3 models. All individuals with their first Lp(a) measurement after the cutoff date were selected for the test data set, and all individuals with measurements before this date were combined into the training data set. Individuals with <1 year, 2 years, or 3 years of data after their index date and with no ASCVD event after their index date were excluded due to incomplete data. For the ML ASCVD short-term risk prediction models, a total of 139 variables were utilized (Supplemental Table 3). Further details on the ML model development and Gradient Boosted Machine hyperparameter combinations (Supplemental Table 4) are included in Supplemental Methods.25
ML ASCVD event prediction model evaluation
The area under the receiver operating characteristic curve (AUC) was calculated for each model on both the derivation (test data) and validation cohorts, as well as the following subcohorts: men vs women, age >50 vs age ≤50 years, Asian, Black, White, Hispanic, primary prevention patients, secondary prevention patients, patients with race/ethnicity data, and patients lacking race/ethnicity data. For each model, 3 operating points were chosen for further analysis: ≥5% predicted risk, ≥25% predicted risk, and ≥50% predicted risk. Additional metrics were reported for these operating points, including sensitivity (SN), specificity (SP), accuracy, and number needed to evaluate (NNE).
Model calibration was also assessed and reported. Feature importance plots were generated for the 1-year ASCVD model to interpret the relative contribution for each variable (Supplemental Table 5, Supplemental Methods).26 To further characterize the specific contribution of Lp(a) to the 1-year ASCVD prediction, the model was retrained after omitting Lp(a) from the input variables. Net reclassification improvement (NRI) for the 1-year ASCVD model with Lp(a) vs without Lp(a) was calculated by determining the proportions of patients that were correctly reclassified as having an event or not experiencing an event across discrete predicted risk thresholds (5%, 10%, and 25%).27 In addition, an integrated discrimination improvement (IDI) was calculated by determining NRIs across the entire continuum of predicted risk.28 The projected number of ASCVD events missed per 1,000 patient-years was calculated based on the change in SN in the 1-year ASCVD model with and without Lp(a) in the validation data set (Supplemental Methods).
Results
Baseline characteristics of derivation and validation study populations
The derivation cohort for this study consisted of 731,983 individuals from a large U.S. claims database with at least one Lp(a) assessment (Figure 1). The median age was 56 years, and 53% of the individuals in the cohort were women. Among the individuals with race/ethnicity data, 60% were White, 9% were Black, 5% were Asian, and 14% were Hispanic. Lp(a) levels in the cohort were asymmetrically distributed and the median was 30 nmol/L (Figure 2A). Women had higher median Lp(a) levels compared to men (32 nmol/L vs 27 nmol/L) (Figure 2B), and Black individuals had the highest median Lp(a) levels (91 nmol/L, IQR: 42-181) across racial and ethnic subgroups (Figure 2C). Lp(a) levels did not correlate with any other standard lipid parameters (Figure 2D). Approximately 20% of the overall cohort had elevated Lp(a) levels (>125 nmol/L or 50 mg/dL). Individuals with elevated Lp(a) had a median Lp(a) level of 193 nmol/L and demonstrated a higher burden of cardiometabolic risk factors and ASCVD-related events and procedures (Supplemental Table 6). In addition, patients with elevated Lp(a) were more likely to be on statins and other forms of lipid-lowering therapy (LLT).
Figure 2.
Lipoprotein(a) Distribution and Correlation With Other Lipid Parameters
(A) Overall distribution of plasma Lp(a) levels across the entire derivation cohort. (B) Plasma Lp(a) levels in derivation cohort stratified by sex. (C) Plasma Lp(a) levels in derivation cohort stratified by race and ethnicity. (D) Correlation of plasma levels Lp(a) with other lipid parameters in derivation cohort. HDL-C = high-density lipoprotein cholesterol; TG = triglycerides; Other abbreviation as in Figure 1.
The validation cohort was derived from the Optum Market Clarity database which included 53,930 individuals with at least one Lp(a) measurement (Table 1). Compared to the derivation cohort, individuals in the validation cohort were older and had a significantly higher prevalence of ASCVD and cardiometabolic risk factors. The median Lp(a) level was higher in the validation cohort (44 nmol/L), and a greater percentage (27.4%) had an elevated Lp(a). Baseline ASCVD was also more prevalent, and a higher proportion of patients were on LLT and cardioprotective medications.
Table 1.
Baseline Characteristics of Patients in Derivation and Validation Cohorts With Lp(a) Measurement
| Derivation: HealthVerity | Validation: Market Clarity | |
|---|---|---|
| (n = 731,983) | (n = 53,930) | |
| Age, y, median (IQR) | 56 (45-64) | 62 (54-70) |
| Female, N (%) | 386,595 (53) | 22,816 (42) |
| Race and ethnicity, N (%) | ||
| White | 185,882 (25) | 43,329 (80) |
| Black | 26,989 (3.7) | 4,074 (7.6) |
| Hispanic | 42,366 (5.8) | 1,618 (3) |
| Asian | 16,435 (2.2) | 2,002 (3.7) |
| Other | 38,671 (5.3) | – |
| Unknown | 464,006 (63) | 4,525 (8.4) |
| Vital signs, median (IQR) | – | 28.1 (24.8-32.2) |
| Body mass index | – | 125 (116-136) |
| Systolic blood pressure | – | 76 (70-82) |
| Diastolic blood pressure | ||
| Family history of CVD, N (%) | 120,841 (17) | 17,048 (32) |
| Medical comorbidities, N (%) | ||
| Hypertension | 337,433 (46) | 35,239 (65) |
| Hyperlipidemia | 492,452 (67) | 47,662 (88) |
| Diabetes mellitus | 154,252 (21) | 15,857 (29) |
| Acute coronary syndrome | 37,662 (5.1) | 11,407 (21) |
| Myocardial infarction | 35,634 (4.9) | 9,454 (18) |
| Coronary artery disease | 122,874 (17) | 31,181 (58) |
| Heart failure | 47,645 (6.5) | 9,728 (18) |
| Ischemic stroke | 24,541 (3.4) | 5,580 (10) |
| Peripheral arterial disease | 41,257 (5.6) | 7,608 (14) |
| Aortic stenosis | 8,237 (1.1) | 2,891 (5.4) |
| Chronic kidney disease | 52,882 (7.2) | 7,994 (15) |
| Systemic lupus erythematosus | 9,386 (1.3) | 817 (1.5) |
| Psoriasis | 22,919 (3.1) | 2,107 (3.9) |
| Rheumatoid arthritis | 21,265 (2.9) | 2,014 (3.7) |
| Inflammatory bowel disease | 10,813 (1.5) | 130 (0.24) |
| HIV | 4,205 (0.57) | 386 (0.72) |
| Laboratory studies, median (IQR) | ||
| Lp(a) (nmol/L) | 30 (11-94) | 44 (15-140) |
| Lp(a) (mg/dL) | 12 (6-37) | – |
| Total cholesterol (mg/dL) | 191 (161-222) | 171 (138-208) |
| LDL cholesterol (mg/dL) | 111 (86-138) | 93 (68-125) |
| Non-HDL cholesterol (mg/dL) | 133 (106-163) | 1 (0-1) |
| Triglycerides (mg/dL) | 102 (73-150) | 105 (74-153) |
| Apo A (mg/dL) | – | 145 (128-166) |
| Apo B (mg/dL) | – | 88 (69-110) |
| Creatinine (mg/dL) | – | 0.91 (0.79-1.08) |
| eGFR | – | 82 (63-96) |
| hs-CRP | – | 1.4 (0.6-3.33) |
| Hemoglobin A1C (%) | – | 5.6 (5.3-6.1) |
| Medications, N (%) | ||
| Statin | 213,874 (29) | 25,192 (47) |
| Ezetimibe | 24,437 (3.3) | 3,756 (7) |
| PCSK9 mAb | 6,046 (0.83) | 1,103 (2) |
| Bempedoic acid | 310 (0.042) | 58 (0.11) |
| Beta-blockers | 128,395 (18) | 16,310 (30) |
| Ace inhibitor/angiotensin II receptor blocker | 150,182 (21) | 14,495 (27) |
| Diuretics | 66,106 (9) | 5,783 (11) |
| Calcium-channel blocker | 83,639 (11) | 8,320 (15) |
| Procedures, N (%) | ||
| CABG | 4,611 (0.63) | 1,547 (2.9) |
| Coronary angiography | 64,792 (8.9) | 17,819 (33) |
| Coronary revascularization | 20,803 (2.8) | 7,286 (14) |
| Revascularization (any) | 43,982 (6) | 14,417 (27) |
| Aortic valve replacement (any) | 336 (0.046) | 464 (0.86) |
CABG = coronary artery bypass graft; eGFR = estimated glomerular filtration rate; HDL = high-density lipoprotein; hs-CRP = high sensitivity C-reactive protein; LDL = low-density lipoprotein; Lp(a) = lipoprotein(a); PCSK9 = proprotein convertase subtilisin/kexin type 9.
Incident ASCVD events in derivation and validation cohorts
During a mean follow-up period of 3.2 ± 2.1 years in the derivation cohort, a total of 15,235 ASCVD events occurred, including 8,427 MIs and 8,051 ischemic strokes (Table 2, Supplemental Figure 1). Over 44% of the MIs observed occurred within the first year of follow-up (0.6% of the overall derivation cohort). By the end of 3 years of follow-up, over 80% of observed MIs in the cohort had occurred. Similar trends were noted for ischemic stroke. In the validation cohort, which had a higher overall burden of ASCVD and cardiometabolic disorders, the mean follow-up was 2.8 ± 2.7 years, and a higher percentage of the validation cohort population experienced an ASCVD event compared to the derivation cohort (6.6% vs 1.2% for MIs, 4.7% vs 1.1% for ischemic strokes). Similar to the derivation cohort, a significant proportion of the ASCVD events occurred early in the study observation period (Table 2, Supplemental Figure 1). For both cohorts, there were significant differences between primary and secondary prevention patients (Supplemental Tables 7 and 8). Individuals without pre-existing ASCVD were younger and more likely to be women, while a family history of ASCVD and cardiometabolic risk factors were more common in secondary prevention patients (Supplemental Table 7). In addition, ASCVD event rates were approximately 4- to 6-fold higher in secondary prevention patients across the derivation and validation cohorts (Supplemental Table 8).
Table 2.
Temporal Trends in ASCVD Events in Derivation and Validation Cohorts
| Clinical Events | Incident Eventsa | Cumulative Event Rate (100 patient-years) | Events in 1st Yeara | Events in 2nd Yeara | Events in 3rd Yeara |
|---|---|---|---|---|---|
| Derivation cohort | |||||
| ASCVD event | 15,235 (2.1) | 0.66 | 6,859 (1.1) | 3,417 (0.8) | 2,049 (0.6) |
| MI | 8,427 (1.2) | 0.36 | 3,749 (0.6) | 1,890 (0.4) | 1,172 (0.3) |
| Ischemic stroke | 8,051 (1.1) | 0.66 | 3,478 (0.6) | 1,823 (0.4) | 1,103 (0.3) |
| Revascularization | 37,919 (5.2) | 1.70 | 27,335 (4.4) | 5,265 (1.2) | 2,554 (0.8) |
| Mortality | 3,752 (0.5) | 0.16 | 1,112 (0.2) | 902 (0.2) | 703 (0.2) |
| Validation cohort | |||||
| ASCVD event | 5,523 (10.2) | 3.24 | 2,372 (6.4) | 817 (3.1) | 618 (3.2) |
| MI | 3,568 (6.6) | 2.03 | 1,440 (3.9) | 564 (2.1) | 397 (2.0) |
| Ischemic stroke | 2,552 (4.7) | 1.41 | 1,063 (2.9) | 331 (1.2) | 295 (1.5) |
| Revascularization | 4,575 (8.5) | 2.72 | 2,126 (5.8) | 730 (2.8) | 510 (2.6) |
ASCVD = atherosclerotic cardiovascular disease; MI = myocardial infarction.
Values are n (%) unless otherwise indicated.
% for the 1-, 2-, and 3-year columns indicate the percentage of patients who had an event in that year as a fraction of individuals with event-free follow-up to that year.
Association of elevated LP(a) with ASCVD events and all-cause mortality
Multivariable Cox proportional hazard models adjusting for age, sex, and race/ethnicity (Figure 3A, top panel) in the derivation cohort demonstrated that elevated Lp(a) levels were independently associated with an increased risk of the primary composite outcome (incident MI and ischemic stroke) (HR: 1.072; 95% CI: 1.059-1.084 per 50 nmol/L higher Lp(a)) and secondary composite endpoint (incident MI, incident ischemic stroke, and revascularization) (HR: 1.090; 95% CI: 1.080-1.10 per 50 nmol/L increment in Lp(a)). When we examined individual components of the composite outcomes, we found the strongest association between elevated Lp(a) and revascularization (HR: 1.130; 95% CI: 1.112-1.141 per 50 nmol/L increase in Lp(a)) (Supplemental Figure 2). Lp(a) was also positively associated with all-cause mortality (HR: 1.041; 95% CI: 1.015-1.068 per 50 nmol/L increment in Lp(a)) (Figure 3A). When the study cohort was stratified based on the presence or absence of pre-existing ASCVD (Supplemental Table 9), similar trends were observed. In primary prevention patients, elevated Lp(a) was independently associated with both the primary and secondary composite outcomes, as well as MI, stroke, and revascularization. Although elevated Lp(a) was independently associated with an increased risk of all-cause mortality in primary prevention patients, this was not observed in individuals with pre-existing ASCVD.
Figure 3.
Association of Lipoprotein(a) With Incident ASCVD Events and All-Cause Mortality
(A) Cox proportional hazard model adjusting for age, sex, race/ethnicity only (top panel), and after adjustment for 36 clinical variables (bottom panel). (B) Primary composite outcome (MI and ischemic stroke, left panel) and secondary composite outcome (MI, ischemic stroke, and revascularization) rates stratified by Lp(a) levels (≥125 nmol/L vs <125 nmol/L). ASCVD = atherosclerotic cardiovascular disease; MI = myocardial infarction; other abbreviation as in Figure 1.
After more rigorous adjustment for an additional 32 variables (Supplemental Table 2), the association of Lp(a) with all the composite and individual ASCVD outcomes remained statistically significant (Figure 3A, bottom panel; Supplemental Figure 2). Consistent with these findings, the event rates were higher in individuals with Lp(a) >125 nmol/L vs Lp(a) <125 nmol/L after propensity score matching with HRs of 1.16 (95% CI: 1.11-1.20) and 1.23 (95% CI: 1.20-1.27) for the primary and secondary composite outcomes, respectively (Figure 3B). Higher event rates were also observed in patients with elevated Lp(a) for all of the individual components including incident MI, ischemic stroke, and revascularization, as well as mortality (Supplemental Figure 3).
Machine learning ASCVD prediction model evaluation
Short-term ASCVD event prediction models were developed using regression and tree-based ML techniques. The derivation cohort was divided into training and testing sets in an approximately 3:1 ratio. A total of 139 clinical variables were used in the final ML models (Supplemental Table 3). The receiver operator curves showed robust discrimination of short-term ASCVD event risk in the derivation cohort. The C-statistic for the 1-year ASCVD risk prediction model was 0.843 with a 95% CI of 0.837 to 0.849. The 2- and 3-year prediction models showed comparable performance with C-statistics of 0.843 and 0.841, respectively (Figure 4A, left panel). The C-statistics for the 1-, 2-, and 3-year ASCVD event prediction models in the validation cohort were similarly high at 0.837, 0.829, and 0.828, respectively (Figure 4A, right panel). The models retained robust discrimination across demographic subgroups including women, patients <50 years and ≥50 years, and different racial and ethnic groups (Table 3). The median C-statistic for the model applied to primary prevention patients only ranged from 0.724 to 0.734 across the 3 models in the derivation cohort, with lower risk discrimination estimates in the validation cohort (0.655-0.675). The ASCVD risk discrimination was superior in patients with established ASCVD with C-statistics of 0.80 across all the models in the validation cohort (Table 3).
Figure 4.
Short-Term ASCVD Event Prediction Machine Learning Model Performance
(A) AUC for derivation (left panel) and validation (right panel) cohorts. (B) Calibration curves in derivation (left panel) and validation (right panel) cohorts for 1-year ASCVD event prediction ML model. ML = machine learning; ROC = receiver-operating characteristic; other abbreviation as in Figure 3.
Table 3.
Model Performance Across Subgroups
| Derivation Cohort |
Validation Cohort |
|||||
|---|---|---|---|---|---|---|
| 1-Y ASCVD Model | 2-Y ASCVD Model |
3-Y ASCVD Model |
1-Y ASCVD Model | 2-Y ASCVD Model |
3-Y ASCVD Model |
|
| Women | 0.851 | 0.850 | 0.854 | 0.851 | 0.836 | 0.834 |
| Men | 0.832 | 0.832 | 0.823 | 0.822 | 0.819 | 0.819 |
| Age <50 y | 0.839 | 0.850 | 0.849 | 0.861 | 0.853 | 0.853 |
| Age >50 y | 0.821 | 0.818 | 0.817 | 0.833 | 0.825 | 0.823 |
| Asian | 0.812 | 0.859 | 0.864 | 0.800 | 0.793 | 0.772 |
| Black | 0.823 | 0.837 | 0.850 | 0.833 | 0.827 | 0.824 |
| White | 0.842 | 0.842 | 0.836 | 0.834 | 0.826 | 0.826 |
| Hispanic | 0.839 | 0.835 | 0.839 | 0.818 | 0.824 | 0.806 |
| Primary prevention | 0.724 | 0.734 | 0.732 | 0.655 | 0.667 | 0.675 |
| Secondary prevention | 0.777 | 0.774 | 0.777 | 0.804 | 0.802 | 0.801 |
Abbreviation as in Table 2.
The calibration of the 1-year model was excellent in the derivation cohort with a slope of 1.05. However, in the validation cohort, a modest degree of risk underestimation was observed for the 1-year model, especially at the higher ranges of predicted ASCVD event risk with a calibration slope of 1.25 (Figure 4B). Similar degrees of model underfitting were noted for the 2- and 3-year ASCVD event ML prediction tools in the validation cohort with calibration slopes of 1.28 and 1.11, respectively (Supplemental Figure 4).
Model accuracy and efficiency
The distribution of predicted ASCVD event risk across all 3 models was asymmetric with right-skewing in both the derivation and validation cohorts (Supplemental Figure 5). As expected, the median 1-year predicted ASCVD event risk in the validation cohort was higher (2.7% vs 0.4%). Approximately 1 in 6 individuals in the validation cohort had a 1-year predicted ASCVD event risk of >10%, and 6.2% had a 1-year predicted risk of >25% (Supplemental Figure 5). The overall accuracy of the models was high across multiple predicted risk thresholds of >5% to >50%, ranging from 89% to 99% (Table 4). The SP of the models for detecting individuals who are likely to experience an acute ASCVD event was also consistently high across all models and risk thresholds, ranging from 0.92 to 0.99 (Table 4). The models showed robust efficiency, especially at higher predicted risks. The NNE to correctly identify one individual with a high predicted short-term CVD risk (>50%) who went on to have an ASCVD event was only 3 patients for the 1-year model. For the 2- and 3-year models, the NNE was only 2 patients (Table 4). The overall accuracy and SP of the models were similar in primary and secondary prevention patients at lower risk thresholds (5%); direct comparisons at higher risk thresholds were not possible due to a lack of cases in the primary prevention group (Supplemental Table 10).
Table 4.
Accuracy and Precision of Models in Derivation Cohort
| Model | Predicted Risk (%) | SN | SP | PPV | Accuracy | NNEa (N) |
|---|---|---|---|---|---|---|
| 1 y | >5 | 0.411 | 0.955 | 0.102 | 0.948 | 10 |
| >25 | 0.041 | 0.998 | 0.250 | 0.987 | 4 | |
| >50 | 0.007 | 0.999 | 0.333 | 0.988 | 3 | |
| 2 y | >5 | 0.508 | 0.932 | 0.235 | 0.916 | 5 |
| >25 | 0.103 | 0.995 | 0.448 | 0.960 | 3 | |
| >50 | 0.010 | 0.999 | 0.583 | 0.961 | 2 | |
| 3 y | >5 | 0.541 | 0.921 | 0.344 | 0.893 | 3 |
| >25 | 0.148 | 0.993 | 0.619 | 0.932 | 2 | |
| >50 | 0.015 | 0.999 | 0.684 | 0.929 | 2 |
PPV = positive predictive value; SN = sensitivity; SP = specificity.
NNE-number needed to evaluate to correctly predict 1 ASCVD event at different predicted risk thresholds.
Contribution of LP(a) to imminent ASCVD risk prediction
Among potentially modifiable CV risk factors used in the models, only hypertension had a higher impact on 1-year ASCVD risk prediction (Figure 5A) than Lp(a). Lp(a) contributed more to short-term ASCVD event prediction than smoking, low-density lipoprotein cholesterol (LDL-C), non-high-density lipoprotein cholesterol, total cholesterol, triglycerides, or diabetes in the 1-year ASCVD prediction model. Overall, Lp(a) was in the top 10% of all input variables in terms of feature importance. Prior ASCVD events, previous revascularization and age emerged as the clinical input variables with the highest overall contribution to the ML model predictions of short-term ASCVD risk over 1-, 2-, and 3-year time frames after Lp(a) testing (Supplemental Table 5).
Figure 5.
Lipoprotein(a) Contribution to 1-Year ASCVD Event Prediction
(A) Potentially modifiable risk factors ranked by feature importance in 1-year ASCVD (left panel), 2-year (middle panel), and 3-year (right panel) event ML prediction models. (B) Integrated discrimination improvement (IDI) plot demonstrating net reclassification improvement (NRI) across 5%, 25%, and 50% risk thresholds. Abbreviations as in Figures 1, 2, and 3.
The contribution of Lp(a) to the performance of the 1-year ASCVD event prediction model was quantified by comparing risk discrimination with vs without Lp(a). Inclusion of Lp(a) in the 1-year ASCVD event model led to a modest improvement in risk discrimination (ΔC-statistic of 0.001 in the derivation cohort and 0.002 in the validation cohort). The NRI varied across different predicted risk thresholds (Figure 5B, Table 5). Overall, the 1-year ASCVD model demonstrated an integrated discrimination improvement (IDI) index of 0.03 with an IDI for events of 0.034 and −0.0037 for nonevents (Figure 5B). Optimal model performance was noted at a predicted threshold of 26% (overall NRI 10.9%, with NRIevent of 13.8% and NRInon-event of −2.9%). The potential population-level impact of omitting Lp(a) from the 1-year ASCVD risk prediction tool was estimated to be 8.6 missed acute ASCVD events per 1,000 patient-years using a 25% predicted risk threshold to delineate high- and low-risk patients (see Supplemental Methods).29
Table 5.
Net Reclassification of Risk With Inclusion of Lp(a) in 1-Year ASCVD Model
| Model Without Lp(a) | Model With Lp(a) | |
|---|---|---|
| Patients who had ASCVD events (N = 1,283) | Risk <5% | Risk >5% |
| Risk <5% | 290 | 28 |
| Risk >5% | 17 | 948 |
| Patients without ASCVD events (N = 18,221) | Risk <5% | Risk >5% |
| Risk <5% | 13,440 | 267 |
| Risk >5% | 299 | 4,215 |
| Patients who had ASCVD events (N = 1,283) | Risk <25% | Risk >25% |
| Risk <25% | 924 | 180 |
| Risk >25% | 12 | 167 |
| Patients without ASCVD events (N = 18,221) | Risk <25% | Risk >25% |
| Risk <25% | 17,299 | 625 |
| Risk >25% | 60 | 237 |
| Patients who had ASCVD events (N = 1,283) | Risk <50% | Risk >50% |
| Risk <50% | 290 | 28 |
| Risk >50% | 17 | 948 |
| Patients without ASCVD events (N = 18,221) | Risk <50% | Risk >50% |
| Risk <50% | 13,440 | 267 |
| Risk >50% | 299 | 4,215 |
A comparison of NRI and IDI with inclusion of Lp(a) in the 1-year ML ASCVD model demonstrated significant differences in primary vs secondary prevention patients (Supplemental Table 11). The NRI and IDI were higher in secondary prevention patients across different risk thresholds. The optimal NRI in secondary prevention patients was 9.2% at a risk threshold of 25%, with an overall IDI of 2.9%. In contrast, the highest NRI in primary prevention patients was 1.2% at a risk threshold of 5% and the overall IDI was 0.0. Given that a significant percentage of patients in the derivation cohort did not have data on race/ethnicity, we performed additional SN analyses. Similar C-statistics values were observed when the 1-year model performance was assessed in patients with vs patients without race/ethnicity data (Supplemental Table 12). Furthermore, the model accuracy and NNE were very similar regardless of whether race/ethnicity data were available (data not shown).
Discussion
Current ASCVD risk assessment strategies focus on 10-year and lifetime risk and anchor on primary prevention even though approximately 25% of all MIs and ischemic strokes occurring annually in the United States are recurrent events.29 Temporal patterns observed in our analysis underscore the importance of forecasting short-term ASCVD event risk in a population tested for Lp(a), with over 40% of ASCVD events occurring within the first year of follow-up. Emerging data also suggest that imminent risk of ASCVD events may be more motivating to patients and providers with regard to timely initiation and titration of preventative medical therapies, as well as behavioral changes.18 Existing prediction tools also rely almost exclusively on traditional risk factors, in spite of growing evidence that nontraditional risk factors like elevated Lp(a) contribute significantly to ASCVD burden. Furthermore, current ASCVD risk prediction models like PREVENT (Predicting Risk of CVD EVENTs) have been trained on population data sets and registries current through 2017 that do not fully capture significant shifts in prevention and treatment strategies that have emerged over the past decade. In this context, there is a strong argument for reframing risk prediction approaches by developing complementary ASCVD event prediction tools, built using more contemporary data, that focus on shorter time horizons and incorporate nontraditional but potentially modifiable risk factors.
In this study, we addressed these gaps by leveraging, to our knowledge, the largest database of patients with Lp(a) testing and longitudinal clinical follow-up to date. We were motivated to investigate the role of elevated Lp(a) in short-term ASCVD event risk given the strong association between Lp(a) and high-risk, noncalcified plaque on coronary cardiac computed tomography angiography.30, 31, 32 We noted an independent association between elevated Lp(a) and ASCVD, with a 7% increase in risk for incident first and recurrent ASCVD events per 50 nmol/L increase when adjusting for age, sex, and race, as was done previously in a large cohort study.6 We built novel ML models that accurately predict imminent risk of acute ASCVD events in patients with and without established ASCVD, providing the framework for more precise and quantitative approaches to short-term ASCVD risk prediction, especially in the secondary prevention setting. The ML models also shed important insight into the role of elevated Lp(a) as a strong predictor of ASCVD events over short-term risk horizons in both primary and secondary prevention settings. Lp(a) emerged as an important variable in the 1-year ASCVD event prediction model, contributing to risk prediction more than any other potentially modifiable risk factor aside from hypertension. Importantly, Lp(a) ranked higher than any other lipid parameter in terms of feature importance across the 1-, 2-, and 3-year ASCVD event prediction models.
The novel short-term ASCVD event prediction models presented here were developed and validated using ∼785,000 patients with contemporary data spanning from 2017 to 2024. The models demonstrate robust discrimination with very high accuracy in both derivation and validation cohorts. The Harrell C-statistic for the short-term risk models match or exceed previously reported discrimination estimates for 10-year ASCVD risk prediction using tools such as the Pooled Cohort Equation and PREVENT equations. Importantly, the short-term ASCVD risk models performed well across age, sex, and multiple racial/ethnic groups (Table 3). Including Lp(a) in the 1-year model improved incident ASCVD risk prediction, with a NRI of 11% using a risk threshold of ∼25%. Given the 1.5 million MIs and ischemic strokes occurring annually in the United States,29 inclusion of Lp(a) in the 1-year ASCVD event model is projected to potentially identify an additional 194,000 individuals each year who may have otherwise been classified as being at a lower risk of having an acute ASCVD event. Coupled with the high predictive accuracy of these short-term risk models (as evidenced by the very low NNE values in Table 4), the model efficiency for capturing these patients and potentially avoiding an imminent ASCVD event is compelling.
Together, these data underscore the importance of measuring Lp(a) and quantifying short-term ASCVD risk as important strategies to implement timely and targeted interventions in the highest risk patients. Currently, several highly effective nonstatin LLTs are available, but several challenges associated with access and cost may limit the impact of these medications in eligible patients. Quantifying short-term ASCVD risk may facilitate identifying the patients most likely to benefit from currently available, advanced nonstatin LLT, as well as future RNA silencing therapeutics in advanced stages of clinical development that potently reduce Lp(a) levels by up to 90%.33
Study limitations
Collectively, these observations suggest the potential for broad applicability of the novel ML risk prediction models presented; however, our study has limitations. Although the models have the versatility to forecast short-term risk in both primary and secondary prevention patients, the predictive tools performed consistently better in secondary prevention patients. The decrement in model performance observed in primary prevention settings is likely attributable to the baseline characteristics of the derivation and validation cohorts, which both had higher proportions of patients with pre-existing ASCVD and cardiometabolic diseases (Table 1). The high prevalence of ASCVD and cardiac risk factors in the cohorts is linked to another notable limitation: the models presented were developed using patients who were tested for Lp(a). Existing literature demonstrates that while overall testing rates for Lp(a) are very low, individuals with high cardiometabolic disease burden are more likely to have an Lp(a) assessment. As such, the models’ utility may be more limited in healthier populations without pre-existing ASCVD. Developing similar ML models which feature Lp(a) as a clinical variable using larger cohorts of patients without pre-existing ASCVD represents an important future direction and strategy to advance short-term ASCVD risk prediction in primary prevention settings.
Deployment of the ML ASCVD event prediction models developed in this study, which employ 139 clinical variables, faces potential technical, strategic, and operational challenges. Implementation of such models may be facilitated by the development of complex application programing interfaces to manage the data collection burden. Future studies focusing on development of parsimonious ML prediction tools with prioritized variables may facilitate broader adoption in a clinical setting and facilitate tracking key parameters such as feature drift and stability postdeployment.
Conclusions
The ML models presented in this study provide novel approaches to assessing short-term ASCVD event risk, which is an important factor for patients in their health care decisions and for health systems and payers for designing effective care delivery strategies. These models incorporate and feature Lp(a) in ASCVD risk over shorter risk horizons of 1 to 3 years. In our analysis, Lp(a) emerged as a significant predictor of short-term ASCVD events, ranking higher than multiple traditional modifiable risk factors such as smoking, LDL-C, and diabetes. The performance of the models across diverse demographic subgroups and data environments suggests potential for broad applicability in providing personalized and tailored ASCVD event prediction in higher-risk primary and secondary prevention populations. Assessing Lp(a) and short-term ASCVD risk may provide novel approaches to identifying patients who may benefit from escalation of lipid-lowering and other preventative therapies.
Funding support and author disclosures
This work was supported by Novartis Pharmaceutical Corporation. This material is the result of work supported with resources and the use of facilities at the VA Palo Alto Healthcare System, Palo Alto, CA. Drs Ward and Kron are employed by Clint. Dr Ward reported owning stocks in Clint. Dr Lozama is employed by Novartis Pharmaceutical Corporation and has received salary support from Novartis. Dr Sandhu has received research funding from Novo Nordisk, Reprieve Cardiovascular, and AstraZeneca outside the submitted work. Dr Khandelwal has received research funding from Novartis, Lilly, American Heart Association, and FH foundation outside the submitted work. Dr Rodriguez has received personal fees from Carta Healthcare, HealthPals, Novartis, Novo Nordisk, Esperion Therapeutics, Movano Health, Kento Health, Inclusive Health, Edwards, Arrowhead Pharmaceuticals, HeartFlow, iRhythm, Amgen, and Cleerly Health outside the submitted work. Dr Dash has received research funding from Novartis and is a paid consultant for Clint. Dr Nallamshetty has received research funding from Novartis and American Heart Association outside the submitted work.
Perspectives.
COMPETENCY IN MEDICAL KNOWLEDGE: Lp(a) is a potent, genetically determined lipid risk factor that independently increases vulnerable plaque burden and the risk of ASCVD. Lp(a) contribution to short-term ASCVD risk is undefined. In large, real-world databases, a significant number of patients tested for Lp(a) have acute ASCVD events over a 1- to 3-year time frame. Novel ML models that predict ASCVD events at 1, 2, and 3 years with robust discrimination and accuracy across diverse patient populations demonstrate that Lp(a) is one of the most important potentially modifiable risk factors in short-term ASCVD event prediction.
COMPETENCY IN PATIENT CARE: High-risk primary and secondary prevention patients should be considered for assessments of Lp(a) and short-term ASCVD risk to facilitate more personalized treatment approaches.
TRANSLATIONAL OUTLOOK 1: Measuring Lp(a) contributes significantly to forecasting short-term ASCVD event risk in predictive ML models.
TRANSLATIONAL OUTLOOK 2: Assessment of ASCVD risk over shorter risk horizons may provide a framework for identifying the highest-risk primary and secondary prevention patients. Evaluation of short-term ASCVD risk may provide complementary predicted patient-trajectory data that can allow for efficient identification of which patients should be prioritized for treatment with advanced preventative therapies that are currently available including proprotein convertase subtilisin/kexin type 9 inhibitors and emerging therapies such as angiopoietin-like protein 3 inhibitors and RNA interference therapies that target Lp(a).
Footnotes
The authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the Author Center.
Appendix
For an expanded Methods section, tables and figures, please see the online version of this paper.
Supplementary Material
References
- 1.Emerging Risk Factors C., Erqou S., Kaptoge S., et al. Lipoprotein(a) concentration and the risk of coronary heart disease, stroke, and nonvascular mortality. JAMA. 2009;302:412–423. doi: 10.1001/jama.2009.1063. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Clarke R., Peden J.F., Hopewell J.C., et al. Genetic variants associated with Lp(a) lipoprotein level and coronary disease. N Engl J Med. 2009;361:2518–2528. doi: 10.1056/NEJMoa0902604. [DOI] [PubMed] [Google Scholar]
- 3.Afshar M., Kamstrup P.R., Williams K., Sniderman A.D., Nordestgaard B.G., Thanassoulis G. Estimating the population impact of Lp(a) lowering on the incidence of myocardial infarction and aortic stenosis-brief report. Arterioscler Thromb Vasc Biol. 2016;36:2421–2423. doi: 10.1161/ATVBAHA.116.308271. [DOI] [PubMed] [Google Scholar]
- 4.Saleheen D., Haycock P.C., Zhao W., et al. Apolipoprotein(a) isoform size, lipoprotein(a) concentration, and coronary artery disease: a mendelian randomisation analysis. Lancet Diabetes Endocrinol. 2017;5:524–533. doi: 10.1016/S2213-8587(17)30088-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Schmidt K., Noureen A., Kronenberg F., Utermann G. Structure, function, and genetics of lipoprotein (a) J Lipid Res. 2016;57:1339–1359. doi: 10.1194/jlr.R067314. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Patel A.P., Wang M., Pirruccello J.P., et al. Lp(a) (Lipoprotein[a]) concentrations and incident atherosclerotic cardiovascular disease: new insights from a large national biobank. Arterioscler Thromb Vasc Biol. 2021;41:465–474. doi: 10.1161/ATVBAHA.120.315291. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Bjornson E., Adiels M., Taskinen M.R., et al. Lipoprotein(a) is markedly more atherogenic than LDL: an apolipoprotein B-based genetic analysis. J Am Coll Cardiol. 2024;83:385–395. doi: 10.1016/j.jacc.2023.10.039. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Nissen S.E., Wolski K., Cho L., et al. Lipoprotein(a) levels in a global population with established atherosclerotic cardiovascular disease. Open Heart. 2022;9:e002060. doi: 10.1136/openhrt-2022-002060. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Reyes-Soffer G., Ginsberg H.N., Berglund L., et al. Lipoprotein(a): a genetically determined, causal, and prevalent risk factor for atherosclerotic cardiovascular disease: a scientific statement from the American heart association. Arterioscler Thromb Vasc Biol. 2022;42:e48–e60. doi: 10.1161/ATV.0000000000000147. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Koschinsky M.L., Bajaj A., Boffa M.B., et al. A focused update to the 2019 NLA scientific statement on use of lipoprotein(a) in clinical practice. J Clin Lipidol. 2024;18:e308–e319. doi: 10.1016/j.jacl.2024.03.001. [DOI] [PubMed] [Google Scholar]
- 11.Kronenberg F., Mora S., Stroes E.S.G., et al. Lipoprotein(a) in atherosclerotic cardiovascular disease and aortic stenosis: a European Atherosclerosis Society consensus statement. Eur Heart J. 2022;43:3925–3946. doi: 10.1093/eurheartj/ehac361. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Kamstrup P.R., Tybjaerg-Hansen A., Nordestgaard B.G. Extreme lipoprotein(a) levels and improved cardiovascular risk prediction. J Am Coll Cardiol. 2013;61:1146–1156. doi: 10.1016/j.jacc.2012.12.023. [DOI] [PubMed] [Google Scholar]
- 13.Willeit P., Kiechl S., Kronenberg F., et al. Discrimination and net reclassification of cardiovascular risk with lipoprotein(a): prospective 15-year outcomes in the Bruneck Study. J Am Coll Cardiol. 2014;64:851–860. doi: 10.1016/j.jacc.2014.03.061. [DOI] [PubMed] [Google Scholar]
- 14.Delabays B., Marques-Vidal P., Kronenberg F., Waeber G., Vollenweider P., Vaucher J. Use of lipoprotein(a) for refining cardiovascular risk prediction in a low-risk population: the CoLaus/PsyCoLaus study. Eur J Prev Cardiol. 2021;28:e18–e20. doi: 10.1177/2047487320938271. [DOI] [PubMed] [Google Scholar]
- 15.Verbeek R., Sandhu M.S., Hovingh G.K., et al. Lipoprotein(a) improves cardiovascular risk prediction based on established risk algorithms. J Am Coll Cardiol. 2017;69:1513–1515. doi: 10.1016/j.jacc.2017.01.017. [DOI] [PubMed] [Google Scholar]
- 16.Wong N.D., Fan W., Hu X., et al. Lipoprotein(a) and long-term cardiovascular risk in a multi-ethnic pooled prospective cohort. J Am Coll Cardiol. 2024;83:1511–1525. doi: 10.1016/j.jacc.2024.02.031. [DOI] [PubMed] [Google Scholar]
- 17.Wong N.D., Zhao Y., Xiang P., Coll B., Lopez J.A.G. Five-year residual atherosclerotic cardiovascular disease risk prediction model for statin treated patients with known cardiovascular disease. Am J Cardiol. 2020;137:7–11. doi: 10.1016/j.amjcard.2020.09.043. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Waldron C.A., van der Weijden T., Ludt S., Gallacher J., Elwyn G. What are effective strategies to communicate cardiovascular risk information to patients? A systematic review. Patient Educ Couns. 2011;82:169–181. doi: 10.1016/j.pec.2010.04.014. [DOI] [PubMed] [Google Scholar]
- 19.Backus B.E., Six A.J., Kelder J.C., et al. A prospective validation of the HEART score for chest pain patients at the emergency department. Int J Cardiol. 2013;168:2153–2158. doi: 10.1016/j.ijcard.2013.01.255. [DOI] [PubMed] [Google Scholar]
- 20.Antman E.M., Cohen M., Bernink P.J., et al. The TIMI risk score for unstable angina/non-ST elevation MI: a method for prognostication and therapeutic decision making. JAMA. 2000;284:835–842. doi: 10.1001/jama.284.7.835. [DOI] [PubMed] [Google Scholar]
- 21.Gazit T., Mann H., Gaber S., Adamenko P. A novel, machine-learning model for prediction of short-term ASCVD risk over 90 and 365 days. Front Digital Health. 2024;6:1485508. doi: 10.3389/fdgth.2024.1485508. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Bush M., Sturmer T., Stearns S.C., et al. Position matters: validation of medicare hospital claims for myocardial infarction against medical record review in the atherosclerosis risk in communities study. Pharmacoepidemiol Drug Saf. 2018;27:1085–1091. doi: 10.1002/pds.4396. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Davidson-Pilon C. Lifelines: survival analysis in Python. J Open Source Software. 2019;4:1317. [Google Scholar]
- 24.Franklin J.M., Eddings W., Austin P.C., Stuart E.A., Schneeweiss S. Comparing the performance of propensity score methods in healthcare database studies with rare outcomes. Stat Med. 2017;36:1946–1963. doi: 10.1002/sim.7250. [DOI] [PubMed] [Google Scholar]
- 25.van Buuren S., Groothuis-Oudshoorn K. Mice: multivariate imputation by chained equations in R. J Stat Software. 2011;45:1–67. [Google Scholar]
- 26.Natekin A., Knoll A. Gradient boosting machines, a tutorial. Front Neurorobot. 2013;7:21. doi: 10.3389/fnbot.2013.00021. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Kerr K.F., Wang Z., Janes H., McClelland R.L., Psaty B.M., Pepe M.S. Net reclassification indices for evaluating risk prediction instruments: a critical review. Epidemiology. 2014;25:114–121. doi: 10.1097/EDE.0000000000000018. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Pencina M.J., D'Agostino R.B., Pencina K.M., Janssens A.C., Greenland P. Interpreting incremental value of markers added to risk prediction models. Am J Epidemiol. 2012;176:473–481. doi: 10.1093/aje/kws207. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Tsao C.W., Aday A.W., Almarzooq Z.I., et al. Heart disease and stroke statistics-2023 update: a report from the American heart association. Circulation. 2023;147:e93–e621. doi: 10.1161/CIR.0000000000001123. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Kaiser Y., Daghem M., Tzolos E., et al. Association of lipoprotein(a) with atherosclerotic plaque progression. J Am Coll Cardiol. 2022;79:223–233. doi: 10.1016/j.jacc.2021.10.044. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Yu M.M., Wang M.L., Wang J.J., et al. Association of lipoprotein(a) levels with myocardial infarction in patients with low-attenuation plaque. J Am Coll Cardiol. 2024;83:1743–1755. doi: 10.1016/j.jacc.2024.03.367. [DOI] [PubMed] [Google Scholar]
- 32.Dai N., Chen Z., Zhou F., et al. Association of lipoprotein (a) with coronary-computed tomography angiography-assessed high-risk coronary disease attributes and cardiovascular outcomes. Circ Cardiovasc Imaging. 2022;15 doi: 10.1161/CIRCIMAGING.122.014611. [DOI] [PubMed] [Google Scholar]
- 33.Tsimikas S. Lipoprotein(a) in the year 2024: a look back and a look ahead. Arterioscler Thromb Vasc Biol. 2024;44:1485–1490. doi: 10.1161/ATVBAHA.124.319483. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.