Predictive Performance of Cardiovascular Risk Scores in Cancer Survivors From the UK Biobank

Background Cardiovascular preventive strategies are guided by risk scores with unknown validity in cancer cohorts. Objectives This study aimed to evaluate the predictive performance of 7 established cardiovascular risk scores in cancer survivors from the UK Biobank. Methods The predictive performance of QRISK3, Systematic Coronary Risk Evaluation 2 (SCORE2)/Systematic Coronary Risk Evaluation for Older Persons (SCORE-OP), Framingham Risk Score, Pooled Cohort equations to Prevent Heart Failure (PCP-HF), CHARGE-AF, QStroke, and CHA2DS2-VASc was calculated in participants with and without a history of cancer. Participants were propensity matched on age, sex, deprivation, health behaviors, family history, and metabolic conditions. Analyses were stratified into any cancer, breast, lung, prostate, brain/central nervous system, hematologic malignancies, Hodgkin lymphoma, and non-Hodgkin lymphoma. Incident cardiovascular events were tracked through health record linkage over 10 years of follow-up. The area under the receiver operating curve, balanced accuracy, and sensitivity were reported. Results The analysis included 31,534 cancer survivors and 126,136 covariate-matched controls. Risk score distributions were near identical in cases and controls. Participants with any cancer had a significantly higher incidence of all cardiovascular outcomes than matched controls. Performance metrics were significantly worse for all risk scores in cancer cases than in matched controls. The most notable differences were among participants with a history of hematologic malignancies who had significantly higher outcome rates and poorer risk score performance than their matched controls. The performance of risk scores for predicting stroke in participants with brain/central nervous system cancer was very poor, with predictive accuracy more than 30% lower than noncancer controls. Conclusions Existing cardiovascular risk scores have significantly worse predictive accuracy in cancer survivors compared with noncancer comparators, leading to an underestimation of risk in this cohort.

2][3] The heightened risk of cardiovascular disease (CVD) in cancer patients is widely recognized and attributed to shared risk factors, the toxicity of cancer therapies, and biologic pathways related to the cancer itself. 4Recent reports highlight a persistently elevated CVD risk extending beyond the duration of cancer treatment into long-term survivorship. 5,6rdio-oncology has emerged as a subspecialty dedicated to the cardiovascular care of cancer patients. 4However, a stronger evidence base is needed to guide clinical care.In clinical practice, CVD prevention strategies are guided by risk prediction tools such as the QRISK3 7 and the Framingham Risk Score (FRS). 8These instruments, developed and validated in the general population, do not include cancerspecific predictors.As such, they may underestimate the risk of CVD in cancer survivors, leading to undertreatment of this vulnerable population.
This study compared the predictive performance of SETTING AND STUDY POPULATION.The UK Biobank is a prospective cohort of over 500,000 participants.UK residents aged 40 to 69 years and living within 25 miles of 1 of 22 assessment centers in urban and rural areas were identified through NHS registers and recruited between 2006 and 2010.Baseline assessment included medical, social, demographic, lifestyle, environmental, and physical parameters. 9dividuals who were unable to consent or complete baseline assessment because of illness or discomfort were not recruited.Extensive health record linkage has been established for the entire cohort, allowing prospective tracking of incident health events.(Supplemental Table 1).The date of the first occurrence and the main cancer site were identified by the first cancer record in any linked database.We began with a set of 25 cancer types, following the work of Strongman et al, 5 and identified subsets with sufficient outcome counts for analysis.The final cancer groups were as follows: any cancer, breast, lung, prostate, brain/central nervous system (CNS), hematologic, Hodgkin lymphoma, and non-Hodgkin lymphoma.Hematologic cancer included all lymphomas, leukemia, multiple myeloma, polycythemia vera, myelodysplastic syndrome, and rarer blood cancers.
Individuals without any record of cancer were considered as potential noncancer comparators.
CARDIOVASCULAR RISK SCORES.We considered 7 risk scores currently used in standard clinical practice (Supplemental Table 2).The QRISK3, 7 FRS, 8 and SCORE2/SCORE-OP 10,11 scores all have a composite endpoint of myocardial infarction, stroke, or cardiovascular mortality, which we labeled CVD 1.We McCracken et al

Cardiovascular Risk Prediction in Cancer Survivors
A U G U S T 2 0 2 4 : 5 7 5 -5 8 8 presented to provide a broader category of CVDs to which cancer survivors may be more susceptible (eg, nonischemic CVDs).
We implemented CHARGE-AF 12 for the risk of AF, PCP-HF 13 to assess the risk of heart failure, and 2 scores (QStroke 14 and CHA 2 DS 2 -VASc 15 ) for the risk of stroke.Although the CHA 2 DS 2 -VASc 15 score was originally developed for estimating the risk of stroke in AF cohorts, its predictive utility has also been demonstrated in patients without AF. 16Cardiovascular risk scores were calculated for each participant based on data available at the baseline visit using equations reported in published sources as detailed in Supplemental Table 2.
ASCERTAINMENT OF COVARIATES.The discriminative performance of each risk score in predicting its intended outcome was assessed in cancer and control groups separately using timedependent area under the receiver operating curve (AUC) in the presence of competing risks. 18,19The AUC reflects the chance that the risk score correctly assigns a higher value to the person with the disease compared with the one without it. 20In addition to AUC, sensitivity (the proportion of true outcomes correctly identified) and balanced accuracy (the average of sensitivity and specificity) were assessed as static performance metrics on the 10-year outcomes.
Although AUC is agnostic to the score cutoff threshold, sensitivity and balanced accuracy are not.
Therefore, for each cancer index-outcome combination, the cutoff threshold was set at the point where balanced accuracy was highest in the noncancer control group (80% of the data) and was then applied to the cancer group.In other words, both cancer and matched controls were scored for accuracy with the same threshold value.Standard errors and CIs for  and 5 along with rarer hematologic conditions for further reference.
In both the cancer and control groups, the average age was 59.4 AE 7.2 years, with 66.5% being women and 33.5% men.Thirty percent were above the UK median Townsend deprivation score, and 9.8% were current smokers.Among the cancer cohort, the prevalence of hypertension, diabetes, and high cholesterol was 32.1%, 5.4%, and 17.8%, respectively.
Participants with past prostate cancer were the oldest, with an average age of 63.7 AE 4.4 years, whereas those with brain/CNS cancers were the youngest, with an average age of 54.9 AE 8.5 years.
PREDICTED RISK AND OBSERVED OUTCOMES.For all cancer index-outcome combinations, the matching procedure produced case and control groups that had near identical risk score profiles, with no significant differences in risk score estimates (Supplemental Figure 3, Supplemental Table 6).Over 10 years of follow-up, cardiovascular events (CVD 2) were observed in 11.5% of participants with a record of any cancer compared with 9.6% of noncancer comparators (Table 1, Supplemental Table 6).
A n y c a n c e r .Within any cancer group, the 10-year cumulative incidence was higher in cancer cases than in matched controls for all outcomes considered (Supplemental Table 7, Figure 1, Supplemental Figure 4).In terms of overall calibration and performance, SCORE2/OP had the best calibration and sensitivity for predicting CVD 1 in both cancer cases and controls, whereas QRISK3 had the best performance for CVD 2 in terms of calibration and AUC in both groups.When reporting for the cancer subgroups, only differences from this pattern will be mentioned.
Across all risk scores and outcomes, participants with cancer had significantly lower AUC than matched controls, even though this difference was small (2-5 percentage points) (Supplemental Table 7, B r e a s t c a n c e r .In the breast cancer group (Supplemental Table 8, Figure 1), there were no significant differences in 10-year cumulative incidence or risk score accuracy between cases and controls for CVD 1 and stroke.Breast cancer cases had slightly higher incidences of CVD 2, AF, and heart failure than controls (<1.5 percentage points).When predicting CVD 2, the AUCs for breast cancer cases were consistently lower than their matched controls for all risk scores considered, with lower balanced accuracy and sensitivity observed with QRISK3, FRS (blood), and SCORE/OP (Figure 3, Supplemental Table 8).SCORE2/OP and QRISK3 were the best performing risk scores in breast cancer survivors (for CVD 1 and CVD 2, respectively), although AUC failed to reach the 0.70 mark.
Lung cancer.In the lung cancer group (Supplemental  Values are mean AE SD, n (%), or median (Q1-Q3).a Physically active is defined as >600 summed metabolic equivalent task minutes per week.Median UK Townsend deprivation index ¼ À0.35 as per 2011 UK Census.CVD 1 ¼ combined endpoint including nonfatal myocardial infarction, nonfatal stroke, or cardiovascular mortality where cardiovascular mortality is defined as any death with a primary cause from International Classification of Diseases-Tenth Revision codes I00-I80.CVD 2 ¼ combined endpoint including everything from CVD 1 plus incident atrial fibrillation, heart failure, nonischemic cardiomyopathies and valvular heart disease.
CNS ¼ central nervous system; CVD ¼ cardiovascular disease; HbA1c ¼ glycated hemoglobin; HDL ¼ high-density lipoprotein.10), there were no significant differences in cumulative incidence rates between cases and controls (Figure 1), and there were no significant differences in AUC between cancer cases and matched controls in predicting the outcomes considered (Supplemental Table 10, Figures 2 to 4).Although SCORE2/OP had the best calibration for predicting CVD 1 in prostate cancer, sensitivity in cancer cases was significantly lower than in matched controls.
Sensitivity for prostate cancer cases was also significantly lower when predicting heart failure with PCP-HF.
B r a i n / C N S c a n c e r .We observed significantly higher rates of CVD 1 and CVD 2 (6.4 and 7.1 percentage points, respectively) in participants with brain/CNS cancer (Central Illustration), even after adjusting for the competing risk of death (Supplemental Table 11).
FRS (blood) provided the best overall performance for predicting CVD 2 in brain/CNS cancer cases, with better calibration and sensitivity than SCORE2/OP and QRISK3.
We also observed a significantly higher rate of stroke in brain/CNS cancer cases (7.5% [95% CI: 4.4%-11.2%]vs 0.7% [95% CI: 0.3%-1.4%]).Although both CHA 2 DS 2 VASc and QStroke discriminated stroke well in the control group, AUC for both scores was H e m a t o l o g i c m a l i g n a n c i e s .Among participants with hematologic cancer (Supplemental Table 12, Figure 1), we observed significantly higher cumulative incidence rates for all outcomes considered, even after adjusting for the competing risk of death.AUC results are from time-dependent analyses accounting for the competing risk of death at the 10-year follow-up.Balanced accuracy and sensitivity are from static performance scoring of 10-year outcomes, with 95% CIs derived from bootstrapping and permutation testing with 1,000 replicates.CVD 1 includes nonfatal myocardial infarction, nonfatal stroke, or cardiovascular mortality, where cardiovascular mortality is defined as any death with a primary cause from International Classification of Diseases-10th Revision codes I00-I80.Abbreviations as in   Abbreviations as in Figures 1 and 2.    (Supplemental Tables 15 to 18).Group-specific cumulative death risk curves are provided in Supplemental Figure 6. of shared risk factors. 5,6evious studies of breast 23 and childhood cancer survivors 24

7
established cardiovascular risk scores in cancer survivors compared with matched noncancer controls from the UK Biobank, considering a range of cardiovascular outcomes and differential relationships by cancer type.METHODS ETHICAL APPROVAL.This study complies with the Declaration of Helsinki.Ethical approval for UK Biobank studies was granted by the NHS National Research Ethics Service on June 17, 2011 (reference 11/ NW/0382) and extended on June 18, 2021 (reference 21/NW/0157).Written informed consent was obtained from all participants.
ASCERTAINMENT OF CANCER STATUS.Cancer status at baseline recruitment was ascertained based on International Classification of Disease codes in linked Hospital Episode Statistics and National Cancer Registration and Analysis Service records created a second enriched composite cardiovascular endpoint by adding incident heart failure, atrial fibrillation (AF), nonischemic cardiomyopathies, and valvular heart disease to CVD 1, which we labeled CVD 2. The CVD 2 outcome does not match the original intended outcome of the tested risk scores but is A B B R E V I A T I O N S A N D A C R O N Y M S AF = atrial fibrillation AUC = area under the receiver operating curve BMI = body mass index CNS = central nervous system CVD = cardiovascular disease FRS = Framingham Risk Score

Figures 2 to 4 ,Figure 4 )
Figures 2 to 4, Central Illustration).These differences were also present in balanced accuracy and sensitivity for all outcomes.The sensitivity of PCP-HF in predicting heart failure was 8 percentage points lower in the cancer group compared with matched controls (73% vs 81%).Cumulative incidence plots (Supplemental Figure 4) confirmed the small but significant differences in cumulative outcome incidence over time.Calibration plots (Supplemental Figure 5) confirm that, in general, SCORE2/OP had the best calibration for CVD 1 in our sample, whereas QRISK3 had the best calibration for CVD 2. Risk score values for CHARGE-AF tended to underestimate the risk for AF, whereas PCP-HF and QStroke tended to overestimate heart failure and stroke outcomes.

J 4
A C C : C A R D I O O N C O L O G Y , V O L .6 , N O . 4 , 2 0 2 McCracken et al A U G U S T 2 0 2 4 : 5 7 5 -5 8 8 Cardiovascular Risk Prediction in Cancer Survivors results showed that the absolute SCORE2/OP value underestimated the observed risk by 15 percentage points.Other risk scores had overall comparable performances in lung cancer cases and matched controls (Figures 2 to 4).P r o s t a t e c a n c e r .In the prostate cancer group (Supplemental Table

FIGURE 1
FIGURE 1 Observed 10-Year Outcome Rates

FIGURE 2
FIGURE 2 Standard CVD Risk Score Predictive Performance in Cancer Groups vs Matched Controls for CVD 1

J 4
A C C : C A R D I O O N C O L O G Y , V O L .6 , N O . 4 , 2 0 2Incidences of CVD 1, AF, and heart failure were higher by 4 percentage points, whereas incidence of CVD 2 was 10.6 percentage points higher than in covariatematched controls.Because of this, even though SCORE2/OP was the best overall score for predicting CVD 1, it significantly underestimated CVD 1 risk by 3.5 percentage points.CVD 2 accuracy for all risk scores and all metrics was significantly poorer for participants with hematologic cancer than their matched controls.This was also the case for AF (CHARGE-AF) and heart failure (PCP-HF) (Figure4, Central Illustration).

FIGURE 3
FIGURE 3 Extended CVD Risk Score Predictive Performance in Cancer Groups vs Matched Controls for CVD 2

J
A C C : C A R D I O O N C O L O G Y , V O L .6 , N O . 4 , 2 0 2 4 Cardiovascular Risk Prediction in Cancer Survivors A U G U S T 2 0 2 4 : 5 7 5 -5 8 8 N o n -H o d g k i n l y m p h o m a .Participants with non-Hodgkin lymphoma (Supplemental Table 13) had a higher incidence of CVD 2 (5.7 percentage points) than in matched controls.In addition, there were several significantly lower values for AUC, balanced accuracy, and sensitivity across the various CVD 2 risk scores.The best performing score in our sample profiling CVD 2 in non-Hodgkin lymphoma was FRS (BMI), although AUC is modest at 0.61.Significant differences were also present in the estimation of heart failure (PCP-HF) in people with non-Hodgkin lymphoma, with significantly poorer

FIGURE 4
FIGURE 4 Risk Score Predictive Accuracy in Cancer Groups vs Matched Controls for Specific Outcomes .

J 4
A C C : C A R D I O O N C O L O G Y , V O L .6 , N O . 4 , 2 0 2 McCracken et al A U G U S T 2 0 2 4 : 5 7 5 -5 8 8 Cardiovascular Risk Prediction in Cancer Survivors AUC, balanced accuracy, and sensitivity than covariate-matched controls (Figure 4, Central Illustration).H o d g k i n l y m p h o m a .Participants with Hodgkin lymphoma had higher rates for CVD 1, CVD 2, AF, and heart failure than controls, with the CVD 2 rate 16.5 percentage points higher (SupplementalTable 14).All the scores we tested had very poor calibration in this cancer group, with FRS (BMI) being the best option with an AUC of 0.67 and risk underestimate of 9.4 percentage points.Cancer cases had significantly lower balanced accuracy and sensitivity in all risk scores predicting CVD 2, which was also the case with heart failure.Risk score performance was even poorer when predicting AF (CHARGE-AF), with significantly lower AUC, balanced accuracy, and sensitivity than matched controls.O t h e r h e m a t o l o g i c m a l i g n a n c i e s .We were underpowered to distinguish differences in analyses among more granular hematologic subtypes CENTRAL ILLUSTRATION The Performance of Cardiovascular Risk Scores in Cancer Survivors 31,534 cancer survivors and 126,136 matched controls from UK Biobank followed for 13.6 years • Risk score performance was worst in patients with brain cancer for stroke prediction and in • et al.JACC CardioOncol.2024;6(4):575-588.The area under the curve (AUC) for the best-performing risk score in cancer survivors.AF ¼ atrial fibrillation; BMI ¼ body mass index; CNS ¼ central nervous system; CV ¼ cardiovascular; CVD ¼ cardiovascular disease; HF ¼ heart failure.
STRENGTHS AND LIMITATIONS.The detailed participant phenotyping and health record linkage in the UK Biobank enabled faithful replication of a range of cardiovascular risk scores and prospective tracking of incident events.However, healthy participant and survival bias may have influenced our sample and observed risks.Further studies in nationally representative cohorts are needed to evaluate the generalizability of our observations.The data set did not permit reliable distinction of individuals who may have received active cancer treatment during the study follow-up period for recurrent or secondary malignancies, which may represent a particularly high-risk cohort.Similarly, cardiovascular risk prediction may be worse in people with past exposure to cardiotoxic therapies; the absence of this information in the current data set precludes evaluation of this hypothesis in the current analysis and represents an important priority for future research.The 94% 10-year survival rate is not consistent with the average 10-year survival of patients with most newly diagnosed cancers among the types included in this paper.This discrepancy suggests potential circumstantial evidence of immortal time bias.It is plausible that this bias favors the performance of the risk scores evaluated in this paper; in other words, these scores may perform even worse in patients with newly diagnosed cancer.CONCLUSIONS Our findings underscore the heightened long-term risk of CVD in cancer survivors, which are independent of shared risk factors and incompletely captured by existing risk scores.The deficits in risk assessment of hematologic cancer survivors have important implications for clinical practice and research.

Table 9 )
, we observed significantly higher 10-year incidences of CVD 2, AF, and heart failure in cancer cases compared with controls, but there were no significant differences in the time-dependent AUC.Although FRS (with BMI) provided the best absolute CVD 2 calibration for lung cancer cases, it provided significantly lower balanced accuracy and sensitivity for CVD 1 and CVD 2 than controls.SCORE2/OP had the highest sensitivity and AUC for discriminating CVD 2 among lung cancer; however, calibration

TABLE 1
Cancer Cases and Matched Controls Characteristics In the absence of dedicated risk scores, the following recommendations may be considered based on our findings when estimating risk in people with a history of cancer: CLINICAL TAKE-HOME MESSAGES: CHOICE OF RISK ESTIMATES.2 -VASc has somewhat better performance than QStroke, both tools had very poor metrics, and their use for stroke prediction in this setting is not recommended based on our results.Heart failure: although PCP-HF is a very wellperforming score in noncancer controls, it had