A risky business: recalibrating the future of cardiovascular risk prediction in cancer survivors

This editorial refers to ‘External validation of the systematic coronary risk evaluation 2 (SCORE2) and SCORE2-older persons in cancer patients’, by M.N. Gynnild et al., https://doi.org/10.1093/eurjpc/zwaf356. People living with cancer, and those who have survived cancer, have an increased risk of cardiovascular (CV) disease.1 This association is only becoming more clinically salient in the modern era of cancer treatment, where many patients are cured or live with cancer as a chronic disease and die of other causes. Cardiovascular disease is the dominant non-cancer cause of death in cancer survivors, and in some populations, it may even surpass cancer-related mortality at an early phase of follow-up.2,3 Elevated CV risk may be contributed to by traditional risk factors but can also be explained by cancer-specific variables like a pro-inflammatory and pro-coagulant state and the effects of cardiotoxic cancer therapies.1,4 In many cases, these risk factors may not be readily apparent and risk assessment can be complicated by heterogeneity across the spectrum of cancer types. Researchers and clinicians have therefore looked to risk prediction tools to guide who requires closer cardiac monitoring and cardio-protection during and after cancer treatment. A distinction must be drawn between different phases of the cancer pathway and CV risk. In the active cancer phase, the most common problem is cancer treatment-related cardiotoxicity including left ventricular dysfunction and heart failure, known as cancer therapy-related cardiac dysfunction, and other common toxicities include cancer therapy-related hypertension, arrhythmias, arterial and venous thrombosis, accelerated atherosclerosis, and pulmonary hypertension. Clinicians are aided in this phase by guideline-recommended monitoring strategies and treatment-specific baseline cardiotoxicity risk tools which have demonstrated benefit.5,6 In the survivor phase, clinicians are concerned with a longer term and more general CV risk encompassing all types of CV events. However, in this phase, there is less specific guidance. A range of existing general CV risk prediction tools, developed in different regions of the world, could be utilized; however, these likely underestimate risk in cancer populations.7 Although the original derivation studies for these tools did not exclude cancer patients, they do not incorporate the variables responsible for the elevated CV risk in this population. Before cancer-specific tools are developed though, is there a way to enhance these well-established, readily available risk tools? The study by Gynnild et al.8 addressed this question. This study utilized data from the HUNT3 registry, a longitudinal population health study in Norway. Participants who were diagnosed with cancer within 4 years of enrolment in the registry were included (1622 in total). The SCORE2 CV risk prediction tool was applied to patient data at the time of enrolment in the HUNT3 study and assumed CV risk was stable from this point until cancer diagnosis.9 SCORE2-OP was applied to the smaller sample of patients aged >70 years of age, providing similar findings. Variables in the SCORE2 assessment include age, male sex, smoking status, systolic blood pressure, diabetes status, and cholesterol levels (total and HDL). SCORE2 outcome data (a composite of MI, stroke, or CV mortality at 10 years) were collected using local and national registries. The performance of the tool was assessed using the expected versus observed (E/O) ratio of the composite outcome (a ratio of the predicted mean risk and observed incidence for the entire sample) as well as the Harrell-C index, a standard discrimination statistic. A ‘recalibration’ was then performed whereby initial expected risk for each participant was modified by a constant incorporating the E/O ratio. A range of cancer types were included with gastrointestinal, prostate, and breast predominating. Median time from study enrolment to cancer diagnosis was 2.8 years. At a median follow-up of 8.8 years, the rate of CV events was 8.8 and 20.2% for participants aged 70 years of age, respectively. The authors state that SCORE2 underestimated the risk of CV events and although this is true for females, for whom the E/O ratio was 0.63, and in males, it was higher at 0.91. Recalibration then improved the performance of SCORE2 as demonstrated by visually better concordance between expected and observed risks on a calibration plot. The C-statistic post recalibration was 0.693 (95% CI 0.643–0.743). A number of sensitivity analyses provide further insight, particularly with regards to temporal changes in the performance of SCORE2. With increasing duration, the C-statistic improved meaning SCORE2 performed better as the time from cancer diagnosis lengthened. Improvements were also seen when SCORE2 was applied solely to patients who had survived 2 years post cancer diagnosis, those with non-metastatic disease, and those with a history of prior cancer. The study by Gynnild et al. helps confirm previous work documenting the underestimation of CV risk in cancer patients by conventional risk prediction tools. There have been other, more extensive studies illustrating this including a UK Biobank study by McCracken et al.7 in 2024. This study assessed the use of seven different CV risk tools in 31 534 cancer survivors and compared these to matched controls without cancer in a 1:4 ratio. The discrimination of all tools in cancer survivors was lower than in controls, although the difference was actually only small. The key contribution to the literature from the Gynnild et al. study is that it proposes a novel method to improve available risk tools which could be useful while we await the future developments discussed in the conclusion of this editorial. Recalibration of existing tools is not ready for clinical use based on this study alone, which is only hypothesis generating. Firstly, the exact magnitude of improvement after recalibration is not clear as only the C-statistic for the recalibrated model was reported, and furthermore, at 0.693, discrimination power would be viewed as moderate only. Secondly, there was also a significant sex disparity; for males, the C-statistic in the recalibrated model was lower, at 0.633. Recalibration may improve discrimination but it remains unknown whether this is a clinically meaningful improvement. The concept also needs to be tested in larger sample sizes given the significant heterogeneity in CV risk in a cancer population. The study illustrated this by finding significant variations in CV risk and the discrimination performance of SCORE2 between cancer types and sex. Furthermore, this recalibration method has only been tested internally and external validation in independent patient cohorts is required. The sensitivity analysis findings should be highlighted as they emphasize the important concept that the timing of risk tool application must be considered. SCORE2 generally performed better in cancer survivors and with increasing time from cancer diagnosis. Cardiovascular risk tools, in their current form, will likely perform better when applied in the cancer survivor phase for a number of reasons. In the active cancer phase, there are significant competing risks and cancer treatment is likely to modify risk significantly but is not included in these tools. Furthermore, SCORE2 was designed to predict a composite of MI, stroke, or CV mortality only, and not the wider range of events patients are at risk of during this phase, most importantly left ventricular dysfunction. Although this study offers a novel solution to an important problem, recalibration of existing CV risk tools is unlikely to be the best long-term solution. What is required is the incorporation of cancer variables into established risk tools or the development of entirely new, cancer-specific scores which can be developed from cardio-oncology registries. Advances have already been made, with an updated QRisk4 score including four different cancer types and in an external validation discrimination was high (0.83 for females and 0.81 for males).10 QRisk4 also outperformed other scores including SCORE2 although it is still very general including cancer diagnosis only and did not include specific cancer therapies. In the artificial intelligence and big data era though, we should be aiming high. The goal should be the creation of personalized and highly accurate risk scores that utilize patient-specific cancer variables including not only cancer type and genomics, but treatment details and even data from cardiac monitoring tests performed during treatment. The number of cancer survivors is only going to increase, underpinning the urgent need for accurate risk profiling tools which can facilitate delivery of personalized cardio-protective strategies. A.R.L. is supported by the Fondation Leducq Network of Excellence in Cardio-Oncology, and the Royal Brompton Cardio-Oncology Centre of Excellence is supported by The Big Heart Foundation.