Risk prediction models for cancer therapy related cardiac dysfunction in patients with cancer and cancer survivors: systematic review and meta-analysis.

Introduction

Cancer therapy related cardiac dysfunction (CTRCD) is a serious complication of cancer treatments. Although anthracyclines and the monoclonal antibody trastuzumab are well known for their association with this toxicity, 1 2 risk also extends to other cytostatic agents, immunotherapies, and molecular targeted therapies. 3 4 5 6 7 The spectrum of CTRCD encompasses a range of manifestations, which in the earliest stages may be reflected by increased levels of cardiac biomarkers (eg, troponin and brain natriuretic peptide (BNP)) or a reduction in global longitudinal strain, progressing to asymptomatic left ventricular systolic dysfunction and potentially culminating in symptomatic heart failure. 7 CTRCD can have profound consequences for patients with cancer and cancer survivors, limiting tolerability to cancer treatment and negatively impacting quality of life and survival. 8 Given the widespread use of antineoplastic agents associated with CTRCD, accurate prediction models are essential to enable targeted screening and monitoring, facilitating earlier diagnosis, timely treatment, and the initiation of preventive measures.

The first clinical cardio-oncology guideline released by the European Society of Cardiology (ESC) in 2022 recommends the use of the Heart Failure Association-International Cardio-Oncology Society (HFA-ICOS) risk assessment tool for patients scheduled to undergo potentially cardiotoxic antineoplastic treatments. 7 This tool was, however, based on expert opinion, 9 and its prognostic value and clinical utility across different settings remain unclear. Multiple other recommendations have been written, including the 2016 American Society of Clinical Oncology guideline for the prevention and monitoring of cardiac dysfunction in adults with cancer, 10 and the 2020 European Society of Medical Oncology consensus recommendations on the management of cardiac disease in patients with cancer. 11 However, these guidelines merely provide a list of patient and treatment related factors that should be taken into consideration to identify patients at increased risk for developing CTRCD, rather than a risk prediction model.

Although several CTRCD risk prediction models have been developed for patients with various malignancies, 12 13 14 15 none has been thoroughly evaluated. In this systematic review, we comprehensively evaluated existing CTRCD risk prediction models and related external validation studies and assessed the performance and risk of bias of the models to determine their reliability to guide personalised screening and monitoring of CTRCD in clinical practice.

Methods

This systematic review is reported according to the Transparent Reporting of multivariable prediction models for Individual Prognosis or Diagnosis for Systematic Reviews and Meta-Analyses (TRIPOD-SRMA) guideline 16 (see supplementary table 1) and the reporting guidelines of the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) (see supplementary table 2), including the extensions for abstracts and literature searches (PRISMA-S) (see supplementary tables 3 and 4). 17 18 The protocol is available online. The protocol underwent some amendments. At the time of protocol registration, we planned to include only models with at least three predictors. During the screening phase, we opted to track models with only two predictors and ultimately included them if they met all other eligibility criteria. Similarly, although the original protocol specified the exclusion of non-English studies, we reviewed these using online translation tools. Of the seven non-English articles identified, none met all inclusion criteria; therefore, no studies were excluded solely based on language.

Search strategy

We conducted a systematic literature search in Medline (Ovid), Embase (Ovid), and the Cochrane Central Register of Controlled Trials to identify relevant studies published from inception of each database to 21 September 2023. We repeated the search with the original search string on 23 August 2024. The search string was generated together with a medical librarian (AM), after establishing a comprehensive set of keywords and subject headings pertaining to cancer and CTRCD (see supplementary table 5). Search filters originally developed and validated to identify prognostic studies using the PubMed interface 19 20 21 were adapted for use within the Ovid interface. The search was restricted to studies in humans and specific publication types, excluding case reports, reviews, editorials, and letters. No restrictions were applied based on language or publication date. Duplicates were removed using an online tool developed in-house ( DedupEndNote ). To identify further relevant studies, we screened reference lists of selected studies, reviews on the topic, and clinical guidelines.

Study selection and eligibility criteria

Two researchers (CG and ML) manually and independently screened study titles and abstracts for eligibility, resolving any discrepancies through discussion. Full text articles were assessed for relevance, with disagreements resolved through consultation with a third researcher (TPAD). Non-English studies were translated using online translation software.

Studies were eligible if they met the following criteria: reported the development, validation, or update of one or multiple prediction models to estimate CTRCD risk in patients with cancer or cancer survivors; the outcome of interest encompassed any form of CTRCD as described in the ESC cardio-oncology guideline, including symptomatic heart failure, any new decline in global longitudinal strain or left ventricular ejection fraction, or new increase in cardiac biomarkers (troponins or BNP) 7 ; CTRCD was either the primary outcome or part of a composite cardiovascular outcome; the study population was primarily composed of patients treated with chemotherapy or targeted treatments, such as protein kinase inhibitors, monoclonal antibodies, or immunotherapy; the intended time of use of the model was either before the initiation of systemic anticancer treatment or in cancer survivors; and the prediction model included at least two predictors. External validation studies were included if they met the listed criteria, regardless of whether the prediction model was originally developed specifically for patients with cancer or for another population (eg, models developed to predict cardiovascular risk in individuals without cancer). Models based on expert opinion, rather than a formal development process, were also considered eligible if externally validated.

We excluded studies that primarily focused on radiation induced cardiotoxicity; evaluated the predictive or incremental value of individual predictors; had a sample size of <100 participants; developed models that relied on the collection of longitudinal data during cancer treatment (eg, changes in left ventricular ejection fraction from baseline) and therefore could not be applied at the initiation of cancer treatment; used a non-nested case-control or cross sectional design; or were solely published on preprint servers.

Data extraction and risk of bias assessment

We used the CHecklist for critical Appraisal and data extraction for systematic Reviews of prediction Modelling Studies (CHARMS) to extract data from selected studies, 22 supplemented by items extracted during a previous systematic review on cardiovascular risk prediction models. 23

For studies describing the development of a prognostic model, we extracted data on study design, study population and sample size, outcome definitions, prediction horizon, duration of follow-up, number and type of predictors incorporated into the model, number of outcome events, modelling method, method of internal validation, performance measures (eg, C statistic, calibration, or overall performance), and model presentation. For data extraction of external validation studies, we utilised a different extraction sheet, including information on the type of external validation performed, whether the validation was conducted by the same investigators who originally developed the model, if the model was adjusted or updated, and performance measures before and after any adjustments or updates. When a study reported validations in multiple independent cohorts, we counted each as a separate external validation. When both the C statistic and the area under the receiver operating characteristic curve were reported, we extracted the C statistic—throughout this study both are referred to as C statistic for consistency.

Outcome definitions were extracted as reported in each study. For each outcome of interest, we collected all available performance metrics. For instance, if a study reported a composite cardiovascular outcome and heart failure separately, we extracted performance measures on both. Similarly, if performance was assessed at multiple time points, all were recorded. If multiple models were developed within a single study, we extracted data for all the models unless the authors explicitly identified one as the model intended for clinical use. In studies that compared multiple machine learning techniques to identify the best performing approach, we only extracted data from the model with the best performance. Two researchers (CG and JG) independently extracted data. Any disagreements were resolved through discussion and by consulting a third researcher (ML).

The Prediction model Risk Of Bias ASsessment Tool (PROBAST) was used to evaluate the risk of bias of retrieved studies, 24 according to the four key domains: participants, predictors, outcome, and analysis. Two researchers (CG and JG) independently conducted the assessments in duplicate, utilising a standardised quality assessment spreadsheet. Each potential source of bias was graded as low, high, or unclear, with a justification provided for each judgement. We evaluated the applicability of studies in relation to the target population defined in the research question of this systematic review. Any disagreements were resolved through discussion, with input from a third researcher (ML) when necessary. We did not assess small study effects or publication bias using funnel plots or statistical tests, as only a few models were validated more than once, and traditional methods are underpowered with the inclusion of fewer than 10 studies.

Data synthesis and statistical analysis

For descriptive purposes, we categorised the included studies based on the age group of the population in which the model was developed or validated: children together with adolescents and young adults (≤39 years) versus older adults (≥40 years). Predictors were combined into related categories (eg, body mass index grouped together with obesity), as displayed in supplementary table 6, and plotted to provide an overview of the most included predictors across all developed models. Likewise, outcome definitions were grouped as either CTRCD, including heart failure, decline in left ventricular ejection fraction or global longitudinal strain, or increase in cardiac biomarkers; or composite cardiovascular outcomes, including other cardiovascular events beyond CTRCD (see supplementary table 7).

Continuous variables were reported as median and interquartile range (IQR) or mean and standard deviation, depending on data distribution. Categorical variables were summarised using absolute counts and percentages. Given the infrequent reporting of calibration measures, we compared the observed outcome proportions from external validation studies of the most validated risk assessment tool (HFA-ICOS) with the expected risk for each risk stratum according to this tool. Confidence intervals (CIs) for proportions were estimated using Wilson’s method. 25 Potential sources of heterogeneity in the performance of this tool were assessed by categorising studies based on the severity of the most frequently reported CTRCD events, as defined by the 2022 ESC cardio-oncology guidelines 7 : studies predominantly reporting mild CTRCD, and studies in which moderate to severe (or symptomatic) CTRCD predominated (see supplementary table 8). From this analysis we excluded those studies with a composite endpoint including other cardiovascular events beyond CTRCD.

Meta-analyses were conducted to assess the performance of the HFA-ICOS tool, with a focus on the C statistic and observed outcome proportions across different risk strata. A pooled C statistic represents a weighted average of this measure across multiple external validation studies, providing an overall estimate of the model’s discrimination across different studies. Observed outcome proportions were also pooled for each risk stratum and compared with expected risks, offering an indirect assessment of calibration. Before pooling, both measures were transformed to the logit scale. For the logit transformed C statistic, we calculated standard errors from reported CIs, when available. If CIs were not reported, we derived standard errors from the reported sample sizes and observed number of events, following established methods. 26 27 28

Random effects meta-analyses were performed using the restricted maximum likelihood estimator for between study variance and the Hartung-Knapp-Sidik-Jonkman adjustment, which accounts for heterogeneity when estimating pooled performance measures and CIs. 28 29 As a sensitivity analysis, we excluded HFA-ICOS external validation studies with high applicability concerns to our research question (according to PROBAST criteria) and pooled C statistics and observed outcome proportions only from studies with low concerns to assess the robustness of our findings. All analyses were conducted in R (version 4.3.2), utilising the metafor, 30 metamisc, 31 and meta 32 packages.

Patient and public involvement

No patients or members of the public were involved in this study, as it was based solely on secondary analysis of data from previously published studies.

Results

Out of 10 935 articles screened by title and abstract, 546 were selected for full text assessment. Among the 546 studies, 42 involved multivariable modelling but were excluded after discussion for not meeting inclusion criteria, with most classified as prognostic factor studies (see supplementary table 9). Ultimately, 54 studies were included, supplemented by two additional studies identified through snowballing—which were not found in our initial search, bringing the total to 56 studies. Among these, 29 studies developed one or more prediction models, 20 conducted external validation, and seven performed both ( fig 1 ). Supplementary table 10 presents a detailed overview of each study included in this systematic review.

The number of studies developing or validating CTRCD prediction models has grown substantially in recent years, with a particular increase in those focusing on external validation ( fig 2 ). In total, 63 prediction models were identified: 51 were developed to predict CTRCD risk in patients with cancer or cancer survivors, nine were originally developed for patients without cancer (n=8) or to predict a non-cardiovascular outcome (n=1), and three were derived from expert opinion rather than a formal model development process. Supplementary table 11 provides a detailed summary of the models that were identified through external validation studies, but for which the original development study did not fulfil the inclusion criteria of this systematic review.

Studies on CTRCD risk prediction model development

Study design and participants

A total of 51 prediction models were identified, of which 67% (34/51) were developed in data derived from adults. Most models originated from cohort studies (45/51, 88%), primarily conducted in North America and Europe. Table 1 shows the key characteristics of the studies.

Among the models developed in adults, breast cancer was the most common malignancy (20/34, 59%), followed by haematological malignancies (6/34, 18%). In contrast, for models developed in children, adolescents, and young adults, the study population often comprised cancer survivors (16/17, 94%) with a diverse range of cancers (14/17, 82%), including haematological malignancies and embryonal tumours ( table 1 ). In paediatric populations, a few well established cohorts were leveraged to develop multiple prediction models. The Childhood Cancer Survivor Study cohort served as the basis for seven CTRCD prediction models developed by two independent research groups. Four of these models were tailored to different age categories, 33 while the remaining three varied in their approach to account for use of anthracycline and radiation (see supplementary table 10). 34

Although prediction horizons were seldom reported, follow-up periods were generally longer in children, adolescents, and young adults (>5 years’ duration), reflecting the intended use of the model in cancer survivors. In contrast, models in adults were typically developed to be applied before cancer treatment, and patients were followed-up during shorter periods (<5 years) ( table 1 ). Anthracycline based chemotherapy was the mainstay of treatment in 44% (15/34) of models developed in adults and in 47% (8/17) of those developed in children, adolescents, and young adults.

Outcomes and modelling methods

Among children, adolescents, and young adults, all models were developed to predict heart failure and other CTRCD surrogates, such as a new increase in cardiac biomarkers and/or a new decline in global longitudinal strain or left ventricular ejection fraction. In adults, a substantial proportion of models (14/34, 41%) were developed to predict a composite outcome, including other cardiovascular events besides CTRCD (see supplementary table 7). The outcome was assessed as a binary variable in 56% (19/34) of models developed in adults and in 41% (7/17) of models developed in children, adolescents, and young adults, without accounting for censoring or competing risks such as cancer related mortality. Logistic regression and survival analysis were used in the development of 78% (40/51) of models ( table 1 ).

Predictors

Developed models included a median of 7 (IQR 5-11) predictors. In children, adolescents, and young adults, 65% (11/17) of developed models included age, sex, anthracycline use and/or dose, and radiation therapy ( fig 3 ). Among models developed for adults, the five most included predictors were age (25/34, 74%), hypertension (21/34, 62%), diabetes (16/34, 47%), use of anthracyclines (15/34, 44%), and chest radiation (13/34, 38%). Supplementary table 12 provides an overview of the main predictors included in each model, grouped into broad domains (eg, demographics, comorbidities, cancer treatment). To illustrate how model complexity relates to discrimination, the corresponding C statistics are also provided.

Models published in cardiology journals (19/51, 37%) often incorporated cardiovascular comorbidities, cardiac imaging variables (eg, baseline left ventricular ejection fraction), and cardiac serum biomarkers, in contrast with models published in oncology journals (23/51, 45%). Figure 3 shows the most included predictors in models published in oncology versus cardiology journals.

Internal validation and reporting of performance measures

A total of 43 models (43/51, 84%) underwent internal validation, with the most employed internal validation methods listed in table 1 and specified for each model in supplementary table 13. The C statistic was the most frequently reported performance metric, available for 94% (48/51) of models. Calibration measures were reported for 27% (14/51) of the developed models, while both calibration and discrimination measures were reported for 25% (13/51) of models ( table 2 ). Supplementary table 13 outlines all reported performance measures.

Model presentation

Models were presented as a sum score (n=21), online calculator (n=9), full regression formula, including intercept and coefficients (n=5), nomogram (n=3), and risk score table (n=1). Additionally, three models presented predictor coefficients but lacked an intercept, while 17 out of 51 models (33%) presented neither regression coefficients nor intercepts (see supplementary table 10).

Studies on external validation of a CTRCD risk prediction model

Twenty seven studies reported 44 external validations of 24 models. Among these, 12 models were either originally developed in populations without cancer (n=8), aimed at predicting non-cardiovascular outcomes (eg, cancer mortality (n=1)), or based on expert opinion without formal development methodology (n=3). Notably, 13 external validations were conducted either in the same study as the model was developed or by the same research group, while seven validations (spanning five models) were performed independently by separate research groups.

Among adults, the HFA-ICOS risk assessment tool was the most frequently validated, with 11 external validations. However, the tool is composed of six different models depending on the type of treatment administered (anthracyclines, HER2 (human epidermal growth factor receptor 2) targeted therapies, vascular endothelial growth factor (VEGF) inhibitors, BCR-ABL (Breakpoint cluster region Abelson murine leukaemia viral oncogene homolog 1) inhibitors, multiple myeloma therapies, and RAF (rapidly accelerated fibrosarcoma) and MEK (mitogen activated protein kinase) inhibitors). Most external validations (n=5) were conducted in patients with breast cancer receiving HER2 targeted therapies, followed by patients with breast cancer receiving anthracyclines (n=3). Table 3 outlines the main characteristics of the HFA-ICOS external validation studies. Among the other models externally validated in adults, one model was validated three times, four were validated twice, and the remaining 16 were validated once (see supplementary table 13). In children, adolescents, and young adults, two models were validated three times and two were validated once, all among survivors of childhood cancer for ≥5 years.

Overall, six external validations (6/44, 14%) reported both calibration and discrimination measures, while seven (7/44, 16%) reported no performance metrics ( table 2 ). Studies that lacked performance measures primarily described the distribution of patients across risk strata and how these categorisations aligned with observed outcomes. Supplementary figure 1 shows the distribution of C statistics across development datasets (with or without internal validation) and external validation samples, and supplementary table 13 provides further details.

Quality assessment

Supplementary tables 14 and 15 provide the assessment of risk of bias and applicability for each model’s development and external validation for each PROBAST domain and for each signalling question, respectively. The overall risk of bias was deemed high for all developed models (n=51) and for 98% (43/44) of external validations, primarily due to concerns in the analysis domain ( fig 4 ). Key factors included inadequate reporting of performance measures (75/95, 79%), poor handling of missing data (72/95, 76%), failure to account for censoring and competing risks (61/95, 64%), and lack of strategies to prevent overfitting (31/51, 61%). Other concerns were related to the handling of categorical and continuous predictors (54/95, 57%), selection of predictors based on univariable analysis (23/51, 45%), and small sample sizes (29/95, 31%). Additionally, 33% (17/51) of developed models and 59% (26/44) of external validations raised concerns about applicability, predominantly in the participants and outcome domains.

Quantitative analysis

HFA-ICOS risk assessment tool

Figure 5 shows the observed event proportions by HFA-ICOS risk category across all validation studies, compared with the expected risk ranges. 9 In most studies the incidence of CTRCD and other cardiovascular events in patients classified as low or medium risk exceeded the 10% threshold typically associated with high risk, indicating that this tool underestimates the actual risk. Similarly, among high risk patients, event incidences exceeded 20%, a level more consistent with very high risk classifications according to this tool. Notably, when studies focusing on CTRCD related events were categorised based on the severity of the predominantly reported events (ie, mild versus moderate to severe CTRCD) (see supplementary table 8), the observed outcome proportions more closely aligned with the expected risks in the studies focusing on more severe events (see supplementary figure 2).

The meta-analysis was restricted to studies validating the model in patients treated with HER2 targeted therapies (n=5), given the limited number of external validation studies for the models of other classes of drugs (ie, anthracyclines, VEGF inhibitors and BCR-ABL inhibitors) ( table 3 ). The pooled observed risks were 12% (low risk group), 15% (medium risk), 25% (high risk), and 41% (very high risk), exceeding the expected risks by <2%, 2-9%, 10-19%, and ≥20%, respectively (see supplementary figure 3). 9 Importantly, three of these studies reported mild CTRCD as the predominant outcome. The pooled C statistic, based on the four studies that reported this metric, was 0.60 (95% CI 0.52 to 0.68) ( fig 6 ) with moderate heterogeneity (I 2 =57.88% (95% CI 0.00 to 97.40), τ 2 =0.0014, H 2 =2.37). Results of a sensitivity analysis limited to studies with low concerns about applicability were similar (see supplementary figures 4 and 5).

Discussion

This systematic review provides a comprehensive summary and quantitative analysis of CTRCD risk prediction models developed for adults and for young people (children, adolescents, and young adults) with cancer and cancer survivors treated with systemic antineoplastic agents. The key findings were: most studies were deemed at high risk of bias; most prediction models were developed in data derived from patients with breast cancer, and other cancer populations were underrepresented in both model development and validation studies; the limited number of external validation studies, combined with substantial heterogeneity between studies and poor reporting of performance measures, hindered thorough assessment of the performance of existing prediction models; the HFA-ICOS risk assessment tool was the most validated, particularly in patients with breast cancer treated with HER2 targeted therapies, but its discriminative performance was suboptimal in this population, with a pooled C statistic of 0.60 (95% CI 0.52 to 0.68); and overall the HFA-ICOS risk assessment tool tended to systematically underestimate the risk of CTRCD, particularly in studies where mild CTRCD events comprised the majority of reported outcomes.

Although a substantial number of CTRCD risk prediction models exist, evidence supporting their clinical use remains limited. We assessed all developed models and 98% (43/44) of external validations as being at high risk of bias, primarily due to methodological concerns within the analysis domain (see supplementary table 14). These findings align with previous systematic reviews of risk prediction models, which have highlighted pervasive methodological shortcomings and poor reporting standards across various clinical domains. 45 46 47

Despite the growing body of research on CTRCD risk stratification, many cancer populations remain underrepresented in both model development and validation. No studies have been conducted in Africa, Australia, or central or South America, and most studies in adult populations focus on breast cancer, likely due to the high incidence and relatively favourable prognosis. However, emerging evidence suggests that, for example, patients with haematological malignancies face a substantially higher cardiovascular risk, particularly heart failure, compared with many other types of cancer, in both the short and the long term. 48 49 50 Historically, the poor survival in these patients may have diverted attention from cardiovascular outcomes, but recent improvements in outcomes underscore the need to include these patient populations in CTRCD risk prediction research. 51 52 In total we identified nine models that were developed and/or validated in haematological malignancies, of which five have been externally validated showing moderate performance (see supplementary table 13). Patients who have undergone a haematopoietic stem cell transplantation have been most studied, with all three available models externally validated. 53 The CARE-BMT (Cardiovascular Registry in Bone Marrow Transplantation) showed the best discrimination (C statistic 0.77, 95% CI 0.72 to 0.83), but no information was provided on calibration. Further independent validations are necessary before this model can be considered the reference tool in this setting.

Among children, adolescents, and young adults, 16 out of 17 (94%) CTRCD prediction models were developed specifically for cancer survivors, of which four underwent external validation (see supplementary table 13). The models by Chow et al, based on easily available predictors (sex, age, and cancer treatment), showed variable performance across validation cohorts (C statistics 0.63 to 0.82; supplementary figure 1). 34 In contrast, the tool developed by Leerink et al, externally validated once, showed good discrimination (C statistic 0.86, 95% CI 0.83 to 0.89) and calibration. 54 Despite this model’s simplicity—comprising just three predictors—it includes baseline left ventricular ejection fraction, which may limit its applicability in some settings. Without further evidence, it is not possible to determine which of these tools performs best, as they were evaluated in studies adopting different outcome definitions and prediction horizons. This may explain why the updated recommendations for cardiomyopathy surveillance in children, adolescents, and young adult cancer survivors from the International Late Effects of Childhood Cancer Guideline Harmonization Group do not include any existing models. 55 Instead, they proposed a risk stratification tool based on cumulative anthracycline and chest radiation doses, which, to the best of our knowledge, lacks formal methodological development and external validation.

Overall, only 12 (24%) of the 51 identified models developed to predict CTRCD have undergone external validation. Of these, seven were validated only once (see supplementary table 13), often within the original study or by the same research group. This raises major concerns about the generalisability of these models to different settings and populations, as well as the potential presence of publication bias and selective reporting. 56 57 Among the models validated more than once, unclear prediction horizons, heterogeneity in outcome definitions, and performance metric reporting further hindered robust evaluation of their accuracy.

The HFA-ICOS risk assessment tool, endorsed by the ESC, 7 is currently the most extensively externally validated tool for risk stratification of cardiovascular toxicity in adults with cancer. Despite its prominence, our systematic review indicates that evidence supporting the clinical use of HFA-ICOS is insufficient. The tool is proposed to estimate the risk of cardiovascular toxicity for six distinct classes of cancer drugs, and as such it comprises six different models. Of the 11 external validation studies identified, most focused on the HFA-ICOS score for HER2 targeted therapies (n=5), followed by anthracyclines (n=3). Notably, no external validation studies were found for multiple myeloma therapies, and validation for patients treated with BCR-ABL or VEGF inhibitors was each limited to a single study ( table 3 ). Additionally, while one study evaluated the tool in patients with melanoma treated with BRAF and MEK inhibitors, we excluded it from this review owing to small sample size (n=63). 58

Across all external validation studies of the HFA-ICOS risk assessment tool (ie, regardless of class of antineoplastic drug), we identified an important mismatch between observed outcomes and expected risks. The tool consistently underestimated the risk of CTRCD and other cardiovascular events, with observed risks exceeding predictions across all risk strata, except the very high risk category ( fig 5 ). This underestimation was most pronounced in studies where mild CTRCD comprised most events (see supplementary figure 2A), suggesting that the HFA-ICOS tool may not be suitable to predict milder forms of CTRCD, unless calibrated for that purpose. Furthermore, it is important to consider that the clinical significance of mild CTRCD has been questioned in recent studies conducted in patients with breast cancer treated with anthracyclines 59 and trastuzumab, 60 showing a low progression rate from mild to more severe forms of CTRCD during follow-up. Considering these findings, prospective cohort studies with sufficient follow-up are crucially needed to clarify the prognostic importance of mild CTRCD.

In studies focusing on moderate to severe or symptomatic CTRCD, the HFA-ICOS tool appears to perform better (see supplementary figure 2B). The largest validation to date included more than a thousand patients (mainly with breast cancer) undergoing treatment with anthracyclines, and reported a C statistic of 0.78 (95% CI 0.71 to 0.84) with good calibration to predict moderate or severe CTRCD. 41 Interestingly, although mild CTRCD was not included in the outcome, 61% of patients fulfilled the definition of this complication, with high incidences across all risk strata (62% in low risk patients, 63% in moderate, 52% in high, and 30% in very high), again illustrating the tool’s poor calibration for such events. No other external validation studies have reported performance measures for the HFA-ICOS tool in anthracycline treated populations, and only a few have assessed its ability to predict more severe outcomes (see supplementary table 8).

Besides the questionable prognostic accuracy of the HFA-ICOS tool, its complexity raises concerns about practicality and cost effectiveness. Depending on the type of cancer systemic treatment, the tool incorporates 16 to 22 predictors—including demographic variables, cardiovascular risk factors, comorbidities, electrocardiographic (ECG) and cardiac imaging variables, serum biomarkers, and use of cancer treatments—some of which may be difficult to obtain, especially in resource limited settings. While the tool may be suitable for use by specialised cardio-oncology clinics, the comprehensive cardiovascular evaluation it requires limits its feasibility in routine oncology practice. Although the ESC guidelines propose a stepwise approach, starting cardiovascular toxicity risk assessment with easily available predictors (such as age, cardiovascular risk factors, and cancer treatment), followed by selective testing of cardiac biomarkers and echocardiography in patients with pre-existing cardiovascular disease or patients deemed at risk of CTRCD, this approach is not further substantiated. 7

Simpler models may provide a practical alternative. For example, one study used an ECG model based on four variables, originally developed to predict sudden death in patients with coronary artery disease to predict CTRCD. 61 The model showed good discrimination (C statistic 0.84, 95% CI 0.77 to 0.92), performing better than many more complex models included in this review (see supplementary figure 1). Similarly, the AI-ECG model, a deep learning model originally trained to detect left ventricular systolic dysfunction from ECG images, 62 showed promising results in predicting left ventricular ejection fraction <40% in a cohort of patients with breast cancer or non-Hodgkin lymphoma (C statistic 0.82, 95% CI 0.74 to 0.89). 63 These studies demonstrate the potential of simple and widely available tools, such as ECGs, to offer practical alternatives to more complex and resource demanding models, especially before referral of patients to specialised cardio-oncology care. Nonetheless, without head-to-head comparisons using the same population and outcomes, it remains unclear which models offer the optimal balance between feasibility and predictive performance.

Strengths and limitations of this review

This study comprehensively summarised existing CTRCD risk prediction models for patients with cancer and cancer survivors undergoing systemic antineoplastic treatment. A previous systematic review focused on cardiotoxicity prediction models in patients with breast cancer 64 identified comparable methodological limitations, including studies with a high risk of bias, and highlighted a widespread lack of external validation. Our review provides a broader and more rigorous synthesis of models developed and validated across diverse cancer populations and systemic treatment regimens. Additionally, our study offers a detailed assessment of the predictive performance of the HFA-ICOS risk assessment tool, including a meta-analysis of its performance in patients with breast cancer receiving HER2 targeted therapies, further advancing the understanding of its prognostic value.

This systematic review also has some limitations. Heterogeneity was present among the studies included in the meta-analyses, particularly in the definition of CTRCD, the use of composite outcomes incorporating other cardiovascular events, and varying follow-up durations. This variability likely reflects the lack of guidance on the intended use of the HFA-ICOS tool, including the outcomes targeted by the models and the intended prediction horizon, which may influence the pooled estimates. 7 9 Furthermore, by focusing exclusively on systemic cancer therapies, we excluded studies on radiation induced cardiotoxicity, which may have reduced the representation of certain malignancies such as lung cancer. Likewise, malignancies treated with hormonal therapy, such as prostate cancer, were excluded, limiting the applicability of our findings to these populations.

Implications and recommendations for future research

Accurate prediction of CTRCD risk carries important clinical implications. Early identification of high risk individuals before the initiation of cancer treatment creates a window of opportunity for the initiation of cardioprotective strategies, can aid in the selection or modification of cancer therapies to regimens associated with less cardiotoxicity, 65 and facilitate closer cardiac monitoring during and after treatment. 66 Evidence suggests that such proactive approaches can reduce the incidence and severity of cardiotoxicity, potentially improving long term cardiovascular outcomes. 67 68 69 70 71 72 Importantly, since referral of all patients to specialised cardio-oncology clinics is not feasible, oncologists have a central role in recognising and assessing cardiovascular risk, ensuring referral to specialised care, when appropriate. Primary care doctors, as key members of long term patient care, play a complementary role and should be alert to the heightened cardiovascular risk in patients with cancer and cancer survivors, in both the short and the long term, to enable timely screening and early diagnosis of cardiovascular disease in this population. This underscores the need for simple, accessible risk stratification tools.

The findings from our study offer important insights for clinical practice and future research. While the HFA-ICOS risk assessment tool is currently recommended for clinical use by the ESC, its limited discriminative ability and poor calibration present challenges for effective decision making. Until further evidence substantiates its performance, caution is warranted when applying this model in routine practice. High quality prospective studies with adequate sample sizes are needed to validate and recalibrate the models included in the HFA-ICOS tool. These studies should ideally use standardised outcome definitions, define prediction horizons tailored to specific malignancies and cancer treatments, and appropriately weight predictors based on their relative contribution to CTRCD risk. Moreover, after these improvements, implementation studies are crucial to evaluate whether the use of this tool improves clinical outcomes, decision making, and cost effectiveness in real world settings.

Certain subgroups of patients with cancer remain substantially underrepresented in the development and validation of CTRCD prediction models, including children, adolescents, and young adults with active cancer, adults with malignancies other than breast cancer, and populations from non-Western regions. While conducting external validation studies across all cancers and cardiotoxic treatments may be infeasible, leveraging large, linked, or pooled datasets from different sources can enable direct comparisons of multiple models simultaneously, 73 74 and assessment of model transportability and generalisability across different populations and clinical settings. This approach can ultimately optimise research efforts and provide a solid foundation for model updating or recalibration tailored to specific patient groups. Moreover, individual participant data meta-analyses with internal-external cross validation are emerging as robust strategies for both developing and validating prediction models while simultaneously assessing their generalisability. 73 75 The growing availability of cardio-oncology registries makes such approaches increasingly feasible. 76 77 78 However, it is of importance that a diverse case mix is included in such registries, encompassing patients of all risk levels rather than focusing solely on groups of patients who are perceived as high risk based on current clinical knowledge.

Lastly, the integration of machine learning approaches offers a promising avenue to enhance CTRCD risk prediction, as these methods have the potential to capture complex relations among many clinical, biochemical, and imaging variables, particularly relevant in oncology patients receiving multiple treatment regimens. However, despite their potential, machine learning models involve important challenges. Their complexity and occasional lack of interpretability can hinder both external validation efforts and clinical uptake. Ensuring transparency, prospective validation, and alignment with clinical workflows will be key steps toward their successful implementation. 79 Additionally, while these models are often developed by engineers and data scientists, their utility in healthcare will be enhanced when developed in close collaboration with clinicians who can help define outcomes rigorously, understand the underlying biases of the data used, and ensure that clinical reasoning remains central to the decision making process.

Conclusion

Over the past decades, the prognosis of many malignancies has steadily improved, and cancer is increasingly regarded as a chronic disease. However, the cardiovascular toxicity associated with many antineoplastic treatments is a growing concern. This systematic review provides a comprehensive overview of available CTRCD risk prediction models. Despite recent efforts to improve CTRCD risk stratification in patients with cancer and cancer survivors, no prediction model has consistently shown reliable performance. The HFA-ICOS risk assessment tool currently proposed by the ESC guidelines, 7 9 is the most extensively validated tool, but we conclude that its prognostic accuracy is suboptimal. To advance the discipline, using standardised outcome definitions and prediction horizons is essential to ensure consistency in external validation studies. Additionally, more rigorous research is needed to validate and refine existing models across various cancer types and to evaluate their impact when implemented in routine clinical practice.

What is already known on this topic

Cancer therapy related cardiac dysfunction (CTRCD) is a major complication of cancer treatment, driven by both patient specific and treatment related factors

Given the high global burden of cancer and the widespread use of antineoplastic agents associated with CTRCD, effective risk stratification models are urgently needed

It remains uncertain which CTRCD risk prediction models show the strongest predictive performance and are best suited for clinical application

What this study adds

Almost all existing CTRCD risk prediction models show a high risk of bias, raising concerns about the reliability of their predictions when applied in routine clinical practice

This study identified underrepresentation of data from patients with malignancies other than breast cancer and a scarcity of externally validated models

The findings of this review highlight the need for greater methodological rigour, standardised outcome definitions and prediction horizons, and external validation of models across diverse cancer types

Extra material supplied by authors

Supplementary information: Tables 1-15 and figures 1-5

Supplementary information: Additional data to share

Contributors: CG and ML conceived the study and developed the protocol, and together with AM, they defined the query for the electronic database search. CG and ML screened studies for eligibility, consulting TPAD in cases of disagreement. CG and JG extracted the data. CG analysed the data. CG and ML wrote the first draft of the manuscript, which was thereafter critically revised by all authors. All authors approved the final version of the submitted manuscript. CG and ML are the guarantors, had full access to the study data, and are responsible for its integrity and the accuracy of the analysis. The corresponding author attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.

Funding: CG was granted a PhD scholarship (reference 2022.11384.BD) from the Foundation for Science and Technology in Portugal. ML is supported by the Dutch Cancer Society (project No 14657) and Dutch Heart Foundation (02-001-2022-0126 Clinical Trials 2022). FWA is supported by University College London Hospitals National Institute for Health and Care Research Biomedical Research Centre and EU Horizon (AI4HF 101080430). None of the funding sources had a role in the design, data collection, analysis, interpretation of results, or decision to publish this work.

Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/disclosure-of-interest/ and declare: no support from any organisation for the submitted work; no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work.

Transparency: The lead author affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned have been explained.

Dissemination to participants and related patient and public communities: The results of this study will be disseminated through presentations at national and international conferences. Results will also be shared through social media channels to reach a broader audience.

Provenance and peer review: Not commissioned; externally peer reviewed.

Ethics statements

Ethical approval

Not required.

Data availability statement

Additional supporting data are available in the supplementary file.

References

1. Swain SM, Whaley FS, Ewer MS. Congestive heart failure in patients treated with doxorubicin: a retrospective analysis of three trials. Cancer 2003;97:2869-79. 10.1002/cncr.11407.

2. Thavendiranathan P, Abdel-Qadir H, Fischer HD, et al. Breast Cancer Therapy-Related Cardiac Dysfunction in Adult Women Treated in Routine Clinical Practice: A Population-Based Cohort Study. J Clin Oncol 2016;34:2239-46. 10.1200/JCO.2015.65.1505.

3. Pai VB, Nahata MC. Cardiotoxicity of chemotherapeutic agents: incidence, treatment and prevention. Drug Saf 2000;22:263-302. 10.2165/00002018-200022040-00002.

4. Ewer MS, Ewer SM. Cardiotoxicity of anticancer treatments. Nat Rev Cardiol 2015;12:547-58. 10.1038/nrcardio.2015.65.

5. Kamphuis JAM, Linschoten M, Cramer MJ, et al. Cancer Therapy-Related Cardiac Dysfunction of Nonanthracycline Chemotherapeutics: What Is the Evidence? JACC CardioOncol 2019;1:280-90. 10.1016/j.jaccao.2019.09.007.

6. Stein-Merlob AF, Rothberg MV, Ribas A, Yang EH. Cardiotoxicities of novel cancer immunotherapies. Heart 2021;107:1694-703. 10.1136/heartjnl-2020-318083.

7. Lyon AR, López-Fernández T, Couch LS, et al. ESC Scientific Document Group . 2022 ESC Guidelines on cardio-oncology developed in collaboration with the European Hematology Association (EHA), the European Society for Therapeutic Radiology and Oncology (ESTRO) and the International Cardio-Oncology Society (IC-OS). Eur Heart J 2022;43:4229-361. 10.1093/eurheartj/ehac244.

8. Cox CL, Rai SN, Rosenthal D, Phipps S, Hudson MM. Subclinical late cardiac toxicity in childhood cancer survivors: impact on self-reported health. Cancer 2008;112:1835-44. 10.1002/cncr.23378.

9. Lyon AR, Dent S, Stanway S, et al. Baseline cardiovascular risk assessment in cancer patients scheduled to receive cardiotoxic cancer therapies: a position statement and new risk assessment tools from the Cardio-Oncology Study Group of the Heart Failure Association of the European Society of Cardiology in collaboration with the International Cardio-Oncology Society. Eur J Heart Fail 2020;22:1945-60. 10.1002/ejhf.1920.

10. Armenian SH, Lacchetti C, Barac A, et al. Prevention and Monitoring of Cardiac Dysfunction in Survivors of Adult Cancers: American Society of Clinical Oncology Clinical Practice Guideline. J Clin Oncol 2017;35:893-911. 10.1200/JCO.2016.70.5400.

11. Curigliano G, Lenihan D, Fradley M, et al. ESMO Guidelines Committee Management of cardiac disease in cancer patients throughout oncological treatment: ESMO consensus recommendations. Ann Oncol 2020;31:171-90. 10.1016/j.annonc.2019.10.023.

12. Ezaz G, Long JB, Gross CP, Chen J. Risk prediction model for heart failure and cardiomyopathy after adjuvant trastuzumab therapy for breast cancer. J Am Heart Assoc 2014;3:e000472. 10.1161/JAHA.113.000472.

13. Abdel-Qadir H, Thavendiranathan P, Austin PC, et al. Development and validation of a multivariable prediction model for major adverse cardiovascular events after early stage breast cancer: a population-based cohort study. Eur Heart J 2019;40:3913-20. 10.1093/eurheartj/ehz460.

14. Kang Y, Assuncao BL, Denduluri S, et al. Symptomatic Heart Failure in Acute Leukemia Patients Treated With Anthracyclines. JACC CardioOncol 2019;1:208-17. 10.1016/j.jaccao.2019.10.008.

15. de Vries S, Haaksma ML, Jóźwiak K, et al. Development and Validation of Risk Prediction Models for Coronary Heart Disease and Heart Failure After Treatment for Hodgkin Lymphoma. J Clin Oncol 2023;41:86-95. 10.1200/JCO.21.02613.

16. Snell KIE, Levis B, Damen JAA, et al. Transparent reporting of multivariable prediction models for individual prognosis or diagnosis: checklist for systematic reviews and meta-analyses (TRIPOD-SRMA). BMJ 2023;381:e073538. 10.1136/bmj-2022-073538.

17. Rethlefsen ML, Kirtley S, Waffenschmidt S, et al. PRISMA-S Group . PRISMA-S: an extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews. Syst Rev 2021;10:39. 10.1186/s13643-020-01542-z.

18. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021;372:n71. 10.1136/bmj.n71.

19. Ingui BJ, Rogers MA. Searching for clinical prediction rules in MEDLINE. J Am Med Inform Assoc 2001;8:391-7. 10.1136/jamia.2001.0080391.

20. Haynes RB, McKibbon KA, Wilczynski NL, Walter SD, Werre SR, Hedges Team . Optimal search strategies for retrieving scientifically strong studies of treatment from Medline: analytical survey. BMJ 2005;330:1179. 10.1136/bmj.38446.498542.8F.

21. Geersing GJ, Bouwmeester W, Zuithoff P, Spijker R, Leeflang M, Moons KG. Search filters for finding prognostic and diagnostic prediction studies in Medline to enhance systematic reviews. PLoS One 2012;7:e32844. 10.1371/journal.pone.0032844.

22. Moons KG, de Groot JA, Bouwmeester W, et al. Critical appraisal and data extraction for systematic reviews of prediction modelling studies: the CHARMS checklist. PLoS Med 2014;11:e1001744. 10.1371/journal.pmed.1001744.

23. Damen JAAG, Hooft L, Schuit E, et al. Prediction models for cardiovascular disease risk in the general population: systematic review. BMJ 2016;353:i2416. 10.1136/bmj.i2416.

24. Wolff RF, Moons KGM, Riley RD, et al. PROBAST Group† . PROBAST: A Tool to Assess the Risk of Bias and Applicability of Prediction Model Studies. Ann Intern Med 2019;170:51-8. 10.7326/M18-1376.

25. Wilson EB. Probable Inference, the Law of Succession, and Statistical Inference. J Am Stat Assoc 1927;22:209-12. 10.1080/01621459.1927.10502953.

26. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982;143:29-36. 10.1148/radiology.143.1.7063747.

27. Newcombe RG. Confidence intervals for an effect size measure based on the Mann-Whitney statistic. Part 2: asymptotic methods and evaluation. Stat Med 2006;25:559-73. 10.1002/sim.2324.

28. Debray TPA, Damen JAAG, Snell KIE, et al. A guide to systematic review and meta-analysis of prediction model performance. BMJ 2017;356:i6460. 10.1136/bmj.i6460.

29. Röver C, Knapp G, Friede T. Hartung-Knapp-Sidik-Jonkman approach and its modification for random-effects meta-analysis with few studies. BMC Med Res Methodol 2015;15:99. 10.1186/s12874-015-0091-1.

30. Viechtbauer W. Conducting Meta-Analyses in R with The metafor Package. J Stat Softw 2010;36:1-48. 10.18637/jss.v036.i03.

31. Debray TP, Damen JA, Riley RD, et al. A framework for meta-analysis of prediction model studies with binary and time-to-event outcomes. Stat Methods Med Res 2019;28:2768-86. 10.1177/0962280218785504.

32. Balduzzi S, Rücker G, Schwarzer G. How to perform a meta-analysis with R: a practical tutorial. Evid Based Ment Health 2019;22:153-60. 10.1136/ebmental-2019-300117.

33. Chen Y, Chow EJ, Oeffinger KC, et al. Traditional Cardiovascular Risk Factors and Individual Prediction of Cardiovascular Events in Childhood Cancer Survivors. J Natl Cancer Inst 2020;112:256-65. 10.1093/jnci/djz108.

34. Chow EJ, Chen Y, Kremer LC, et al. Individual prediction of heart failure among childhood cancer survivors. J Clin Oncol 2015;33:394-402. 10.1200/JCO.2014.56.1373.

35. Battisti NML, Andres MS, Lee KA, et al. Incidence of cardiotoxicity and validation of the Heart Failure Association-International Cardio-Oncology Society risk stratification tool in patients treated with trastuzumab for HER2-positive early breast cancer. Breast Cancer Res Treat 2021;188:149-63. 10.1007/s10549-021-06192-w.

36. Cronin M, Crowley A, Davey MG, et al. Heart Failure Association-International Cardio-Oncology Society Risk Score Validation in HER2-Positive Breast Cancer. J Clin Med 2023;12:1278. 10.3390/jcm12041278

37. Liu X, Tao L, Wang M, Li H, Xu W. ABSDELL Model: Development and Internal Validation of a Risk Prediction Model of LVEF Decline in Breast Cancer Patients Treated With Trastuzumab. Clin Breast Cancer 2023;23:23-31. 10.1016/j.clbc.2022.10.010.

38. Suntheralingam S, Fan CS, Calvillo-Argüelles O, Abdel-Qadir H, Amir E, Thavendiranathan P. Evaluation of Risk Prediction Models to Identify Cancer Therapeutics Related Cardiac Dysfunction in Women with HER2+ Breast Cancer. J Clin Med 2022;11:847. 10.3390/jcm11030847

39. Tini G, Cuomo A, Battistoni A, et al. Baseline cardio-oncologic risk assessment in breast cancer women and occurrence of cardiovascular events: The HFA/ICOS risk tool in real-world practice. Int J Cardiol 2022;349:134-7. 10.1016/j.ijcard.2021.11.059.

40. Di Lisi D, Madaudo C, Faro DC, et al. The added value of the HFA/ICOS score in the prediction of chemotherapy-related cardiac dysfunction in breast cancer. J Cardiovasc Med (Hagerstown) 2024;25:218-24. 10.2459/JCM.0000000000001589.

41. Rivero-Santana B, Saldaña-García J, Caro-Codón J, et al. Anthracycline-induced cardiovascular toxicity: validation of the Heart Failure Association and International Cardio-Oncology Society risk score. Eur Heart J 2025;46:273-84. 10.1093/eurheartj/ehae496.

42. Fernando F, Andres MS, Claudiani S, et al. Cardiovascular events in CML patients treated with Nilotinib: validation of the HFA-ICOS baseline risk score. Cardiooncology 2024;10:42. 10.1186/s40959-024-00245-x.

43. Stefanini B, Tovoli F, Trevisani F, et al. Italian Liver Cancer (ITA.LI.CA) group . Prediction of cardiovascular risk in patients with hepatocellular carcinoma receiving anti-angiogenic drugs: lessons from sorafenib. Intern Emerg Med 2024;19:1151-60. 10.1007/s11739-024-03578-8.

44. Shibata T, Nohara S, Morikawa N, et al. Cardiovascular adverse events and prognosis in patients with haematologic malignancies and breast cancer receiving anticancer agents: Kurume-CREO Registry insights. Eur J Prev Cardiol 2023;30:1941-9. 10.1093/eurjpc/zwad210.

45. Collins GS, Mallett S, Omar O, Yu L-M. Developing risk prediction models for type 2 diabetes: a systematic review of methodology and reporting. BMC Med 2011;9:103. 10.1186/1741-7015-9-103.

46. Collins GS, Omar O, Shanyinde M, Yu L-M. A systematic review finds prediction models for chronic kidney disease were poorly reported and often developed using inappropriate methods. J Clin Epidemiol 2013;66:268-77. 10.1016/j.jclinepi.2012.06.020.

47. Damen JAAG, Hooft L, Schuit E, et al. Prediction models for cardiovascular disease risk in the general population: systematic review. BMJ 2016;353:i2416. 10.1136/bmj.i2416.

48. Mitchell JD, Laurie M, Xia Q, et al. Risk profiles and incidence of cardiovascular events across different cancer types. ESMO Open 2023;8:101830. 10.1016/j.esmoop.2023.101830.

49. Strongman H, Gadd S, Matthews A, et al. Medium and long-term risks of specific cardiovascular diseases in survivors of 20 adult cancers: a population-based cohort study using multiple linked UK electronic health records databases. Lancet 2019;394:1041-54. 10.1016/S0140-6736(19)31674-5.

50. Paterson DI, Wiebe N, Cheung WY, et al. Incident Cardiovascular Disease Among Adults With Cancer: A Population-Based Cohort Study. JACC CardioOncol 2022;4:85-94. 10.1016/j.jaccao.2022.01.100.

51. Zhang N, Wu J, Wang Q, et al. Global burden of hematologic malignancies and evolution patterns over the past 30 years. Blood Cancer J 2023;13:82. 10.1038/s41408-023-00853-3.

52. Sant M, Minicozzi P, Mounier M, et al. EUROCARE-5 Working Group . Survival for haematological malignancies in Europe between 1997 and 2008 by region and age: results of EUROCARE-5, a population-based study. Lancet Oncol 2014;15:931-42. 10.1016/S1470-2045(14)70282-7.

53. Vasbinder A, Catalan T, Anderson E, et al. Cardiovascular Risk Stratification of Patients Undergoing Hematopoietic Stem Cell Transplantation: The CARE-BMT Risk Score. J Am Heart Assoc 2024;13:e033599. 10.1161/JAHA.123.033599.

54. Leerink JM, van der Pal HJH, Kremer LCM, et al. Refining the 10-Year Prediction of Left Ventricular Systolic Dysfunction in Long-Term Survivors of Childhood Cancer. JACC CardioOncol 2021;3:62-72. 10.1016/j.jaccao.2020.11.013.

55. Ehrhardt MJ, Leerink JM, Mulder RL, et al. Systematic review and updated recommendations for cardiomyopathy surveillance for survivors of childhood, adolescent, and young adult cancer from the International Late Effects of Childhood Cancer Guideline Harmonization Group. Lancet Oncol 2023;24:e108-20. 10.1016/S1470-2045(23)00012-8.

56. Steyerberg EW, Harrell FE, Jr. Prediction models need appropriate internal, internal-external, and external validation. J Clin Epidemiol 2016;69:245-7. 10.1016/j.jclinepi.2015.04.005.

57. Siontis GCM, Tzoulaki I, Castaldi PJ, Ioannidis JPA. External validation of new risk prediction models is infrequent and reveals worse prognostic discrimination. J Clin Epidemiol 2015;68:25-34. 10.1016/j.jclinepi.2014.09.007.

58. Glen C, Adam S, McDowell K, et al. Cardiotoxicity of BRAF/MEK Inhibitors: A Longitudinal Study Incorporating Contemporary Definitions and Risk Scores. JACC CardioOncol 2023;5:628-37. 10.1016/j.jaccao.2023.04.004.

59. Mecinaj A, Gulati G, Ree AH, et al. Impact of the ESC Cardio-Oncology Guidelines Biomarker Criteria on Incidence of Cancer Therapy-Related Cardiac Dysfunction. JACC CardioOncol 2024;6:83-95. 10.1016/j.jaccao.2023.10.008.

60. Oristrell G, Burcet G, Valente F, Escrivá-De-Romaní S, Arumí M, Ferreira-González I. Treatment of mild asymptomatic cardiotoxicity in early-stage HER 2-positive breast cancer. Is it justified? Rev Esp Cardiol (Engl Ed) 2023;76:480-3. 10.1016/j.rec.2022.11.014.

61. Chatterjee NA, Tikkanen JT, Panicker GK, et al. PREDETERMINE Investigators . Simple electrocardiographic measures improve sudden arrhythmic death prediction in coronary disease. Eur Heart J 2020;41:1988-99. 10.1093/eurheartj/ehaa177.

62. Sangha V, Nargesi AA, Dhingra LS, et al. Detection of Left Ventricular Systolic Dysfunction From Electrocardiographic Images. Circulation 2023;148:765-77. 10.1161/CIRCULATIONAHA.122.062646.

63. Oikonomou EK, Sangha V, Dhingra LS, et al. Artificial Intelligence-Enhanced Risk Stratification of Cancer Therapeutics-Related Cardiac Dysfunction Using Electrocardiographic Images. Circ Cardiovasc Qual Outcomes 2025;18:e011504. 10.1161/CIRCOUTCOMES.124.011504.

64. Kaboré EG, Macdonald C, Kaboré A, et al. Risk Prediction Models for Cardiotoxicity of Chemotherapy Among Patients With Breast Cancer: A Systematic Review. JAMA Netw Open 2023;6:e230569. 10.1001/jamanetworkopen.2023.0569.

65. Wang S-Y, Long JB, Hurria A, et al. Cardiovascular events, early discontinuation of trastuzumab, and their impact on survival. Breast Cancer Res Treat 2014;146:411-9. 10.1007/s10549-014-3029-0.

66. Cardinale D, Colombo A, Bacchiani G, et al. Early detection of anthracycline cardiotoxicity and improvement with heart failure therapy. Circulation 2015;131:1981-8. 10.1161/CIRCULATIONAHA.114.013777.

67. Akpek M, Ozdogru I, Sahin O, et al. Protective effects of spironolactone against anthracycline-induced cardiomyopathy. Eur J Heart Fail 2015;17:81-9. 10.1002/ejhf.196.

68. Avila MS, Ayub-Ferreira SM, de Barros Wanderley MR, Jr, et al. Carvedilol for Prevention of Chemotherapy-Related Cardiotoxicity: The CECCY Trial. J Am Coll Cardiol 2018;71:2281-90. 10.1016/j.jacc.2018.02.049.

69. Bosch X, Rovira M, Sitges M, et al. Enalapril and carvedilol for preventing chemotherapy-induced left ventricular systolic dysfunction in patients with malignant hemopathies: the OVERCOME trial (preventiOn of left Ventricular dysfunction with Enalapril and caRvedilol in patients submitted to intensive ChemOtherapy for the treatment of Malignant hEmopathies). J Am Coll Cardiol 2013;61:2355-62. 10.1016/j.jacc.2013.02.072.

70. Cardinale D, Colombo A, Sandri MT, et al. Prevention of high-dose chemotherapy-induced cardiotoxicity in high-risk patients by angiotensin-converting enzyme inhibition. Circulation 2006;114:2474-81. 10.1161/CIRCULATIONAHA.106.635144.

71. Cardinale D, Ciceri F, Latini R, et al. ICOS-ONE Study Investigators . Anthracycline-induced cardiotoxicity: A multicenter randomised trial comparing two strategies for guiding prevention with enalapril: The International CardioOncology Society-one trial. Eur J Cancer 2018;94:126-37. 10.1016/j.ejca.2018.02.005.

72. Vaduganathan M, Hirji SA, Qamar A, et al. Efficacy of Neurohormonal Therapies in Preventing Cardiotoxicity in Patients with Cancer Undergoing Chemotherapy. JACC CardioOncol 2019;1:54-65. 10.1016/j.jaccao.2019.08.006.

73. Riley RD, Ensor J, Snell KIE, et al. External validation of clinical prediction models using big datasets from e-health records or IPD meta-analysis: opportunities and challenges. BMJ 2016;353:i3140. 10.1136/bmj.i3140.

74. McCracken C, Condurache D-G, Szabo L, et al. Predictive Performance of Cardiovascular Risk Scores in Cancer Survivors From the UK Biobank. JACC CardioOncol 2024;6:575-88. 10.1016/j.jaccao.2024.05.015.

75. Debray TP, Riley RD, Rovers MM, Reitsma JB, Moons KG, Cochrane IPD Meta-analysis Methods group . Individual participant data (IPD) meta-analyses of diagnostic and prognostic modeling studies: guidance on their use. PLoS Med 2015;12:e1001886. 10.1371/journal.pmed.1001886.

76. Teske AJ, Moudgil R, López-Fernández T, et al. Global Cardio Oncology Registry (G-COR): Registry Design, Primary Objectives, and Future Perspectives of a Multicenter Global Initiative. Circ Cardiovasc Qual Outcomes 2023;16:e009905. 10.1161/CIRCOUTCOMES.123.009905.

77. López-Sendón J, Álvarez-Ortega C, Zamora Auñon P, et al. Classification, prevalence, and outcomes of anticancer therapy-induced cardiotoxicity: the CARDIOTOX registry. Eur Heart J 2020;41:1720-9. 10.1093/eurheartj/ehaa006.

78. Lapirow D, La Gerche A, Toro C, et al. The Australia and New Zealand Cardio-Oncology Registry: evaluation of chemotherapy-related cardiotoxicity in a national cohort of paediatric cancer patients. Intern Med J 2021;51:229-34. 10.1111/imj.14719.

79. Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI Consortium . FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ 2025;388:e081554. 10.1136/bmj-2024-081554.

Associated Data

Supplementary Materials

Supplementary information: Tables 1-15 and figures 1-5

Supplementary information: Additional data to share

Data Availability Statement

Additional supporting data are available in the supplementary file.