Artificial Intelligence to Enhance Precision Medicine in Cardio-Oncology: A Scientific Statement From the American Heart Association

Abstract

Artificial intelligence is poised to transform cardio-oncology by enabling personalized care for patients with cancer, who are at a heightened risk of cardiovascular disease due to both the disease and its treatments. The rising prevalence of cancer and the availability of multiple new therapeutic options has resulted in improved survival among patients with cancer and has expanded the scope of cardio-oncology to not only short-term but also long-term cardiovascular risks resulting from both cancer and its treatments. However, there is considerable heterogeneity in cardiovascular risk, driven by the nature of the malignancy as well as each individual’s unique characteristics. The use of novel therapies, such as targeted therapies and immune checkpoint inhibitors, across multiple cancer groups has also broadened the populations among which cardiotoxicity has become an important consideration of therapy. Therefore, the ability to understand and personalize cardiovascular risk management in patients with cancer is a key target for artificial intelligence, which can deduce and respond to complex patterns within the data. These advances necessitate an overview of established biomarkers of risk, spanning advanced imaging, diagnostic testing, and multi-omics, the evidence supporting their use, and the proven and proposed role of artificial intelligence in refining this risk to attain greater precision in risk prediction and management in cardio-oncologic care.

Cardio-oncology is a burgeoning field, with its growth driven by the rising incidence of cancer, the emergence of several therapeutic agents that have improved survival, and the short- and long-term cardiovascular risks resulting from both cancer and its treatments.¹ However, the underlying pathophysiologic processes that drive cardiovascular risk vary across individuals and represent a complex interplay between the specific malignancy and its treatment, as well as each individual’s variable susceptibility to these factors.² There have also been advances, such as targeted therapies and immune checkpoint inhibitors (ICIs), for which cardiotoxicity is a key consideration of therapy,³ but those at risk have been challenging to identify. Therefore, the ability to understand and personalize cardiovascular risk management in patients with cancer is a key target for artificial intelligence (AI), bolstered by the ability to leverage large swaths of data and identify complex patterns. These advances necessitate an overview of established biomarkers of risk, the evidence supporting their use, and the proven and proposed role of AI in refining this risk to attain greater precision in risk prediction and management (Figure 1).

Figure 1. Leveraging artificial intelligence and machine learning in cardio-oncology for precision care.

*National and international cardio-oncology data registries created by cardiology and oncology specialty societies (eg, American Heart Association, American Society of Clinical Oncology, other broad specialty groups). AI indicates artificial intelligence; BNP, brain natriuretic peptide; CMR, cardiac magnetic resonance imaging; CT, computed tomography; CV, cardiovascular; EHR, electronic health record; GLS, global longitudinal strain; LLM, large language model; LVEF, left ventricular ejection fraction; ML, machine learning; and PET, positron emission tomography.

NOVEL DIAGNOSTIC AND BIOMARKER DATA

Advanced Imaging Techniques in Cardio-Oncology

In cardio-oncology, cardiovascular imaging as a marker of disease is an accepted standard for the early detection, confirmatory diagnosis, and monitoring of potential cardiotoxic change. Imaging is widely available, is noninvasive, and can identify several pathophysiologic processes (eg, muscle dysfunction, inflammation, thrombosis), making it an excellent resource for monitoring patient responses to an increasingly diverse set of chemotherapeutic agents.⁴ Techniques with well-established data include echocardiography, cardiac magnetic resonance imaging (CMR), nuclear imaging, and cardiac computed tomography.^4,5 Because most clinical trials have used echocardiography or CMR, we focus on these modalities below.

In both echocardiography and CMR, measures commonly used to monitor cardiotoxicity include left ventricular ejection fraction (LVEF), left ventricular global longitudinal strain, and diastolic function assessment.⁴ CMR also allows parametric T1 and T2 mapping and provides higher accuracy in assessing cardiac structure and function than echocardiography.⁶ In an investigation of 136 patients with suspected ICI-associated myocarditis, elevation in native T1 on CMR was highly predictive of future major adverse cardiac events.⁷ However, in the SUCCOUR trial (Strain Surveillance of Chemotherapy for Improving Cardiovascular Outcomes), global longitudinal strain was equivocal in the prediction of future heart failure events among patients treated with anthracyclines and therapies targeting human epidermal growth factor receptor 2, respectively.⁸

In addition, measurements of left atrial remodeling and vascular resistance or stiffness represent increasingly available options for ascertaining long-term risk after the initiation of cancer treatment. For example, in 2 studies of patients with hematologic malignancy treated with the tyrosine kinase inhibitor ibrutinib, abnormal left atrial strain and size by echocardiography, as well as increased native T1/T2 by CMR, were highly predictive of future major adverse cardiac events and other adverse events.^9,10 Table 1 lists uses of cardiac imaging and their future role in precision medicine.

Table 1.

Imaging Biomarkers in Cardio-Oncology: Uses and Potential Integration Into Multi-Omics and Personalized Medicine

Modality	Uses		Future directions
	Measurements	Supporting evidence in cardio-oncology
Echocardiography	LVEF Strain Diastology Valve dysfunction	LVEF decline is a marker of cancer treatment–related HF11 Early cardioprotection based on LVEF by echocardiography may reduce HF ratesafter cancer therapy12 Change in GLS precedes HF from anthracyclines, taxanes, and trastuzumab5 VHD after radiotherapy may lead to increased morbidity and mortality13	Standardize measurements Establish evidence for use in cardiovascular toxicities arising from new cancer therapies Establish cardiovascular risk thresholds for established biomarkers in clinical use Evaluate larger cohorts with integrated cardiovascular and cancer phenotyping Clarify role of radiomics in identifying early cardiac disease or progression (eg, coronary plaques, inflammatory myocardial changes) Establish novel imaging biomarkers (eg, vascular biomarkers, myocardial perfusion, myocardial bioenergetics) Define how imaging biomarkers respond to therapeutics
Cardiac MRI	LVEF Strain T1/T2 mapping LGE Amyloid	LGE and T2-weighted STIR in ICI myocarditis14 T1/T2 mapping as prognostic biomarkers in ICI myocarditis7 Native T1, ECV, and LGE predict survival outcomes15 and can guide treatment in cardiac amyloidosis16
Coronary CT	CAC Luminal stenosis CT-FFR	Deep learning (AI)–based algorithm for CAC after chest radiotherapy predicted future acute coronary events17 The presence of CAC among patients with breast cancer (before and after chemotherapy) predicts future cardiovascular events18 CCTA imaging captures increased rates of substantial CAD after chest radiotherapy and chemotherapy, including CAD requiring intervention19-21
Nuclear imaging	Myocardial perfusion Coronary flow reserve Amyloid	SPECT imaging demonstrates early myocardial perfusion abnormalities after chest radiotherapy22 Tc-99m PYP scintigraphy imaging distinguishes AL from ATTR cardiac amyloidosis23 Tc-99m PYP imaging predicts survival in ATTR amyloidosis24

AI indicates artificial intelligence; AL amyloid, light chain amyloidosis; ATTR, transthyretin amyloidosis; CAC, coronary artery calcium; CAD, coronary artery disease; CCTA, coronary computed tomography angiography; CT, computed tomography; ECV, extracellular volume fraction; FFR, fractional flow reserve; GLS, global longitudinal strain; HF, heart failure; ICI, immune checkpoint inhibitor; LGE, late gadolinium enhancement; LVEF, left ventricular ejection fraction; MRI, magnetic resonance imaging; SPECT, single-photon emission computed tomography; STIR, short tau inversion recovery; Tc-99m PYP, technetium-99m pyrophosphate; and VHD, valvular heart disease.

Biomarkers and Multi-Omics Approach for Care Personalization

Traditional Biomarkers

Conventional cardiovascular blood biomarkers, such as cardiac troponins and brain natriuretic peptides (BNPs), have been evaluated in several studies with regard to their predictive value for subsequent cancer therapeutics–related cardiac dysfunction,²⁵ also referred to as cardiotoxicity. Early studies in patients with aggressive or advanced malignancies treated with high-dose cytotoxic chemotherapy demonstrated an association between troponin levels and subsequent cardiotoxicity.^26,27 In contrast, more recent studies integrating newer high-sensitivity troponin assays with BNP measurement in patients treated with anthracyclines and human epidermal growth factor receptor 2–targeted therapies for early-stage breast cancer have inconsistently demonstrated an association of these biomarkers with cancer therapeutics–related cardiac dysfunction, albeit in small patient cohorts.^28,29 Overall, randomized clinical trials that specifically study the use of combination biomarker assays (BNPs and troponins) in patients with cancer on treatment are lacking. In addition, set cut points to define cardiotoxic troponin levels are yet to be defined, and greater standardization across assays is needed.²⁵ Nonetheless, in day-to-day practice involving the general population of patients with mainly low baseline cardiac risk, these assays are most useful for their high negative predictive value, which ranges from 84% to 100%, depending on the cut point.³⁰ Cardiac troponins and BNPs should also be interpreted in the context of comorbid conditions, including renal dysfunction, lung disease, arrhythmias, and high blood pressure.²⁵

In the same vein, evidence supporting the use of biomarker-guided initiation of cardioprotective therapies is not substantial. For example, a strategy of serial troponin measurement to guide the initiation of enalapril for cardioprotection did not yield any significant benefit compared with the universal initiation of enalapril before chemotherapy in a randomized clinical trial.³¹ However, the study was limited by its small sample size (total n=276), relatively low cumulative total anthracycline dose (maximum ≈240 mg/m² doxorubicin equivalent) with resultant low risk of cardiotoxicity, and lack of uniformity in the troponin assays (4 different types of troponin assays) tested. Initiation of candesartan and carvedilol based on high-sensitivity troponin I did not prevent left ventricular dysfunction among patients with breast cancer and lymphoma in a recent prospective, multicenter randomized clinical trial.³² In the study, which also was limited by its small size, nearly one-third of patients discontinued the prescribed cardioprotective medications, and there were differences in anthracycline doses (mean doxorubicin equivalent ≈240 mg/m² in the cardioprotection versus ≈300 mg/m² in the standard care group). Furthermore, the incidence of cardiotoxicity in patients taking anthracycline was low in both trials regardless of treatment group.

In general, the high negative predictive value of these biomarkers supports their use as screening tools for cardiotoxicity. With the available evidence, in clinical practice, these biomarkers can be used to rule out cardiotoxicity, with further testing with imaging to rule in disease among patients with higher levels of BNP and troponin during cancer treatment. The European Society of Cardiology Guidelines on Cardio-Oncology recommend baseline measurement of cardiac biomarkers in all patients treated with potentially cardiotoxic agents and during treatment in select patient groups.³³

Traditional cardiac biomarkers have proven effective in the diagnosis and prognosis of ICI-associated myocarditis. In a prospective longitudinal study of patients undergoing treatment with ICIs, serial measurement of high-sensitivity troponin I detected ICI-associated myocarditis in both symptomatic and asymptomatic patients and showed potential feasibility for early detection and mitigation of adverse cardiovascular events.³⁴ Furthermore, in a prospective cohort of patients diagnosed with ICI-associated myocarditis, relative elevation in cardiac troponin T was prognostic of increased incidence of major adverse cardiac events.³⁵

Development and testing of additional biomarkers, as well as generation of adequate cut points with reproducible and accurate measurements in clinical practice, could be aided by AI for image analysis and adequate timing and thresholds for biomarker analysis. As such, in the short term, more reliable and reproducible imaging and blood biomarkers for cardiotoxicity assessment could be included, the validation and reliable deployment of which could be aided by AI. Moreover, AI and machine learning (ML) strategies can enable transformation in diagnostic and prognostic modeling in these domains through triangulation of disparate data streams, as has been successful in other clinical domains.³⁶

Genomics

In cardio-oncology, genomic studies have largely focused on defining single nucleotide polymorphisms associated with anthracycline cardiotoxicity, both in childhood cancer survivors and in adults. Candidate single nucleotide polymorphisms have been identified in several mechanistic pathways implicated in the pathogenesis of anthracycline cardiotoxicity based on preclinical models (Table 2).³⁷ Genetic variants in cardiomyocyte structural proteins that overlap with other forms of cardiomyopathy, such as titin-truncated variants, have also been associated with anthracycline cardiomyopathy in several small cohorts. In general, the use of unbiased genetic approaches to discover new causal pathways has been limited by small sample sizes and the lack of consensus for defining cardiotoxicity across studies. As a result, whereas there is potential to incorporate genomics into cardiotoxicity risk prediction, the evidence defining patients at risk for cardiotoxicity is insufficient to incorporate this pathway into routine clinical Cardio-Oncology care.

Table 2.

Blood Biomarkers in Cardio-Oncology: Uses and Potential Integration Into Multi-Omics and Personalized Medicine

Modality	Uses		Future directions
	Measurements	Supporting evidence in cardio-oncology
Conventional biomarkers	Cardiac troponins BNP	hsTnI in HF from anthracyclines, taxanes, and trastuzumab5 BNP, troponin, and other risk factors in predicting cardiotoxicity TnI as prospective screening biomarker for early ICI myocarditis34	Standardize measurements Establish evidence for use in cardiovascular toxicities arising from new cancer therapies Establish cardiovascular risk thresholds for established biomarkers in clinical use Evaluate larger cohorts with integrated cardiovascular and cancer phenotyping Establish shared biomarkers of cancer and cardiovascular risk Determine preclinical models that better recapitulate cardiovascular toxicity in patients Examine technologic advances to enable high-throughput use at lower cost Perform high-dimensional characterization of pathogenic cell subsets in cardiotoxicity
Genomics	Variants in cardiac structural proteins Variants identified by GWAS CHIP	TTNtvs in chemotherapy-induced cardiomyopathy38 RARG in anthracycline cardiomyopathy39
Transcriptomics	Bulk vs scRNA-seq TCR sequencing	T cells specific for α-myosin in ICI myocarditis40 Peripheral immune cell subset characterization in ICI myocarditis41
Proteomics	LC/MS vs ELISA Aptamer-based proteomics (eg, SomaScan) Antibody-based proteomics (eg, CITE-seq, Olink) Phosphoproteomics	Hemopexin in anthracycline cardiotoxicity42 CSK in ibrutinib-associated atrial fibrillation43
Metabolomics	Intermediary metabolism Lipidomics	TCA cycle metabolites in anthracycline cardiotoxicity44

BNP indicates brain natriuretic peptide; CHIP, clonal hematopoiesis of indeterminate potential; CITE-seq, cellular indexing of transcriptomes and epitopes by sequencing; CSK, C-terminal Src kinase; ELISA, enzyme-linked immunosorbent assay; GWAS, genome-wide association study; HF, heart failure; hsTnI, high-sensitivity troponin I; ICI, immune checkpoint inhibitor; LC/MS, liquid chromatography/mass spectrometry; scRNA-seq, single-cell RNA sequencing; TCA, tricarboxylic acid; TCR, T-cell receptor; and TTNtv, titin-truncating variant.

Transcriptomics

Transcriptomic profiling using RNA sequencing or single-cell RNA sequencing (scRNA-seq) has provided novel insights into the biologic mechanisms of cardiotoxicity. Whereas bulk RNA sequencing provides an average expression level for each gene transcript within a sample, scRNA-seq maps RNA transcripts to individual cells, allowing for tissue profiling on a single-cell resolution in heterogeneous disease samples.⁴⁵ Although more costly than bulk RNA sequencing or scRNA-seq, spatial transcriptomics allows for the visualization and quantification of RNA transcripts while preserving the overall cellular architecture in histologic tissue sections,⁴⁶ providing useful information on cell–cell spatial relationships relative to their transcriptomics. When combined with other single-cell multi-omics technologies, such as time-of-flight mass cytometry, T-cell receptor sequencing, and cellular indexing of transcriptomes and epitopes by sequencing, scRNA-seq and spatial transcriptomics can be a powerful tool to gain insights into the cellular and transcriptomic landscape of cardiotoxicities.⁴⁵ The application of these novel technologies was demonstrated in a recent study examining the immune landscape of ICI myocarditis.⁴¹ The combination of scRNA-seq, single-cell T-cell receptor sequencing, cellular indexing of transcriptomes and epitopes by sequencing, and time-of-flight mass cytometry revealed pathogenic T-cell populations in samples collected from patients with ICI myocarditis, along with data from a paired mouse model, providing new insights into the immune cell subsets involved in ICI cardiotoxicity. Transcriptomics can also provide a useful adjunct tool in mechanistic preclinical studies, as demonstrated by recent works investigating T-cell antigenicity and macrophage subsets in ICI-associated myocarditis.⁴⁰ Spatial transcriptomics would potentially add valuable mechanistic insights into various cardiotoxicities.

Proteomics and Metabolomics

New high-throughput proteomics technologies enable the measurement of several hundreds or thousands of proteins in <100 μL of plasma. These have been applied to pilot cohorts of patients treated with cardiotoxic cancer therapies, including anthracyclines⁴² and the protea-some inhibitor carfilzomib.⁴⁷ Targeted metabolite profiling using liquid chromatography/mass spectrometry-based techniques has been used to provide insights into metabolic abnormalities in patients with breast cancer with or without anthracycline cardiac toxicity.⁴⁴ In that study, early changes in the tricarboxylic acid cycle metabolites differentiated subsequent cardiotoxicity development, and patients with cardiotoxicity exhibited pronounced alterations in purine and pyrimidine metabolism. Future research should validate these pilot -omics study findings in larger patient cohorts with serial blood sampling and cardiovascular phenotyping after cancer treatment.

AI AND ML IN CARDIO-ONCOLOGY

AI-Based Risk Prediction Models

In the context of this scientific statement, the literature uses the terms AI and ML interchangeably, referring to algorithms designed to infer complex relationships between elements of various data streams to achieve a certain goal or accomplish a task. Advances in AI have allowed for the analysis of integrated clinical data, including blood biomarkers, ECG, echocardiography, computed tomography, and CMR, to predict the risk for cardiovascular disease (CVD) in patients with cancer. The ability of AI to accurately analyze conventional measures such as LVEF/global longitudinal strain and BNP/troponin and identify novel predictive biomarkers of cardiotoxicity, such as other laboratory tests, CMR, cardiac computed tomography (including coronary artery calcium), and positron emission tomography, can be powerful tools to aid clinicians in identifying at-risk cohorts in which to optimize cardioprotective strategies.^48,49 There are several recent examples of the implementation of this strategy of using clinical and imaging data to develop risk prediction models.

By using large-scale longitudinal patient data (eg, laboratory tests, symptoms, demographics, echocardiography variables) from 1 study, classification models were trained for 6 types of cardiovascular outcomes in 4309 patients with cancer, providing proof of concept for successful cardiovascular risk stratification in oncology patients.⁵⁰ Another study reported a prediction model that included breast cancer treatment–related risk factors, such as previous anthracycline or radiation exposure, and conventional risk factors, such as hypertension and diabetes, for predicting major adverse cardiac events at up to 7 years.⁵¹ Another risk prediction model trained with 30 286 low-dose computed tomography scans from patients with lung cancer successfully identified patients with high CVD mortality risk, which outperformed existing models.⁵² An integrative and comprehensive risk prediction model, which included cancer risk factors, cardiovascular risk factors, and blood biomarkers, such as troponin and BNP, also performed well for cardiotoxicity prediction, with risk discrimination in the 90% range for multiple cohorts.⁵³ A primer for key ML/AI-specific metrics is available from Stevens and colleagues,⁵⁴ and a glossary of key terms relevant to both AI and multiomics included in the study is provided in Table 3.

Table 3.

Glossary of Key Terms in Artificial Intelligence and Precision Medicine

Domain	Term	Definition
Artificial intelligence	Artificial intelligence	A term for a set of technologies that describe the ability of a machine to perform tasks that are typical of human intelligence, such as parsing language and addressing complex problems
	Generative AI	A class of AI that specifically includes a set of algorithms capable of generating new content, such as images, text, or any other data elements
	Machine learning	A series of algorithms that can iteratively learn and improve from simple or complex data to make predictions about new scenarios
	Neural network	A class of algorithms that contain layers of interconnected neurons that sequentially learn from data and outputs of preceding nodes
	Convolutional neural network	A class of neural networks that are adept at learning features of data using specialized mathematic functions (called kernels) and are important in image, signal, and text processing
	Deep learning	Deep learning is a class of machine learning that includes many hidden layers of interconnected neurons able to learn from complex data streams
	Natural language processing	Machine learning/AI technology that enables a machine to process and interpret human language for reproducible inference
	Large language models	A class of natural language processing models that leverage deep learning and advanced AI to demonstrate general purpose comprehension of language and text generation
Precision medicine	Precision medicine	Integration of genetics and related technologies with clinical and environmental factors to target subgroups of patients for specific therapies
	Genomics	The study of all genes in an organism and their interindividual variability
	Transcriptomics	The study of all RNA transcripts expressed in a sample
	Proteomics	Large-scale analysis of all proteins in a sample
	Metabolomics	Large-scale analysis of all metabolites in a sample
	Single-cell RNA sequencing	The study of all RNA transcripts expressed in a specific cell type
	Single nucleotide polymorphisms	Genomic variation at a single base position in a DNA sequence
	Cellular indexing of transcriptomes and epitopes by sequencing	A technology that integrates single cell transcriptomics with antibody-based cell surface protein analysis
	T-cell receptor sequencing	A technology that permits identification and characterization of specific T cells and their clones
	Time-of-flight mass cytometry	A technology that uses metal isotope–labeled antibodies to detect multiple proteins in individual cells

AI indicates artificial intelligence.

The combination of genetic sequencing data and existing clinical data can provide a powerful resource for ML-based risk models. Exome sequencing in 289 childhood cancer survivors with a history of anthracycline exposure identified differentially enriched variants associated with reduced LVEF, which improved the predictive power of a risk prediction model using these genetic and clinical predictors.⁵⁵ This and other examples in the pipeline show promise for the combination of novel biomarker data with existing clinical data in predicting risk.

However, it is important to account for the possibility of bias in these studies due to selected populations, small sample sizes, lack of consistency in handling missing data, differences in the appropriateness of statistical approaches, and the potential for overfitting. In a systematic review of risk prediction models for cardiotoxicity of chemotherapy among patients with breast cancer, only 1 of 6 studies used external validation.⁵⁶ The emergence of data sets focusing on cardiotoxicity with enough clinical detail for adequate training of ML algorithms will advance the field further.

AI-Enabled Analysis of Cardiac Imaging

Imaging-based biomarkers can, and many do, play a role in the detection and monitoring of cardiac effects of cancer treatments. AI could help improve the speed, reproducibility, and scale of deploying existing imaging biomarkers, and would also aid in more rapid development and testing of new ones (Figure 1).

A principal imaging biomarker is the measurement of LVEF. Measuring LVEF requires several time-consuming measurements of the left ventricle that can be poorly reproducible in practice.⁵⁷ Several companies now offer AI-assisted LVEF measurement, with a randomized clinical trial suggesting benefits in such AI-based assessment of LVEF.⁵⁸ Research and testing of automated LV strain, diastolic function, myocardial perfusion, and other measures useful in cardio-oncology are ongoing.^59,60

AI-based technology could further advance the field of cardio-oncology with new image-based biomarkers. These include a class of tools that identify features that otherwise are not observable by human experts,⁶¹ although these classes of biomarkers are limited in specific cardio-oncology applications. With appropriate data sets, including imaging, cardiovascular outcomes, and cancer diagnostic, treatment, and outcomes information, many different biomarkers (including composite biomarkers leveraging multimodal imaging) can be studied.^62,63 Data sets must be of sufficient quality and size, but growing research suggests they need not be extremely large if they are designed while accounting for diversity.⁶⁴ Advancements in AI algorithms, such as semi-supervised and self-supervised learning,⁶⁵ as well as foundation models,⁶⁶ will decrease the need for painstaking manual image labeling and thereby expedite research and development.

Natural Language Processing for Clinical Data Extraction and Decision Support

Natural language processing recently emerged as a mechanism for supplementing discrete data from electronic health records (EHRs) with more enriched content from unstructured sources such as clinical notes, which contain a wealth of patient information.⁶⁷ Whereas the reliability of unstructured data can be questionable, using it in combination with codified data produces robust insight into patient history.⁶⁸ Unstructured data may also supplement sources used as inputs into AI models, especially when abstracted into discrete data elements.⁶⁹ The rapid rise of generative AI—specifically large language models—promises to make analyses of large data sets easier, faster, and more efficient, thereby traversing large amounts of unstructured data that exceed human capacity, and in a format to which humans are accustomed. These models lack accuracy, reliable prioritization of information, and regulation,⁷⁰ and yet have demonstrable successes in select medical use cases, such as passing standardized medical examinations with high scores.⁷¹ Large language model use cases are expanding, particularly in areas where a wealth of unstructured data is a considerable advantage (eg, public health, personalized medicine, health care administration, mental health, training, education).⁷² However, specific cardio-oncology implementations have yet to be documented in published literature. The expectation of these endeavors would be a more thorough evaluation of patient records, including clinical notes, which may reveal more information about the patient at risk for cardiotoxicity. Generative AI and large language models could extract and convert relevant cardiotoxicity data from EHRs (eg, erratic sources with missing elements, data ranges, or unstructured formats) into meaningful structured insights for research and clinical use. They can also extend the ability of clinicians to interact directly with diverse electronic data sources using smart chatbots that enable clinicians to draw such precision care insights directly without the need for advanced programming.⁷² There is also a potential for natural language processing to help identify opportunities for collaboration between oncologists and cardiologists through flagging high-risk patients in need of referral to cardio-oncology care before administration of oncologic treatment regimens.⁵³ Such flags are especially relevant in higher-risk individuals receiving potentially cardiotoxic therapy.

AI-Driven Drug Discovery and Development

The possibility of using AI to aid drug discovery and development has emerged in the past few years.⁷³ The initial promise of efforts such as IBM Watson’s use to identify altered RNA-binding proteins in amyotrophic lateral sclerosis underscored the potential of AI to aid in therapeutic discovery.⁷⁴ In addition, ML algorithms are being integrated into every aspect of drug development, from screening and predicting efficacy and toxicity to accelerating product development and the manufacturing pipeline to shorten the time to market.⁷⁵ However, these early efforts must be paired with experimental approaches for validation. In the field of cardio-oncology, AI holds promise for predicting the toxicity of existing drugs and for the discovery of therapeutic approaches to treat toxicity. This remains an exciting future direction for upcoming years.

INTEGRATING AI-DRIVEN CARE PERSONALIZATION INTO CARDIO-ONCOLOGY CLINICAL PRACTICE

The dearth of integrative technologies to embed AI into clinical practice has produced limited results in cardio-oncology. However, developments are ongoing on multiple fronts, particularly in CVD risk assessment, analyzing ECG outputs as predictive instruments for CVD, and applying AI to empower personalized approaches to care.

The Role of AI-Enabled ECG Algorithms in Detecting Cardiotoxicity

Several studies have demonstrated the use of data automation in generating clinical insight from ECGs using AI algorithms such as convolutional neural networks.^76-78 Associative studies between diagnosis and potential outcome indicate that ECG features early during cancer treatment may have specific patterns suggestive of subsequent mortality and CVD complications; for instance, in a study of life-threatening arrhythmias and mortality in 125 cases of ICI-associated myocarditis from 49 institutions across 11 countries and 50 cases of posttransplant rejection as controls from a single US institution.⁷⁸

Outside the context of cardio-oncology, AI-enabled ECG algorithms can detect and predict reduced LVEF <40%,^79,80 myocardial ischemia,⁸¹ predisposition to atrial fibrillation while in normal sinus rhythm,⁸² and a host of structural cardiac disorders.⁷⁶ These studies based their methodology on using signal data, with tens to hundreds of thousands of patients included in their development. More recent developments, using similarly large populations, enable the detection of LVEF reduction from ECG images directly,⁶⁷ or ECGs available from portable or wearable devices.^83,84 A few recent studies have extended these directly in oncologic populations, such as predicting atrial fibrillation risk among 754 patients with leukemia using ECGs.⁸⁵ Other applications have enabled the detection of cancer therapeutics–related cardiac dysfunction, specifically LV systolic dysfunction and heart failure risk, in patients treated with anthracyclines and trastuzumab across 2 distinct studies with 3364 and 889 patients.^86,87

There are also examples of small, focused studies to address specific drug side effects, such as applying convolutional neural networks to quantify ECG alterations induced by sotalol to enhance the prediction of drug-induced torsades de pointes and diagnosis of congenital long QT syndrome.⁸⁸ Moving forward, similar ECG analysis algorithms may be applied to analyze cardiac conditions in patients with cancer and can be used as inputs for detecting cardiotoxicity.⁴⁹

Precision Medicine, Cardio-Oncology, and AI

Rather than a one-size-fits-all approach to medicine, the precision medicine model customizes health care (including medical decisions, treatments, preventive strategies, practices, or products) to one or more subgroups of patients, by incorporating individual genetic, environmental, and experiential variability (Figure 2).⁸⁹ Developments in oncology, cardiovascular precision medicine, and technology platforms highlight the emerging promise of personalized medicine to predict cardiotoxicity.⁹⁰

Figure 2. Data streams and suitability of artificial intelligence applications for cardio-oncologic care.

The schematic presents the broad range of data streams accessible for artificial intelligence (AI) applications. These have been presented across multiple axes, spanning the availability of these data in clinical practice, direct availability at the point of care, the biologic context behind their use, and the availability of AI tools in practice that leverage these data. The schematic provides a gradation along these axes, with the width of the bars proportional to the state of a given data stream across that axis. Blood biomarkers include troponins, brain natriuretic peptides, and other novel markers. Advanced cardiovascular (CV) imaging includes cardiac magnetic resonance imaging, computed tomography, and positron emission tomography.

As discussed previously, whereas measurable genetic factors are associated with cardiotoxicity risk,³⁷ they are yet to be sufficiently validated to be used in clinical practice. Biomarkers derived from molecular mechanisms of cardiotoxicity can be combined with clinical data to develop predictive algorithms.⁹¹ Indeed, most research efforts are focused on a more inclusive measure of cardiotoxicity stemming from multiple clinical and genetic factors evaluated as part of the big picture of a patient’s health, including common biomarkers pointing to potential risks of heart failure and cardiomyopathy.⁹² This potential is even more relevant with targeted and immune-based therapies, wherein the incidence of other cardiac events may be lower (ie, myocarditis, ventricular arrhythmias), but where cardiotoxicity development could be life-threatening or fatal.^93,94 Other studies combine outcomes of clinical factors, including patient history, imaging-based cardiac function, and multi-omics, into a more comprehensive risk assessment algorithm.⁹⁵

Some of the latest efforts are centered around large amounts of data generation to identify variants in genes, proteins, and metabolites. By combining bioinformatics tools with clinical phenotyping, studies have uncovered specific genetic or molecular patterns associated with cardiotoxicity risk.⁹⁶ Clinical observations can be combined with clinical variables, phenotypes,⁹⁷ EHR data, and data from multi-omic studies, including proteomics, genomics, epigenomics, transcriptomics, metabolomics, and immunomics, to generate comprehensive discrete data for AI analysis and subsequent risk prediction.

Integration of EHR-Based Algorithms and Registries Into Patient Care

As cardio-oncology continues to move into mainstream care of patients in the community, there is a growing need for informatics and analytic support for this burgeoning subspecialty, signified by implementations in commercial EHRs.⁹⁸ Numerous early research programs point to the value of EHR-based clinical decision support systems. Approaches vary from implementation of an EHR-based patient registry at the national level to aid clinical care and research⁹⁹ to building a registry for quality improvement, and from identifying care gaps¹⁰⁰ to producing various algorithms to assess CVD risk in eligible oncology patient cohorts.⁵³ The initial premise of these approaches is fundamentally the same: building programmatic rules to filter eligible oncology patients with relevant diagnosis codes and a record of chemotherapy, radiation therapy, targeted therapy, or immunotherapy regimens in the patient history, with the ultimate goal of improving patient care. Such tools can also be used to identify patients at increased risk of remote CVD due to cancer-directed therapies, such as women with early-stage breast cancer receiving estrogen-deprivation therapies. Using the precision medicine tools within the existing infrastructure would ultimately enhance collaboration between the specialties of cardiology and oncology, and assist these clinicians in the identification, management, and prevention of CVD among patients with cancer at highest risk of serious or fatal CVD.

Barriers to Integrating EHR-Based Algorithms Into Patient Care

Some barriers to the rapid development of AI and ML in cardio-oncology include erratic data sources with missing elements, difficulty in converting data from unstructured sources (such as clinical notes) or formats (such as LVEF ranges [eg, LVEF 45%–50%]), and disparate systems or sources involved with integrating relevant cardiotoxicity data. As discussed, these could potentially be addressed using large language model/GenAI technology. In parallel, the definition of cardiotoxicity continues to evolve, representing a moving target for developers of algorithms that rely on dynamic end points governing cardiotoxicity prediction. Another challenge is the presentation of the output from complex algorithms to a user-friendly format to be used by clinicians with limited time and level of trust for many lines of computer codes, also known as “algorithmic bias.”¹⁰¹ Therefore, it is essential that the synthesis of complex data is interpretable to the clinical end user and allows them to identify when the models may be less reliable.¹⁰²

The development and implementation of AI in medicine is a multidisciplinary field requiring buy-in and cooperation among several stakeholders, some of whom are clinicians and researchers.⁶² Creating precision medicine solutions that meet the needs of patients and health care professionals in cardio-oncology will require ongoing collaboration among not only clinicians and researchers but also industry, regulatory, and reimbursement entities.

ETHICAL CONSIDERATIONS AND CHALLENGES

The potential applications of AI in cardio-oncology are tremendous. However, there are unique ethical and legal considerations. These primarily center around security, data privacy and informed consent, algorithmic fairness and bias, and ensuring equitable access to technology (Table 4).⁶³ Nearly 1000 health care data breaches of ≥500 medical records have been reported to the US Department of Health and Human Services since 2022.¹⁰⁵ For instance, Google’s Project Nightingale AI platform leveraged data from health facilities in 21 US states without informed consent, although within the guidance of the Health Insurance Portability and Accountability Act. The platform raised a high degree of public concern about data privacy, resulting in legal action in 2019.^103,106 Similarly, in 2017, the UK Information Commissioner’s Office ruled that the Royal Free National Health Service Foundation Trust was in breach of the UK Data Protection Act of 1998 when it provided the personal data of >1.5 million patients to Google DeepMind for clinical safety testing of an application that aimed to assist with the diagnosis of acute kidney injury.^107,108 Furthermore, data from COVID-19 registry and other sources have identified race and sex bias in AI algorithms, stemming from limited and baseline biased raw data sets, which can potentially affect their downstream use, leading to biased outcomes.¹⁰⁹ These incidents highlight the vulnerable nature of AI systems and the need for oversight. Because most AI systems require patient data to inform models that more accurately predict outcomes, challenges around data privacy and equity will need to be revisited on a continuous basis. In 2023, the American Heart Association put forth principles regarding data sharing, patient privacy, and equity in the era of big data research.⁶³

Table 4.

Challenges in the Application of Artificial Intelligence–Based Systems in Cardio-Oncology and Beyond

Challenges	Examples	Potential solutions
Security	Nearly every year there is at least 1 large data breach affecting >1 million patients in the United States, including many patients with cancer or cardiovascular disease.103	Implement cybersecurity technology measures Protect mobile devices Strengthen network security Implement subnet wireless networks
Data privacy and informed consent	Every year there are multiple informed consent violations, affecting thousands of patients (ranging from unintentional laptop violations to questionable use of large patient data sets by unaffiliated for-profit companies), with resulting legal action.103	Require consent for data sharing at the time of index care initiation (eg, outpatient clinic) Create uniform definitions of low- and high-risk data
Algorithmic fairness and bias	A health care cost–focused AI system incorrectly identified some patients from underrepresented racial and ethnic groups as having lower likelihood for major cardiac events (eg, cardiovascular, cancer risks) due to historically income gap–driven lower use of health care resources.104	Use independent AI algorithm assessments to identify and reduce bias Create and validate AI data sets using diverse or truly representative population data
Ensuring equitable access to technology	Rural counties with low broadband Internet availability have less access to cardiovascular care and experience poorer outcomes.48	Establish public policy initiatives to enhance broadened access to Internet and telehealth services

AI indicates artificial intelligence.

The broad application of several AI-based algorithms in cancer and CVD has been limited by the relative absence of diverse and inclusive data sets for training (Table 2).¹⁰⁴ In cardio-oncology, these gaps may lead to ineffective prediction of disease despite increased integration of multi-omic approaches, which themselves have been largely derived from populations of European ancestry.¹¹⁰ To address these biases, it is imperative that training data sets be inclusive of underrepresented populations to ensure maximal scalability, accuracy, and clinical use. These challenges present a clear opportunity to dynamically improve the trajectory of care through equitable development and application of AI-based systems in cardio-oncology and beyond.

FUTURE DIRECTIONS AND RESEARCH OPPORTUNITIES

The emergence of AI and ML has the potential to improve cardiovascular management in patients with cancer; however, further research is needed to incorporate the implementation of these novel technologies into clinical practice (Table 1).

A key challenge is the heterogeneity in previous research regarding patient populations, cancer phenotypes, CVD risk profiles, definitions of treatment-related cardiotoxicity, biomarker assays, and imaging modalities, which coalesce to hinder the ability to compare research findings across studies.⁴⁸ Much of the previous research has focused on cardiac dysfunction defined by changes in LVEF; as cancer therapeutics evolve, definitions should reflect the range of potential cardiotoxic effects and pathophysiologic mechanisms.^111-113 In addition, large prospective cohort studies are needed to strengthen the evidence supporting existing measures while allowing for adequate power to identify novel markers and multi-omic targets. Cohort studies should incorporate deep cardiac and cancer phenotyping to identify and validate potential shared biologic markers, given the high degree of shared risk factors between CVD and cancer.¹¹⁴ Furthermore, ML- and AI-based approaches could be integrated with polygenic risk scores to predict susceptibility to cardiotoxicity.

Due to the high cost of multi-omic measurements and the need for large data sets, the development of a national cardio-oncology database registry could alleviate many of these limitations and facilitate the development of appropriate and representative ML algorithms to further the application of AI in cardio-oncology research and patient care. In the process, it is important to recognize that a substantial portion of cancer care is provided in community hospitals that may not have the technologic or financial resources to participate in national registries. It is therefore crucial for such a national database to be purposeful in accurately representing the diverse patient population treated at these institutions and thereby mitigate selection bias in the creation of such registries. The American Heart Association is well-poised to lead these efforts with available funding resources and appropriate expertise to advance the field of cardio-oncology.

As ML models are developed, careful consideration is needed to ensure training data sets have adequate racial and ethnic representation to promote health equity.^1,115 Moreover, in addition to ethical and secure development, there is a need to enable the consistent use of fair and reliable algorithms with full awareness of patients.¹¹⁶ As further multi-omic data are generated, scaling and cost strategies should be considered for meaningful implementation into clinical practice. Multi-omic measurements require specialized equipment, trained personnel, large data storage, and clinician education, which are lacking.⁴⁸ ML models should be monitored and re-trained as necessary to reflect the rapidly changing landscape of cardio-oncology. Clinical trials will be needed to test the clinical use of AI and ML models in predicting and diagnosing cardiovascular outcomes in patients with cancer. This is particularly important with the broadening scope of AI and its application across the full spectrum of discovery to implementation in cardiovascular care.¹¹⁶

CONCLUSION

Advances in AI can accelerate the achievement of necessary precision in cardio-oncologic care, addressing the numerous underlying factors that drive the risk of an individual with cancer. The emergence of novel therapies is a boon for patients with cancer, and adequately addressing the cardiac risk of these treatments can enable their long-term health and well-being. The growth in data streams with complementary information is now matched by technology that can adequately leverage these data, with key applications of AI exemplifying and forecasting a future with AI-powered and personalized cardio-oncologic care.

The typical patient with cancer with CVD is complex in myriad ways linked to the type of cancer, cancer stage, treatment combination and modality, cardiovascular risk factors, and biomarkers linked to the disease, with an endless assortment of disease risks. In cardio-oncology, AI can aid in clinical efficiency for the complex population of patients with CVD and cancer, recognize targeted therapies and drug–drug interactions, and potentially identify future areas of research focus based on existing data sets. This is a call to action for national organizations tailored to CVD and cancer on the need for a modern nationwide registry for patients with cancer with CVD that encompasses these novel data streams. By doing so, we have the potential to leverage technology to more accurately address the cardiovascular morbidity and mortality associated with cancer treatment, thereby enabling patients with cancer to achieve longer and healthier lives.