Artificial Intelligence to Enhance Precision Medicine in Cardio-Oncology

Authors, Journal, Affiliations, Type, DOI

• Authors: Rohan Khera, Aarti H. Asnani, Jacob Krive, Daniel Addison, Han Zhu, Alexi Vasbinder, Matthew R. Fleming, Rima Arnaout, Pedram Razavi, Tochukwu M. Okwuosa; on behalf of AHA Cardio-Oncology and Data Science and Precision Medicine Committees
• Journal: Circulation: Genomic and Precision Medicine
• Type: AHA Scientific Statement
• Date: April 2025
• DOI: 10.1161/HCG.0000000000000097

Overview

This 2025 AHA Scientific Statement provides a comprehensive framework for applying AI and machine learning to precision medicine in cardio-oncology. It surveys the full biomarker landscape — advanced imaging (echo, CMR, cardiac CT, nuclear), traditional biomarkers (troponins, BNP), and multi-omics (genomics, transcriptomics, proteomics, metabolomics) — describing how AI can integrate these data streams into individualized cardiotoxicity risk prediction and management. Key practical AI applications covered include automated LVEF measurement, ECG-based cardiotoxicity detection algorithms, and NLP for EHR data extraction. The statement calls for a national cardio-oncology registry and flags critical ethical challenges including algorithmic bias, health equity, and data privacy.

Keywords

Artificial intelligence, cardio-oncology, machine learning, precision medicine, cardiotoxicity, multi-omics, biomarkers

Key Takeaways

Advanced Imaging in Cardio-Oncology

• Core imaging modalities: echocardiography (LVEF, GLS, diastolic function, valve), CMR (T1/T2 mapping, LGE, ECV), coronary CT (CAC, CCTA, CT-FFR), nuclear imaging (myocardial perfusion, Tc-99m PYP).
• ICI myocarditis — CMR: Native T1 elevation was highly predictive of future MACE in 136 patients with suspected ICI myocarditis; LGE and T2-weighted STIR are also diagnostic.
• SUCCOUR trial: GLS-guided cardioprotection was equivocal for predicting future HF events in anthracycline/HER2-targeted therapy patients — GLS did not outperform standard care.
• Ibrutinib cardiotoxicity: Abnormal LA strain/size by echo and elevated native T1/T2 by CMR were highly predictive of future MACE and adverse events in patients with haematologic malignancy on ibrutinib.
• AI-based CAC scoring: Deep learning algorithm for CAC from chest radiotherapy images predicted future acute coronary events.
• CCTA: Captures increased rates of CAD requiring intervention after chest radiotherapy and chemotherapy.

Traditional Biomarkers (Troponins, BNP)

• Cardiac troponins and BNP have been evaluated for predicting CTRCD; evidence is inconsistent in more recent studies using high-sensitivity assays in early-stage breast cancer patients.
• High negative predictive value (84–100%): Most useful to rule out cardiotoxicity; no established cancer-specific cut points.
• Biomarker-guided cardioprotection: Serial troponin to guide enalapril initiation showed no benefit over universal enalapril (RCT, n=276, limited by small size and low anthracycline dose). Candesartan/carvedilol guided by hsTnI also did not prevent LV dysfunction (prospective multicenter RCT).
• ICI myocarditis: Serial hsTnI can detect ICI myocarditis in both symptomatic and asymptomatic patients; relative elevation in cardiac troponin T prognostic for MACE.
• ESC guidelines recommend baseline cardiac biomarker measurement in all patients receiving potentially cardiotoxic agents.

Multi-Omics Biomarkers

Genomics

• SNPs associated with anthracycline cardiotoxicity identified in several pathways (Table 2); limited by small sample sizes and lack of consensus cardiotoxicity definition.
• TTNtv: Titin-truncating variants associated with anthracycline cardiomyopathy (similar to DCM overlap).
• RARG: Variant associated with anthracycline cardiomyopathy in multiple cohorts.
• CHIP: Clonal haematopoiesis of indeterminate potential — listed as a genomic biomarker.
• Insufficient validation for routine clinical use in predicting cardiotoxicity.

Transcriptomics

• Single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics map immune cell subsets in ICI myocarditis.
• scRNA-seq + TCR sequencing + mass cytometry: Identified pathogenic T-cell populations (α-myosin–specific T cells) in ICI myocarditis.
• Spatial transcriptomics would add mechanistic insights into cardiotoxicity distribution within cardiac tissue.

Proteomics

• High-throughput proteomics (SomaScan, Olink, CITE-seq) applied to pilot cohorts receiving anthracyclines and carfilzomib.
• Hemopexin: Identified as a biomarker in anthracycline cardiotoxicity.
• CSK (C-terminal Src kinase): Identified in ibrutinib-associated atrial fibrillation.

Metabolomics

• Targeted LC/MS metabolite profiling in breast cancer patients with/without anthracycline cardiotoxicity.
• Early changes in tricarboxylic acid (TCA) cycle metabolites differentiated subsequent cardiotoxicity development.
• Patients with cardiotoxicity showed pronounced alterations in purine and pyrimidine metabolism.

AI-Based Risk Prediction Models

• Classification models trained on 4,309 cancer patients (laboratory, symptoms, demographics, echo) for 6 CV outcomes — proof of concept for CV risk stratification in oncology.
• Breast cancer model: included treatment-related (anthracycline/radiation) and conventional (hypertension, DM) risk factors; predicted MACE at up to 7 years.
• Lung cancer CT scan model: 30,286 low-dose CT scans; predicted CVD mortality risk, outperformed existing models.
• Comprehensive integrative model (troponin, BNP, cancer + CV risk factors): risk discrimination in the 90% range across multiple cohorts.
• Exome sequencing (289 childhood cancer survivors, anthracycline): differentially enriched variants improved LVEF-decline prediction when combined with clinical predictors.
• Major limitation: Systematic review of cardiotoxicity prediction models for breast cancer found only 1 of 6 studies used external validation — generalisability uncertain.

AI-Enabled Cardiac Imaging Analysis

• Automated LVEF measurement: randomised clinical trial showed benefits for AI-based LVEF assessment.
• Research ongoing for automated LV strain, diastolic function, myocardial perfusion.
• Foundation models and semi-supervised/self-supervised learning reduce need for manual image labelling.
• AI can identify novel imaging biomarkers not observable by human experts; multimodal composite biomarkers possible with appropriate datasets.

NLP and Large Language Models for Clinical Data Extraction

• NLP supplements discrete EHR data with unstructured clinical note content for cardiotoxicity risk.
• LLMs can extract structured cardiotoxicity data from EHR (erratic data, missing elements, unstructured formats).
• Smart chatbots can enable clinicians to query diverse electronic data sources directly.
• Potential to identify high-risk patients for cardio-oncology referral before oncologic treatment.
• No specific cardio-oncology LLM implementations yet documented in published literature (as of 2025).

AI-Driven Drug Discovery

• ML algorithms integrated into drug screening, toxicity prediction, and manufacturing pipeline.
• AI holds promise for predicting toxicity of existing drugs and discovering therapeutic approaches to treat cardiotoxicity.
• Currently a future direction; experimental validation is required alongside AI predictions.

AI-Enabled ECG Algorithms for Cardiotoxicity Detection

• Convolutional neural networks used to analyse ECG signals for cardiotoxicity.
• ICI myocarditis: ECG features early during cancer treatment predictive of life-threatening arrhythmias and mortality (125 ICI myocarditis cases + 50 transplant rejection controls, 49 institutions, 11 countries).
• LV systolic dysfunction/HF risk in anthracycline/trastuzumab patients: AI-ECG detected cancer therapeutics-related cardiac dysfunction in studies of 3,364 and 889 patients respectively.
• AF prediction in leukemia: ECG-based AI predicted AF risk in 754 patients with leukemia.
• Outside cardio-oncology, AI-ECG detects LVEF <40%, myocardial ischaemia, AF during sinus rhythm, and structural cardiac disorders.
• Convolutional neural networks applied to quantify ECG alterations from sotalol — enhanced prediction of drug-induced torsades de pointes and LQTS diagnosis.

Integration of EHR-Based Algorithms and Registries

• EHR-based clinical decision support systems: patient registry at national level for research + care; quality improvement; care gap identification; CVD risk algorithms.
• Programmatic rules filter eligible oncology patients with relevant diagnosis codes + chemotherapy/RT/targeted/immunotherapy records.
• Tools can identify women with early-stage breast cancer on estrogen-deprivation therapy at increased remote CVD risk.

Barriers to AI Integration

• Erratic data with missing elements; difficulty converting unstructured/range data; disparate data systems.
• Evolving cardiotoxicity definitions — moving target for algorithm developers.
• Algorithmic output must be interpretable to clinicians ("algorithmic bias" = opaque outputs).
• Requires multidisciplinary buy-in across clinicians, researchers, industry, regulators, and payers.

Ethical Considerations

• Multi-omics datasets largely derived from populations of European ancestry — limits scalability across diverse populations.
• AHA 2023 principles for data sharing, patient privacy, and equity in big data research.

Future Directions

• National cardio-oncology database registry needed: alleviates multi-omic cost and small dataset limitations; must represent community hospital populations to avoid selection bias.
• Large prospective cohort studies with deep cardiac and cancer phenotyping.
• ML + polygenic risk scores for cardiotoxicity susceptibility prediction.
• Clinical trials to test AI/ML models for predicting and diagnosing CV outcomes in patients with cancer.
• Continuous model re-training as cardio-oncology landscape evolves.

Limitations of the Document

• Scientific Statement rather than guideline; no formal GRADE or COR/LOE methodology.
• Evidence base is heterogeneous: varied populations, inconsistent cardiotoxicity definitions, small cohorts, limited external validation.
• Most AI models reviewed are proof-of-concept with no prospective clinical validation.
• Multi-omics literature largely derived from European-ancestry populations — generalisability limited.
• Rapidly evolving field; some cited applications may be superseded quickly.

Key Concepts Mentioned

• concepts/Cardio-Oncology — overarching discipline framework; AI as future priority
• concepts/Cancer-Therapy-Related-CV-Toxicity — ICI myocarditis imaging, biomarker-guided cardioprotection, multi-omics
• concepts/AI-in-Cardio-Oncology — primary concept page for this source
• concepts/Late-Gadolinium-Enhancement — ICI myocarditis T1/T2 mapping
• concepts/HFA-ICOS-Risk-Stratification — baseline risk stratification context
• concepts/Torsades-de-Pointes — sotalol ECG AI quantification for TdP prediction
• concepts/Drug-Induced-Arrhythmia — drug-induced ECG changes and AI prediction

Key Entities Mentioned

• entities/Atrial-Fibrillation — ibrutinib-induced AF; AI-ECG AF prediction in leukemia
• entities/DCM — TTNtv overlap with anthracycline cardiomyopathy
• entities/Heart-Failure — CTRCD endpoint; AI HF risk prediction

Wiki Pages Updated

• wiki/sources/ai-cardiooncology-aha-2025.md — created
• wiki/concepts/AI-in-Cardio-Oncology.md — created
• wiki/concepts/Cardio-Oncology.md — updated (AI precision medicine section)
• wiki/concepts/Cancer-Therapy-Related-CV-Toxicity.md — updated (multi-omics, AI ECG, biomarker-guided therapy RCTs)
• wiki/concepts/Late-Gadolinium-Enhancement.md — updated (ICI myocarditis CMR data)
• wiki/wikiindex.md — updated
• wiki/sourceindex.md — updated