Predicting Tolerance to Anthracycline Chemotherapy Using Electrocardiogram-Based Artificial Intelligence in Sarcoma

Mayo Clinic Proceedings: Digital Health 2025 AI 8 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Anthracyclines, Cardiotoxicity, and the Need for Smarter Monitoring in Sarcoma

Anthracycline-based chemotherapy, principally doxorubicin, is the cornerstone first-line systemic treatment for most bone and soft tissue sarcomas. Regimens such as doxorubicin/ifosfamide and doxorubicin/cisplatin/mitomycin achieve the best response rates in localized and metastatic disease, but they carry a well-characterized risk of cardiac damage. Cumulative anthracycline doses above 200 mg/m2 are associated with meaningful increases in long-term cardiovascular toxicity, and sarcoma treatment regimens frequently exceed 400 mg/m2, placing patients firmly in a high-exposure category.

Why cardiac monitoring matters in this population: Anthracycline-induced cardiomyopathy manifests as a decline in left ventricular ejection fraction (LVEF), which can progress to clinical heart failure if undetected. Standard practice in modern sarcoma trials incorporates dexrazoxane as a cardioprotectant and serial echocardiography to track LVEF changes. However, repeat echocardiography is resource-intensive, not universally available, and adds cost and logistical burden for patients already on demanding treatment schedules. A cost-effective, readily available alternative screening tool would be clinically valuable.

Performance status as a surrogate for treatment tolerance: Oncologists currently use ECOG performance status or Karnofsky performance status to estimate a patient's likelihood of tolerating cancer-directed therapies. These instruments are useful but imprecise: they are limited by interobserver variability and cannot predict organ-specific toxicities such as anthracycline-related cardiomyopathy. They provide no information about the cardiac reserve of an individual patient, which is the dominant factor determining whether cumulative drug exposure will translate into symptomatic cardiac dysfunction.

This paper, from Mayo Clinic Rochester and published in Mayo Clinic Proceedings: Digital Health in 2025, asks whether AI-enabled ECG analysis can fill this gap. The ECG is universally available, inexpensive, and already collected in most oncology workflows. Two AI-ECG capabilities are tested: prediction of biological age from ECG morphology, and detection of reduced LVEF from ECG waveforms. The study population is 40 sarcoma patients treated with anthracyclines between 2006 and 2020.

TL;DR: Sarcoma treatment routinely delivers anthracycline doses above 200 mg/m2 (often exceeding 400 mg/m2), creating substantial cardiotoxicity risk. This retrospective Mayo Clinic study tests whether AI-ECG predictions of biological age and LVEF can predict chemotherapy tolerance in 40 sarcoma patients treated from 2006 to 2020, as a low-cost alternative to serial echocardiography.
Pages 2-3
How AI Extracts Age and Ejection Fraction Predictions from an ECG Tracing

The AI-ECG platform used in this study was developed and validated at Mayo Clinic using convolutional neural networks (CNNs) trained on large datasets of ECG tracings paired with known patient outcomes. Two distinct prediction models are relevant here. The first predicts the patient's "biological age" from ECG morphology, producing a numerical age estimate that reflects the electrical and structural health of the heart rather than chronological age. The second predicts the probability of a low LVEF, outputting a continuous score where higher values indicate a greater likelihood of ejection fraction below 50% or 35%.

The biological age concept: Biological age predicted from an ECG represents how "old" the heart appears electrically, capturing subclinical changes in conduction, repolarization, and myocardial health that accumulate with cardiovascular disease, metabolic comorbidities, and aging. In the general population, large discrepancies between AI-ECG-predicted age and chronological age (more than 7.4 years older) have been associated with adverse cardiovascular outcomes. In patients with Lamin A/C (LMNA) mutations, accelerated ECG aging rates have been documented using this framework. The hypothesis in this study is that patients whose hearts appear biologically older than their chronological age will tolerate anthracyclines less well.

The low-EF prediction: The AI-ECG low EF model was developed to screen for reduced systolic function from a standard 12-lead ECG. It has previously demonstrated strong performance for detecting LVEF below 50% in general cardiology populations, and a separate publication documented its ability to identify newly abnormal LVEF after anthracycline therapy in a different cancer cohort. In breast cancer and non-Hodgkin lymphoma patients, the low EF predictor tracked treatment-induced LVEF declines, raising the possibility that it could serve as a non-invasive monitoring tool during anthracycline exposure in sarcoma.

ECG aging rate, a novel metric computed in this study, is defined as the ratio of (change in AI-ECG age before versus after treatment) divided by (change in chronological age over the same interval). A ratio greater than 1 indicates accelerated cardiac aging relative to elapsed calendar time, and a ratio less than 1 indicates slower-than-expected aging. Cumulative anthracycline dose is hypothesized to drive accelerated ECG aging in susceptible patients.

TL;DR: Two CNN-based AI-ECG models are applied: a biological age predictor (flagging patients whose hearts appear greater than 7.4 years older than their chronological age) and a low-LVEF probability score (validated against echocardiography-based EF). A novel ECG aging rate metric, defined as the ratio of AI-ECG age change to chronological age change during treatment, is introduced to capture anthracycline-driven cardiac aging.
Pages 3-4
Patient Selection, Data Collection, and Statistical Design

The study enrolled adults (age 18 or older) diagnosed with sarcoma and treated with any anthracycline at Mayo Clinic Rochester between December 2006 and October 2020. Eligibility required an ECG obtained within 1 year prior to treatment initiation. When multiple ECGs were available, the study selected the ECG closest to treatment start, unless the interval between that ECG and the pretreatment echocardiogram exceeded 180 days, in which case the ECG closest to the echocardiogram was chosen. This pairing rule ensured that the AI-ECG predictions reflected cardiac status proximate to the echocardiogram reference standard.

Data abstracted: The electronic medical record review captured patient demographics, cancer histology and stage, treatment regimens with lifetime cumulative anthracycline doses, chemotherapy tolerance (defined as rates of dose reductions, treatment delays, and early discontinuation), cardiac comorbidities (hypertension, diabetes, heart failure, myocardial infarction, obesity, hyperlipidemia), and echocardiogram reports. ECGs within 1 year of treatment completion were also collected for post-treatment analysis, enabling calculation of longitudinal AI-ECG aging rates. The study was approved by the Mayo Clinic Institutional Review Board with HIPAA waiver (ID 23-007295).

Echocardiographic endpoints: Echocardiograms were selected within 1 year of treatment initiation or discontinuation. LVEF was abstracted from each pre- and post-treatment echocardiogram. Global longitudinal strain (GLS) was also recorded per American College of Cardiology guidelines for monitoring chemotherapy-related LV systolic dysfunction: GLS less than -18% is normal, -16 to -18% is borderline, and greater than -16% (less negative) is abnormal. A valid ECG-echocardiogram pair required both studies within 180 days of each other.

Statistical methods: Welch two-sample t-tests with unequal variance compared means between groups. Binomial logistic regression, adjusted for biological age, modeled chemotherapy outcomes against AI-ECG age deviation. Odds ratios for accelerated aging groups used Fisher exact tests given small cell sizes. Kaplan-Meier curves with Cox proportional hazards regression compared overall survival (OS) and relapse-free survival (RFS) between the ECG aging groups. All analyses were performed in R version 4.3.2.

TL;DR: Retrospective single-center cohort (Mayo Clinic Rochester, 2006-2020) of 40 adult sarcoma patients with anthracycline treatment and a pre-treatment ECG within 1 year. Primary outcomes were chemotherapy dose reductions, treatment delays, and early discontinuation. Secondary outcomes included pre/post-treatment LVEF change and concordance between AI-ECG low-EF prediction and echocardiography. Valid ECG-echo pairs required studies within 180 days. Analysis used logistic regression, Fisher exact tests, and Cox proportional hazards models.
Pages 4-6
Cohort Demographics: A Heterogeneous Sarcoma Population with High Cardiovascular Comorbidity Burden

Forty patients met eligibility criteria. The cohort was predominantly male (27 patients, 67.5%), with a median age at diagnosis of 56.5 years (range 18-76 years, mean 51.7 plus or minus 17.5 years). The age range captures both younger patients with bone sarcomas (osteosarcoma, Ewing sarcoma) and older patients with soft tissue subtypes. Most patients (33 patients, 82.5%) had localized disease at treatment initiation, while 7 (17.5%) had metastatic disease.

Histological distribution: Osteosarcoma was the most common subtype (12 patients, 30%), followed by undifferentiated pleomorphic sarcoma (UPS, 6 patients, 15%), Ewing sarcoma (4 patients, 10%), myxofibrosarcoma (3 patients, 7.5%), sarcoma NOS (3 patients, 7.5%), chondrosarcoma (2 patients, 5%), leiomyosarcoma (2 patients, 5%), and pleomorphic rhabdomyosarcoma (2 patients, 5%). Rarer subtypes, each representing a single patient (2.5%), included alveolar soft part sarcoma, myxoid liposarcoma, pleomorphic sarcoma, pleomorphic liposarcoma, spindle cell sarcoma, and synovial sarcoma. This distribution reflects the referral patterns of a major sarcoma center and is broadly representative of the histological landscape in adult sarcoma.

Cardiovascular comorbidities: The cohort carried a substantial cardiovascular comorbidity burden. Hypertension was present in 22 patients (55%), obesity (BMI greater than 30) in 20 patients (50%), hyperlipidemia in 17 patients (42.5%), diabetes in 6 patients (15%), prior heart failure in 2 patients (5%), and prior myocardial infarction in 2 patients (5%). These comorbidities are themselves independent risk factors for anthracycline cardiomyopathy, making the cohort a clinically meaningful group for testing cardiotoxicity prediction tools.

Treatment and cardioprotection: The two most common anthracycline regimens were doxorubicin/ifosfamide (15 patients, 37.5%) and doxorubicin/cisplatin/mitomycin (12 patients, 30%). The mean cumulative anthracycline dose was 215 mg/m2 (plus or minus 140.5 mg/m2), with a median of 180 mg/m2 and a wide range from 30 to 525 mg/m2. Only 4 patients (10%) received dexrazoxane: three for prior anthracycline exposure or high cumulative dose concerns, and one for a history of heart failure with reduced EF that had recovered before treatment. Radiation therapy was given to 28 patients (70%) and surgery to 34 patients (85%), consistent with multimodality management of localized sarcoma.

TL;DR: 40 sarcoma patients, 67.5% male, median age 56.5 years (range 18-76). Most common histologies: osteosarcoma (30%), UPS (15%), Ewing sarcoma (10%). High cardiovascular comorbidity burden: hypertension 55%, obesity 50%, hyperlipidemia 42.5%. Mean cumulative anthracycline dose 215 mg/m2, only 10% received dexrazoxane cardioprotection. Median pretreatment LVEF was 64%, post-treatment 57.5%.
Pages 6-8
AI-ECG Biological Age Does Not Predict Chemotherapy Tolerance

The distribution of the difference between AI-ECG-predicted age and chronological age was approximately normally distributed around zero, indicating that on average the AI model was not systematically biased toward overestimating or underestimating age in this cohort. However, a strong and significant relationship emerged between the AI-ECG age gap and baseline chronological age (P less than 0.001): younger patients tended to be predicted as biologically older, while older patients tended to be predicted as younger. This regression-to-the-mean phenomenon in AI-ECG age prediction has been observed in prior publications using the same algorithm, and the study authors controlled for chronological age in all tolerance analyses.

No relationship with chemotherapy tolerance: After adjusting for chronological age, having a higher AI-ECG-predicted age prior to treatment (appearing biologically older) was not associated with subsequent need for dose reductions (OR 0.25, P=0.45), treatment delays (OR 1.20, P=0.91), or early discontinuation of chemotherapy (OR 1.0, P greater than 0.99). These null results were consistent across all three tolerance endpoints, strongly suggesting that pretreatment biological age deviation does not capture the physiological factors driving poor chemotherapy tolerance in sarcoma patients.

ECG aging rate and dose trends: Twenty-four patients (60%) had post-treatment ECGs within 1 year of treatment completion, enabling longitudinal aging rate calculations. Unlike pretreatment AI-ECG age, the ECG aging rate through treatment showed a meaningful but not statistically significant trend toward higher cumulative anthracycline dose in patients with accelerated ECG aging (ratio greater than 1, P=0.06). This suggests that the cardiac electrical signature of anthracycline exposure may be detectable in the ECG aging trajectory, even if pretreatment age deviation is not predictive.

No survival difference by aging rate: Neither overall survival (P=0.98) nor relapse-free survival (P=0.67) differed between patients with accelerated versus decelerated ECG aging rates, as shown by Kaplan-Meier analysis. While the cohort is small and these analyses are underpowered, the absence of a signal is notable. ECG aging rate, in the current sample size, is not associated with oncologic outcomes, though its potential as a monitoring biomarker for cardiac tolerance remains under investigation.

TL;DR: Pretreatment AI-ECG age deviation does not predict dose reductions (OR 0.25, P=0.45), treatment delays (OR 1.20, P=0.91), or early discontinuation (OR 1.0, P greater than 0.99) after adjusting for chronological age. Longitudinal ECG aging rate trended with cumulative anthracycline dose (P=0.06) in 24 patients with paired ECGs, but neither OS (P=0.98) nor RFS (P=0.67) differed by aging group. The null tolerability findings may reflect that noncardiac toxicities dominate treatment-limiting adverse events in this cohort.
Pages 8-10
AI-ECG Ejection Fraction Prediction: Strong Screening Performance and Longitudinal Concordance

The most clinically significant finding of this paper is the performance of the AI-ECG low LVEF prediction model. Using a previously established threshold of 0.256, the AI-ECG low EF predictor demonstrated a sensitivity of 100% and a specificity of 94% for identifying patients with an echocardiography-measured LVEF below 50% prior to anthracycline treatment. For the more severe threshold of LVEF below 35%, no patients in the cohort had this finding at baseline, so sensitivity could not be calculated, but the specificity was 92%.

What these numbers mean clinically: A sensitivity of 100% means the AI-ECG did not miss a single patient with reduced LVEF (no false negatives). A specificity of 94% means that 94 out of every 100 patients with normal LVEF would correctly screen negative, with only 6% undergoing unnecessary follow-up imaging. For a pre-treatment screening tool intended to flag patients who should not receive cardiotoxic therapy without additional workup, near-perfect sensitivity is the priority metric, and the AI-ECG model achieves this. The specificity performance also limits unnecessary echocardiogram referrals to a manageable rate.

Longitudinal EF tracking after treatment: Six patients had paired ECGs and echocardiograms both before and after anthracycline treatment, allowing direct comparison of AI-ECG EF score changes with echocardiographic LVEF changes. The average change in echocardiographic LVEF across these 6 patients was -6.5%. Four patients had less than 10% decline in LVEF and showed no meaningful change in their AI-ECG low EF prediction score. The 2 patients who experienced a clinically significant LVEF drop of greater than 10% were precisely the 2 patients whose AI-ECG low EF prediction score increased by more than 0.2 units after treatment. No false alarms or missed declines occurred in this small series.

Context from parallel cancer populations: This concordance pattern mirrors results from breast cancer and non-Hodgkin lymphoma cohorts where the same AI-ECG algorithm was applied. A 2024 publication in the European Journal of Preventive Cardiology demonstrated that the AI-ECG low LVEF prediction increased in breast cancer patients who developed anthracycline-induced cardiotoxicity. A 2024 Nature Communications paper showed AI-ECG could predict chemotherapy-induced cardiotoxicity from baseline ECGs across cancer types. The sarcoma data, while small, extend this pattern to a new and particularly high-risk disease context.

TL;DR: AI-ECG achieved 100% sensitivity and 94% specificity for detecting pre-treatment LVEF below 50%, and 92% specificity for LVEF below 35%. In 6 patients with paired pre/post-treatment ECG-echo data, the 2 patients with greater than 10% LVEF decline were exactly the 2 with AI-ECG low-EF score increases greater than 0.2. The 4 patients with stable EF showed stable AI-ECG scores. These results parallel findings in breast cancer and lymphoma cohorts.
Pages 10-11
Methodological Constraints That Shape Interpretation

Small retrospective cohort with selection bias: The most fundamental limitation is the sample size of 40 patients from a single institution. This cohort was further constrained by requiring an ECG within 1 year of treatment start, a criterion that introduces selection bias because ECGs are not routinely obtained prior to sarcoma chemotherapy. Patients who happened to have an ECG available may be systematically different (more medically complex, referred for cardiac evaluation, older, or with more cardiovascular comorbidities) from the broader sarcoma population, limiting generalizability.

Low cumulative dose relative to standard sarcoma regimens: The mean cumulative anthracycline dose in this cohort was 215 mg/m2 (median 180 mg/m2). This is substantially below the 450 mg/m2 that patients with advanced soft tissue sarcoma typically accumulate over the course of treatment. Higher cumulative doses are associated with greater cardiotoxicity risk, and the relatively low average exposure in this study may have attenuated the detectable signal for both EF decline and ECG aging. Results might differ meaningfully in a cohort receiving higher cumulative doses over longer treatment courses.

AI-ECG age algorithm trained on a general, not healthy, population: The underlying CNN for age prediction was developed on a general clinical patient population, which contains individuals with subclinical cardiovascular disease and various comorbidities. This means the predicted biological age already incorporates some degree of illness burden, potentially blunting its discriminative ability in a cancer population where comorbidities are common and variable. The regression-to-the-mean phenomenon (younger patients predicted older, older patients predicted younger) observed in this study has also been noted in other applications of this algorithm and complicates interpretation of individual predictions.

Small longitudinal sample for EF tracking: Only 6 patients had valid paired ECG-echocardiogram data before and after treatment. While the concordance within this tiny sample was perfect (2 of 2 declines detected, 4 of 4 stable cases confirmed), no statistical inferences can be drawn from 6 patients. The 180-day pairing window also means that unmeasured changes in clinical status between the ECG and echocardiogram could introduce noise into the comparison. The absence of left-chest radiation in the 2 patients who declined EF is noted, but confounding from other treatment components cannot be excluded.

TL;DR: Key limitations include the small n=40 single-center retrospective design with selection bias from ECG availability requirement, a mean cumulative anthracycline dose of only 215 mg/m2 (well below the 450 mg/m2 typical in advanced sarcoma), an AI-ECG age model trained on a general (not cancer-specific) population, and only 6 paired ECG-echo cases for the longitudinal EF tracking analysis. Results should be considered hypothesis-generating rather than definitive.
Pages 11-12
Clinical Implications and the Path Toward Broader Validation

AI-ECG as a low-cost screening layer: The most actionable near-term implication of this study is the use of AI-ECG as a pre-treatment cardiac screening tool in sarcoma patients. Given the 100% sensitivity and 94% specificity for detecting LVEF below 50% from a standard ECG, incorporating AI-ECG analysis into pre-chemotherapy workup could identify patients requiring echocardiographic confirmation before proceeding to anthracyclines. This would create a stratified screening workflow: AI-ECG for all patients, full echocardiography reserved for those flagging positive, potentially reducing the volume of baseline echocardiograms without sacrificing sensitivity for reduced EF detection.

Monitoring during treatment: The longitudinal concordance between AI-ECG low-EF score changes and echocardiographic LVEF changes in the 6 paired cases suggests that serial AI-ECG monitoring during anthracycline treatment could serve as a between-echocardiogram signal for LVEF decline. Rather than performing echocardiograms at every cycle, routine ECGs (already standard for many chemotherapy regimens for QT monitoring) could be analyzed in real time with AI, triggering a formal echocardiogram only when the low-EF prediction score rises. This would reduce imaging burden while potentially catching cardiotoxicity earlier than current fixed-interval monitoring protocols.

ECG aging as a systemic marker of treatment effect: The trend between accelerated ECG aging and higher cumulative anthracycline dose (P=0.06) opens an interesting question about whether ECG aging rate could track the systemic biological impact of cancer-directed therapy more broadly. Several studies have documented accelerated aging at the molecular level (p16INK4a expression, DNA methylation clock acceleration, epigenetic aging biomarkers) in patients receiving chemotherapy and radiotherapy. The AI-ECG aging rate could represent an accessible clinical proxy for these molecular aging processes, providing a metric that is obtainable in any clinical setting without specialized laboratory assays.

What the next studies need: Definitive validation requires a prospective, multicenter study with a larger sarcoma-specific cohort, standardized ECG collection at predefined timepoints, complete echocardiographic follow-up, and uniform anthracycline regimens. Capturing patients who receive cumulative doses in the 400-600 mg/m2 range is essential for testing the limits of the AI-ECG's predictive ability. Integration with circulating biomarkers (troponin, BNP) and patient-reported outcomes would further characterize the AI-ECG's place in a multimodal cardio-oncology monitoring framework for this particularly high-risk disease setting.

TL;DR: The authors propose AI-ECG as a pre-treatment screen (replacing or triaging baseline echocardiography) and as a between-cycle monitoring signal during anthracycline therapy. The trending ECG aging rate association with cumulative dose (P=0.06) motivates future work on ECG aging as a proxy for treatment-related biological aging. Definitive validation will require prospective multicenter studies with complete echocardiographic follow-up in patients receiving full-dose sarcoma anthracycline regimens (400+ mg/m2).