Before starting a new tyrosine kinase inhibitor (TKI) or antibody-drug conjugate (ADC) or opening a trial arm, oncologists often rely on cardiac screening to determine whether the treatment is safe.
In oncology, cardiac assessments are often a rate-limiting step.
But what if that ECG predicted risk, flagged structural heart disease, and helped oncologists make faster, better treatment decisions?
That is now upon us. Artificial intelligence (AI) is not just interpreting ECGs or automating echocardiography worksheets; it's redefining how we approach cardiovascular clearance in oncology.
Here, we explore how AI innovations in cardiology are creating new opportunities in oncology care, research, and equity.
Cardio-Clearance Is Now a Clinical Bottleneck
In oncology, patients require numerous cardiac assessments to determine whether they can tolerate a new treatment.
Patients undergo heart rate monitoring via QTc surveillance before receiving oral TKIs like lorlatinib or osimertinib, have baseline ejection fraction assessments for HER2-positive regimens such as trastuzumab-deruxtecan, and have cardiac clearance evaluations prior to procedures like biopsies or tumor resections.
These cardiac assessments can delay care, affect access and equity, and slow eligibility screening for clinical trials. For instance, patients can face long wait times for treatment due to high demand for cardiology testing, or those in rural settings may have no or limited local access. Insurance approvals and prior authorizations are yet other hurdles that disproportionately affect patients with limited or no coverage.
Now with AI, we may be able to transform these rate-limiting steps into opportunities for earlier detection, decentralized screening, and smarter triage.
Here's how.
Foundation Model for the ECG: Precision From the Pocket
ECG-AI models are an important tool to help streamline cardiac assessments in oncology and speed up eligibility screening for clinical trials.
A new Electrocardiogram Foundation Model, for instance, looks particularly promising. This model, trained on over 10 million 12-lead tracings, has demonstrated strong generalizability and adaptability.
To put its performance in context, the key metric used here is AUROC (area under the receiver-operating characteristic curve). The AUROC is a single 0-to-1 score summarizing the test's overall discriminating power, with an AUROC of 0.5 representing a coin-flip, while 1.0 is flawless.
Using that standard, the Foundation Model achieves AUROC values of at least 0.95 across 80 different diagnoses, including QT prolongation and electrolyte abnormalities. It maintains accuracy on single-lead wearable devices, catching AFib (AUROC 0.93) and prolonged QT (AUROC 0.90). It can also be fine-tuned with just 500 local samples to predict outcomes such as 30-day heart failure admissions, and support dozens of downstream tasks, from rhythm detection to long-term risk prediction.
Outside of cardiology, this ECG model holds great relevance in oncology. The model can provide pretherapy safety checks by quickly screening QTc risk in TKI candidates, do prescreening for trial by identifying ineligible patients earlier, and help decentralize care by enabling ECG review from a smartwatch or home patch.
PanEcho and EchoNext: Echocardiography That Writes Itself
PanEcho, a deep learning system trained on more than 1.2 million echocardiogram videos, can also streamline cardio assessments.
PanEcho can autopopulate 39 parameters from a transthoracic echo. It yields a median AUROC of 0.91 across 18 endpoints, including ejection fraction, valve disease, and pulmonary pressures.
This model remains resilient even in handheld, point-of-care settings, with an AUROC of 0.85 for detecting aortic stenosis. For ejection fraction estimation, it reaches a mean absolute error (MAE) of 4%-5%, a performance comparable to interreader variability.
Another option, EchoNext, also had a strong AUROC -- 0.85 for detecting composite structural heart disease -- and was validated prospectively in the DISCOVERY trial, where AI flagged structural heart disease in 73% of high-risk patients without prior imaging.
There are many ways that EchoNext can be applied to oncology:
* Immediate ejection fraction reads before HER2 therapy or anthracyclines
* Immunotherapy myocarditis screening in community or emergency settings
* Enables real-time cardio screening in just-in-time or hybrid trial models
* Enables early detection where echocardiography access is limited
* Reduces disparities by fostering stable performance across sex, race, and clinical contexts (publicly released weights and dataset for transparency and equitable model development)
* Accelerates trial enrollment through prescreening for cardiac eligibility upstream
Below, I dissect what this means through the lens of Moravec's Paradox (machines excel at what clinicians find hard, and vice versa) and offer a possible playbook for putting these tools to work.
What's next in the oncologist's AI playbook to improve cardio screening?
Final Beat: Smarter Screening, Sooner Treatment
As cancer therapies grow more precise, so must our support systems. Imagine using AI to pre-clear patients before they even see a cardiologist. To reclassify risk, not just react to it. These tools are not only possible -- they're here.
A good reminder to never forget the patient attached to the waveform. The heart of AI is promising, but it still needs a physician's soul. We must ensure they work for all, not just for some. Inclusive AI, transparently built and equitably deployed, will be a foundational element in the next generation of oncology care. That sounds like music to my ears. How cool would that be?
Let's keep listening -- to our patients, to each other, and yes, to the heart of AI.
Side Note: AI on My Finger (and Yours)
Before we leave the cardiology suite, a quick pulse check from my own wrist-to-ring ecosystem. I've been wearing a smart ring for the past year; the tiny bioimpedance sensor inside quietly tracks my heart-rate variability, overnight SpO 2 , and microarousals, then nudges me the next morning if yesterday's late-night charting session shaved an hour off REM. That same ring logs my VO 2 estimate during interval runs -- data that now feed directly into my hospital's remote-monitoring dashboard.
This isn't vanity tech; it's part of a fast-maturing family of heart-wearable sensors -- optical PPG watches, single-lead ECG patches, BP-by-bioimpedance rings, even sweat sensors laced into compression shirts.
Together they're turning continuous physiologic streaming into raw material for the very AI models we've discussed today. (Stay tuned for a future column on wearables and AI.)
Thoughts? Drop me a line at [email protected]. Let's keep the conversation -- and the circulation -- going.
Arturo Loaiza-Bonilla, MD, MSEd, is the co-founder and chief medical AI officer at Massive Bio, a company connecting patients to clinical trials using artificial intelligence. His research and professional interests focus on precision medicine, clinical trial design, digital health, entrepreneurship, and patient advocacy. Dr Loaiza-Bonilla serves as Systemwide Chief of Hematology and Oncology at St. Luke's University Health Network, where he maintains a connection to patient care by attending to patients 2 days a week.
