Treatment-Related Cardiotoxicity Risk Prediction for Breast Cancer Patients with a Multimodal AI-based System

Anti-tumor therapies (e.g., anthracyclines, biologic, and immunotherapy) are the standard treatments for most breast cancer patients, and dramatically improve survival rates. However, up to 20% of breast cancer patients treatment with these therapies will develop some form of serious cardiovascular disease (CVD), such as heart failure, myocarditis, or cardiac arrhythmia. Emerging data suggest CVD is now the leading non-malignant cause of death during or after the treatment stage of breast cancer. If clinicians can accurately predict treatment-related cardiotoxicity during the treatment stage or shortly post-treatment, they can then adjust their anti-tumor therapy selection or prescribe interventions that reduce the risk of serious CVD. However, our preliminary data show that clinicians face several challenges in predicting these serious events, including the unpredictable nature of cardiotoxicity, the lack of specific clinical guidelines, and the low sensitivity of diagnostic tests. Recent advances in artificial intelligence and machine learning (AI/ML) have shown great potential to support clinicians in making early decisions about CVD. For example, our and other predictive AI/ML models that were trained and evaluated on cancer patients’ historical electronic health record (EHR) data to predict the risk of a future CVD onset, have yielded promising results. Yet, these existing AI/ML models for risk prediction have inherit limitations, as their predictors are limited only to sparse EHR data; whereas clinicians in practice rely on additional modalities for key signals such as patient-reported symptoms (e.g., discomfort in breathing) from non-clinical settings such as at home, and change in anti-tumor treatment (e.g., increase in dose). To date, these signals are entirely missing in existing risk prediction models for CVD of breast cancer patients. Our primary objectives are: to complement EHR data with multimodal data collected using digital health technologies and to develop a better AI/ML risk prediction model for breast cancer treatment-related cardiotoxicity based on the EHR data and the multimodal data. In aim 1, we will recruit stakeholders, design and develop a human-centered smart system with wearable sensors and a conversational large language model (LLM) chatbot for passive collection of multimodal data. In aim 2, we will recruit breast cancer patients in active treatment, and will collect heart rate variability (HRV) using wearable devices and capture the symptoms potentially related to cardiotoxicity using the LLM chatbot in the home setting. In aim 3, we will first create a deidentified dataset with complete visits and long-term EHR history for patients with breast cancer, corresponding chemotherapy treatments, and potential cardiotoxicity risk. Then, we will develop early treatment-related cardiotoxicity prediction model with large-scale sparse EHRs, and finetune the cardiotoxicity prediction model with additional small-scale dense signals. In aim 4, we will update the human-centered AI system to include the dynamic cardiotoxicity risk score in the physician-facing dashboard and evaluate the updated home-based smart system with physicians. Our findings are expected to innovate and improve treatment-related cardiotoxicity risk prediction and facilitate clinical decision making.