Stress Estimation in Elderly Oncology Patients Using Visual Wearable Representations and Multi-Instance Learning

A team of medical researchers has developed a novel method for estimating psychological stress in elderly breast cancer patients using wearable technology and artificial intelligence, as detailed in a recent paper published on the arXiv preprint server. The study, conducted across multiple clinical centers, aims to address the critical gap in continuous stress monitoring within cardio-oncology by analyzing data from smartwatches and ECG sensors instead of relying solely on traditional patient questionnaires. This research represents a significant advancement in personalized cancer care, particularly for older patients undergoing treatment that can affect heart health. The research focuses on the CARDIOCARE cohort, a multicenter group of elderly breast cancer patients. Current clinical practice typically assesses psychological stress through intermittent patient-reported outcome measures (PROMs), which provide snapshots rather than continuous data and are rarely incorporated into systematic cardiotoxicity surveillance programs. The researchers hypothesized that physiological data collected continuously from wearables could provide more objective, frequent, and integrated stress measurements, which is crucial since psychological stress is a known clinically relevant factor that can influence cardiovascular outcomes in cancer patients. Methodologically, the team employed multi-instance learning, a machine learning technique particularly suited for analyzing complex, sequential data from wearables. They transformed raw data streams from two devices: a smartwatch tracking physical activity and sleep patterns, and a chest-worn electrocardiogram (ECG) sensor monitoring heart activity. By converting these multimodal data streams into visual representations, the algorithm could identify patterns correlating with patients' self-reported stress levels. This approach allows the system to learn from bags of instances (multiple data points over time) rather than single measurements, making it particularly effective for the variable nature of physiological signals. The implications of this research are substantial for the growing field of cardio-oncology, which addresses the intersection of cancer treatment and cardiovascular health. By enabling continuous, objective stress monitoring, clinicians could potentially identify patients at higher risk for cardiotoxic effects from cancer therapies earlier and implement timely interventions. Furthermore, this wearable-based approach reduces reliance on subjective patient recall and can provide data even between clinical visits. While further validation is needed, this study demonstrates a promising pathway toward integrating digital health technologies into routine oncology care to improve outcomes for vulnerable elderly populations.