Multimodal Wearable Devices and Machine Learning: A New Approach for Monitoring Frailty in Elderly Cancer Patients

For elderly cancer patients, frailty status affects treatment tolerance and prognosis, but traditional assessments can only be performed in clinics and cannot capture functional changes between two visits. This study proposes an innovative solution: using a multimodal wearable system consisting of smartwatches (monitoring activity and sleep) and chest-strap ECG monitors (recording HRV), combined with an attention-based multiple instance learning algorithm, to continuously monitor changes in frailty status among elderly breast cancer patients between visits, opening up new possibilities for personalized care.

Elderly cancer patients have limited physiological reserves and multiple comorbidities. Frailty status is closely associated with increased chemotherapy toxicity and higher risk of postoperative complications. However, traditional assessment tools (such as the Fried phenotype and Rockwood index) require completion by professionals in medical institutions, and their frequency is limited by visit schedules, leading to failure to timely capture functional changes between visits, which may pose a risk of delayed treatment adjustments.

Data Collection: Smartwatches record behavioral indicators such as step count, activity intensity, and sleep duration; chest-strap ECG monitors record HRV (an indicator of autonomic nervous system function).

Data Organization: A patient-time window bag structure is adopted to organize data by patient and follow-up time points (3/6 months), addressing irregularities and variable lengths.

Core Technology: Attention-based Multiple Instance Learning (MIL): Each modality (activity, sleep, HRV) is equipped with a dedicated encoder. The attention mechanism enhances interpretability and automatically learns the contribution of key instances to functional changes.

Data were from the CARDIOCARE multicenter study (elderly breast cancer patients). Evaluation indicators included the FACIT-F Fatigue Scale and grip strength test, using leave-one-patient cross-validation.

Key Results: Grip strength change prediction (3-month balanced accuracy: 0.68/F1=0.67; 6-month:0.70/F1=0.69); FACIT-F change prediction (3-month:0.59/F1=0.58;6-month:0.64/F1=0.63).

Ablation Study: Smartwatch data contributed the most; HRV alone had weak effects but improved performance when combined.

This solution enables continuous monitoring to capture functional fluctuations that traditional assessments cannot detect; supports early warning and timely intervention to avoid adverse events; aids personalized treatment decisions (adjusting plans based on real-time status); applicable to telemedicine and home care, reducing hospital visits, especially beneficial for patients with mobility issues.

Limitations: Limited sample size (only breast cancer patients), patient wear compliance affects data quality.

Future Directions: Expand sample size to more cancer types; explore more complex models (e.g., Transformer) and multimodal data; address clinical integration issues (such as workflow and team training).

This study combines digital health technology with geriatric oncology, using consumer-grade devices + machine learning to address monitoring blind spots. With the growing aging population, this framework is expected to become a standard configuration for elderly oncology care, extending the perception capabilities of medical teams and improving patients' treatment experience and quality of life.