Frailty Estimation in Elderly Oncology Patients Using Multimodal Wearable Data and Multi-Instance Learning
Summary: arXiv:2604.06985v1 Announce Type: cross
As the population ages, the incidence of cancer in elderly patients continues to rise, presenting unique challenges in treatment tolerance and outcomes. Frailty and functional decline have emerged as critical factors affecting the management of older oncology patients. Traditional methods of assessing frailty typically involve infrequent visits to healthcare facilities, leading to potential gaps in monitoring changes in patients’ health status. In an innovative approach, researchers have proposed a multimodal wearable framework aimed at estimating frailty-related functional changes in elderly breast cancer patients participating in the multicenter CARDIOCARE study.
Innovative Multimodal Wearable Framework
The proposed framework integrates free-living smartwatch data, capturing physical activity and sleep features, alongside ECG-derived heart rate variability (HRV) metrics obtained from a chest strap. This data is organized into patient-horizon bags that correspond to follow-up assessments at month 3 (M3) and month 6 (M6). The key innovation lies in an attention-based multiple instance learning (MIL) model that adeptly fuses irregular, multimodal data instances, accommodating real-world missing data and weak supervision.
- Smartwatch Features: The smartwatch collects crucial data on physical activity levels and sleep patterns, which are integral to assessing frailty.
- ECG-derived HRV: Heart rate variability provides insights into cardiovascular health, contributing valuable information about the patient’s overall well-being.
- Attention-based MIL Model: This model utilizes modality-specific multilayer perceptron (MLP) encoders to process variable-length and partially missing longitudinal data.
Performance and Evaluation
To evaluate the model’s effectiveness, researchers employed a subject-independent leave-one-subject-out (LOSO) approach. The results demonstrated promising accuracy rates in predicting functional changes:
- Handgrip Strength:
- M3: Balanced accuracy/F1 of 0.68 +/- 0.08/0.67 +/- 0.09
- M6: Balanced accuracy/F1 of 0.70 +/- 0.10/0.69 +/- 0.08
- FACIT-F:
- M3: Balanced accuracy/F1 of 0.59 +/- 0.04/0.58 +/- 0.06
- M6: Balanced accuracy/F1 of 0.64 +/- 0.05/0.63 +/- 0.07
These findings indicate that the full multimodal model is adept at capturing changes in frailty-related functional metrics over time. Further analysis revealed that smartwatch-derived activity and sleep data provided the most substantial predictive information, while HRV measurements offered complementary insights when integrated with smartwatch data streams.
Conclusion
This research underscores the potential of wearable technology and advanced machine learning techniques in transforming the assessment of frailty among elderly oncology patients. By facilitating continuous monitoring and providing timely insights, this multimodal framework could lead to improved patient management and treatment outcomes in a vulnerable population. As technology continues to evolve, further studies will be essential to validate these findings and refine the approach for broader clinical application.