Development of an interpretable machine learning-based intelligent system of exercise prescription for cardio-oncology preventive care: A study protocol

Tianyu Gao, Hao Ren, Shan He, Deyi Liang, Yuming Xu, Kecheng Chen, Yufan Wang, Yuxin Zhu, Heling Dong, Zhongzhi Xu, Weiming Chen, Weibin Cheng, Fengshi Jing, Xiaoyu TaoFrontiers in Cardiovascular Medicine 2023 · DOI

Abstract: This study protocol (ChiCTR2300077887) outlines the development of an AI-driven system designed to prescribe personalized exercise for cancer patients. Published in Frontiers in Cardiovascular Medicine, the research aims to bridge the gap between oncology and cardiology by utilizing machine learning to mitigate the cardiovascular side effects of cancer treatment through standardized, interpretable preventive care.

Introduction: The Cardio-Oncology Challenge

Cancer and cardiovascular disease (CVD) remain the leading causes of global mortality. These conditions frequently intersect; cancer treatments, while life-saving, often induce "cardio-toxicity"—direct damage to the heart and vascular system. Approximately 20% of cancer patients suffer from cardiovascular comorbidities, making preventive care essential.

Physical exercise is a proven intervention to reduce these risks. However, creating personalized "exercise prescriptions"—specific plans detailing frequency, intensity, and duration—is a labor-intensive process requiring specialized clinical exercise physiologists (CEPs). In many regions, including China, there is a severe shortage of these specialists. This study proposes an automated, intelligent system to fill this gap, prioritizing transparency and interpretability to ensure clinical trust.

Methodology and Timeline

The study is structured as a multi-year, retrospective machine learning modeling cohort study with three distinct phases:

  1. Expert Standardization (2023–2024): To ensure high-quality "ground truth" labels for the AI, participating doctors are trained to become certified CEPs. This ensures the initial prescriptions used for training meet rigorous clinical standards.
  2. Data Acquisition (2025–2026): The project will recruit 600 Stage I cancer patients and survivors (aged 18–55). Participants will receive expert-guided prescriptions and wear smart sensors to track adherence and physiological responses over a two-year period.
  3. Modeling & Interpretability (2027–2028): The team will evaluate both standard interpretable models (e.g., Logistic Regression, Decision Trees) and high-performance "black-box" models (e.g., CNNs, XGBoost). To bridge the gap between performance and transparency, SHAP (SHapley Additive exPlanations) will be employed to explain model outputs.

Data Acquisition and Variables

The system relies on a novel longitudinal dataset encompassing a wide array of clinical and lifestyle metrics. These inputs are categorized to provide a holistic view of the patient's health status.

Comprehensive list of categories and variables used for model training.

As shown in the table above, the model integrates demographics, cancer-specific information (type, treatment, medication), lifestyle habits (smoking, diet, sleep), and physical fitness metrics such as cardiorespiratory fitness (CRF) and muscle strength. Crucially, CVD-related items like blood pressure variability and 24-hour ECG data are included to monitor cardiac health directly.

The Intelligent System Architecture

The core of the project is an intelligent recommendation system that transforms complex patient data into actionable exercise plans. The system is designed not just to provide a recommendation, but to justify it.

Demonstration of the intelligent system showing model inputs, AI processing, and interpretable outputs.

As illustrated in the system demonstration, the AI model processes multi-dimensional inputs to generate an output consisting of (a) Frequency, (b) Duration, and (c) Intensity. The "Interpretable" dashboard is a key feature, utilizing SHAP values to show the importance of various features. For instance, a doctor can see that a specific recommendation was 27% influenced by the patient's age and 15% by their diet, allowing for informed clinical oversight.

Key Insights and Innovations

  • The Interpretability Mandate: In healthcare, accuracy alone is insufficient. By integrating SHAP, the system provides a "reasoning" layer, allowing clinicians to understand why a specific intensity was recommended, which is vital for patient safety and adherence.
  • Wearables as Truth-Seekers: Traditional studies often suffer from self-reporting bias. This protocol uses wearable sensors to capture real-time physical activity, ensuring the model is trained on objective adherence data rather than subjective patient accounts.
  • Engineering for Attrition: Recognizing the high dropout rates in cancer-related trials (estimated at 47%), the researchers plan to utilize few-shot learning algorithms. These techniques allow the model to remain effective even if the final cohort size is smaller than initially projected.

Limitations and Future Outlook

While promising, the current protocol is limited to Stage I cancer patients. Future iterations will need to address the increased complexities of Stage II and III patients, where treatment side effects are more severe. Additionally, while internal validation is the first step, the system will eventually require external validation in diverse clinical environments to ensure global applicability.

The ultimate goal is to move toward a dynamic optimization system that adjusts exercise prescriptions in real-time as a patient’s fitness levels or health status evolve during their recovery journey.

Summary updated

Abstract

http://www.chictr.org.cn, identifier ChiCTR2300077887.
BibTeX
@inproceedings{gao2023development,
  author = {Tianyu Gao and Hao Ren and Shan He and Deyi Liang and Yuming Xu and Kecheng Chen and Yufan Wang and Yuxin Zhu and Heling Dong and Zhongzhi Xu and Weiming Chen and Weibin Cheng and Fengshi Jing and Xiaoyu Tao},
  title = {{Development of an interpretable machine learning-based intelligent system of exercise prescription for cardio-oncology preventive care: A study protocol}},
  booktitle = {Frontiers in Cardiovascular Medicine},
  year = {2023},
  doi = {10.3389/fcvm.2022.1091885},
}
Figure