Using Graph Representation Learning to Predict Salivary Cortisol Levels in Pancreatic Cancer Patients.

Area of Science:

• Endocrinology
• Oncology
• Machine Learning

Background:

• Cortisol is a crucial glucocorticoid hormone impacting immune function and potentially influencing tumor growth.
• Monitoring cortisol levels aids cancer treatment decisions, but current methods (serum/saliva samples) are impractical.
• There is a need for non-invasive, convenient methods to track cortisol levels over time.

Purpose of the Study:

• To develop and evaluate a predictive modeling process for salivary cortisol levels using passively sensed actigraphy data.
• To compare the efficacy of graph representation learning methods against traditional feature engineering for cortisol prediction.
• To assess the feasibility of using wearable sensor data for longitudinal cortisol monitoring in cancer patients.

Main Methods:

• Utilized passively sensed actigraphy data as input for predictive models.
• Employed graph representation learning techniques, including Graph2Vec, FeatherGraph, GeoScattering, and NetLSD.
• Compared graph representation learning models with machine learning models using handcrafted feature engineering.

Main Results:

• Preliminary results from 10 pancreatic cancer patients indicate superior performance of graph representation learning models.
• Machine learning models incorporating graph representation learning outperformed those with handcrafted features in predicting salivary cortisol levels.
• Actigraphy data combined with graph representation learning shows promise for non-invasive cortisol monitoring.

Conclusions:

• Graph representation learning applied to actigraphy data provides a viable and effective method for predicting salivary cortisol levels.
• This approach offers a practical, non-invasive alternative to traditional sampling methods for monitoring cortisol in clinical settings, particularly in oncology.
• Further research with larger cohorts is warranted to validate these findings and optimize the predictive models for clinical application.