Bayesian cross-individual transfer learning for optimising predictive ratio CUSUM monitoring scheme
针对糖尿病前期个体历史血糖数据不足导致传统控制图误报率高的问题，提出一种贝叶斯跨个体迁移学习方法，提升预测比CUSUM控制图的鲁棒性，并给出超参数调优和相似性评估的实用指导。
Recent advancements in wearable devices have led to research in diagnosing and identifying pre-diabetes and diabetes based on analysing continuous glucose monitoring data. Statistical process control is a widely used and valuable monitoring technique for detecting variability magnitudes of a continuous monitoring process. However, obtaining sufficient historical reference glucose data for individuals with pre-diabetes in Phase I is challenging, which leads to high false alarm rates with traditional control charts due to misidentifying in-control and out-of-control patterns. This paper introduces a novel Bayesian cross-individual process monitoring approach utilising transfer learning techniques to enhance the performance of the predictive ratio CUSUM control chart. Our proposed method showcases enhanced robustness in both simulated and real-world scenarios. Additionally, we provide practical guidance on assessing similarities between medical metrics and offer insights on fine-tuning hyperparameters and expanding volatility metrics based on simulation and practical results. The monitoring approach is universal and can be flexibly applied beyond the medical field to the industrial field.