Presentation Information
[N-2-16]Error Propagation Analysis of Recursive Blood Glucose Prediction Using an Echo State Network
◎Yifan Geng1, Chenghao Wang1, Takeaki Yajima1 (1. Kyushu Univ.)
Keywords:
Blood Glucose Prediction,Echo State Network,Error Propagation,Frequency-Domain Analysis
Stable blood glucose prediction is important for early detection of hypoglycemia and hyperglycemia. This study analyzes error propagation in recursive multi-step blood glucose prediction using an Echo State Network (ESN) from the frequency-domain perspective. In the prediction process, the one-step-ahead output is repeatedly used as the input for the next step to predict future glucose values. To quantify error amplification, the DC component of the prediction error at each prediction horizon is removed, and the amplitude spectrum is obtained using FFT. The spectrum is divided into low-, mid-, and high-frequency components. The results show that the low-frequency component remains relatively stable as the prediction horizon increases, whereas the mid-frequency component and especially the high-frequency component increase markedly at longer horizons. This suggests that recursive prediction accumulates errors related to short-period fluctuations and noise, and may emphasize unstable dynamic components inherent in blood glucose time series. The analysis shows that frequency-band decomposition can reveal error amplification that is difficult to capture using RMSE alone.