A novel blood glucose time series prediction framework based on a novel signal decomposition method

• Published in: Chaos, solitons and fractals solitons and fractals, 2022-11, Vol.164, p.112673, Article 112673
• Main Authors: Zhu, Ting, Wang, Wenbo, Yu, Min
• Format: Article
• Language: English
• Subjects: Blood glucose prognosis; Ensemble empirical mode decomposition; Fractal dimension; Hypoglycemic warning; Kernel-based extreme learning machine; Sparrow search algorithm
• ISSN: 0960-0779; 1873-2887
• DOI: 10.1016/j.chaos.2022.112673

Unstable blood glucose levels will bring about a series of complications, and even threaten our life in severe cases. Therefore, accurately predicting and controlling blood glucose has become an indispensable part of diabetes treatment. For the sake of boosting the accuracy of blood glucose prediction, a novel time-frequency signal analytical technology called improved ensemble empirical mode decomposition based on fractal dimension (FEEMD) is first proposed to decompose the nonlinear and non-stationary raw blood glucose series to obtain intrinsic mode functions (IMFs) and residual components (Rs) in different frequency bands. Then, in order to overcome the problem that the kernel extreme learning machine (KELM) is easy to fall into the local optimum, the sparrow search algorithm (SSA) is introduced to further enhance the robustness and stability of KELM model. Ultimately, the predictions of each optimized KELM are fused to achieve the final prognostic results. Through the experimental results of three patients, the following contributions can be summarized: (a) FEEMD exploits the box-counting dimension to remove abnormal signals in the original sequence, which guarantees the completeness of decomposition and obtains better decomposition results; (b) SSA is exploited to optimize the KELM, which can significantly boost the prediction effects; (c) When compared with the other seven benchmark models, the proposed framework in this paper is more suitable for predicting blood glucose and has higher prediction accuracy; (d) This paper also verifies that the proposed architecture can effectively reduce the false negative rate and increase the overall alarming accuracy in early warning of hypoglycemia. •The improved ensemble empirical mode decomposition which is based on fractal dimension (FEEMD) method is first proposed.•The effectiveness of the model proposed is further verified by comparing with the other seven models for experimental analysis.•The sparrow search algorithm is used to optimize the KELM, which can efficiently enhance the prediction performance.•The intergration of FEEMD method and SSA-KELM both improved the prediction accuracy and extended the prediction time.•The FEEMD-SSA-KELM model proposed can timely and effectively predict the occurrence of early hypoglycemia events.