Research on multi-parameter fusion non-invasive blood glucose detection method based on machine learning.

OBJECTIVE

Traditional blood glucose testing methods have several disadvantages, such as high pain and poor acquisition continuity. In response to these shortcomings, we propose a multi-parameter fusion non-invasive blood glucose detection method that combines machine learning and photoplethysmography (PPG) signal feature parameter analysis.

MATERIALS AND METHODS

This method uses the signal validity check process based on the correlation operation to test and calculate PPG data. It, then, respectively applies the bootstrap aggregation algorithm and the random forests algorithm to establish two non-invasive blood glucose detection models that comprehensively predict blood glucose data.

RESULTS

Experimental comparative analysis showed that the accuracy of the detection model based on the random forests algorithm is superior. The correlation coefficient of the obtained blood glucose prediction set is 0.972, the mean square error is 0.257, and the relative error is less than ± 20%.

CONCLUSIONS

Relative error in blood glucose prediction meets the national standards in China. Meanwhile, the results of the Clarke Error Grid Analysis indicate that the non-invasive blood glucose testing method proposed in this study meets clinical accuracy requirements.