The invention discloses a sub-tree adaptive weighted
pruning blood glucose prediction method based on Extra Tres regression, and is applied to the field of bioelectrical impedance blood glucose prediction modeling. The method comprises the following steps: step 1, acquiring a
data set containing
human body bioelectrical impedance data and corresponding blood glucose values, and dividing the
data set into a
training set, a
verification set and a
test set; step 2, training an Extra Tres regression model based on the
training set, extracting the prediction output of each sub-tree in the model, splicing the sub-tree prediction output with the overall prediction output, and constructing a sub-tree prediction
result set; step 3, utilizing a mapping relation between a sub-tree prediction result in a
verification set and a real blood glucose value, introducing a Bayesian
ridge regression model to carry out probability modeling on sub-tree prediction contribution, and obtaining prediction
weight distribution corresponding to each sub-tree; 4, a
pruning criterion is constructed according to statistical characteristics of the prediction weight, and self-adaptive
pruning is carried out on the sub-trees with low prediction contribution or high redundancy; and step 5, performing integrated calculation based on prediction output of the sub-trees reserved after pruning to obtain a final blood glucose prediction result. According to the method,
adaptive learning and pruning control of the sub-tree weight are realized through probabilistic modeling of the sub-tree prediction contribution, the redundancy of the blood glucose prediction model is reduced, and the modeling stability and generalization performance of the blood glucose prediction model in a bioelectrical impedance scene are improved.