Prediction-model-building method, state prediction method and devices thereof
A prediction-model-building method includes: generating a plurality of pieces of augmented training data after a piece of training data with a missing item is determined; constituting a training dataset comprising pieces of training data without any missing item and the plurality of pieces of augmented training data; and building a prediction model based on the training dataset by a machine-learning algorithm. Similarly, a state prediction method includes: generating a plurality of pieces of estimation augmented data after a piece of source data with a missing item is determined; inputting the said plurality of pieces of estimation augmented data into a prediction model; and outputting a predicted state based on the output of the prediction model. Devices for performing the above methods are also disclosed.
1 . A prediction-model-building method, comprising:
receiving a plurality of pieces of training data, with each of the plurality of pieces of training data comprising a plurality of training feature items and with each of the plurality of training feature items configured to record a training eigenvalue;
determining whether each of the plurality of training feature items of each of the plurality of pieces of training data records the training eigenvalue or not;
determining a training feature item without the training eigenvalue as a training missing item, determining a piece of training data comprising the training missing item as a defective piece of training data, generating a plurality of training augmented values corresponding to the training missing item of the defective piece of training data and, based on the plurality of training augmented values, generating a plurality of pieces of augmented training data corresponding to the defective piece of training data;
constituting a training dataset at least comprising the plurality of pieces of augmented training data and pieces of training data without the training missing item; and
training a prediction model by a machine-learning algorithm based on the training dataset, and storing or outputting the prediction model,
wherein the plurality of pieces of augmented training data is a plurality of pieces of first augmented training data, and, before constituting the training dataset, the building method of prediction model further comprises:
duplicating the pieces of training data without the training missing item by a ratio corresponding to a number of the plurality of training augmented values to additionally generate a plurality of pieces of duplicated training data, wherein the plurality of pieces of duplicated training data serves as a plurality of pieces of second augmented training data;
wherein constituting the training dataset comprises: constituting the training dataset with the plurality of pieces of first augmented training data, the plurality of pieces of second augmented training data and the pieces of training data without the training missing item.
2 . The prediction-model-building method according to claim 1 , further comprising: performing a validation stage to determine whether the prediction model is updated or not after generating the prediction model, with the validation stage comprising:
receiving a plurality of pieces of validation data, with each of the plurality of pieces of validation data comprising a plurality of validation feature items and wherein each of the plurality of validation feature items is configured to record a validation eigenvalue;
determining whether the plurality of validation feature items of each of the plurality of pieces of validation data record the validation eigenvalue or not;
determining a validation feature item without the validation eigenvalue as a validation missing item, determining a piece of validation data comprising the validation missing item as a defective piece of validation data, generating a plurality of validation augmented values corresponding to the validation missing item of the defective piece of validation data and, based on the plurality of validation augmented values, generating a plurality of pieces of augmented validation data corresponding to the defective piece of validation data;
constituting a validation dataset at least comprising the plurality of pieces of augmented validation data and pieces of validation data without the validation missing item among the plurality of pieces of validation data; and
inputting the validation dataset to the prediction model to obtain a comparison result and selectively updating the prediction model based on the comparison result.
3 . The prediction-model-building method according to claim 1 , wherein the plurality of training feature items comprises at least one non-measured feature item not obtained by measurement, and determining whether the plurality of training feature items of each of the plurality of pieces of training data records the training eigenvalue or not comprises:
determining whether the non-measured feature item records the training eigenvalue or not.
4 . The prediction-model-building method according to claim 3 , wherein the non-measured feature item is associated with a body feature and the prediction model is used to generate a prediction result associated with at least one of the following states:
central sleep apnea (CSA), hypopnea, respiratory arousal and obstructive sleep apnea (OSA).
5 . The prediction-model-building method according to claim 1 , wherein generating the prediction model by the machine-learning algorithm based on the training dataset comprises: generating the prediction model by a clustering algorithm serving as the machine-learning algorithm.
6 . The prediction-model-building method according to claim 1 , wherein generating the prediction model by the machine-learning algorithm based on the training dataset comprises: generating the prediction model by a linear regression algorithm serving as the machine-learning algorithm.
7 . The prediction-model-building method according to claim 1 , wherein generating the prediction model by the machine-learning algorithm based on the training dataset comprises: generating the prediction model by a support vector machine algorithm serving as the machine-learning algorithm.
8 . The prediction-model-building method according to claim 1 , wherein generating the plurality of training augmented values corresponding to the training missing item of the defective piece of training data comprises:
obtaining a plurality of training eigenvalues corresponding to the pieces of training data without the training missing item among the plurality of pieces of training data;
classifying the training eigenvalues to obtain a plurality of classified representative values; and
setting the plurality of classified representative values as the plurality of training augmented values;
wherein the training eigenvalues are respectively recorded in the training feature items of the plurality of pieces of training data, and the training feature item corresponds to the training missing item.
9 . The prediction-model-building method according to claim 1 , wherein generating the plurality of training augmented values corresponding to the training missing item of the defective piece of training data comprises:
determining an acquisition range of the training missing item based on at least one training eigenvalue of the defective piece of training data, and wherein the at least one training eigenvalue is recorded in at least one training feature item other than the training missing item of the defective piece of training data; and
selecting a plurality of numerical values from the acquisition range as the plurality of training augmented values.
10 . A state prediction method comprising:
receiving a piece of source data, with the piece of source data comprising a plurality of estimation feature items and with each of the plurality of estimation feature items configured to record an estimation eigenvalue;
determining whether each of the plurality of estimation feature items records the estimation eigenvalue or not;
determining the estimation feature item without the estimation eigenvalue as an estimation missing item, generating a plurality of estimation augmented values corresponding to the estimation missing item and, based on the plurality of estimation augmented values, generating a plurality of pieces of augmented estimation data;
inputting the plurality of pieces of augmented estimation data to a prediction model built according to the prediction-model-building method of claim 1 to generate a plurality of initial estimation results; and
determining a state predicting result associated with the piece of source data from the plurality of initial estimation results and storing or outputting the state predicting result.
11 . The state prediction method according to claim 10 , wherein determining a state predicting result associated with the piece of source data from the plurality of initial estimation results comprises: taking a mode of the plurality of initial estimation results as the state predicting result.
12 . The state prediction method according to claim 10 , wherein determining a state predicting result associated with the piece of source data from the plurality of initial estimation results comprises: taking an average the plurality of initial estimation results as the state predicting result.
13 . The state prediction method according to claim 10 , wherein the plurality of initial estimation results is associated with at least one of the following states:
central sleep apnea (CSA), hypopnea, respiratory arousal and obstructive sleep apnea (OSA).
14 . The state prediction method according to claim 10 , wherein generating a plurality of estimation augmented values corresponding to the estimation missing item comprises:
determining an acquisition range of the estimation missing item based on at least one estimation eigenvalue of the piece of source data, and wherein the at least one estimation eigenvalue is recorded in at least one estimation feature item other than the estimation missing item of the piece of source data; and
selecting a plurality of numerical values as the plurality of estimation augmented values from the acquisition range.
15 . A prediction-model-building device comprising:
a memory storing a machine-learning algorithm; and
a processor coupled to the memory to access the machine-learning algorithm and performing the prediction-model-building method according to claim 1 .
16 . The prediction-model-building device according to claim 15 , wherein the processor is disposed in a server, and the server is internet-connected to the memory.
17 . The prediction-model-building device according to claim 15 , wherein the processor and the memory are disposed in a server.
18 . A state prediction device comprising:
a memory storing a prediction model; and
a processor coupled to the memory to access the prediction model and performing the state prediction method according to claim 10 .
19 . The state prediction device according to claim 18 , wherein the processor is disposed in a server, and the server is internet-connected to the memory.
20 . The state prediction device according to claim 18 , wherein the processor and the memory are disposed in a server.