	US 12,278,012 B2
	System and method for detection of impairment in cognitive function
Kup Sze Choi, Kowloon (HK); and Xiao Shen, Kowloon (HK)
Assigned to The Hong Kong Polytechnic University, Hong Kong (CN)
Filed by The Hong Kong Polytechnic University, Kowloon (HK)
Filed on Mar. 9, 2021, as Appl. No. 17/196,626.
Prior Publication US 2022/0293266 A1, Sep. 15, 2022
Int. Cl. G06N 3/08 (2023.01); G06N 3/045 (2023.01); G16H 10/60 (2018.01); G16H 50/20 (2018.01); G16H 50/30 (2018.01); G16H 50/70 (2018.01)

CPC G16H 50/20 (2018.01) [G06N 3/045 (2023.01); G06N 3/08 (2013.01); G16H 10/60 (2018.01); G16H 50/30 (2018.01); G16H 50/70 (2018.01)]

13 Claims

1. A machine learning method for predicting whether a specified subject is at high risk of developing cognitive impairment based upon a data record for the specified subject by automatically classifying the subject into a first class associated with a first predicted risk of cognitive impairment or a second class associated with a second predicted risk of cognitive impairment; wherein the method comprises:

acquiring a plurality of subjects' records, wherein each record comprises a first data set of measured data and a second data set including results for two or more health assessment questionnaires and a label indicating that a subject of each record of the plurality of subjects' records belongs to the first class or the second class;

automatically classifying subjects of the plurality of subjects into the first class or the second class according to the label;

training a first neural network and a second neural network together on the plurality of subjects' records by generating representations thereof by iteratively;

(i) generating by the first neural network a first representation of the first data set of a selected subject's record from the plurality of subjects' records; and

(ii) generating by the second neural network a second representation for the second data set of the selected subject's record; and

(iii) concatenating the first and the second representations together and using the concatenated first and second representations as inputs to a first classifier configured for assigning the selected subject's record to either the first class or the second class;

(iv) updating trainable parameters of the first and second neural networks and the first classifier by confirmation with a classification made according to the label for the subject's record;

predicting the risk of cognitive impairment for the specified subject by using the first classifier to evaluate a concatenated vector of the representations generated by the trained first and second neural networks respectively by assigning the specified subject by the first classifier to the first class or the second class;

including a cost sensitive learning weighting to increase sensitivity of detection when using the first classifier for evaluating the concatenated representation of profile data and health assessment data and for updating the training parameters of the first and second neural networks;

wherein the cost sensitive learning weighting w_iassociated with i^thsubject is calculated according to:

w_i=m_rⁿ/m_r^dif y_i=1

and w_i=1 if y_i=0

wherein:

m_rⁿand m_r^dare respectively numbers of normal cases and numbers of high-risk cases in plurality of subjects' records used for training the first network and second neural network; and

y_iis a ground-truth label of the training sample i;

where y_i=1 if the ground-truth label of the training sample i is high-risk; and

where y_i=0 if the ground-truth label of training sample i is normal.