It is found that unsupervised transfer learning improves the generalization of machine learning classification based on brain function image data in schizophrenics

In the field of clinical research, machine learning has been widely used to optimize brain image data analysis and establish predictive models to classify schizophrenia patients The evaluation of generalization is an important step in the performance evaluation of prediction model, but there are few clinical studies on this aspect In order to solve this problem, Chen Chuqiao, a researcher in the laboratory of neuropsychology and Applied Cognitive Neuroscience (NaCN), a Key Laboratory of mental health, Institute of psychology, Chinese Academy of Sciences, and an international cooperator have carried out a special study on the generalization of machine learning classification of schizophrenics based on resting state MRI Among them, the researchers used the internal verification method and the external verification method to evaluate the generalization within the center and across the center respectively In this study, 51 schizophrenics and 51 healthy controls were recruited as the main data set, 34 schizophrenics and 27 healthy controls as the validation set, and all the rest state MRI data and structure image data were collected First, the researchers evaluated the generalization within the center in the main database and got a 0.73 accuracy rate Then, taking the main database as the training set, the prediction model obtained in the training set is generalized across the center to the performance of the verification set, and the accuracy of 0.56 is obtained (not significant after the replacement test) Finally, considering the poor performance of cross center generalization, the researchers update the unsupervised learning algorithm based on the extra unlabeled data, and get the accuracy of 0.70 The findings indicate that more research is needed to promote the application of machine learning across multiple databases in the future At the same time, we found the results of the optimal classification of transfer learning in this study, and emphasized the importance of incorporating sample related factors into the establishment of cross sample and central prediction model All in all, this study suggests that the current results based on single machine learning classification should be carefully interpreted The research is supported by the national key research and development program, the National Natural Science Foundation, the Beijing Science and technology leading talent program, and the Key Laboratory of mental health, Institute of psychology, Chinese Academy of Sciences The article has been published online in human brain m app ing.