Research On Semi-Supervised Learning-based Cognitive Impairment Assisted Diagnosis

Cognitive impairment refers to impairment in memory,learning,language and other cognitive aspects.Depending on the degree of the disease,it can be classified as mild cognitive impairment(MCI)and Alzheimer’s disease(AD).AD is a central neurodegenerative disease commonly seen in old age,which indicates a deteriorating and ultimately fatal condition over time.As a transitional stage between Cognitive Normal(CN)and AD,MCI is difficult to diagnose in the early stages of cognitive disorders because its daily manifestations are highly similar to those of normal aging.For cognitive disorders,early intervention and treatment is the key to prevent further deterioration of the condition.Therefore,early diagnosis of cognitive impairment disorders is of great importance.However,assisted diagnosis of cognitive disorders using deep learning models often uses more expensive and highly invasive biomarker data such as Magnetic Resonance Imaging(MRI)and Cerebrospinal Fluid(CSF),which have more limitations such as high time and price costs.Moreover,most of studies on such problems use fully supervised learning methods,which require each data sample to have its correct diagnosis as a label.However,labeling data can still be a timeconsuming and challenging task for physicians,especially when the size of the data is large.Therefore,there is still room for improvement in methodological and predictive performance in the field of AI-assisted diagnosis of cognitive disorders.Based on the above problems,this paper proposes a new semi-supervised learning algorithm to diagnose the degree of cognitive impairment,which is trained using noninvasive and easily available neuropsychological test data and a small number of diagnostic results as real labels.The algorithm uses a dual encoder mechanism to complete the perturbation of the model,introduces difference regularization so that the two encoders learn different feature representations as much as possible,and combines consistency regularization with pseudo-labeling.819 subjects from the Alzheimer’s Disease Neuroimaging Initiative database are included in this study,including 188 AD,402 MCI and 229 CN.This paper first performs feature selection in 7neuropsychological tests to derive the 15 features most associated with diagnostic outcomes.Next,a deep network model with semi-supervised learning is constructed based on the proposed algorithm.In this study,the diagnoses of 60 and 120 subjects are randomly selected as the true labels of the training samples and the model is extensively experimented.The experimental results show that the model proposed in this paper has the best accuracy and stability on AD,MCI and CN tri-classification compared with other semi-supervised learning models.Based on the above model,this paper proposes another MCI conversion prediction method based on attention mechanism and semi-supervised learning.The algorithm uses two attention mechanisms to complete the perturbation of the model,weakening the dual encoder mechanism.In this paper,a binary classification model is constructed based on this algorithm and trained using AD and CN data.The test set uses the baseline MCI data with their diagnostic results for different follow-up periods as the true labels.The experimental results show that the proposed model outperforms other semisupervised learning models in the task of predicting MCI conversion within 3 years,which could provide clinical insights to physicians when predicting the progression of cognitive impairment.