Objectives:To investigate the predictive effect of regional homogeneity (ReHo) from resting-state functional magnetic resonance imaging (rs-fMRI) in acute phase on memory function of chronic phage in ischemic stroke patients and the effects of residual learning (REL) on the predictive performance of machine learning models.Methods:From June 2019 to June 2021, rs-fMRI data of one-week after stroke (acute phase) were collected from 35 first-time ischemic stroke patients, and their memory scores were assessed by the Rey auditory verbal learning test (RAVLT) at 6 months after stroke (chronic phase). Using ReHo from rs-fMRI data in acute phase of ischemic stroke patients, the support vector regression (SVR) and the REL-based SVR (REL-SVR) were constructed to predict the patients’ memory scores at 6 months after stroke, and the performance of the two models was compared using Pearson correlation coefficient.Results:Based on the ReHo from acute phase, the correlation coefficient between the predicted values and the true scores from the SVR model was r=0.524, P=0.001, while the correlation coefficient obtained by the REL-SVR model was r=0.671, P<0.001. Brain regions with relatively higher weights such as Temporal_Pole_Mid_R (weight value: 1.03), Temporal_Mid_R(weight value: 1.03), Temporal_Inf_R (weight value: 1.03), Occipital_Mid_R (weight value: 0.57), Frontal_Mid_L (weight value: 0.32), Frontal_Sup_Medial_L (weight value: 0.53), SupraMarginal_L (weight value: 1.54), Calcarine_L (weight value: 0.65), Lingual_L (weight value: 0.58), Cuneus_L (weight value: 0.65), Precuneus_L (weight value: 0.83), cerebellum(weight value>1.0) made larger contributions to the prediction model. Conclusions:ReHo in the acute-phase can effectively predict memory in the chronic phase of ischemic stroke patients. Furthermore, REL can improve the performance of the traditional SVR model and achieve higher predictive accuracy.