Objective To analyze the influencing factors for post⁃stroke fatigue (PSF) in elderly patients with ischemic cerebral stroke in acute stage, and to establish a risk prediction model based on random forest model and Logistic regression model. Methods Elderly patients with ischemic cerebral stroke in 2 hospitals were selected as training set (304 cases) or validation set (102 cases), respectively. The general data, detection results of serological indices at admission were collected, and the Fatigue Severity Scale (FSS) was used to evaluate patients' fatigue. The univariate analysis in the training set was performed, and serological indices with statistical significance were enrolled into the Lasso regression model for screening prediction variables, and then the risk prediction model of PSF in elderly patients with ischemic cerebral stroke in acute stage was established by using the random forest model and Logistic regression model, respectively. Employing the data from the validation set, the prediction efficiency of the two models was evaluated through receiver operating characteristic curve, calibration curve, Hosmer⁃Lemeshow goodness of fit test, and decision curve analysis. Results A total of 5 prediction variables with non⁃zero coefficient were finally screened out, concerning C⁃reactive protein (CRP), homocysteine (Hcy), total cholesterol, fasting blood glucose, cystatin C, respectively. The random forest model ranked the importance of prediction variables as CRP, Hcy, total cholesterol, fasting blood glucose, and cystatin C according to the decreasing average value of Gini coefficient. The Logistic regression model revealed that the elevations of fasting blood glucose, total cholesterol, CRP, cystatin C, and Hcy levels could increase PSF risk in elderly patients with ischemic cerebral stroke in acute stage. The results of external validation indicated that areas under the curve, accuracy, sensitivities, specificities, positive prediction values, and negative prediction values between the random forest model and Logistic regression model were 0.905 and 0.885, 0.814 and 0.755, 0.830 and 0.855, 0.800 and 0.787, 0.780 and 0.826, and 0.846 and 0.696, respectively. The calibration degree of the two models was good, and the net benefit rate was relatively high in the threshold range of 0.08-0.86. Conclusion Fasting blood glucose, total cholesterol, CRP, cystatin C, and Hcy levels are closely related to PSF occurrence in elderly patients with ischemic cerebral stroke in acute stage. The random forest model and Logistic regression model established based on the aforementioned serological indices for predicting PSF occurrence in elderly patients with ischemic cerebral stroke in acute stage exert favorable values, therein the random forest model exhibits relatively favorable prediction efficiency, whereas the Logistic regression model can directly explain the influence degree of variables on the risk of PSF occurrence, which can be used as a supplement to the random forest model.

无