Objective To investigate effect of plasma ammonia on β-amyloid(Aβ)of CSF and cognitive function in mild Alzheimer's disease(AD).Methods A total of 108 mild AD patients admitted to our hospital and 47 volunteers(HC group)were studied,the cognitive function were measured and plasma ammonia were detected.The mild AD patients were divided into normal ammonia group(ADA-group)and hyperammonemia group(ADA + group)by the plasma ammonia level.The level of Aβ42,Aβ40,total tau and phosphorylated tau protein in CSF of AD patients were detected,and the results were compared.Results Compared with HC group,the MMSE and Montreal Cognitive Assessment(MoCA)in the ADA-group and ADA + group were significantly lower(all P＜0.001).The MoCA of ADA + group was significantly lower than that of ADA-group(P＜0.001).The plasma ammonia of ADA + group were significantly higher than those of ADA-group and HC group(all P＜0.001).The levels of Aβ42 and Aβ40 in ADA + group were significantly lower than those of ADA-group(t =2.29,P =0.024;t =2.72,P =0.008).In ADA + group,the plasma ammonia level was negatively correlated with the MoCA score(r =-0.47,P＜0.001)and level of Aβ42 in CSF(r =-0.63,P＜0.001).The level of Aβ42 played a partial mediating role between the plasma ammonia level and cognitive function,and mediating effect accounted for 45.94%of the total effect.Conclusion The plasma ammonia level is elevated in some AD patients,high plasma ammonia level may reduce the Aβ42 level in CSF,contributes to Aβ42 deposition which forms the amyloid plaques,ultimately aggravates cognitive impairment.

Objective:To evaluate the effect of deep learning based on DWI and fluid attenuated inversion recovery (FLAIR) to construct a prediction model of the onset time in acute stroke.Methods:A total of 324 cases of acute stroke with clear onset time, from January 2017 to May 2020 in Nanjing First Hospital, were retrospectively enrolled and analyzed. The patients were divided into a training set of 226 patients and a test set of 98 patients according to the complete randomization method using a 7∶3 ratio, and the patients were divided into ≤ 4.5 h and >4.5 h according to symptom onset time in each group. The acute infarction areas on DWI and the corresponding high signal area on FLAIR were manually outlined by physician. Using the InceptionV3 model as the basic model for image features extraction, the deep learning prediction model based on single sequence (DWI, FLAIR) and multi sequences (DWI+FLAIR) were established and verified. Then the area under curve (AUC), accuracy of human readings, single sequence model and multi sequence model in predicting the acute stroke onset time from imaging were compared.Results:DWI-FLAIR mismatch was found in 94 cases (94/207) of patients with symptom onset time from imaging ≤ 4.5 h, while in 28 cases (28/117) of patients with symptom onset time from imaging >4.5 h. ROC analysis showed that the AUC of DWI-FLAIR mismatch in predicting acute stroke onset time from imaging was 0.607, and the accuracy was 60.2%. The prediction model of deep learning based on single sequence showed that the AUC of FLAIR was 0.761 and the accuracy was 71.4%; the AUC of DWI was 0.836 and the accuracy was 81.6%. The AUC of predicting stroke onset time based on the multi-sequence (DWI+FLAIR) deep learning model was 0.852, which was significantly better than that of manual identification ( Z = 0.617, P = 0.002), FLAIR sequence deep learning model ( Z = 2.133, P = 0.006) and DWI sequence deep learning model ( Z = 1.846, P = 0.012). Conclusion:The deep learning model based on DWI and FLAIR is superior to human readings in predicting acute stroke onset time from imaging, which could provide guidance for intravenous thrombolytic therapy for acute stroke patients with unknown onset time.