Study On The Diagnosis And Prognosis Of Autoimmune Encephalitis Based On Machine Learning Combined With Clinical And MRI Feature

DIAGNOSIS OF AUTOIMMUNE ENCEPHALITIS BASED ON CLINICAL AND MRI FEATURES AND COMPARISON WITH HERPES SIMPLEX ENCEPHALITISObjective:This study investigated the early diagnosis of autoimmune encephalitis（AE）based on clinical and MRI characteristics,and differentiated from herpes simplex virus encephalitis（HSVE）.Material and Methods:（1）A total of 134 AE patients and 160 HSVE patients diagnosed in the department of Neurology of our hospital from October 2012 to January 2020 were collected.All patients underwent plain MRI and enhanced cranial MRI examination within 1 week after admission.Cerebrospinal Fluid（CSF）,blood test,electroencephalogram（EEG）,tumor screening and treatment of patients in the two groups were analyzed and compared.（3）To observe and compare the differences in the signal,distribution and enhancement mode of the lesions on conventional head MRI between the two groups.（4）The MRI manifestations and tumor incidence of patients with different subtypes of AE were collected,and the commonalities and differences of different subtypes of AE were summarized.Results:（1）The prodromal symptoms of AE group were very similar to those of HSVE,but the patients with HSVE were more likely to have headache and fever in the early stage of onset,and the difference was statistically significant（P<0.05）;Seizures,involuntary movement and memory disorders were more common in AE patients,and the differences were statistically significant（P<0.05）.The days of admission and the time from symptom onset to diagnosis in AE patients were longer than those in HSVE group,and the differences were statistically significant（P<0.05）.（2）The incidence of tumor in AE group was significantly higher than that in HSVE group,and the difference was statistically significant（P<0.05）.The incidence of ovarian teratoma was the highest in AE patients,and all of them were associated with anti-NMDAR patients.The second highest incidence of small cell lung cancer,mostly anti-GABA_BR encephalitis complicated.Among the 6 patients（4.48%）with positive multiple antibodies,3 were positive for anti-NMDAR antibody combined with other antibodies,and 3 were complicated with tumor（1 ovarian tumor,2 small cell lung cancer）.（3）In clinical treatment,most AE patients received first-line immunotherapy,and 76.12%of patients had improved symptoms at discharge.91.25%of HSVE patients received antiviral therapy,and some of them also received immunotherapy.88.75%of HSVE patients had improved symptoms upon discharge.（4）The intracranial pressure and white blood cell count of cerebrospinal fluid in the HSVE group were significantly higher than those in the AE group,and the differences were statistically significant（P<0.05）.Blood routine examination results showed that the white blood cell count of HSVE patients was lower than that of AE patients,the difference was statistically significant（P<0.05）.EEG results showed that the proportion of epileptic discharge in HSVE group was significantly lower than that in AE group,and the difference between the two groups was statistically significant（P<0.05）.（5）Comparison of MRI features showed that:（1）MRI findings of AE patients showed extensive distribution of lesions,which may occur in a wide range of gray and white matter areas including frontal,temporal,parietal,occipital lobe and hippocampus.In the anti-NMDAR encephalitis group,non-marginal lobe was the dominant lesion,while in the non-anti-NMDAR encephalitis group,marginal lobe including hippocampus was the dominant lesion,and the enhanced lesion was more in the anti-NMDAR encephalitis patients after enhanced scan.（2）The MRI of HSVE patients showed unilateral asymmetric abnormal signals in the medial temporal lobe,mostly involving the insula,cingulate gyrus or the deep frontal lobe.DWI may have limited diffusion and slightly high signal,and linear or gyrus-like enhancement may occur,and some patients could see signs of bleeding.（3）Compared with HSVE patients,AE patients were more likely to be associated with hippocampal lesions,although the difference was not statistically significant,and AE was also found to be more likely to have memory impairment,the difference was statistically significant（P<0.05）.Compared with gray matter,white matter is more affected in AE and HSVE patients.Conclusions:Compared with the patients in the AE group,the clinical symptoms of the patients were more complex and the positive rate of laboratory indicators was less.MRI examination could reflect part of the differences in image characteristics.Combined with the early clinical and image manifestations,part of the early and correct diagnosis could be made,which has certain significance for guiding clinical treatment.STUDY ON THE DIAGNOSIS OF AUTOIMMUNE ENCEPHALITIS BASED ON DEEP LEARNING COMBINED WITH MULTI-PARAMETER MRIObjective: To establish and verify an algorithm for early diagnosis of autoimmune encephalitis（AE）using deep learning（DL）method and multi-parameter MRI features.Material and Methods:（1）160 patients with acute AE,77 herpes simplex virus encephalitis（HSVE）patients and 188 age-and sex-matched healthy control（HC）subjects diagnosed in the department of Neurology of our hospital from January 2012 to December 2020 were collected.They were randomly divided into training set,validation set and internal test set in a ratio of 3:1:1.Fifteen AE patients,17 HSVE patients,and 20 HC patients from another hospital were collected as an external validation set.（2）All patients underwent brain MRI scan within 1 week after admission,including T₂-weighted imaging（T₂WI）,T₁-weighted imaging（T₁WI）,T₂-fluid attenuated inversion recovery sequence（T₂-FLAIR）and diffusion weighted imaging（DWI）sequences.（3）Five DL models were established based on single or combined four MRI sequences（T₁WI/T₂WI/T₂-FLAIR/DWI）,and the data sets were classified as AE, HSVE or HC,and reader experiments was performed by three radiologists.（4）The area under the ROC curve,AUC,accuracy,sensitivity,specificity,negative predictive value（NPV）,positive predictive value（PPV）,and F1 scores were used to evaluate the early diagnosis value of AE.The diagnostic performance was compared with that of radiologists（P< 0.05 was considered statistically significant）.Results:（1）In the internal test set,for the diagnosis of AE,the AUCs of the DL models trained with T₁ WI and T₂ WI sequences were 0.710 and0.716,respectively,which were slightly lower than the AUCs of the DL models based on T₂-FLAIR or DWI sequences,which were 0.806 and0.816,respectively.Similarly,in the external test set,the AUCs of the DL models based on T₁ WI and T₂ WI sequence training were lower than those of the DL models based on T₂-FLAIR or DWI sequence.（2）All DL models based on the single sequences have similar performance in HSVE diagnosis.The AUCs of models trained with T₁ WI,T₂WI,T₂-FLAIR and DWI were0.807,0.783,0.811 and 0.817,respectively.Among all single-sequence models,HC consistently has the highest classification performance,with an AUC of 0.85 and above.（3）In the internal test set,the performance of the fusion model for AE,HSVE and HC was significantly higher than those of the single sequence DL models,with AUC of 0.828,0.884 and 0.899,respectively.The model showed the same high diagnostic performance in the external validation set,with AUC of 0.831,0.882 and 0.892 for AE, HSVE and HC,respectively.（4）Compared with performance of radiologists’ evaluation,fusion model also showed higher diagnostic performance,the accuracy of radiologists and fusion model was 72% and83%,respectively,and the results were statistically different（P < 0.05）.Conclusions: DL models trained with multiple clinical routine MRI sequences can identify AE early and differentiate them from HSVE and HC with high diagnostic performance and may be superior to radiologists’ diagnostic level.This may provide a new diagnostic method for early diagnosis of acute encephalitis and guide clinical strategies.PROGNOSTIC PREDICTION OF ANTI-NMDAR ENCEPHALITIS BASED ON UNIVARIATE AND MULTIVARIATE LOGISTIC REGRESSION MODELSObjective: To study the risk factors affecting poor prognosis in patients with N-methyl-D-N-methyl-D-aspartate receptor（NMDAR）encephalitis,and to establish a clinical model for early prediction of disease prognosis by univariate and multivariate logistic regression.Material and Methods:（1）Clinical and brain MRI data of 139 acute anti-NMDAR patients diagnosed in the department of Neurology of our hospital from October 2012 to January 2021 were collected,and they were randomly divided into a training set of 97 patients and an internal test set of42 patients in a 7:3 ratio.In addition,87 anti-NMDAR encephalitis patients were collected from another local hospital and randomly divided into a training set of 61 patients and a test set of 26 patients.（2）All patients underwent brain MRI scan within 1 week after admission,including T₂-weighted imaging（T₂WI）,T₁-weighted imaging（T₁WI）,T₂-fluid attenuated inversion recovery sequence（T₂-FLAIR）and diffusion weighted imaging（DWI）sequences.Clinical data included demographic data,clinical manifestations,complications,admission to ICU,physical examination,laboratory examination,electroencephalogram（EEG）,electrocardiogram（ECG）,recurrence,and treatment plans.（3）Univariate logistic regression was used to analyze the clinical factors influencing the adverse prognosis of anti-NMDAR encephalitis,and multivariate logistic regression was used to include all clinical variables with P < 0.05.In order to find the independent predictors of adverse prognosis of anti-NMDAR encephalitis,a clinical prognostic model was established.Results:（1）Among 139 anti-NMDAR patients,the incidence of clinical symptoms of mental behavior abnormalities,motor disorders,cognitive dysfunction,decreased consciousness and speech disorders in the poor prognosis group was significantly higher than that in the good prognosis group,and the difference was statistically significant.Meanwhile,the incidence of prodrome in 14 patients（41.18%）was significantly higher than that in the good prognosis group（34.29%）,and the difference was statistically significant.（2）Among other clinical features,there were significant differences in ICU admission,tracheotomy,pyramid sign and initial m RS score between the poor and good prognosis groups（P< 0.05）.（3）In addition,18 cases（52.94%）of 34 patients with poor prognosis had recurrence,which was significantly higher than that of the good prognosis group（15.24%）,and the difference was statistically significant.The time from onset to initiation of treatment was significantly longer in the anti-NMDAR encephalitis group with poor prognosis,and the difference was statistically significant.（4）In terms of treatment,32.35% of anti-NMDAR patients in the poor prognosis group did not receive immunotherapy,while only 8 cases（7.62%）in the good prognosis group did not receive immunotherapy,the difference was statistically significant.Cerebrospinal Fluid（CSF）,blood test,EEG,ECG,and routine MRI were not associated with poor prognosis in the laboratory.（6）Multi-factor logistic regression analysis showed that the clinical model had high performance,with AUC of 0.840,accuracy of 0.905,specificity of0.914 and sensitivity of 0.857 in the internal test set.The AUC of the clinical model in the external test set was 0.837,slightly lower than that in the internal test set,and its sensitivity and specificity were 0.840 and 0.818.Conclusions: Clinical models can be used as a non-invasive auxiliary diagnostic tool for early prediction of anti-NMDAR encephalitis,and can help provide more early personalized treatment for patients who need more active monitoring.PROGNOSTIC PREDICTION OF ANTI-NMDAR ENCEPHALITIS BASED ON MACHINE LEARNING COMBINED WITH MULTI-PARAMETER MRIObjective: To explore the value of prognostic prediction for anti-N-methyl-D-aspartate receptor（NMDAR）encephalitis using two machine learning models（including deep learning（DL）model and radiomics model）,which trained with multi-parameter MRI.Material and Methods:（1）Clinical and brain MRI data of 139 acute anti-NMDAR patients diagnosed in the department of Neurology of our hospital from October 2012 to January 2021 were collected,and they were randomly divided into a training set of 97 patients and an internal test set of42 patients in a 7:3 ratio.In addition,87 anti-NMDAR encephalitis patients were collected from another local hospital and randomly divided into a training set of 61 patients and a test set of 26 patients.（2）All patients underwent brain MRI scan within 1 week after admission,including T₂-weighted imaging（T₂WI）,T₁-weighted imaging（T₁WI）,T₂-fluid attenuated inversion recovery sequence（T₂-FLAIR）and diffusion weighted imaging（DWI）sequences.（3）5 DL models and 5 radiomics models trained with single or combined 4 MRI sequences（T₁WI/T₂WI/T₂-FLAIR/DWI）were used to predict the prognosis of anti-NMDAR encephalitis in the acute phase.（4）The area under the ROC curve（AUC）and accuracy were used to evaluate the prediction performance of each model,and the paired t-test was used to compare the prediction value of each model（P < 0.05 was considered statistically significant）.Results:（1）For the DL model,the AUC of T₁ WI,T₂WI and T₂-FLAIR sequences was 0.721,0.747 and 0.77,respectively,which was lower than the AUC of DWI sequence（0.805）.The accuracy and sensitivity of the DL model trained with T₁ WI,T₂WI and T₂-FLAIR sequences were lower than that trained with DWI sequence.The DL model trained with combined 4 MRI sequences showed higher performance than all single-sequence models,with an AUC of 0.845 and an accuracy of 0.857,and the difference was statistically significant.（2）For the radiomics model,the performance of the radiomics model based on single sequence was satisfactory,with AUCs of 0.773（T₁WI）,0.786（T₂WI）,0.823（T₂-FLAIR）and 0.803（DWI）.In addition,the sensitivity and specificity of radiomics model trained with single-sequence MRI were higher and greater than 0.79 and 0.88,respectively.The radiomics model trained with the combined multi-parameter MRI had higher performance than all single-sequence models,with an AUC of 0.889 and an accuracy of 0.857.（3）For the fusion model,in the internal test set,the fusion model combined with DL and radiomics features had a very high prognostic performance,with AUC and accuracy of 0.963 and 0.976,respectively,much higher than the DL and radiomics models trained by combined sequences.Similarly,in the external test set,the AUC of the fusion model was 0.927 and the accuracy was0.880,which was significantly higher than the DL and radiomics models trained by the combined sequence,and the difference was statistically significant.Conclusions: Machine learning models trained with multi-parameter MRI features can be used as a non-invasive computer-assisted tool for early predict the prognosis of anti-NMDAR encephalitis.