Use and Safety of Immunotherapeutic Management of N-Methyl-d-Aspartate Receptor Antibody Encephalitis

Key Points Question What are the most effective treatments for N-methyl-d-aspartate receptor (NMDAR) antibody encephalitis? Findings In this meta-analysis of individual patient data including 1550 cases, treatment factors at first event that were significantly associated with good functional outcome 12 months from disease onset included first-line treatment with therapeutic apheresis alone, corticosteroids in combination with intravenous immunoglobulin (IVIG), or corticosteroids in combination with IVIG and therapeutic apheresis, while lack of immunotherapy within 30 days of disease onset was significantly associated with poor outcome. Rituximab and long-term IVIG use were significantly associated with nonrelapsing disease course. Meaning Separate treatment factors are associated with functional outcomes and relapsing disease biology in those with NMDAR antibody encephalitis.

N -methyl-D-aspartate receptor (NMDAR) antibody encephalitis (NMDARE) is the most common autoimmune encephalitis, predominantly affecting children and young adults. Management is challenging owing to the frequently severe disease in the acute phase, often requiring prolonged hospitalization, the variable and unpredictable outcomes, and the possibility of relapses. 1 Some patients recover fully, but many experience long-term cognitive and psychiatric problems, with significant effects on education, employment, and quality of life. 2 Overall, immunotherapy improves functional outcome 1,3-5 and reduces relapses, 1,6,7 but many therapeutic questions remain incompletely answered, including the effects of specific immunotherapies on outcome and risk of relapses and the role of second-line and maintenance long-term immunotherapies.
While a number of reviews have been published, [8][9][10][11][12][13] to our knowledge, no definite guiding data or treatment guidelines are available. Treatment strategies are heterogeneous, especially regarding second-line and maintenance immunotherapy. 14,15 Here, we have performed a systematic literature review and evidence synthesis of all published patients with NMDARE with available individualized immunotherapy data, toward 4 main aims: (1) mapping the use and safety of immunotherapies; (2) identifying early predictors of poor outcome and risk of relapse; (3) evaluating changes in immunotherapy use and disease outcome over the 14 years since first reports of NMDARE; and (4) providing an assessment of the Anti-NMDAR Encephalitis One-Year Functional Status (NEOS) score. 16 Methods A systematic search in PubMed was conducted from inception to January 1, 2019 (M.N.), with the following search keys: (anti-N-methyl-D-aspartate receptor encephalitis) OR (N-methyl-D-aspartate antibody encephalitis) OR (anti-NMDAR encephalitis) OR (anti-NMDA receptor encephalitis) OR (NMDA receptor encephalitis) OR (anti-N-methyl-D-aspartate receptor antibody encephalitis). Publications were eligible if they (1) included patients with NMDARE with positive NMDAR antibodies in serum and/or cerebrospinal fluid (CSF) and (2) provided individual patient data on immunotherapy. Patients with NMDARE preceded by central nervous system infection (ie, herpes simplex virus encephalitis) and large cohorts where individual patient details were not available 1 were omitted. This study followed the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guideline.

Study Definitions
Abnormal investigation findings were defined as presence of diffuse slow, disorganized, or epileptic activity or extreme delta brush on electroencephalography (EEG) 17 ; presence of CSF pleocytosis of 5 cells/uL or more or intrathecal oligoclonal bands 16,17 ; or presence of increased T2 or fluidattenuated inversion recovery parenchymal signal intensity or contrast enhancement on brain magnetic resonance imaging (MRI). Neurological severity in the acute phase and outcome at last follow-up were assessed via the modified Rankin Scale (mRS) score. 18 When not reported in the original article, mRS score was retrospectively assigned following review of adequate clinical data provided. Good outcome was defined as a final mRS score of 0 to 2 assigned within 12 months of disease onset (inferring mRS score of 0 to 2 at 12 months) and poor outcome as a final mRS score of 3 to 5 assigned after 12 months from disease onset (inferring mRS score of 3 to 5 at 12 months) or mRS score of 6 (death from NMDARE) at any time. Patients with a final mRS score of 0 to 2 assigned after 12 months or a final mRS score of 3 to 5 assigned before 12 months were excluded, as mRS score at 12 months could not be inferred. Monophasic course was the absence of relapse 24 months or more from disease onset.
First-line immunotherapy included corticosteroids, intravenous immunoglobulin (IVIG), and/or therapeutic apheresis. Second-line immunotherapy included rituximab and/or cyclophosphamide. Long-term (6 months or more) maintenance immunotherapy included monthly pulses or daily oral corticosteroids, monthly IVIG, rituximab redosing, and/or steroid-sparing agents (oral mycophenolate mofetil, azathioprine, or methotrexate). Early treatment was defined as initiation of immunotherapy within 30 days from first recognizable disease feature. 16,19,20 Adverse events to immunotherapy were categorized according to the National Institutes of Health Common Terminology Criteria for Adverse Events (CTCAE) version 5.0, with a focus on severe events (grade 3 to 5). Changes in immunotherapy use and disease outcome with time were primarily analyzed over 2 epochs: before and after 2013 (Titulaer et al 1 ; eMethods 1 in the Supplement). Additional post hoc analyses over 6 epochs (with cutoffs at 2-year intervals) were performed for rituximab use, relapse rate, and functional outcome only.

NEOS Score Validation
The NEOS score is a tool for prediction of 1-year functional status in patients with NMDARE determined from 5 patient characteristics, each scoring 1 point. 16 One of the 5 (no clinical improvement after 4 weeks of treatment) was not usually specified in published reports, so we instead applied a modified NEOS (mNEOS) score including 4 items, scoring 1 point each: intensive care unit (ICU) admission, no treatment within 30 days of symptom onset, abnormal findings on brain MRI, and CSF white blood cell count greater than 20 cells/uL. mNEOS scores were calculated for patients with available data in both the imputed and nonimputed data sets, with good and poor 1-year functional status operationalized per the study definitions above. The strength of association between mNEOS score and 1-year functional status was assessed using the Cuzick-Wilcoxon nonparametric test for trend 21 ; patients with scores of 0 to 1 and 3 to 4 were pooled together to avoid small groups. 16

Statistical Analysis
To prepare the data set for multivariable logistic regression, missing values in 35 predictor variables were imputed using hot-deck imputation with univariate-estimated weighting (eMethods 2 in the Supplement). 22 Target variables (functional outcome, relapsing disease course) were not imputed. After imputation, worst mRS score in the acute phase was binarized, age was recoded as 4 categorical variables, and 5 variables were removed, resulting in a final set of 33 predictor variables (eTable 1 in the Supplement). Patients with complete data in one or both target variables were entered into the respective logistic regression models, implemented with Statsmodels version 0.12.2 in Python version 3.6 (Python Software Foundation). Two-tailed P values less than .05 were regarded as significant. Three approaches were used. (1) Multivariable regression, in which the full available data set was presented in a single model. (2) Bootstrapped multivariable regression, in which random subsets (folds) of the data set were iteratively presented, to assess the stability and variability of the results. Cross-validation was run with 50-50 train-test split in bootstrap (n = 10 000) using class-balanced pseudorandomization (eMethods 2 in the Supplement). Odds ratios (ORs), 95% CIs, and P values were derived from the distributions of the regression coefficients over the 10 000 folds. (3) Bootstrapping with cross-validation, in which 20% of cases were held out of training (40%) and testing (40%), to show generalization of the methods to new data and provide a fair assessment of the prediction accuracy that can be reached on unseen patients.

Factors at First Disease Event Associated With Relapsing Disease Course
A total of 410 patients were included in the model predicting relapse (182 with relapse). In the bootstrapped logistic regression model ( Figure 2B), the only factor significantly associated with increased odds of relapsing disease was adolescent age (OR, 2.18; 95% CI, 1.18-4.15; P = .01). Two treatment factors were associated with nonrelapsing disease: use of rituximab as second-line treatment (OR, 0.17; 95% CI, 0.05-0.42; P < .001) and use of maintenance IVIG for 6 months or more from first event (OR, 0.16; 95% CI, 0.07-0.33; P < .001). Results from the nonbootstrapped multivariable regression are provided in eTable 7 in the Supplement. Accuracy for prediction of relapsing disease course in the cross-validation cohort was 63.4%.

Discussion
To our knowledge, this is the most comprehensive evidence synthesis to date focusing on immunotherapy in NMDARE, including individual treatment data from 1550 patients. Factors associated with good functional outcome (mRS score of 0 to 2) were adolescent age and first-line treatment with either therapeutic apheresis alone, corticosteroids with IVIG, or corticosteroids with IVIG and therapeutic apheresis. Factors associated with poor functional outcome (mRS score of 3 to 6) were infant or older adult age, ICU admission, extreme delta brush pattern on EEG, lack of immunotherapy within 30 days of onset, and IVIG treatment for 6 months or more. By contrast, relapsing disease was associated with adolescent age, and monophasic disease was associated with rituximab use or IVIG for 6 months or more.   While our data broadly validate the NEOS score, our different but complimentary approach was to evaluate the association of individual immunotherapies with outcome while controlling for a broad range of demographic and clinical characteristics, including age. Infants (younger than 2 years) and older adults (65 years and older) experienced the worst outcomes of NMDARE, with 3.6-fold and 3.8-fold, respectively, increased odds of poor outcome, while adolescents (aged 12 to 19 years) had 2.6-fold increased odds of good outcome compared with those aged 20 to 65 years at disease onset. This increased vulnerability to insult at the extremes of age may relate to differences in synaptic NMDAR composition in both the developing and aging brain. 23,24 Among the other clinical characteristics, ICU admission was the only significant independent predictor of poor outcome (2-fold increased odds), consistent with previous studies. 1,16,20 Abnormal MRI and EEG findings and inflammatory CSF have been variably reported in association with poor outcome of NMDARE. 16,[25][26][27][28] Here, we show that the extreme delta brush pattern, an interictal EEG abnormality highly specific for NMDARE, is independently associated with 2.6-fold increased odds of poor outcome, 29 as observed, albeit not statistically significant, in the first article describing it. 30 Although first-line immunotherapy (corticosteroids, IVIG, therapeutic apheresis) is widely used, limited and contrasting evidence exists regarding the best efficacy safety profile of different first-line treatment combinations (eTable 8 in the Supplement). [31][32][33][34][35] We found that therapeutic apheresis alone (5.6-fold increased odds of good outcome) or first-line treatment options used in combination (2.7-fold increased odds with corticosteroids and IVIG; 2.8-fold increased odds with corticosteroids, IVIG, and therapeutic apheresis) were effective in NMDARE, providing support for a pragmatic approach to selection of first-line therapies guided by adverse effect profile and patient acceptability. Importantly, the only approach to first-line treatment associated with worse outcome was deferral of treatment: lack of immunotherapy within 30 days of disease onset, which occurred in 363 of 728 patients (49.9%) in the total literature review cohort, was associated with 2.7fold increased odds of poor outcome, consistent with the findings of Titulaer et al 1 and other studies. 3,5,16,20,25,28,36,37 Regarding second-line immunotherapies, this evidence synthesis showed a striking association of rituximab administration with monophasic course, with 5.9-fold reduced odds of relapse after 24 months or more follow-up, also confirmed across all the major age groups in univariate analyses (eTable 9 in the Supplement). Our review also showed the emerging use of escalation second-line therapies, such as intravenous/ intrathecal methotrexate, subcutaneous/intravenous bortezomib, and intravenous tocilizumab, in a very limited subset of patients (eTable 4 in the Supplement), insufficient for inclusion in our multivariable modeling. [38][39][40][41][42][43][44][45][46][47][48][49][50][51][52][53][54] Their efficacy and safety in NMDARE warrant further investigation. 55 We were not able to assess the effect of timing of second-line immunotherapy in our multivariable models; however, post hoc univariate analysis showed that receiving earlier initiation of second-line immunotherapy (within 60 days of disease onset) was associated with 7.1-fold reduced odds of poor outcome compared with later initiation, despite similar disease severity at nadir (eTable 10 in the Supplement).
We found great heterogeneity in the use of and choice of agents for maintenance immunotherapy. In multivariable modeling, IVIG use for 6 months or more was associated with 10.3-fold increased odds of poor functional outcome (poor outcome at last follow-up in 10 of 24 patients [41%] in this group, reported across 17 articles) and yet 6.3-fold decreased odds of relapsing disease. These observations comprise a very small subgroup of cases within the analyzed group. Therefore, con- sidering the potential for publication bias of smaller cohorts favoring more atypical clinical features and immunotherapeutic responsivity, caution should be used in generalizing these findings more widely. A dissociation of functional outcome from relapse risk was also observed in the adolescent group (2.6-fold increased odds of good functional outcome yet 2.2fold reduced odds of monophasic disease), perhaps reflecting gene-environment interactions operating at this age to promote neuroinflammatory biology. 56 In our analysis of changes in immunotherapy use with time, the key finding was an increase in the use of rituximab  (Figure 3). However, overall functional outcomes did not improve with time. This is consistent with our finding (different from Titulaer et al 1 ) that second-line treatments were not significantly associated with functional outcome, but other factors may also be important; patients in the early epoch tended to fit more closely the typical patient with NMDARE, ie, young female adults with ovarian teratomas, whereas in the later epoch, we observed expansion of the recognized phenotype with significantly more men and children, lower tumor prevalence, lower symptom burden, and more atypical presentations, such as demyelination, alongside nonsignificant findings of later hospitalization and less prompt immunotherapy initiation (eTable 2 in the Supplement). These less typical patients might be recognized and treated later and thereby incur worse functional outcomes, despite similar or even less severe disease at nadir. Nevertheless, other metrics do suggest some improvements in management, such as a reduction in the proportion of patients requiring 60 days or more of hospitalization (125 of 210 [59.5%] vs 100 of 126 [79.4%] in the early epoch). Across the whole cohort, immune treatments were well tolerated, with treatment-related severe adverse events (CTCAE grades of 3 to 5) reported in only 3% and tending to occur mostly in bedbound patients in the ICU (eTable 5 in the Supplement), 57 although adverse events, especially minor ones, were likely greatly underreported and hence not presented in this analysis.

Limitations
The main limitations of our study include the retrospective nature of the data and inclusion of case reports that are susceptible to biases, such as reporting patients with worse disease or atypical features compared with the general population with NMDARE; this could also be a factor in the observed lack of improvement in functional outcomes across temporal epochs. Since only patients with individually reported treatment data were considered, different to previous reviews, 20,58 some of the major published cohorts 1 were not included but were instead used as a comparison (eTable 11 in the Supplement). Although extensively used, the mRS score may be too coarse to capture subtle differences in neuropsychiatric and cognitive symptoms at follow-up. 59 As mRS score was usually only reported or determined at nadir of the acute illness and final follow-up (rarely at 12 months exactly), our assignments of functional outcome at 12 months were operationalized estimates, justified by the well-documented typical trajectory of recovery in NMDARE. 1,13 Data collected were inherently limited by heterogeneous availability, hence hot-deck imputation, a robust method for handling missing data in large data sets, was used to enable multivariable analysis. 22 Although this method generates clinically plausible values (by constraining imputation to values already present in the database), it does not guarantee complete extinction of bias, as implicit assumptions are required in the choice of metric to match donors to recipients. While our multivariable modeling overall accounts for the contribution of each variable to the final predictions, statistical power may be limited when predictor variables are highly correlated (eg, 57.1% of patients receiving cyclophosphamide also received rituximab). We did not design the main model to test for specific interactions between predictors but afterwards evaluated this in a set of additional models and did not find any significant interactions between immunotherapy and severity or demographic factors in their association with functional outcome (eTable 12 in the Supplement).

Conclusions
This comprehensive and focused literature review on immunotherapy in NMDARE establishes a clear role for early immunotherapy and timely escalation to second-line treat-ment, particularly with rituximab. Importantly, we found that clinical factors that are associated with functional outcome often are not associated with relapsing disease biology. This should prompt important reevaluation of pragmatic treatment paradigms and physician preference to treat more severely affected patients more aggressively. Additionally, this synthesis of real-world data is an important means to direct relevant clinical and research questions, serving as the basis for the development of evidence-based optimal management and treatment guidelines of NMDARE.