This study was a retrospective analysis of data from patients prospectively included in the NMOSD cohort at the Medical Center of Southwest China (Department of Neurology, West China Hospital, Sichuan University, China) between January 2015 and January 2022. Patients were included in this study based on the following criteria: (1) NMOSD diagnosis according to the 2015 diagnostic criteria (4) and (2) AQP4-IgG tested using a cell-based assay. Patients that met the following criteria were excluded: (1) incomplete clinical data, (2) follow-up time of less than 6 months, (3) MOG-IgG seropositivity, (4) uncertain treatment records, and (5) refusal to be followed up.

All patients diagnosed with NMOSD were registered in our database, and data on the following characteristics were collected: sex, age at onset, presentation of the first attack, concomitant autoimmune disease (including but not limited to systemic lupus erythematosus, inflammatory bowel disease, Sjogren’s syndrome, and rheumatoid arthritis), and maintenance therapy. Face-to-face or video call follow-ups were performed every 3–6 months, and features and presentation of each relapse were recorded. AQP4-IgG serology was detected using a commercial cell-based assay (CBA) (EUROIMMUN AG, Luebeck, Germany) (5, 6) MOG-IgG serology was tested by CBA as previously reported (7). Re-examination of AQP4-IgG was performed after 3–6 months for double-seronegative patients, and the Expanded Disability Status scale (EDSS) was assessed by a trained physician.

Patients who received immunosuppression therapy (IST) were treated with azathioprine (AZA) 2–3 mg/kg/day, mycophenolate mofetil (MMF) 1000–1500 mg/day, or rituximab (RTX) 1000 mg every 6 months, for pediatric patients and elders with weight less than 45 kg, the dose of MMF was adjusted according to the weight (20 mg/kg/day); the dose of RTX was adjusted according to the body surface area (375mg/m²), the bods surface area was calculated using Mosteller formula (weight [kg] x height [cm]/3600)½ (8).

Disease duration was defined as the interval between the date of first attack and the date of the final follow-up; time before IST was defined as the interval between the date of first attack and the date of initiation of IST, or the disease duration if patients were not treated with immunosuppressor; severe attack was defined as visual score ≥ 5 or EDSS ≥ 6, (the visual score was accessed using Snellen chart at 5 meters, a visual score of 5 was defined as: the worse eye with maximal visual acuity [corrected] less than 20/200 [0.1]) (9, 10); relapse was defined as a new worsening neurological function lasting more than 24 h in the absence of other identifiable causes, or newly found lesions confirmed by MRI, and occurring more than 30 days after a previous attack; active relapse was defined as more than 2 relapses with 12 months (11). Annualized relapse rate (ARR, number of relapses per patient-year) was calculated based on the recorded information; the first attack at onset was not included in the calculation of ARR. Visual disability was defined as a visual score of ≥ 5 points for more than 6 months (9).

Categorical variables are presented as frequencies (percentages). Continuous variables are presented as mean ± standard deviation (SD), and if not normally distributed, (Kolmogorov–Smirnov test was applied to verify the normality of distribution of continuous variables), as median (interquartile range, IQR). The mean differences between the two groups were analyzed using unpaired t-tests, while the median differences were evaluated using Mann-Whitney U tests. Categorial variables were compared using Fisher’s exact test.

Penalized likelihood-based methods have drawn much attention recently (12, 13). Because the variables are numerous, to avoid a complex model, predictor selection was performed using the Lasso regression algorithm and 10-fold cross-validation (14, 15). We also performed an exploratory Cox regression analysis, and variables associated with significant changes (p<0.2) in the univariable Cox regression were further analyzed using multivariable Cox regression. Specifically, sex as a demographic variable was included in multivariable Cox regression. A Schoenfeld residual test was performed to evaluate the possible violation of the proportional hazards assumption. To improve the statistical power, treatment with AZA, MMF, and RTX was combined with IST, treatment with oral glucocorticoids (GCs) only, and no therapy was combined with non-IST.

Prediction models were generated according to predictor subsets selected via different methods and displayed as nomograms. The C-index was applied to evaluate the discrimination of the models and was visualized via time-dependent curves. Calibration curves were created to evaluate the calibration ability of the models, and decision curves analysis (DCA) was performed to evaluate the clinical usefulness. To internally validate the results, we utilized the 0.632 bootstrap method (16) with 1000 resamples and calculated the mean overoptimism and corrected C-index, and also created calibration curves. Finally, the risk scores of all patients were calculated according to different models, and cut-off points were generated to recognize high-risk patients via receiver operating characteristic (ROC) analysis. Kaplan–Meier curves and log-rank tests were applied to detect the discrimination ability of cut-off points.

Restricted cubic splines (RCS) with four knots at 5^th, 35^th, 65^th, and 95^th centiles were used to flexibly model the association between onset age and the risk of visual disability, analysis of variance (ANOVA) was used to test the statistical significance of non-linear correlation.

All statistical analyses were performed using R software (version 4.0.3; R Foundation for Statistical Computing, Vienna, Austria; http://www.r-project.org/) with rms, ggDCA, survivalROC, glmnet, and survival packages. Statistical significance was set at p-values of < 0.05.

This study was approved by the Ethics Committee of Sichuan University (approval number: 2018 SHEN 29). Written informed consent was obtained from all patients.

Of 914 patients registered in our database between January 2015 and June 2022, 274 were excluded, and 640 patients were ultimately included in this study. The inclusion and exclusion flows of participants are shown in Figure 1 .

Table 1 summarizes the demographic and clinical characteristics of the study population. The cohort was characterized by several demographic features. The mean onset age was 38 ± 14 years (range 8 to 76 years), and 87% of the population was female. The majority of the cohort were Han Chinese, while 24 (3.8%) patients were Tibetans. The median follow-up time was 68 months. Seventy-one (11%) patients had concomitant autoimmune diseases, and 572 (89%) patients were AQP4-IgG seropositive. A total of 452 (70.6%) patients received maintenance IST during follow-up, and 188 (29.4%) patients did not receive maintenance therapy.

One hundred and sixty-one (25.2%) patients developed visual disabilities. The differences in sex, onset age, ethnicity, and comorbidity were not statistically significant between disabled and non-disabled patients. Compared to those who did not develop visual disability, the disabled patients had a higher ratio of optic neuritis as the presentation of onset (p <0.001) and a higher proportion of AQP4-IgG seropositive (p=0.017). Most (86%) of patients with non-visual disabilities received IST, while the percentage was significantly lower (26%) in patients with visual disabilities (p <0.001). Patients with visual disability showed a higher ratio of suffering a severe attack at the first onset and experiencing more frequent relapses before IST initiation.

We additionally analyzed visual outcomes in different age groups, the results are seen in Supplementary Table 1 . The highest proportion of patients who developed visual disability was observed in patients aged < 20 years, and that proportion decreased in the following 2 age groups, and then increased in the last 2 age groups, although the differences among groups were not statistically significant. RCS ( Supplementary Figure 1 ) suggested a non-linear correlation between onset age and risk of visual disability, the S-shaped curve suggested a decrease in risk of visual disability over age, and the risk increased after onset age > 30 years, and finally stay flat in patients over 55 years (p for non-linear association = 0.024).

Fourteen candidate variables were included in the selection, and the event per variable (EPV) was 11.5 (14/161), suggesting that the estimation of average risk was reliable (17, 18).

Four predictors were selected using Lasso regression: ON onset (Lasso’ β=0.52), ARR before therapy (Lasso’ β =0.15), IST (Lasso’ β=-1.58), and initial severe attack (Lasso’ β=0.19) ( Figures 2A, B ).

We performed an exploratory univariate Cox regression and a subsequent multivariate Cox regression. The effects of the predictors are summarized in Table 2 . Patients with older age at onset had a higher risk of developing visual disability (hazard ratio [HR]=1, 95% confidence interval [CI] 1-1.02, p=0.045); ON onset was an extremely strong predictor for visual disability (HR=2.91, 95% CI 1.99-4.24, p<0.001); compared to AQP4-IgG seronegative patients, the seropositive patients had significantly higher risk of visual disability (HR=2.98, 95% CI 1.48-5.98, p=0.002); maintenance IST decreased the risk of visual disability by 91% (HR=0.09, 95% CI 0.07-0.15, p<0.001); however, clinical features before IST initiation still predict the visual outcome: initial severe attack and higher ARR before IST were both identified as risk factors for visual disability.

Subsequently, two prediction models were generated and presented as nomograms (model 1, based on variables selected by Lasso; and model 2, based on variables selected by multivariable Cox regression). The nomograms for 1-, 3-, and 5-year visual-disability-free probability were depicted in Figure 3 . Lasso identified ON onset, higher ARR before IST initiation, initial severe attack, and no IST as predictors of visual disability. Three additional predictors were identified by multivariable Cox regression: male sex, older age at onset, and seropositive AQP4-IgG.

Discrimination ability was evaluated using 0.632 bootstraps with 1000 resamplings. Model 1 showed an overoptimism of 0.0034 and a corrected C-index of 0.88. In contrast, model 2 showed an overoptimism of 0.0074 and a corrected C-index of 0.88. The time-dependent C-index curves ( Figure 4 ) showed a decrease in the C-index over time but the model remained robust with C-index of 0.85. The C-index of Model 1 was relatively higher in the previous two years and reversed in the last three years. Calibration curves demonstrated good agreement between the predicted and actual 1-, 3- and 5-year visual-disability-free probabilities in both models ( Figures 5A–F ). DCA demonstrated that if the threshold probability of patients is between 0.1 to 0.9, both models presented superior net benefits for predicting visual disability compared to treat-all or treat-non strategy ( Figure 6 ).

In both models, the risk scores of each patient were calculated, and an optimal cut-off point for distinguishing between patients with low and high risk in 1 year was generated using ROC analysis. Patients were divided into high-risk and low-risk groups according to their scores. Kaplan-Merier curves ( Figure 7 ) showed good discrimination between those with high and low risk of visual disability. In both models, the cutoff point was 89.5 (the corresponding 1-year visual disability probability according to the nomogram is 0.15, and the corresponding net benefit according to DCA is 0.138) and 147.6 (the corresponding 1-year visual-disability probability according to nomogram is 0.15, and the corresponding net benefit according to DCA is 0.136), respectively.