Three biomedical repositories (PubMed, Web of Science, Embase) were systematically queried from inception through October 23, 2024, limited to English publications. The keywords used included terms such as “myelitis,” “prediction model,” “predictive factors,” “predictors,” “model,” and “scoring.” The complete search strategy is detailed in Table 1 . Our systematic review employed the PICOTS framework per CHARMS guidelines (20), which standardizes review objectives, search methodology, and eligibility criteria (21). Additionally, we utilized the CHARMS-PROBAST integrated Excel template developed by B. M. Fernandez-Felix and colleagues to streamline evidence extraction and bias appraisal for clinical prediction models (22). Core systematic review components include:

Two investigators (Huang JJ, Zhang ZZ) independently conducted selection. Initial title/abstract screening determined eligibility. Full-text articles underwent inclusion/exclusion assessment, with duplicate removal. Disagreements were resolved through tripartite consensus deliberation (Huang JJ, Zhang ZZ, Wu BQ).

Two researchers independently evaluated search results. Full-text eligibility assessments and discrepancies were resolved through deliberation or third-reviewer arbitration.

Data were documented via standardized Excel template (22), categorized as:

Data extraction was performed by one researcher, and all data were subsequently verified by another researcher to ensure consistency and accuracy, with necessary corrections made as needed.

Study bias risk and applicability were evaluated using the Prediction Model Risk of Bias Assessment Tool (PROBAST) (23). Two investigators (Huang JJ, Zhang ZZ) conducted independent appraisals. PROBAST provides a critical appraisal framework for prediction model studies (development/validation/updating), comprising 20 signaling queries across four domains: Participants, Predictors, Outcomes, Analysis. Each query accepts five responses: “Yes,” “Probably Yes,” “No,” “Probably No,” or “No Information.” Domain-level bias was deemed high-risk when ≥1 query received “No”/”Probably No.” Overall low bias risk required all domains achieving low-risk status.

Validation model AUC values underwent meta-analysis in RStudio (https://posit.co/download/rstudio-desktop/). Heterogeneity was quantified via I² index and Cochran’s Q test, with 25%/50%/75% thresholds denoting low/moderate/high heterogeneity (24). Effect model selection (fixed/random) depended on heterogeneity magnitude. Publication bias was assessed using Egger’s test (P>0.05 suggesting minimal bias) (25).

Figure 1 delineates the systematic review/meta-analysis search flowchart, presenting comprehensive search methodology and outcomes.

As of October 23, 2024, the initial search across the three databases yielded a total of 1,892 records. In the first stage, screening titles and abstracts resulted in the selection of 33 articles that were relevant to the study topic (at this stage, articles that clearly did not develop prediction models or focused solely on risk factors were excluded, and some borderline articles proceeded to the next evaluation stage). In the second stage, applying the inclusion and exclusion criteria led to 31 articles being selected for further evaluation. During this evaluation phase, 2 studies were excluded after discussion and assessment because they did not constitute prediction models in the strict sense and had fewer than two predictors. In the third stage, after removing 20 duplicate records, the final review included 11 studies and 11 models.

Details regarding the literature, model names, and key algorithmic applications of the 11 included studies are summarized in Table 2 ; the study designs, validation methods, sample sizes, data processing, and comprehensive analytical techniques are detailed in Table 3 . These studies, published between 2003 and 2024, include six multicenter collaborative studies, four single-center studies, and one study with unclear institutional reporting. Methodologically, the majority of studies (n=7) adopted a retrospective design, two incorporated both retrospective and prospective designs, and the remaining two consisted of one case-control study and one cross-sectional study. The study populations predominantly involved patients with neuromyelitis optica (NMO) (8 studies), three of which also included cohorts with multiple sclerosis (MS), and one included patients with myelin oligodendrocyte glycoprotein antibody-associated disease (MOGAD). The other studies focused on transverse myelitis (2 studies) and severe myelitis (1 study). The sample sizes across all studies ranged from 66 to 640 participants.

Among all the included studies, six focused on predictive diagnostic models, while five were prognostic models. Six studies utilized multivariate logistic regression algorithms to develop their models, and one employed Cox regression. Notably, Dean M. Wingerchuk et al. combined both multivariate logistic regression and Cox regression algorithms. Additionally, Wenqin Luo et al. integrated LASSO regression with Cox regression for modeling, Laura Cacciaguerra et al. (52) applied a deep learning convolutional neural network algorithm, Jacqueline Palace et al. adopted a joint modeling strategy, and Laura Clarke et al. utilized both machine learning and scoring algorithms in their modeling approaches.

The most frequently occurring predictor across the models was gender (featured in 6 models), followed by age (in 4 models). Other common predictors included number of relapses, central cord lesion, and long-segment transverse myelitis (LETM) (each appearing in 3 models). Predictors with relatively high frequency also included aquaporin-4 antibody (AQP4) serostatus (antibody titer), treatment duration, type of treatment regimen, spasticity, brainstem lesion, cervical cord lesion, and occurrence of optic neuritis (each present in 2 models).

Among the included studies, only three conducted both internal and external validation, while five performed internal validation only. Two studies did not carry out external validation, and their reporting on internal validation was unclear in the articles. Additionally, one study performed neither internal nor external validation (see Table 3 ).

Among the 11 included studies, the sample sizes used for training and testing the models are detailed in Table 3 . Only three studies described their methods for handling missing data, all of which employed direct exclusion, while the remaining eight studies did not report any approach for addressing missing values. Just four studies utilized sampling techniques: Bootstrap Resampling, Data Augmentation, Case-control Matching, and Stratified Random Sampling were employed respectively. Only one study (Wenqin Luo et al. (33),) implemented 10-fold cross-validation.

Additionally, seven studies applied dimensionality reduction techniques: Lucia Campetella and Liang Wang both used Statistics-Driven Variable Screening; Wenqin Luo applied LASSO Regression; Laura Cacciaguerra et al. (52) utilized a Convolutional Neural Network for feature reduction; Jingzi ZhangBao adopted a Stepwise Procedure; Laura Clarke conducted an Extensive Literature Review to select variables; and Fan Yang first screened variables with p < 0.05 in univariate analysis before incorporating significant predictors into multivariate logistic regression for further selection.

Furthermore, only Wenqin Luo et al. and Laura Cacciaguerra et al. (52) explicitly reported the use of hyperparameter tuning in their studies, while Laura Clarke et al. did not provide clear information on whether hyperparameter optimization was performed.

Of the 11 studies included, the core performance metrics used and their performance evaluations are presented in Table 4 . The model from Jingzi ZhangBao et al. demonstrated excellent discriminative ability (AUC 0.94), making it an outstanding diagnostic tool. The model developed by Laura Clarke et al. showed superior comprehensive performance (high precision and high F-score) and was significantly better than existing standards. The model by Fan Yang et al. had acceptable discriminative ability (AUC 0.77), but exhibited excellent calibration and clinical utility, making it a reliable prognostic risk prediction tool. In terms of methodological rigor and model performance, the models from Wenqin Luo et al. and the Laura Cacciaguerra (31) et al. are considered reliable, as they employed stringent internal validation and a comprehensive set of evaluation metrics. The respectable performance of earlier studies (e.g., Dean M.Wingerchuk et al.) should be interpreted with caution due to their higher risk of overfitting. The models by Laura Cacciaguerra (2022) et al. and Lucia Campetella et al. demonstrated exceptional discriminative ability (AUC > 0.9); the former is designed for classification tasks, and the latter for diagnostic purposes. The model from Liang Wang et al. showed moderate discriminative ability but excellent calibration and clinical utility, which is crucial for risk prediction models. The study by Jacqueline Palace et al. focused on interpreting risk factors rather than predicting individual outcomes; thus, its performance is reflected in the precision and significance of its effect estimates.

Of the 11 included studies, only one [Lucia Campetella et al. (35)], provided a model directly accessible to the public. The remaining studies did not offer prediction models that could be directly downloaded or used online. Among these, four studies had models whose construction details or scoring rules were fully published in their respective papers, but researchers, clinicians, or peers wishing to use these models must manually extract the information from the articles and implement them independently. There are no readily downloadable code or tools available, making the process highly cumbersome and error-prone, so these models are considered “theoretically applicable.” The other six studies presented models that are highly complex, lacking core parameters or code, rendering them usable only as research instruments and thus “unusable” (see Table 5 ).

In summary, while the majority of the studies have made valuable methodological contributions and pointing the way for future clinical practice, but their primary contribution lies in generating new medical knowledge rather than serving as direct diagnostic or treatment tools. From the perspective of clinical applicability, they have not yet become mature tools that are “truly ready for direct use by others.”

The PROBAST-based quality assessment revealed substantial methodological limitations. Only a single study (9.1%) exhibited low overall risk of bias. Conversely, six investigations (54.5%) demonstrated high bias risk, signaling deficiencies in development or validation methodology, while four studies (36.4%) presented unclear risk due to insufficient reporting transparency. Within the Participant domain, four studies (36.4%) incurred high bias risk principally because of inadequately specified inclusion/exclusion criteria (26–29), and two studies (18.2%) showed unclear risk owing to incomplete descriptions of data sources or participant selection methods (30, 31). Regarding Predictors, one study (9.1%) was rated high risk, potentially reflecting non-blinded assessment (28), while six studies (54.5%) held unclear risk as they omitted reporting quality control procedures for predictor measurement—a lapse potentially attributable to retrospective designs (26, 28–30, 32–34). For Outcomes, seven studies (63.6%) carried unclear risk, none having documented whether blinding was maintained between outcome and predictor evaluations (26, 28–30, 32–34). Analysis concerns were pronounced: six studies (54.5%) manifested high bias risk. Specific issues encompassed deficient handling of missing data (n=4 studies) (26, 28, 31, 35), exclusive use of apparent validation without further testing (n=5 studies) (27–29, 31, 35), problematic conversion of continuous variables to categorical formats (n=2 studies) (26, 27), and reliance on single-split internal validation alone (n=1 study) (31). An additional study (9.1%) exhibited unclear analytical risk due to unreported statistical details (30). ( Table 6 , Figure 2 ).

Applicability risk assessment revealed two studies (18.2%) with high overall risk, contrasting with nine (81.8%) demonstrating low applicability concerns. Domain-specific evaluation indicated high risk in Participants for one investigation (9.1%) due to notably limited sample size (29). Within Predictors, two studies (18.2%) received high-risk ratings stemming from suboptimal predictor measurement timing (29, 33). Conversely, all eleven studies exhibited universally low applicability risk in the Outcomes domain ( Table 6 , Figure 3 ).

Some of the included studies did not explicitly address comorbidity-related bias during the modeling process. These potential biases may lead to unmeasured confounding factors, limiting the accuracy and generalizability of the study results and models. The evaluation identified four studies with high bias, three with moderate bias, and four with low bias (see Table 7 ). Among these, studies with low bias clearly documented autoimmune or other underlying comorbidities and incorporated them into the analysis, demonstrating better bias control. Studies with moderate bias systematically excluded patients with severe comorbidities, leading to an underestimation of the negative impact of comorbidities on adverse outcomes, or presented risks of unmeasured comorbidity-related bias, or exhibited complex interactions between comorbidities and the disease under investigation. Studies with high bias either did not report or systematically adjust for comorbidities.

Synthesis of validation models proved feasible for eight studies ( Table 8 ), as insufficient methodological reporting precluded inclusion of others. Although the investigation by Clarke et al. employed multiple modeling methodologies (27), all derived from a common cohort. Consequently, only objectively developed machine learning models underwent meta-analysis. A random-effects model yielded a pooled AUC of 0.83 (95% CI: 0.75–0.91). Substantial heterogeneity emerged (I² = 97.4%, p < 0.0001). Publication bias analysis via Egger’s test indicated non-significance (z ≈ -0.47; intercept ≈ 0.91, p = 0.656), corroborated by visual inspection of funnel and forest plots ( Figures 4 , 5 , Supplementary Figures 1–2 ).