This study was registered in the International Prospective Register of Systematic Reviews under the registration number CRD42024609424. It adheres to the 2020 PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [12], and the full 2020 PRISMA checklist is provided in Checklist 1.

A systematic search was conducted across 8 databases: PubMed, Embase, Cochrane Library, Web of Science, SinoMed, China National Knowledge Infrastructure, Wanfang Database, and VIP Database. The search covered the inception of each database up to November 2, 2024. Boolean operators (AND and OR) were used to combine Medical Subject Headings and free-text keywords, and detailed search strategies for all databases are provided in Multimedia Appendix 1.

The inclusion and exclusion criteria were systematically structured based on the population, intervention, comparator, outcome, and study design (PICOS) framework, with additional relevant domains (eg, language) incorporated, as summarized in Textbox 1.

Initially, 1 researcher (ZB) imported all retrieved records into EndNote X9 (Clarivate) software for deduplication. The deduplicated records were independently reviewed by 2 researchers (YS and ZS) through titles, abstracts, and keywords to identify studies that met the inclusion criteria. The same researchers reviewed the full texts of potentially eligible studies to finalize inclusion based on the criteria. At each stage, another researcher (LH) cross-checked the results to ensure consistency. Discrepancies were resolved through discussions involving all 3 researchers until a consensus was reached.

Data from the included studies were independently extracted by 2 researchers (YS and ZS) and documented in a Microsoft Excel spreadsheet. Data extraction and organization were performed using Microsoft Excel (version 2024). Extracted data included basic study information such as the title, first author, publication year, study region, and so on.

Another researcher (CJ) cross-checked the extracted data with the original reviewers. Disagreements were addressed through discussion among the 3 researchers until mutual agreement was achieved.

Two researchers (YL and JA) independently evaluated the quality of the included studies using the Prediction Model Risk of Bias Assessment Tool (PROBAST). This tool assesses bias risk and applicability in prediction model studies across 4 domains: participants, predictors, outcomes, and analysis. It also includes 20 signaling questions designed to pinpoint potential biases in study design, data analysis, and reporting. Additionally, PROBAST assesses the applicability of models within specific clinical contexts. Each study’s risk of bias was categorized as “high risk,” “low risk,” or “unclear risk” [13]. Differences were addressed through discussion involving a third researcher (HS).

The synthesis of data and statistical analysis were conducted using Stata 17 (StataCorp LLC) software. Meta-analyses of performance metrics were conducted using the meta package and the metagen function, with subgroup analyses comparing internal and external validation datasets. A random effects model was used to estimate pooled effect sizes with 95% CI, and heterogeneity was evaluated using conventional metrics. Funnel plots and the Egger test were used to assess potential publication bias. Sensitivity analyses were performed by excluding individual studies one at a time to evaluate the stability of the results and the impact of each study on the pooled effect sizes.

A total of 2307 studies were identified from the databases. After removing 985 duplicates (985/2307, 42.70%), 1322 studies (1322/2307, 57.30%) remained for screening. During the preliminary screening, we excluded 312 studies (312/1322, 23.6%) that did not include patients with UC, 865 studies (865/1322, 65.4%) unrelated to prediction models, 74 studies (74/1322, 5.6%) focusing on diagnostic prediction models, and 41 (41/1322, 3.1%) studies focusing on conference abstracts. Consequently, 30 studies (30/1322, 2.3%) were chosen for a full-text review. After independent verification and discussion, all 30 studies were ultimately included in the analysis (Figure 1).

The studies included in the analysis were published between 2015 and 2024, involving a total of 7452 patients with UC. Among these, 12 studies (12/30, 40%) focused on patients with moderate-to-severe UC, and 6 (6/30, 20%) exclusively targeted patients with severe UC. Geographically, the largest number of studies was conducted in China (11/30, 37%), followed by Japan (4/30, 13%). With respect to study design, 6 studies (6/30, 20%) were prospective, 22 (22/30, 73%) were retrospective, and 2 (2/30, 7%) used a mixed prospective-retrospective approach. Ten were single-center studies (10/30, 33%), whereas 20 involved multiple centers (20/30, 67%).

Key objectives of the included UC prognostic models included predicting treatment efficacy and response, assessing surgery-related risks, forecasting disease progression or relapse, and exploring molecular biomarkers or disease mechanisms. Treatment efficacy and response prediction was the most common objective (18/30, 60%), focusing on interventions such as tumor necrosis factor-alpha (TNF-α) inhibitors (eg, infliximab and adalimumab), vedolizumab, ustekinumab, tofacitinib, fecal microbiota transplantation, leukocyte apheresis, and the Chinese herbal medicine Wuwei Kushen Enteric–coated capsules. Surgery-related risk assessments (5/30, 17%) primarily aimed to predict pouchitis, postoperative complications, and long-term surgical outcomes. Disease progression and relapse prediction (6/30, 20%) focused on risks of acute severe UC, relapse probabilities, and remission likelihoods (Multimedia Appendix 2).

Most studies (19/30, 63%) did not specify how missing data were handled. Among the remaining studies, the most common approach was direct deletion (6/30, 20%), whereas others used random forest (1/30, 3%), k-nearest neighbors (1/30, 3%), mean imputation (2/30, 7%), and multiple imputations (1/30, 3%). Data splitting into training and validation sets was the most common method, with a 7:3 split used in 6 studies. Predictor selection methods included regression techniques (13/30, 43%), such as logistic regression (6/30, 20%), LASSO regression (4/30, 13%), Cox regression (1/30, 3%), and elastic net regularization regression (1/30, 3%). Machine learning methods were used in 6 studies (6/30, 20%), comprising random forest (5/30, 17%) and CatBoost (1/30, 3%).

Common predictors across studies were age, C-reactive protein (CRP), sex, albumin, erythrocyte sedimentation rate, platelet count, hemoglobin, disease extent or severity, endoscopic findings, and Mayo scores. Logistic regression was the most frequently used modeling method (12/30, 40%), followed by neural networks (8/30, 27%) and random forest (3/30, 10%). Cross-validation was used in 17 studies (17/30, 57%), with 5-fold cross-validation being the most common. Bootstrap validation was performed in 9 studies (9/30, 30%). 12 studies (12/30, 40%) provided calibration curves, 3 (3/30, 10%) included decision curve analysis (DCA), and 12 (12/30, 40%) developed decision-support tools. Most studies adequately discussed their limitations, which included small sample sizes, single-center designs, retrospective biases, and lack of external validation (Multimedia Appendix 3).

We systematically reviewed 30 studies on UC prognostic models. Most studies focused on predicting treatment outcomes, disease progression or relapse, and surgery-related risks, especially the effectiveness of biologic therapies such as TNF-α inhibitors. Logistic regression was the most commonly used method for predictor selection and model construction because of its simplicity, interpretability, and ability to provide insights into the relationships between predictors and outcomes [44]. Common predictors across most UC prognostic models included age, CRP, sex, albumin, erythrocyte sedimentation rate, platelet count, hemoglobin, disease extent/severity, endoscopic findings, and Mayo scores. Prior research has indicated a strong link between age and UC prognosis [45], whereas CRP has been recognized as an effective, noninvasive biomarker for evaluating treatment response in patients with UC [46]. Most studies did not clearly report how missing data were handled, and few developed decision support tools (eg, calculators or nomograms). Small sample size was a common limitation across the majority of studies.

The meta-analysis reported pooled AUCs of 0.84 (95% CI 0.77‐0.92) for overall validation, 0.83 (95% CI 0.70‐0.96) for internal validation, and 0.87 (95% CI 0.78‐0.95) for external validation. A notable and somewhat unexpected finding was that models evaluated through external validation demonstrated a higher pooled AUC than those assessed internally. This contradicts conventional expectations, as external validation typically yields lower predictive performance due to increased heterogeneity and real-world variability. Several factors may account for this discrepancy. First, further analysis of the 4 studies (4/30, 13%) in the external validation subgroup revealed several shared characteristics. Some used advanced modeling techniques, such as convolutional neural networks and artificial neural networks, which may have contributed to improved predictive performance. Additionally, the outcomes predicted by these models, such as drug sustainability or short-term treatment response, were inherently more predictable, which may have facilitated higher accuracy. Second, publication bias may have contributed to the observed results. Studies reporting poor performance in external validation may have remained unpublished or underreported, thereby inflating the pooled estimates. Third, the number of studies reporting external validation AUCs was relatively small (4/30, 13%), which may have introduced statistical uncertainty and led to overestimation of model performance. Moreover, according to the PROBAST risk of bias assessment, all 4 studies (4/30, 13%) were rated as having a high overall risk of bias. Therefore, although the pooled AUC from external validation appears high, this result should be interpreted with caution due to the limited number of studies, the high risk of bias, and the potential for selective reporting. Regarding sensitivity and specificity, the mean values for internal validation were 0.814 and 0.761, respectively, and for external validation, 0.830 and 0.757. It is important to note that these values were derived using simple unweighted averages due to the inconsistent reporting of 95% CIs across studies. Consequently, this approach does not account for variations in sample size or the precision of estimates, which may introduce bias. However, significant heterogeneity was observed across the included studies, particularly for internal validation. Sensitivity analysis indicated that studies with small sample sizes may have overestimated model performance due to overfitting. This finding underscores the importance of prioritizing high-quality training data and conducting appropriate sample size estimations during model development [47]. Some studies reported calibration curves and DCA, indicating the potential utility of these models in specific clinical scenarios. However, only a minority of studies comprehensively reported performance metrics (eg, 95% CIs for AUC, sensitivity, and specificity), limiting the comparability and robustness of the results.

The PROBAST tool–based quality assessment showed that the majority of studies carried a high risk of bias, especially in aspects concerning participant selection and model analysis. Many studies were limited to specific UC subgroups (eg, moderate-to-severe cases or patients receiving specific treatments), reducing the external applicability of the results. Additionally, some studies did not report sample size calculations or follow-up durations, further restricting the generalizability of the models. This limitation may lead to insufficient evaluation of model performance and reduce the credibility of the results. The lack of standardized approaches for predictor selection and the absence of explicit methods for handling missing data were notable limitations. The frequent use of direct deletion to address missing data may have introduced sample bias. Although 14 studies (14/30, 47%) conducted external validation, most validation datasets were small and lacked multicenter data. Only 2 studies (2/30, 7%) conducted both internal and external validation, raising concerns about the reliability and generalizability of current UC prognostic models in external settings. One study, which focused on a specific population, was identified as having a high risk of limited applicability [14]. However, it demonstrated a low risk of bias, underscoring its potential for future validation in targeted populations.

Artificial intelligence has attracted growing attention in the diagnosis and management of UC. Previous research has primarily concentrated on diagnostic models. For example, Jahagirdar et al [48] conducted a meta-analysis evaluating convolutional neural network–based algorithms for predicting endoscopic severity, whereas Puga-Tejada et al [49] examined artificial intelligence performance in detecting histological remission. However, despite their promise for real-time assessment, these studies primarily address current activity and overlook long-term outcomes.

In contrast, our study is the first to systematically review and meta-analyze prognostic prediction models for UC, focusing on outcomes, such as treatment response, relapse, disease progression, and the need for surgery. We used the PROBAST tool to evaluate model quality rigorously and synthesized key performance measures, including the AUC, sensitivity, and specificity. Furthermore, we assessed model calibration, decision-support tools, and the completeness of reporting. Unlike previous reviews, our study highlights the predictive utility and clinical relevance of prognostic models, providing valuable insights for advancing precision medicine in UC.

Future studies should prioritize the complete and transparent reporting of performance metrics, such as AUC, sensitivity, specificity, and their 95% CIs, to enable more reliable and meaningful evaluations of model utility. Additionally, researchers should incorporate appropriate sample size calculations and clearly report follow-up durations to improve methodological rigor and ensure reproducibility [50]. Addressing these limitations will enhance the overall quality, transparency, and clinical applicability of predictive models for UC.

In light of the growing significance of prediction models in UC prognosis, we recommend that future research adhere to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis reporting guidelines and adopt the 9-step framework for developing and validating clinical prediction models. These frameworks emphasize best practices, including clearly defining the target population and intended end users, selecting high-quality data sources, properly handling missing data, exploring alternative modeling approaches, and rigorously assessing model performance through both internal and external validation [51,52]. For instance, when defining the target population (step 1 of the 9-step framework), researchers should clearly distinguish between patients with newly diagnosed UC and those with long-standing disease, as their clinical trajectories, risk profiles, and treatment responses may differ significantly. During model development, step 4 (handling missing data) is particularly relevant in UC research, where laboratory and endoscopic variables are frequently incomplete. Therefore, appropriate statistical approaches, such as multiple imputation, should be used in place of listwise deletion, which can introduce bias and reduce statistical power. For step 8 (external validation), it is crucial to evaluate model performance using datasets from distinct geographic regions or health care systems. For example, a model developed in a tertiary care center in East Asia could be externally validated using a population-based registry from Europe or North America. This approach helps assess the model’s generalizability across diverse UC populations. Furthermore, in alignment with Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis recommendations, future studies should ensure transparent reporting of all aspects of model development, including predictor selection, handling of missing data, and strategies to address overfitting. Reporting should also include calibration plots, discrimination metrics (eg, AUC or C-index), and their 95% CIs to support comprehensive and interpretable performance evaluation.

This study has several limitations. First, substantial heterogeneity existed among the included studies in terms of study design, population characteristics, modeling approaches, and outcome definitions, which may have affected comparability and introduced variability into the pooled estimates. Second, external validation was limited, with most models relying on small or single-center datasets, thereby restricting their generalizability. Third, many studies did not consistently report key performance metrics, such as 95% CIs for AUC, sensitivity, and specificity, limiting the ability to critically evaluate and compare model performance. Fourth, missing data were often poorly addressed, with many studies using complete-case analysis or listwise deletion, which increases the risk of bias and reduces statistical power. Finally, we only included studies published in English or Chinese and searched 8 major databases, which may have introduced language bias and led to the omission of relevant studies from other languages, sources, or the grey literature.

This study identified 30 prognostic prediction models for UC, encompassing 7452 patients, with most studies conducted in China and Japan. The majority of studies were retrospective and multicenter, primarily aimed at predicting therapeutic responses, particularly to TNF-α inhibitors, such as infliximab and adalimumab, as well as estimating the risks of surgery, disease progression, or relapse. Logistic regression was the most frequently used method for predictor selection and model development, with commonly used predictors, including age, CRP, albumin, hemoglobin, disease extent, and Mayo scores. The meta-analysis revealed a pooled AUC of 0.84 (95% CI 0.77-0.92). However, notable limitations were identified, including a high risk of bias in most studies, insufficient external validation, and restricted generalization due to small sample sizes and subgroup-specific designs. These findings underscore the need for future research to prioritize robust model development, rigorous external validation, and enhanced generalization to support broader clinical application. Ultimately, well-developed and validated prognostic models hold the potential to guide personalized treatment strategies, enhance clinical decision-making, and improve outcomes for patients with UC in real-world settings.