This study was conducted in accordance with the Preferred Reporting Item for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (8). The PRISMA checklist can be seen from Supplementary Table 1. This study was registered in the Prospective Register of Systematic Reviews (PROSPERO; CRD420251054364).

Articles published between the establishment of the databases and May 2025 were retrieved in PubMed, Web of Science, Embase and the Cochrane Library databases. We constructed the retrieval strategy based on the PICO principle (population, intervention, control, and outcomes). Population: patients with HZ. Intervention: establish ML model for predicting PHN. Control: the gold standard for diagnosing PHN. Outcomes: The predicted result, including true positive (TP), false positive (FP), true negative (TN) and false negative (FN). The research on search strategies is summarized in Supplementary Table 2.

The parameters of this systematic review, as defined by the PICOS framework, were as follows:

Inclusion criteria: 1. All the included literatures were published in English. 2. The included study adopted the clear definitions of HZ and PHN. The population was patients diagnosed with HZ, and the model endpoint was set as the onset of PHN. 3. Modeling was carried out using at least one ML method. 4. Reporting data that can infer TP, FP, TN and FN of the model.

Exclusion criteria: 1. Duplicate studies, non-English publications, and studies with missing or non-convertible data were excluded. 2. Studies that only reported the corresponding risk factors of PHN using logistic regression (LR) but did not model for prediction were excluded.

Titles and abstracts of potentially eligible studies were screened by two independent researchers (Zheng Lin and Hongfei Wang), and the disagreement were resolved by the third independent researcher (ChenXi Ma). Subsequently, the full text of these studies was systematically assessed to further confirm whether they met the inclusion criteria.

For the publications included in the analysis, two independent researchers (Zheng Lin and Wang Hongfei) systematically collected the following information: 1. basic information about the study, including authors, publication year, country, study design, sample size, and diagnostic criteria for HZ and PHN. 2. Information related to the models, including the dataset, model type, and predictors. 3. The performance indicators of the model, including the area under the curve (AUC), sensitivity, specificity, TP, FP, TN and FN.

The Prediction Model Risk of Bias Assessment Tool (PROBAST) is suitable for evaluating the risk of bias and applicability of the original studies for the development or validation of multivariate diagnostic/prognostic models (9). Two independent researchers (Zheng Lin and Wang Hongfei) used the PROBAST to evaluate each included study, and the disagreement were resolved by the third independent researcher (ChenXi Ma).

All analyses in this study were done based on Stata 14.0 (Stata Corporation, Texas, United States) and R 4.4.2 (R Foundation for Statistical Computing, Vienna, Austria).

The main statistical measures used in meta-analysis were sensitivity and specificity. Draw the Summary Receiver Operating Characteristic (SROC) curve summarized by the sensitivity and specificity of each study. The diagnostic value of the ML is reflected through the Fagan diagram and the distribution scatter diagram and evaluated by positive likelihood ratio (PLR) and negative likelihood ratio (NLR).

I^2 was used to evaluate the heterogeneity level of the included studies. When the I^2 is less than 50%, the fixed-effect model is selected. When I^2 was greater than or equal to 50%, the random effects model is selected. If the heterogeneity among studies is significant, potential sources of heterogeneity are explored through sensitivity analysis and meta-regression. Further subgroup analyses were conducted on studies that adopted different dataset types, model types, and types of predictors. Publication bias was evaluated using funnel plots and Deeks’ test.

According to our search strategy, a total of 5,183 relevant literatures were retrieved. After eliminating 829 duplicate literatures, a preliminary screening was conducted on the remaining 4,354 literatures. By reviewing the titles and abstracts, 4,279 literatures that did not meet the inclusion criteria were excluded. Subsequently, a full-text search and evaluation were further conducted on the remaining 75 literatures. Among them, 47 articles were excluded because the research results were not applicable to the topic of this study, 10 articles were excluded because ML methods were not used for modeling, 3 articles were excluded because the research subjects did not meet the requirements, and another 1 article was excluded because the data was unavailable. Ultimately, 14 literatures were included in the meta-analysis (10–23). The process of literature retrieval is shown in Figure 1.

Among the 14 included studies, 6 were prospective cohort studies, and the remaining 8 were retrospective cohort studies or cross-sectional studies. These studies involve four countries: Germany, Japan, South Korea and China. Among them, only one study was published in 1998, while the rest were all published after 2019. The total sample size included 16,514 patients with HZ and 2,726 patients with PHN. Regarding the definition of PHN, two studies defined it as pain that persisted for 1 month after the rash healed, 11 studies defined it as pain that persisted for 3 months after the rash healed, while one study did not clarify the definition of PHN it used. The basic information of this research can be found in Supplementary Table 3.

A total of 41 model performance metrics were generated in 14 studies. Among them, 24 were developed based on the training set, 16 were verified through the internal validation set, and only 1 model completed the external validation. In terms of the frequency of use of ML methods, LR (11 times) is the most common, followed by random forest (7 times), linear regression (5 times), support vector machine (5 times), gradient boosting (4 times), artificial neural network (3 times), K-nearest neighbor (2 times), and Bayesian layering (2 times). The top three in terms of the frequency of use of predictors were age (34 times), tumor (21 times), and NRS/VAS score (19 times). The summary of the characteristics of these models is presented intuitively in Figure 2.

After PROBAST’s assessment (Supplementary Table 4), all 14 studies had a certain degree of risk of bias, but most of the studies showed good applicability. Furthermore, all the studies have certain deficiencies in statistical analysis. Among them, 12 studies did not report the calculation process of the required sample size, 4 studies did not report the handling method of missing data, 7 studies did not evaluate the model according to the PROBAST standard, and 11 studies did not consider the risk of model overfitting. Besides, three studies also show obvious bias tendencies in the selection of predictors.

We synthesized the performance of 41 models using the random effects model. The results (Figure 3) showed that the sensitivity of ML in predicting PHN was 0.81 (95% confidence interval (CI): 0.74–0.86), and the I^2 was 98.74% (p < 0.001). The specificity was 0.84 (95% CI: 0.79–0.88), and the I^2 was 99.63% (p < 0.001). Furthermore, the SROC curve (Figure 4) showed that the AUC value of ML for predicting PHN was 0.90 (95% CI: 0.87–0.92).

The Fagan plot (Figure 5) shows that HZ patients determined as positive by the ML model have an 80% probability of developing PHN, while the probability of HZ patients determined as negative by the model developing PHN is only 20%.

The distribution scatter diagram (Figure 6) shows that the PLR of ML for predicting PHN is 5.17 [95%CI: 3.94–6.76], with a NLR of 0.22 [95%CI: 0.17–0.30].

We conducted a detailed subgroup analysis, the subgroups included the comparison between the training set and the validation set, the comparison between LR and other ML methods, the comparison between clinical indicators and clinical indicators with serum indicators and omics analysis, the comparison between prospective cohort and retrospective cohort. Additionally, a subgroup was defined where PHN was defined as postherpetic pain lasting more than 3 months after HZ (Table 1).

The results of subgroup analysis showed that on the training set, the predictive performance of the ML model for PHN was sensitive (0.81 [95% CI: 0.70–0.89]), specific (0.88 [95% CI: 0.81–0.92]), and AUC (0.92 [95% CI: 0.89–0.94]; Supplementary Figure 1). On the validation set, the predictive performance of the ML model for PHN was sensitivity (0.81 [95% CI: 0.75–0.86]), specificity (0.78 [95% CI: 0.71–0.84]), and AUC (0.87 [95% CI: 0.83–0.89]; Supplementary Figure 2).

The results of subgroup analysis showed that the predictive performance of the LR model for PHN was sensitive (0.84 [95% CI: 0.78–0.88]), specific (0.73 [95% CI: 0.65–0.80]), and AUC (0.86 [95% CI: 0.83–0.89]; Supplementary Figure 3). The predictive performance of the other ML model for PHN was sensitivity (0.78 [95% CI: 0.66–0.87]), specificity (0.90 [95% CI: 0.86–0.93]), and AUC (0.93 [95% CI: 0.90–0.95]; Supplementary Figure 4).

The results of subgroup analysis showed that the predictive performance of the model constructed using clinical indicators for PHN was sensitive (0.82 [95% CI: 0.72–0.89]), specific (0.82 [95% CI: 0.70–0.90]), and AUC (0.89 [95% CI: 0.86–0.91]; Supplementary Figure 5). The predictive performance of the model constructed by comprehensively using clinical indicators, serum indicators, and omics analysis for PHN was sensitivity (0.81 [95% CI: 0.72–0.87]), specificity (0.85 [95% CI: 0.80–0.89]), and AUC (0.90 [95% CI: 0.87–0.92]; Supplementary Figure 6).

The results of subgroup analysis showed that the predictive performance of the model constructed using prospective cohort for PHN was sensitive (0.87 [95% CI: 0.82–0.90]), specific (0.69 [95% CI: 0.61–0.76]), and AUC (0.88 [95% CI: 0.84–0.90]; Supplementary Figure 7). The predictive performance of the model constructed by comprehensively using retrospective cohort for PHN was sensitivity (0.78 [95% CI: 0.68–0.85]), specificity (0.88 [95% CI: 0.83–0.91]), and AUC (0.90 [95% CI: 0.88–0.93]; Supplementary Figure 8).

The results of subgroup analysis showed that the predictive performance for PHN of the model constructed using studies which is defined PHN as pain persisting for more than 3 months after HZ was sensitive (0.81 [95% CI: 0.73–0.87]), specific (0.85 [95% CI: 0.80–0.89]), and AUC (0.90 [95% CI: 0.87–0.93]; Supplementary Figure 9).

The sensitivity test (Supplementary Figure 10) indicated that even after re-meta-analysis after removing any one study, the results remained robust.

Meta-regression analysis (Figure 7) indicates that model type, dataset type and predictor types are the main sources leading to heterogeneity.

The results of the Deeks’ test showed that there was a certain publication bias in this study (p < 0.01). Funnel plot analysis further indicates that there is asymmetry in the literature distribution on both sides of the regression line (Figure 8).

This meta-analysis integrates the existing evidence and covers 41 models generated from 14 studies. The results show that the use of ML to predict PHN has high efficiency. The comprehensive sensitivity and specificity of ML are both greater than 0.8, which is more effective than single indicators such as age and pain score or the neuropathic pain scale (24, 25). This indicates that the precision medicine strategy based on ML has potential value in managing PHN.

Among the studies included in the meta-analysis, age, pain score and tumor history were the most commonly used indicators. These three indicators have been repeatedly verified as key factors closely related to PHN (26, 27). Therefore, when developing PHN prediction models in the future, we suggest that at least these three indicators should be included in the initial screening indicator library.

Furthermore, the currently developed PHN models generally regard the pain score of patients at their first visit as an important predictor. However, the pain trajectory of HZ patients may also be of great significance for the prediction of PHN. The results of a prospective study based on the community population verified the potential value of pain trajectories in the prediction of PHN (28). The latent category trajectory model, as a method of ML, can divide heterogeneous populations into several homogeneous patterns or categories, thereby effectively describing the pain trajectories in the HZ patient group and further being used for the risk prediction of PHN (29). Therefore, we suggest that more attention be paid to the pain trajectory of patients with HZ in future research and clinical practice.

In the subgroup analysis, no significant performance differences were observed between the models constructed based on clinical indicators and those constructed in combination with clinical indicators, serum indicators or omics analysis (AUC 0.89 [95% CI: 0.86–0.91] vs. 0.90 [95% CI: 0.87–0.92]). Some indicators, although they may improve the performance of the model to a certain extent, are rarely adopted in clinical practice. This prompts us to further explore how to achieve an effective balance between model performance optimization and the complexity of predictors. The achievement of this goal requires more high-quality research for in-depth exploration.

Among the included literature, LR is widely used as the main ML method. This might be attributed to its relatively simple structure, which is easy to understand and more acceptable to medical professionals (30). A meta-analysis summarized the differences in predictive efficacy between LR and other ML methods, and pointed out that LR has similar predictive performance to other ML methods (31). In this study, we found that in the field of predicting PHN, the predictive efficacy of the LR model was generally comparable to that of other ML methods. Among them, the LR model shows higher sensitivity (0.84 [95% CI: 0.78–0.88] vs. 0.78 [95% CI: 0.66–0.87]), while other ML models demonstrate better specificity (0.90 [95% CI: 0.86–0.93] vs. 0.73 [95% CI: 0.65–0.80]). Therefore, on the premise of fully explaining the model, the performance advantages of other ML models over LR models in the field of predicting PHN remain a highly controversial topic.

Ensemble techniques typically integrate multiple ML algorithms through methods such as bagging, boosting, stacking, and voting, thereby balancing the strengths and weaknesses of ML (32). We suggest that if the constructed PHN ML models have their own advantages and disadvantages in terms of performance indicators, the ensemble learning method can be attempted to be adopted to integrate the advantages of different ML algorithms, thereby further improving the overall performance of the models.

In the subgroup analysis of the meta-analysis, we observed that the performance of the ML model on the training set was significantly higher than that on the test set (AUC 0.92 [95% CI: 0.89–0.94] vs. 0.87 [95% CI: 0.83–0.89]). This result indicates that the currently published PHN prediction models may have a certain degree of overfitting risk. Meanwhile, in the validation set, the specificity of predicting PHN decreased significantly (0.88 [95% CI: 0.81–0.92] vs. 0.78 [95% CI: 0.71–0.84]). This means that when applied to external datasets, the PHN prediction model has a certain false positive rate. Furthermore, most studies do not adopt external validation methods, which leads to considerable uncertainty in the performance evaluation of these models on external data. Therefore, in the future development process of PHN models, we suggest comprehensively applying strategies such as external validation, cross-validation, and resampling. These methods have been widely proven to significantly enhance the generalization ability of the model (33).

Besides, we observed that the models constructed based on prospective cohorts performed worse than those based on retrospective cohorts (AUC 0.88 [95% CI: 0.84–0.90] vs. 0.90 [95% CI: 0.88–0.93]). This might be because the risk of bias in retrospective data is higher, and thus the risk of overfitting in the models built from retrospective data is also higher. Therefore, we suggest that future research should, as far as possible, adopt the strategy of prospective cohort studies.

We have noticed that in the process of model development, many researchers often face the problem of balancing sensitivity and specificity (34). For instance, in the development of the PHN ML model, a study from South Korea reported that their model demonstrated low sensitivity (<0.6) and high specificity (>0.9) (23). In fact, based on the ROC curve drawn in this study, researchers were fully able to achieve a balance between sensitivity and specificity by adjusting the cut-off value. To solve this problem, we suggest that in the future when developing PHN prediction models, at least three cutoff values should be reported to meet different clinical needs. The first one is the balanced cutoff value, which is used to determine whether PHN occurs. The second one is called the positive cutoff value. This cutoff value should ensure that the sensitivity is higher than 90% without overly focusing on specificity. The third one is called the negative cutoff value. This cutoff value should ensure that the specificity is higher than 90% without overly focusing on the sensitivity. By introducing these three cut-off values, more comprehensive support can be provided for clinical decision-making.

In recent years, numerous advanced pain management techniques, such as spinal cord stimulation and stellate ganglion block, have demonstrated efficacy in preventing the onset of PHN (35–37). When integrated with efficient machine learning models, these interventions can generate synergistic effects. HZ patients identified by machine learning models as being at high risk for developing PHN can be prioritized for advanced pain management strategies to reduce the likelihood of PHN occurrence, whereas those classified as low-risk can be managed with conventional treatment approaches. This integrated approach enables optimal patient care in a cost-effective manner.

However, despite the good performance demonstrated by current PHN prediction models, their applicability to external datasets remains uncertain due to the absence of external validation, non-standardized reporting formats, and limited sample sizes used in model development. We suggest that future development of PHN prediction models should involve at least two independent cohorts and strictly follow the TRIPOD guidelines to ensure standardized reporting of predictive models (38).

Despite rigorous search and systematic evaluation, this meta-analysis still has several limitations that cannot be completely avoided. Firstly, the heterogeneity among the included studies was relatively high. The difference between ML models and traditional diagnostic tests lies in the different built-in parameters and cutoff values of each model. ML is more sensitive and more susceptible to the influence of data quality, which leads to significant heterogeneity among ML models. The meta-regression indicated that this might stem from differences in datasets, model selection, and predictor definitions. In the future, this issue can be addressed by adopting a more rigorous standardized reporting framework for predictive models and using larger sample sizes to improve model quality. Secondly, some studies did not strictly follow the standard norms of the prediction model reports, and the vast majority of studies lacked external validation (39). Therefore, a cautious attitude should be maintained when interpreting the performance of these studies. Finally, the included studies showed significant publication bias, which might lead to a certain degree of overestimation of the conclusions regarding ML performance in this study. More high-quality research is urgently needed in the future to further verify and supplement the current findings.