We used 2 types of data sources for this study. Synthetic data were based on the “Colorectal Cancer Clinical Library Artificial Dataset” (the synthesized data is archived on the National Cancer Data Center [NCDC] portal) [20]. This dataset is a high-quality synthetic resource that replicates the structure and statistical characteristics of real-world medical data. It was developed based on real-world patient information collected from 3 leading Korean medical institutions specializing in cancer (Samsung Medical Center, Yonsei University Severance Hospital, and National Cancer Center) using the real-world time-series generative adversarial network (RTSGAN) technology [21]. Hospital data comprised clinical information of patients who underwent colorectal cancer surgery at the Hwasun Jeonnam University Hospital in South Korea. This dataset is derived from real patient treatment experiences and medical records, containing detailed clinical indicators and postoperative follow-up results. Study participants included in the synthetic and hospital data were selected based on the following inclusion and exclusion criteria. Inclusion criteria were as follows: (1) patients diagnosed with colorectal cancer between 2009 and 2014, and (2) patients undergoing surgery. The exclusion criterion was patients with missing survival information.

The primary outcome of this study was 7-year overall survival data. The synthetic dataset included 31 variables, while the hospital dataset contained 48 variables. In total, 12 variables common to both datasets were selected as the input variables for model training. The baseline model was intentionally restricted to 12 shared variables common to both datasets to allow for a methodologically controlled comparison between the hospital and synthetic datasets. This feature-constrained design is intended to minimize confounding factors that may arise due to differences in variable availability and information content [22,23]. The final 12 variables were sex, age, stage (0, I, II, III, or IV), adenocarcinoma histology, vascular invasion, lymph node invasion, perineural invasion, tumor location (colon, junction, or rectum), preoperative carcinoembryonic antigen (log-transformed), postoperative chemotherapy (FOLFOX, FOLFIRI, GELOX, or other), postoperative radiation therapy, and number of resected metastatic lymph nodes.

All categorical variables were encoded as one-hot. Age was used as a binary indicator by interval (<20, 20‐29, …, 80‐89, ≥90). Input variables or outcomes with missing values were excluded from the analysis, based on the principles of complete case analysis.

To minimize the domain gap between synthetic and hospital data, we applied a domain adaptation technique, a form of transfer learning. Domain adaptation broadly refers to solving the problem of performance degradation caused by the difference in distribution between the source and target domains. In this study, we implemented domain adaptation through parameter transfer. It was first trained on synthetic data (source domain) and then fine-tuned on hospital data (target domain) using a low learning rate and regularization. This approach allows the model to adapt to hospital-specific characteristics using limited real-world data while retaining generalizable patterns learned from large-scale synthetic data. In this study, we did not apply feature-level alignment or adversarial domain adaptation techniques. Instead, we focused on parameter-level adaptation, which is particularly well-suited to tabular clinical data, where explicit feature alignment across domains is difficult, and privacy constraints may limit access to joint feature distributions [24-26]. Furthermore, to determine the presence of a domain gap, we performed the Kolmogorov-Smirnov test for continuous variables and chi-square (or Fisher exact test, where expected cell counts were less than 5) independence tests for categorical variables.

In this study, synthetic data were used for pretraining and internal optimization, while hospital data were used for domain adaptation, threshold selection, probability calibration, and independent final evaluation. Therefore, the synthetic data were split into source-train or source-validation data based on the year of diagnosis, and the hospital data were sorted by the year of diagnosis and split into target-train, target-validation, or target-test using a 6:2:2 3-partition (Figure 1).

Synthetic data were split into source-train and source-validation sets based on the year of diagnosis and were used for pretraining and internal optimization. Hospital data were chronologically divided into target-train, target-validation, and target-test sets using a 6:2:2 ratio and were used for domain adaptation, threshold selection, probability calibration, and independent final evaluation. The pretrained model was fine-tuned on target-train data through domain adaptation, with threshold selection and calibration on target-validation, and independent final evaluation on the held-out target-test set. Four model variants were evaluated: a baseline model (Hospital → Hospital), a zero-shot model (Synthetic → Hospital), a domain adaptation model (Synthetic → Hospital), and an ensemble model combining predictions from multiple models.

According to the National Cancer Information Center, the 5-year relative survival rate for colorectal cancer is 75% [27]; however, the outcome measured in this study showed similarly high survival expected at 7 years. To address this class imbalance, sampling techniques—including no sampling, random undersampling (RUS) [28], and the synthetic minority oversampling technique combined with edited nearest neighbors (SMOTEENN) [29]—were applied to the source and target training datasets. To preserve the true prevalence during evaluation, no sampling technique was applied to the validation and test datasets [30].

We used light gradient boosting machine (LightGBM) and extreme gradient boosting (XGBoost). Specifically, LightGBM offers fast learning, effective handling of imbalanced data through is_unbalance/weight, and supports efficient stopping and fine-tuning. XGBoost reliably controls overfitting and imbalanced data with L1/L2 normalization and scale_pos_weight, and is suitable for normalization fine-tuning with the xgb_model load. Furthermore, it has a high training speed to compare various combinations quickly [31,32]. Considering our settings, moderate class imbalance, domain shift, and tabular features, these 2 algorithms are natural choices.

For domain adaptation, fine-tuning was performed by first using a model trained on synthetic data as an initial model and then training additional trees using hospital data. In this process, the previously trained tree structure was kept fixed without modification, and the adaptation to the hospital data was done by adding new trees. In order to preserve the knowledge learned from the synthetic data as much as possible and prevent overfitting due to domain shift, a low learning rate and strong L1/L2 normalization were applied in the fine-tuning phase.

We configured four model variants as follows:

Each model was optimized for 2 objectives—area under the precision-recall curve (AUPRC) and F₁-scores. On the target validation set, hyperparameter tuning, early stopping, and threshold selection were performed by maximizing each metric.

In this study, 48 combinations were tested as follows: algorithm (LightGBM, XGBoost)×sampling technique (no-sampling, RUS, SMOTEENN)×model (baseline, domain adaptation, zero-shot, ensemble)×optimization objective (AUPRC, F₁).

The performances of the domain adaptation and baseline models were compared in the primary analysis. To objectively evaluate the domain adaptation effect, zero-shot and ensemble models were included as comparators. The main evaluation metrics are the AUPRC and the Brier score; for each experimental combination, we determined absolute values on the test set and changes relative to the baseline (ΔAUPRC and ΔBrier score). The Δ values presented in this study are not the arithmetic differences between point estimates. They represent the mean difference in performance between models, estimated by bootstrapping the test set 2000 times. This approach was adopted to provide more robust estimates of performance differences. The 95% CIs were derived from the 2.5th and 97.5th percentiles of the bootstrap distribution, and 2-sided P values were obtained as twice the minimum of the proportion of bootstrap resamples with a difference of zero or less and the proportion with a difference of zero or greater. The AUPRC was chosen to measure discriminative performance in the presence of class imbalance. In situations of class imbalance, the precision-recall approach further highlights the limitations of receiver operating characteristic–based evaluation [33]. The Brier score was included to assess whether the predicted probability matches the observed outcome. Models with similar discriminative power can have significant differences in probability accuracy and real-world interpretability; therefore, a comprehensive evaluation of clinical prediction models should consider both discriminative power and degree of calibration [34]. The AUROC, precision, recall, F₁, F₂, accuracy, Matthews correlation coefficient, and specificity were reported as secondary metrics.

Data preprocessing and extraction were performed using SAS (version 9.4; SAS Institute), while all subsequent analyses were conducted in Python (version 3.8.10; Python Software Foundation) within a controlled computational environment ensuring reproducibility. All analyses in this study were performed in a controlled computational environment with fixed software versions, an object-oriented pipeline structure, and systematic logging to ensure reproducible experiments for multiple model combinations [35].

Data management:

Python environment:

This retrospective study was approved by the Institutional Review Board of Chonnam National University Hwasun Hospital (IRB file CNUHH-2024‐113). The ethics committee of Chonnam National University Hwasun Hospital waived the requirement for informed consent because the study involved a retrospective review of anonymized medical data. All data used in the analysis were fully deidentified before access by the research team, and no personally identifiable information was available. Data were stored on secure, access-restricted systems to ensure privacy and confidentiality. No financial or other compensation was provided to participants, as this study involved secondary analysis of existing clinical data. The manuscript and all supplementary materials do not contain any images or information that could lead to the identification of individual participants.

In this study, synthetic data comprising 161,410 records—54,311 from the National Cancer Center, 52,767 from Samsung Medical Center, and 54,332 from Yonsei University Medical Center—were used for pretraining. During preprocessing, 44,106 records were excluded due to missing values in essential covariates. After applying selection and exclusion criteria and data preprocessing, the final dataset for survival model prediction included 30,683 patients. The external validation dataset from the Hwasun Jeonnam National University Hospital included 5490 surgical patient records, of which 2170 were used in the analysis. A flowchart of the participant selection process is illustrated in Figure 2.

Distribution comparison of the 2 datasets (Table 1) showed statistically significant differences for all variables (P<.05). For patient characteristics, the hospital data showed higher age (median 67, IQR 59-74 y vs median 63, IQR 55-70 y; P<.001) and a slight but statistically significant difference in sex (P=.04). The prevalence of adenocarcinoma (histology) was significantly higher in the hospital data than in the synthetic data (97.2% vs 55.6%, P<.001), and perineural invasion (29.4% vs 0.7%, P<.001) and vascular invasion (19.2% vs 9.4%, P<.001) were higher in hospital data. Stage distribution differed significantly (P<.001), with a higher proportion of stage III patients in the hospital cohort, which comprised predominantly high-risk patients. A clear domain shift was observed between synthetic and hospital data, necessitating the use of domain adaptation techniques to ensure model performance in real-world settings.

The synthetic data were split into source training (20,520 patients) and source validation (10,163 patients). Hospital data were split into target training (1386 patients), target validation (409 patients), and target tests (375 patients).

Synthetic data were obtained from 3 Korean institutions (161,410 records in total) and reduced to 30,683 patients after applying selection, exclusion, and preprocessing steps for model pretraining. External validation was performed using a dataset from the Hwasun Jeonnam National University Hospital, which included surgical patient data (5490 records), of which 2170 patients were included in the final analysis.

The experimental results are presented in Tables 2-5. Detailed performance results for all model combinations optimized using AUPRC and F₁-score are provided in Multimedia Appendix 1 and Multimedia Appendix 2, respectively. The AUPRC for models trained on synthetic data and evaluated on hospital data ranged from 0.2706 to 0.5391—domain adaptation (0.4124‐0.5391), zero-shot (0.2706‐0.4171), and ensemble (0.3919‐0.5161). In comparison, the hospital-trained baseline model showed an AUPRC of 0.3576‐0.4663. The low absolute performance of AUPRC is likely the result of a complex interplay of various factors, including moderate class imbalance, a limited feature set, and domain differences between synthetic and hospital data [36]. Therefore, we used changes relative to the baseline to evaluate the net improvement effect of external validation.

The domain adaptation model showed ΔAUPRC>0 in 8 out of 12 combinations, with significant improvement over the baseline model observed in 4 combinations. The largest improvement was observed in the XGBoost+RUS+F₁ optimization setting (ΔAUPRC=+0.1474, 95% CI +0.0523 to +0.2387; P<.001; AUPRC=0.5391). For calibration performance, 11 out of 12 combinations showed ΔBrier score<0, with 8 of these demonstrating statistically significant improvement. Thus, the predicted rates of the domain adaptation model matched the actual occurrence rates, even in the external hospital distribution.

The ensemble model, serving as a comparator group, showed ΔAUPRC>0 in 7 out of 12 combinations, with statistically significant improvement over the baseline model observed in 3 combinations. The largest improvement was observed in the XGBoost+RUS+AUPRC optimization (ΔAUPRC=+0.1258, 95% CI +0.0478 to +0.2034; P=.002; AUPRC=0.5060). Calibration performance showed ΔBrier score<0 in 9 out of 12 combinations, with 8 being statistically significant, demonstrating results similar to those of the domain adaptation model.

In contrast, the zero-shot model consistently showed poor discrimination performance. Only 2 of 12 combinations exhibited ΔAUPRC>0, and none of these improvements were statistically significant. The remaining 10 combinations showed no improvement, including 3 with statistically significant degradation. For the ΔBrier score, 6 out of 12 combinations showed performance decline, with 2 combinations being significant.

Although absolute performance was constrained by the limited number of shared features, certain model combinations significantly outperformed the baseline. The performance of the zero-shot model was limited and inconsistent, and there was no statistically significant improvement. In some combinations, performance actually dropped significantly. In contrast, models that directly applied domain adaptation (domain adaptation models) or models using domain adaptation (ensemble models) showed performance improvements across various combinations. Statistically significant performance improvements were observed in some settings. This confirms the clinical value of synthetic data when the distribution of target data is leveraged.

This study evaluated the feasibility and limitations of survival prediction models using specific synthetic data (NCDC synthetic dataset) in a privacy-constrained environment. After pretraining the model using large-scale synthetic data, we applied domain adaptation using single-center clinical data. The results showed that domain adaptation using real clinical data can partially compensate for the performance observed in the synthetic data-based zero-shot model used in this study. This suggests that synthetic data alone cannot replace real clinical data in the setting of this study, but when combined with limited real data, it can be used as a supplementary resource to improve model transfer performance.

Previous studies have mainly focused on domain adaptation between multicenter real-world clinical data, and relatively few studies have evaluated the transfer performance to real-world clinical data after using synthetic data as the main training resource. This study differs from previous studies in that we used generated and publicly available NCDC synthetic data for pretraining and applied a strategy to calibrate it with single-institution data. As reported in previous studies, synthetic data may not sufficiently reproduce the distribution of rare or clinically important variables depending on how they are generated. We show empirically in an external validation setting that this limitation can lead to distributional differences between synthetic data and real-world hospital data, which can manifest as poor zero-shot model performance. This demonstrates the potential of synthetic data-based pretraining learning strategies, as well as the limitations due to data quality and domain differences, with concrete examples.

This study had several limitations. First, there are limitations regarding the quality and distributional representativeness of the synthetic data used in this study. Significant discrepancies were observed in the prevalence of perineural invasion (PNI) between the NCDC synthetic data and hospital data. PNI is a well-established prognostic factor in predicting cancer survival. In addition, the synthetic dataset used in this study is a publicly available dataset, and a significant number of records were excluded from the analysis due to missing essential covariates. The NCDC synthetic data is RTSGAN-based and publicly available, and the distributional skew and missingness issues are consistent with the nature of RTSGAN-based synthetic data, where the goal is to preserve global statistical distribution. Previous studies have repeatedly reported underrepresentation of rare or clinically important variables as a limitation of synthetic data generation models [37]. It has also been reported that missingness in synthetic data may not be a newly generated problem, but rather a reflection of missingness mechanisms in the original clinical data or structural limitations of the generation model [38,39]. In this study, we did not artificially correct for these distributional skewnesses and missingnesses but rather performed the analysis while reflecting the realistic characteristics of the available synthetic data. Distributional differences in key prognostic factors and sample loss due to missingness may have partially contributed to the poorer predictive performance observed in the zero-shot transition setting. Additionally, differences in the distribution of major clinical variables were also observed between the synthetic dataset and the hospital dataset. In particular, there was a significant difference in the distribution of histological subtypes. While 97.2% of the hospital cohort was classified as adenocarcinoma, only 55.6% of the synthetic cohort was classified as adenocarcinoma, with the remainder consisting of nonadenocarcinoma subtypes, such as mucinous carcinoma and carcinoid tumors. Differences were also observed in postoperative chemotherapy patterns. These included regimen changes, combination or sequential therapies, and institution-specific treatment patterns not predefined, reflecting heterogeneous real-world treatment practices. These inconsistencies may have contributed to the observed performance differences and limitations in transferability between datasets. Second, the results of this study do not represent a direct performance comparison with an optimized clinical prediction model that uses all clinical variables included in the hospital data. The hospital data–based baseline model used in this study is a feature-constrained baseline model for comparison that is limited to 12 common variables to maintain commonality with synthetic data. Therefore, the performance improvement of domain adaptation models observed in this study should be interpreted as a validation of the methodological efficiency of synthetic data-based pretraining learning and model transfer strategies under the same input constraints, not as a claim of clinical superiority over local models that use all clinical information. Third, the external validation of this study was conducted using data from a single health care organization, which does not sufficiently validate its applicability to a variety of health care settings. Multicenter validation is needed to confirm the transferability and domain adaptation of synthetic data in settings with different institutional sizes, geographic characteristics, and patient populations. In addition, while the approximately 80:20 survival-to-death ratio observed for 7-year survival prediction is not an extreme level of class imbalance, the limited number of variables and domain differences between synthetic and hospital data, combined with the single-institution validation environment, may have formed an upper bound on model performance. Considering these characteristics, we chose AUPRC as the main evaluation metric to more appropriately reflect the discriminative performance of the model in an unbalanced clinical event prediction environment, while simultaneously using the Brier score to assess the degree of calibration of the prediction probability, which is important in clinical risk prediction. However, despite this choice of metrics, the absolute performance of the models remained in a limited range. In fact, the absolute performance of the model based on synthetic data ranged from an AUPRC of 0.3070 to 0.5165 and an F₁-score of 0.3391 to 0.5132, while the baseline model using only hospital data had a similarly wide performance range with an AUPRC of 0.3975 to 0.458 and an F₁-score of 0.4231 to 0.4751. This suggests that the performance limitations observed in this study may be the result of a combination of structural constraints of single-site validation, variable limitations, and domain differences, rather than simple class imbalance. Fourth, while age can be used as a continuous variable, we binned it in this study to increase distributional stability and mitigate nonlinear effects between synthetic and hospital data. This approach was intended to simplify interactions between variables in the tree-based model and reduce sensitivity to distributional differences between domains. However, this can be interpreted as a limitation of this study in that some continuous information may be lost.

Future work should extend analyses using public synthetic data, such as this study, to more systematically evaluate the applicability and limitations of synthetic data-based pretraining learning and model transfer strategies using real-world clinical data. In particular, the impact of distributional mismatches or missing characteristics of clinically important rare variables on model performance should be quantitatively analyzed, and data usage and transfer strategies that can mitigate these issues should be further explored. Furthermore, by introducing a quality assessment framework that considers how the missingness mechanisms of the original clinical data are reflected in the synthetic data, it is expected that the information loss that may occur when using synthetic data can be interpreted more clearly. Furthermore, external validation using multicenter real-world data will validate the generalizability of the synthetic data-based transfer learning approach presented in this study, and studies extending it to other diseases and prediction tasks will contribute to a broader examination of the conditional use of synthetic data.

In this study, we define NCDC synthetic data not as a stand-alone learning resource, but as an auxiliary resource for pretraining learning that can be used in a privacy-preserving environment and demonstrate that it can improve prediction performance when combined with real-world clinical data through domain adaptation. In the setting of this study, the results clearly highlight the limitations of synthetic data-based zero-shot models, while confirming that limited real-world data plays a key role in model transfer. These results suggest that synthetic data are not a substitute for independent clinical application but can be used as a complement to increase methodological efficiency when combined with real-world clinical data with limited sample size and variable composition. This supports that the purpose of this study was not to replace real-world clinical decision-making, but to evaluate the methodological feasibility of synthetic data-based transfer learning strategies. Recent studies have emphasized that reproducibility and computational efficiency are also important quality criteria for medical AI research involving large-scale synthetic data and iterative model evaluation [35]. From this perspective, the fixed software environment and systematic experimental pipeline design of this study can be interpreted as a foundation for ensuring the sustainability and reproducibility of future synthetic data-driven clinical AI research.