We thank Therán-León et al. for their interest in our study and appreciate their constructive comments on aspects of its methodology. We agree that machine learning-based predictive models should be cautiously interpreted, especially when they are developed from retrospective, single-center, small cohorts. Our work was conceived as an exploratory, hypothesis-generating pilot study, whose primary aim was to evaluate the feasibility of integrating routinely obtained clinicopathologic variables into a preliminary model for stratifying the risk of distant recurrence in right-sided colon cancer.1

We share the concerns they expressed about the risk of overfitting and the size of the sample. However, it is worth noting that our analysis was not an empiric comparison of multiple machine-learning algorithms, but rather the development of a simplified random forest model based on 4 clinically interpretable variables readily available in routine practice. Because it was a retrospective pilot study, a previous sample size calculation was not performed, and the size of the sample was determined by the cases that met the eligibility criteria during the study period. We recognize that the contemporary criteria for determining sample size in predictive models, such as those proposed by Riley et al., are essential for planning confirmatory studies and reinforce the exploratory interpretation of our findings.2

To reduce the risk of statistical optimism, the final model was limited to four clinically relevant and easily reproducible predictors: age, sex, lymphovascular invasion, and the number of lymph nodes evaluated. Likewise, internal evaluation strategies were employed, including class balancing through upsampling, 5-fold cross-validation, and learning curve analysis.1 Nevertheless, we realize that such measures reduce, but do not eliminate, the risk of overfitting, particularly in the presence of a limited number of events. Therefore, the results should not be interpreted as definitive evidence of immediate clinical applicability, but rather as an initial approach that requires external confirmation.

We also consider adherence to the TRIPOD+AI statement a relevant observation. In preparing our manuscript, we followed the STROBE guidelines for observational studies, as stated in the Methods section of the original article.1 However, we agree that the TRIPOD+AI provides a more specific and comprehensive framework for describing predictive models developed through regression or machine-learning methods, and clearly, should be adopted in future reports and validation studies on the model.3 This includes a detailed description of data preprocessing, missing data management, a class-balancing strategy, analytic flow reproducibility, and when possible, code availability. In our study, patients with incomplete clinicopathologic data were excluded, but in confirmatory studies, it will be necessary to carefully characterize the missing data pattern and assess its potential repercussions on the performance of the model.

We also agree that discrimination, measured through the area under the ROC curve (AUC), sensitivity, specificity, and F1 score, does not exhaust the assessment of a model’s clinical utility. Evaluation of the calibration, including the calibration intercept and slope, as well as that of the net benefit, measured through the decision curve analysis (DCA), are necessary steps before considering any clinical implementation.4 Because ours was a pilot cohort with a low absolute number of events, and a total of 13 recurrences, those analyses would have resulted in inaccurate and potentially unstable estimates.1 Consequently, they should be central components of methodological evaluation in larger cohorts.

Finally, we fully agree that no predictive model should be transferred to clinical practice without external validation in geographically and temporally independent cohorts, in accordance with the methodological principles described by Ramspek et al.5 This limitation was explicitly acknowledged in our manuscript and we believe that the next logical step is multicenter evaluation, ideally with the participation of various Latin American centers.1 This would enable not only the overall performance of the model, but also its calibration, reproducibility, clinical utility, and performance across different clinical subgroups and populations.

In conclusion, we appreciate the comments by Therán-León et al., which complement and strengthen the interpretation of our study. We reiterate that our model should be regarded as a preliminary, hypothesis-generating tool. Its primary value lies in proposing an approach based on clinically accessible variables for individual risk stratification. The model must be refined and externally validated before its application can be generalizable.