Response to ‘Further Detail Concerning the Deep Learning Model for Mortality After Total Gastrectomy’

We deeply appreciate the thoughtful comment from Dr., Goto and colleagues regarding our article entitled “Further Detail Concerning the Deep Learning Model for Mortality After Total Gastrectomy.” We are very grateful for your interest in our work on deep learning (DL) mortality prediction after gastrectomy. In this study, model accuracy was evaluated using receiver operating characteristic (ROC) curves and their area under the curve (AUC). However, for highly imbalanced data, precision-recall (PR) curves are certainly more informative. We initially calculated PR curves during exploratory analyses but did not include them in the final manuscript because the interpretation of PR-based metrics is strongly influenced by outcome prevalence. Since prevalence differed across cohorts, direct comparison could be misleading and reduce interpretability for the intended clinical readership. Given that the primary aim was to compare overall discriminative performance across multiple settings, we therefore focused on AUC-ROC, a well-established and prevalence-independent metric. The calibration plot is shown in figure 4 of the original paper [1](Figure 1). The horizontal axis represents the predicted probability of the event, and the vertical axis indicates the observed event rate. As noted, the observed mortality remains close to the overall event rate even at high predicted probabilities, particularly on the right-hand side of the figure. This likely reflects the rarity of the outcome and the small number of observations in the high-risk range, leading to substantial uncertainty. Accordingly, predicted probabilities in this range may not accurately represent absolute risk. Nevertheless, the overall monotonic trend indicates that the model remains useful for risk stratification rather than precise absolute risk estimation. This study aimed to evaluate whether predictive performance could be improved by applying DL techniques. The National Clinical Database (NCD), as the largest clinical registry in Japan, represents a uniquely rich resource whose potential may not be fully captured by traditional analytical approaches. Accordingly, we explored the applicability of DL to this dataset, but did not observe performance exceeding previous reports. Our findings suggest that, for tabular clinical data, methodological modifications alone may yield limited gains, and that substantial improvement may require more fundamental advances beyond conventional approaches. As discussed above, applying DL to tabular data does not necessarily leverage its full strengths, and improvements in predictive accuracy are not guaranteed. As noted, boosting-based methods have demonstrated strong performance in this domain, and in some cases, random forests or even simpler models such as support vector machines may outperform DL approaches [2]. Model performance in this context is fundamentally influenced by the characteristics of the dataset, including variable structure and signal-to-noise ratio. To better leverage the potential of deep learning, incorporating data modalities beyond tabular data may be necessary. Therefore, although benchmarking against other machine learning methods is important, comprehensive comparisons were beyond the scope of this study and are left for future work. We thank you again for your constructive comments on our study. We hope that our responses have adequately addressed your concerns. Ryosuke Fukuyo: conceptualization, investigation, visualization, writing – original draft. Hiroyuki Yamamoto: writing – review and editing, project administration. Masanori Tokunaga: writing – review and editing, investigation, project administration. Yusuke Kinugasa: conceptualization, writing – review and editing, supervision, project administration. Yoshihiro Kakeji: conceptualization, writing – review and editing, supervision, project administration. Ken Shirabe: conceptualization, supervision, writing – review and editing, project administration. The authors have nothing to report. The authors have nothing to report. Hiroyuki Yamamoto is affiliated with the Department of Healthcare Quality Assessment at the University of Tokyo. This department is a social collaboration department supported by grants from the National Clinical Database, Intuitive Surgical Sarl, Johnson & Johnson K.K., and Nipro Co. Yusuke Kinugasa is an associate editor member of Annals of Gastroenterological Surgery. Yoshihiro Kakeji and Ken Shirabe are editorial board members of Annals of Gastroenterological Surgery. The other authors declare no conflicts of interest. Further Detail Concerning the Deep Learning Model for Mortality After Total Gastrectomy https://doi.org/10.1111/ags3.70155.