From static thresholds to dynamic trends: Reassessing serum calcium in anastomotic leakage prediction

Abstract

Kang et al published a study recently in the World Journal of Gastroenterology introducing an interpretable machine learning model to predict anastomotic leakage after rectal cancer surgery, highlighting postoperative serum calcium as a key predictive feature. While this represents a significant advancement, we argue that reliance on a static calcium threshold may limit clinical applicability. We advocate for a dynamic, trajectory-based assessment of serum calcium levels across perioperative time points, using modeling approaches such as time-series regression or mixed-effects models. Furthermore, the model’s robustness could be improved by incorporating systemic inflammation and nutritional indices such as C-reactive protein, procalcitonin, the neutrophil-to-lymphocyte ratio, and the systemic immune-inflammation index, supported by recent prospective studies. Finally, generalizability remains a concern, warranting broader validation and clearer clinical deployment strategies. By addressing these aspects, the model’s clinical translation and decision-making impact could be substantially enhanced.

Key Words: Anastomotic leakage; Serum calcium; Machine learning; Dynamic monitoring; Systemic inflammation; Clinical implementation

Core Tip: This letter evaluates Kang et al’s machine learning model predicting anastomotic leakage post-rectal surgery. It emphasizes the need for dynamic serum calcium modeling, integration of inflammatory and nutritional biomarkers (e.g., C-reactive protein, procalcitonin, neutrophil-to-lymphocyte ratio, systemic immune-inflammation index), and broader validation strategies. Specific modeling suggestions are provided to enhance the model’s translational potential and alignment with real-world clinical settings.

TO THE EDITOR

We read with great interest the multi-center study by Kang et al[1], which developed a novel extreme gradient boosting-based predictive model using serum calcium to predict anastomotic leakage after rectal cancer resection. We commend the authors for their contribution in constructing a multi-center interpretable model with practical clinical implications. This study impressively integrates perioperative clinical features with machine learning, to address one of the most challenging complications after rectal resection. Particularly, the incorporation of Shapley Additive Explanations values offers clarity to the decision-making process of the model, making it more transparent and adaptable in clinical settings[2]. Among the 10 variables included, the identification of postoperative serum calcium ion concentration as a key predictive feature is especially noteworthy. This adds to the growing evidence of the biological importance of calcium in anastomotic healing, possibly through its roles in coagulation, tissue regeneration, and antimicrobial activity. However, we would like to offer several constructive perspectives to enhance the model’s robustness and clinical applicability.

Prioritizing dynamic serum calcium trends over static thresholds

The identification of a postoperative serum calcium cutoff (< 2.06 mmol/L) as a significant risk factor is indeed a key finding. Nonetheless, perioperative calcium levels are affected by numerous confounders, such as fluid resuscitation, surgical stress, and hemodilution, which may lead to transient fluctuations. Therefore, relying solely on a single time-point value may not fully capture the patient’s true physiological risk state.

We suggest incorporating time-series calcium monitoring to track calcium trajectory trends over multiple perioperative time points. However, as postoperative calcium levels may be influenced by calcium supplementation protocols, simple delta calculations (e.g., ΔCalcium = post-operation - pre-operation), which overlook temporal dynamics and inter-individual heterogeneity, may inadequately reflect intrinsic physiological changes. Instead, modeling trajectories using analytical approaches such as time-series regression or mixed-effects models across at least three perioperative timepoints (e.g., preoperative, post-operative day 1, and postoperative day 3) may help mitigate treatment-related confounding and provide more physiologically meaningful insights.

Expanding feature scope: Inflammation, nutrition, and immune indices

Several biomarkers like C-reactive protein, procalcitonin, and neutrophil-to-lymphocyte ratio have shown strong associations with anastomotic leakage in recent studies[3-5]. These indicators capture systemic responses that may not be reflected by calcium alone. In particular, systemic immune-inflammation index (SII), calculated from platelets, neutrophils, and lymphocytes, was already partially acknowledged by the authors through the preoperative platelet count. Recent prospective evidence further supports SII’s value in gastrointestinal oncology. For instance, a study by Ding et al[6] demonstrated that a combined SII prognostic nutritional index score could predict chemotherapy response and progression-free survival in patients with locally advanced gastric cancer undergoing neoadjuvant treatment. This highlights the broader relevance of immune-nutritional indices like SII in perioperative gastrointestinal risk modeling. A complete inclusion of SII might further enhance model granularity[7].

Model generalizability and clinical translation

The authors’ initiative to externalize validation is laudable; however, performance degradation (area under the curve drop to 0.703) indicates that overfitting or institutional heterogeneity may still be at play. Future prospective validation across more diverse cohorts and inclusion of omitted clinical variables (e.g., tumor distance from dentate line) could help generalize the model[8]. Moreover, the developed interface is a promising translational step. We encourage integrating this tool into clinical workflows, ideally in real-time with electronic medical record linkage and alerts.

CONCLUSION

Kang et al’s study represents a forward-thinking approach in surgical artificial intelligence[1]. By optimizing the interpretation and modeling of serum calcium as a dynamic rather than static biomarker, and by integrating systemic inflammatory and nutritional metrics, this model can reach greater clinical precision and applicability. We look forward to further iterations of this work.