Machine learning for operative time prediction in robotic ventral hernia repair patients: Perspective on predictive model development and validation

Abstract

Machine learning (ML) is increasingly applied to predict operative time in surgery, yet its role in robotic ventral hernia repair (VHR) remains unexplored. Brucchi and colleagues recently developed ML models, Random Forest, Gradient Boosting, and Ridge Regression, to predict operative time in 208 patients undergoing elective robotic VHR, achieving modest performance with Random Forest (R² = 0.22, MAE = 38 min). While this proof‑of‑concept study represents a novel application, several methodological considerations warrant attention before clinical translation. These include the absence of comparison with traditional statistical methods, the omission of surgeon‑level temporal data and learning curve dynamics, the lack of external validation, and the gap between SHAP‑based interpretability and actionable clinical utility. This Matters Arising offers a perspective on these considerations and proposes directions for future refinement, including incorporation of surgeon‑specific metrics, benchmarking against conventional regression models, external validation, and integration into clinical workflows, to enhance the predictive accuracy and clinical applicability of ML‑based operative time prediction in abdominal wall surgery.

To the editor

We read with interest the article by Brucchi and colleagues on machine learning for operative time prediction in robotic ventral hernia repair [1]. While the study offers a valuable proof‑of‑concept, several methodological considerations warrant attention. We offer a perspective on these considerations and propose directions for refinement.

Missing comparison with traditional statistical methods

The study by Brucchi and colleagues evaluates three machine learning models for predicting operative time in robotic ventral hernia repair, yet a notable methodological gap is the absence of a direct comparison with traditional statistical approaches [1]. The authors compared Random Forest, Gradient Boosting, and Ridge Regression against each other, but did not include a multivariable linear or log‑linear regression model as a benchmark, an omission particularly significant given the modest number of features (nine variables) and limited sample size (n = 208). A narrative review of perioperative ML applications concluded that performance improvements achieved by ML models are often promoted without adequate comparison to traditional statistical techniques, preventing informed assessment of predictive benefit [2]. A systematic review comparing ML models with conventional risk scores in adult cardiac surgery found that while ML models showed improved or comparable performance, the magnitude of improvement varied considerably, with AUC differences ranging from 0.006 to 0.42 [3]. Common methodological limitations included retrospective designs, single‑center cohorts, and limited external validation, all applicable to the present study. The Random Forest model achieved an R² of only 0.22, explaining just 22% of the variance in operative time [1]. This does not rule out the possibility that a well‑specified log‑linear or polynomial regression model, incorporating non‑linear terms for BMI and hernia size, could yield superior or equivalent performance with greater interpretability [2]. We suggest that future studies should routinely include traditional statistical models as benchmarks to determine whether the added complexity of ML confers a meaningful advantage over conventional methods.

Surgeon experience and learning curve

The study by Brucchi and colleagues spans a 4.5‑year period (October 2020 to April 2025), during which two robotic platforms, Da Vinci Xi and Hugo™ RAS, were used [1]. Yet the model treats surgeon identity and experience as absent features, ignoring the substantial temporal improvements in performance that occur as surgeons progress along the learning curve. This omission is particularly consequential given that operative efficiency improves earlier than complication‑adjusted outcomes, with proficiency thresholds for robotic VHR reported at 26–49 cases depending on technique, and prior experience substantially shortening the learning curve [4]. The temporal dynamics of skill acquisition are well documented across robotic procedures. In a single‑surgeon series of robotic‑assisted transabdominal retromuscular repair using the Hugo™ RAS platform, console time decreased significantly from 177.5 min in the first tertile to 138.5 min in the third (p = 0.005), with dissection time improving from 69.8 to 45.7 min (p = 0.003) [5]. CUSUM analysis identified proficiency thresholds of 34 and 37 operations for console and dissection times respectively. The fact that the present study used both Da Vinci and Hugo platforms, each with distinct docking procedures, instrument articulation, and learning curves, without accounting for these differences introduces substantial unmeasured variability. Several studies have demonstrated that surgeon case volume is one of the strongest predictors of operative time, with high‑volume surgeons achieving significantly shorter durations even in complex robotic cases [6, 7]. Risk‑adjusted CUSUM analyses further show that operative time stabilises after approximately 30–32 cases when patient‑related factors are accounted for [7]. The absence of surgeon‑level temporal data in the present study likely contributes substantially to the unexplained variance (R² = 0.22) observed in the model. We suggest that future studies should incorporate longitudinal surgeon‑specific metrics, such as cumulative case volume per platform, learning curve phase indicators, and individual step durations, to capture the dynamic nature of surgical performance and improve predictive accuracy.

Beyond internal cross-validation

The study by Brucchi and colleagues employs 5‑fold cross‑validation for model development, yet this internal validation approach cannot substitute for external validation on independent datasets, particularly given the single‑center retrospective design and limited sample size of 208 patients [1]. The authors characterise their work as a proof‑of‑concept and acknowledge that current accuracy is insufficient for clinical implementation. However, this raises a critical question: what standards should guide the development and reporting of surgical prediction models to facilitate meaningful translation? A systematic appraisal of surgical prediction models using the TRIPOD guideline found that median compliance across 120 published studies was only 53.8%, with no study exceeding 85% adherence [8]. Models using conventional statistical methods demonstrated higher compliance (56%) than machine learning‑based models (50%, p = 0.002), suggesting that the added complexity of ML may paradoxically reduce reporting quality. Items such as handling of missing data, reporting performance measures with confidence intervals, and explaining model use were particularly poorly reported [8, 9]. External validation is essential for assessing model generalisability. Suzuki and colleagues developed a random forest‑based operative time prediction system for robot‑assisted radical prostatectomy using data from two institutions, achieving a normalised RMSE of 0.107 in internal validation and 0.148 in external validation, demonstrating that performance can decline substantially when applied to independent cohorts [10]. The present study lacks external validation entirely, and as a single‑center cohort, its findings may not generalise to other institutions with different patient populations, surgical approaches, or practice patterns [1]. We suggest that future studies should adhere to established reporting guidelines such as TRIPOD, prioritise external validation on independent cohorts, and transparently report model performance with confidence intervals to facilitate meaningful interpretation and clinical translation.

The interpretability‑accuracy trade‑off

The study by Brucchi and colleagues employs SHAP analysis to enhance interpretability, identifying mesh position (45.0%), hernia size (30.7%), and BMI (14.2%) as dominant predictors of operative time [1]. However, a gap remains between statistical interpretability and actionable clinical utility, particularly given the model’s modest R² of 0.22. With an MAE of 38 min relative to a median operative time of 90 min (relative error ~ 42%), the model offers limited utility for precise scheduling. The minimal contribution of clinically intuitive features such as obesity status (0.4%) further illustrates that SHAP interpretation requires careful consideration of feature encoding. A survey of barriers to ML implementation in surgery identified limited clinician trust, workflow integration challenges, and unclear economic value as key obstacles [11]. Without prospective validation and workflow integration, SHAP analysis alone cannot bridge the gap between prediction and clinical decision‑making. Future studies should directly compare ML predictions with surgeons’ clinical estimates and incorporate decision‑curve analysis.

Conclusion

The study by Brucchi and colleagues offers a valuable proof‑of‑concept for ML‑based operative time prediction in robotic ventral hernia repair, yet several methodological considerations warrant attention. The absence of comparison with traditional statistical methods, omission of surgeon‑level temporal data, reliance on internal cross‑validation without external validation, and gap between SHAP‑based interpretability and clinical utility collectively limit the model’s readiness for implementation. Future studies should incorporate surgeon‑specific metrics, benchmark against traditional approaches, prioritise external validation, and embed predictions within clinical workflows.

Author contributions

J.R. and A.M.M. conceptualized the commentary and developed the main thematic structure. D.K. and A.B. contributed to the literature review and critical analysis of the original study. S.N. and N.J. participated in data interpretation and manuscript revision. D.N. and F.A. contributed to the methodological analysis and provided critical feedback on the machine learning framework. All authors reviewed and approved the final manuscript for submission.

Funding

None.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethical approval

Not applicable.

Consent for publication

All authors approved the final manuscript and the submission to this journal.

Informed and patient consent

Not applicable.

Competing interests

The authors declare no competing interests.