The team’s argument grew out of a real and humbling experience. They analysed the Spanish EVEREG registry, a national cohort of inguinoscrotal hernia repairs that included almost 750 operations. On paper, that sounds like a dream dataset for machine learning. In practice, postoperative complications occurred in only 52 patients. The researchers explored a formidable arsenal of algorithms, including random forests, support vector machines, neural networks, k-nearest neighbours and elastic-net regression. To their surprise, none demonstrated clinically meaningful superiority over the others. That unexpected flatness across radically different algorithmic architectures prompted a question that many surgical AI researchers rarely stop to ask: are we sometimes demanding more from machine learning than our data can provide?

The core insight is deceptively simple but technically profound. Prediction models do not learn from patient numbers alone; they learn from informative outcome events. Machine learning algorithms extract patterns from examples, and the examples that matter for surgical prediction are the patients who actually experienced the outcome of interest, whether that is a mesh infection, a hernia recurrence, a major surgical site occurrence or severe postoperative morbidity. In abdominal wall surgery, and especially in conditions such as inguinal hernia repair, these outcomes may be uncommon. A registry containing hundreds of patients can therefore contain remarkably little information for predictive modelling, because the effective sample size depends not on the total cohort but on the number of events, the candidate predictors and the intended model complexity.

This distinction has real consequences for how registries should be used. A national registry may be excellent for epidemiology or quality assurance while remaining insufficiently informative for prediction modelling. The EVEREG registry represented an excellent national cohort, but only a small proportion of patients experienced postoperative complications, so machine learning had relatively few patients with complications from which to learn stable and reproducible patterns. This does not diminish the value of surgical registries, the authors stress, but it reminds the field that a good registry is not necessarily a good dataset for prediction. The same registry can be a superb epidemiological instrument and a weak predictive one.

Modern algorithms can detect complex relationships, but they cannot generate biological information that is absent from the data. When clinically relevant outcomes are rare, increasing computational complexity cannot replace genuinely observed events. Importantly, this limitation is not exclusive to machine learning: if key clinical variables are missing, or if too few outcome events are available, any predictive analysis will be limited, whether it uses machine learning or conventional statistical methods. Consequently, different machine learning algorithms may show similar performance simply because the dataset does not contain enough information to improve prediction further. The algorithms are all hitting the same ceiling, and the ceiling is the data itself.

The authors also draw a technically important line between class imbalance and lack of information. When complications are rare, datasets are imbalanced, and researchers often reach for techniques such as oversampling, undersampling or SMOTE, the Synthetic Minority Oversampling Technique, which synthesises new minority-class examples to facilitate model optimisation. But these balancing techniques cannot replace new independent patients with real outcome events. Synthetic examples are mathematical constructions, not clinical observations, and recent methodological work has shown that resampling strategies can distort the calibration and discrimination of clinical risk prediction models. Future advances in abdominal wall surgery prediction will therefore depend more on richer clinical variables, improved follow-up and collaborative registries than on increasingly sophisticated balancing techniques.

Scarcity of outcome events also shapes validation strategy. Conventional train-test splits, which set aside a chunk of data for final testing, may waste valuable information when events are few. Bootstrap procedures and repeated cross-validation often provide more efficient internal validation in these settings, squeezing more learning out of every available event. Model evaluation, meanwhile, should not be judged only by accuracy or the area under the receiver operating characteristic curve. In imbalanced datasets, calibration, sensitivity, positive predictive value and precision-recall curves frequently provide more meaningful assessments of clinical usefulness. A model that looks impressive on AUROC can be badly miscalibrated in the risk ranges where clinicians actually need to make decisions. Ultimately, the authors argue, prediction models should be judged by their capacity to improve decision-making rather than by statistical novelty alone.

There is a broader lesson here for the entire field of clinical prediction. Current evidence indicates that machine learning does not consistently outperform logistic regression across clinical prediction studies; a well-known systematic review found no performance benefit of machine learning over logistic regression for clinical prediction models. Rather than competing approaches, regression and machine learning should be viewed as complementary tools. Logistic regression remains valuable because it is simple and easy to interpret, whereas machine learning models are worthwhile when their additional complexity leads to a clearer improvement in prediction. The authors also caution against confusing predictive modelling with causal inference: variables identified as influential by machine learning are not necessarily causal determinants of postoperative complications, so clinical interpretation remains essential.

Perhaps the principal message of the article is that inconclusive machine learning analyses deserve greater recognition. Publication bias favours successful algorithms, whereas studies reporting modest performance are often overlooked. Yet these investigations frequently provide equally valuable methodological information, because they define the current limits of prediction. The authors’ conclusion from the EVEREG experience is not that machine learning failed; rather, the registry contained insufficient information to support reliable prediction of uncommon complications. This should not discourage the use of machine learning, but it should help researchers design better studies and recognise when prediction is realistically possible. A model with limited predictive performance should not necessarily be considered a failed study; it may simply show that the available data are not yet sufficient for reliable prediction.

Looking ahead, the authors call for a next generation of abdominal wall surgery registries that focus not only on increasing patient numbers but also on collecting more complete and clinically relevant information, using consistent definitions, improving follow-up and increasing collaboration between centres. These developments are likely to contribute more to future prediction performance than replacing one algorithm with another. Reporting and quality frameworks such as TRIPOD+AI and PROBAST+AI provide an excellent methodological foundation for this evolution, setting standards for how prediction models that use regression or artificial intelligence methods should be reported and assessed. Machine learning, in their view, does not simply reveal what algorithms can learn; it also reveals what our registries are currently unable to teach, and it helps us understand the limitations of our own data. Neutral machine learning results should therefore not be regarded as failed studies. They are mirrors held up to the data, showing surgeons exactly where the boundary lies between what prediction can achieve today and what better registries might make possible tomorrow.

Subject of Research: Machine learning prediction of rare postoperative outcomes in abdominal wall surgery registries

Article Title: Large registries, rare outcomes and neutral/inconclusive machine learning results in abdominal wall surgery

Article References: Verdaguer-Tremolosa, M., Mojal, S., Rodrigues-Gonçalves, V., Martínez-López, M. P., & López-Cano, M. (2026). Large registries, rare outcomes and neutral/inconclusive machine learning results in abdominal wall surgery. Journal of Abdominal Wall Surgery, 5, Article 17526. https://doi.org/10.3389/jaws.2026.17526

Image Credits: AI Generated

DOI: 10.3389/jaws.2026.17526

Keywords: machine learning, artificial intelligence, abdominal wall surgery, hernia repair, surgical registries, rare outcomes, prediction models, class imbalance, SMOTE, logistic regression, TRIPOD+AI, calibration