Adoption Ahead of the Evidence: AI Arrives in Metabolic and Bariatric Surgery

The most consequential sentence in the newest guidance on artificial intelligence in weight-loss surgery sits in the background section, before any results. Adoption of AI across the surgical pathway, the paper states, “is outpacing specialty-specific guidance on safe, effective, and equitable implementation.”

That paper is a surgeon-led narrative review from the Emerging Technology and Procedures Committee of the American Society for Metabolic and Bariatric Surgery, published in Surgery for Obesity and Related Diseases on 29 September, ahead of print. It is not a trial and not a pooled meta-analysis; it is a synthesis of published literature organised by perioperative domain, with recommendations developed by committee consensus. Read as such, it is less a demonstration of what algorithms can do than an argument about who gets to decide that they work, and on what evidence.

What the committee credits

The review is generous where the literature allows. Before surgery, it notes, machine-learning models outperform conventional risk indices at predicting complications and readmission, and can forecast long-term weight trajectories. During the operation, computer-vision systems recognise procedural steps in sleeve gastrectomy and Roux-en-Y gastric bypass with accuracy comparable to expert human annotators, while AI-enabled video and virtual-reality platforms support objective skills assessment. After discharge, wearable-based remote monitoring and predictive models may pick up early deterioration, weight regain or the return of a comorbidity. Administratively, ambient scribes, prior-authorisation automation and operating-room scheduling are already reducing clerical load.

That inventory is followed immediately by the qualifier the committee keeps returning to: most of this evidence remains retrospective and single-centre. The three domains are not equivalent in maturity — a model that matches experts on a recorded case library and a tool that changes what happens to a patient on a Tuesday morning are different claims, and the review treats them as such.

A model looking for an external test

What a single-centre model looks like in practice is visible in Obesity Surgery from 21 September. Investigators at a university hospital in Guangzhou built five machine-learning models on 621 patients who had a laparoscopic sleeve gastrectomy between 2021 and 2025 and completed one year of follow-up, using only routinely collected preoperative variables. The Boruta selection algorithm narrowed the field to seven predictors — body mass index, fasting plasma glucose, waist circumference, age, LDL cholesterol, white blood cell count and estradiol. A random forest performed best, with an area under the curve of 0.804 (95% CI 0.728–0.881) and 79.6% accuracy on an internal test set, and the authors packaged it as a web tool and inspected it with SHAP values to show which variables drove the predictions.

It is a competent piece of work with a clear ceiling, and the authors name it: the study is retrospective, from one centre, split seven-to-three into training and test sets but never validated externally. The outcome is also narrower than the phrase “weight-loss surgery success” suggests — an optimal clinical response at one year, defined on weight, not on diabetes control, complications or durability. “Further external validation is warranted,” the paper concludes, which in this literature is the standard polite form of “not yet.”

Where the gap actually is

The broader field reads the same way. A structured narrative review of AI across obesity care, published in Frontiers in Endocrinology on 3 September and reporting a search of five databases through 1 June 2026, reached a blunt destination: within the eligible evidence it examined, no externally validated, prospectively implemented AI prescribing system had demonstrated improved patient outcomes. The strongest weight and metabolic results came from multicomponent digital, automated or hybrid-care programmes — studies in which, the authors stress, no AI component was independently evaluated, and in which HbA1c was often the clinical context and endpoint. Those findings, they write, support digital care delivery but “should not be interpreted as evidence of an AI-specific therapeutic effect in obesity.”

In metabolic and bariatric surgery specifically, the same review lists the plausible uses — risk prediction, operative workflow analysis, readmission stratification, weight-trajectory modelling, digital follow-up — and then notes that most studies remain retrospective or internally validated, and rarely assess calibration, actionable thresholds or prospective deployment. Calibration is the unglamorous part: a model can rank patients correctly and still misstate the absolute risk it hands to a clinician.

This site has covered one of the better recent examples, a seven-variable random forest trained on 511 patients and tested on 300 more from another institution to predict anaemia a year after surgery, which outperformed ten clinicians on a retrospective comparison and was still, explicitly, not a care pathway. Good discrimination on a held-out dataset is the beginning of a clinical tool, not the end of one.

Four risks, eight recommendations

The committee is equally specific about failure modes. Large language models can hallucinate. Automation bias and deskilling can erode the human judgement the tools are meant to support. Algorithmic bias can encode inequity already present in the training data. Data privacy obligations follow every wearable and every ambient scribe into the room.

Against those, it sets eight operational recommendations covering clinical role, validation specific to metabolic and bariatric surgery, equity auditing, documentation, governance, technical safety, workflow integration and learning health systems. The conclusions are conditional almost to the point of being a checklist: AI can enhance the safety, efficiency and scalability of this surgery when it is clinician-led, rigorously validated, equitable and transparent — and prospective, multicentre evaluation is needed before broad adoption.

Nothing in the recommendation set tells a surgical team what to buy. That is the point. The document regulates the claim rather than the software, which is what a specialty can do while the evidence is still thin.

What would settle it

The design the committee keeps gesturing at is not novel. It is prospective, multicentre, and specified in advance: a defined cohort followed forward, endpoints fixed before the data arrive, an algorithm locked before it is tested. The LABS study, funded by NIDDK, is the reference example of that architecture in this field — a multicentre prospective cohort built to measure outcomes across a broad, non-selected surgical population — and the LABS archive on this site documents the design. This site is an independent archive and observatory; it is not affiliated with NIDDK, and it reports on that research rather than speaking for it.

The prediction tools now arriving at the clinic door, from weight-loss forecasting to the discharge-readiness models that decided the shape of our recent review of same-day discharge, are the same species the committee wants tested prospectively. A validated score is not a changed decision, and the last decade of observational enthusiasm in this field — where almost every association that looked decisive at publication looked ordinary five years later — is the reason the caution is written the way it is.