AI-guided decisions on a temporary ileostomy in rectal cancer surgery: what a 2026 randomized trial showed — and the caveat its authors put in plain sight

Rectal cancer surgery often comes with a smaller second decision that patients don't always hear about until they are deep into pre-op consent: whether to create, at the same time, a temporary diverting ileostomy. NHS guidance on bowel cancer treatment notes that some people having lower bowel cancer surgery will need a temporary stoma at the time of the operation, with closure planned later once the bowel has had time to heal [1]. The NHS ileostomy page explains what such a stoma actually is — a section of small bowel brought through an opening in the abdomen — and frames the temporary version as one created to route stool away from a section of bowel that needs time to heal, intended to be reversed in a separate later operation [2].

The reason a surgeon might create one is straightforward in principle. Anastomotic leakage — the dreaded complication where the new bowel join breaks down — is the event everyone is trying to avoid, and diverting stool can blunt its consequences if it happens. The reason this is harder than it sounds is that not every patient needs one, and the criteria for deciding who does have, for years, been clinical judgement more than standardised rules.

In May 2026, Nature Communications published a randomised controlled trial that tried to put a number on that judgement. The trial tested a machine-learning tool called the Risk-Guided Temporary Ileostomy Decision system — RTID for short — against ordinary surgeon discretion. Eight hundred and seventy-two adults with stage I–III rectal cancer undergoing anterior resection were randomised one-to-one to either surgeon-discretion or RTID-guided decisions, with 750 patients (368 control, 382 RTID) in the final analysis. The study was registered on ClinicalTrials.gov in August 2021 as NCT04999007 [3]. That last detail — prospective registration before the trial began — is the kind of housekeeping that quietly raises the credibility of what comes next.

The headline result is on the rate of stoma formation. In the RTID arm, 18.6% of patients ended up with a temporary diverting ileostomy, compared with 40.5% in the surgeon-discretion arm — less than half the rate, with a p-value below 0.001. The rate of what the authors term "unnecessary" stoma formation — that is, a temporary ileostomy created in a patient who, in the event, did not go on to leak — fell similarly: 17.7% in the RTID group versus 41.3% in the control group, again with a p-value below 0.001 [4].

Two things in those numbers are worth slowing down for. The first is that in both arms most temporary ileostomies turned out, in retrospect, to be ones the patient could have done without — the leak the stoma was guarding against never came. That is the gap the tool is aiming at. The second is that the tool reduced the total number of stomas mostly by cutting back on those unnecessary ones; the absolute number of patients who actually needed a diverting ileostomy was small in both arms, and a numerical increase in necessary use in the RTID arm (55.6% versus 10.0% among patients who later developed a leak) did not reach statistical significance, with a p-value of 0.057 [4].

The safety side is where the caveat lives. Anastomotic leak — the very complication the diverting stoma exists to manage — was the co-primary safety outcome. In this trial, the leak rate was almost identical between groups: 2.4% in the RTID arm and 2.7% in surgeon-discretion, with a p-value of 0.753 [5]. On the face of it, that is a reassuring number: the AI-guided arm did not buy its lower stoma rate by accepting more leaks. The authors then add the sentence that should travel with that result: the study was underpowered to formally test non-inferiority for anastomotic leak [5]. In plain language, the trial was sized to show whether the tool reduces stoma formation; it was not sized to formally prove that the leak rate stays the same. The leak numbers look comparable, but a small absolute number of events is a noisy instrument for detecting a rare safety signal, and the authors are explicit that this question is not yet closed.

It is worth being clear about what kind of evidence one trial is. A randomised controlled trial is the strongest study design available for testing whether an intervention causes a difference in outcomes — and this is a multi-centre, prospectively registered RCT in a high-impact general-science journal. That puts it on solid ground for what it claims. It is still, however, one trial, conducted in a defined set of centres, on a specific population having a specific operation. The authors describe the tool as one that supports — note the verb: not replaces — temporary-ileostomy decision-making, without a compromise in safety in this trial [6]. A single positive RCT is how a question gets opened up for wider testing in the real world, not how it gets closed.

For someone living the rectal-cancer decision — and there are few harder weeks — three things from this study are worth carrying with you, kept small on purpose. First, whether a temporary diverting ileostomy is part of your operation has never been a simple yes-or-no answer, and there has been growing recognition in the field that the default towards forming one may have led many people into a stoma and a closure surgery they did not, in retrospect, need. Second, this 2026 RCT is the first registered randomised evidence we are aware of that a structured risk-prediction tool can roughly halve that rate without an apparent rise in leaks in this trial. Third — and this is the part the authors put in plain sight — the same trial was not sized to formally prove safety equivalence, and a single trial is the start of an evidence trail, not the end. We're a curated information desk, not a surgical-planning service — for what this means for your case, the question belongs with the colorectal team who knows it.