AI Flags Infection Risk the Moment Surgery Ends
Postoperative infection is one of the most common complications after surgery.
It extends hospital stays, raises costs, and can lead to serious harm.
Most risk scores rely on facts known before the operation, so they cannot reflect what happens on the table.
Researchers at the University of Bern built a model that closes this gap by reading the patient’s vital signs from the operation itself.
The Problem With Pre-Surgery Risk Scores
Standard calculators use age, existing conditions, and procedure type.
Those inputs are fixed before the first incision, so they cannot change as the operation unfolds.
Operating rooms record vital signs continuously, yet that data is rarely used to estimate what comes next.
What the Model Reads
The team analyzed 10,719 operations performed at Bern University Hospital between 2014 and 2022.
Infection developed after surgery in 504 of those cases, which is 4.7 percent.
The model tracked blood pressure, heart rate, oxygen saturation, temperature, and end-tidal CO₂ across each procedure.
It did not use raw signals but converted them into trends, swings, and measures of how unstable each signal was.
Those intraoperative features carried the largest share of the model’s predictive weight, at 44.4 percent, ahead of patient factors at 42.1 percent.
How Well It Performed
The risk score is produced at skin closure, before any postoperative lab results exist.
A model using only pre-surgery information reached an AUROC of 0.81, while adding the intraoperative data raised it to 0.88.
Adding lab values from the second day after surgery lifted it only to 0.90, so most of the gain came from the operation itself.
A reduced version using just oxygen saturation, temperature, and heart rate scored about 0.87.
That result suggests the approach could work where invasive monitoring is not routine, although the study did not test those settings.
The authors describe this as moving the prediction window forward by up to 48 to 72 hours compared with lab-based assessment.
Higher oxygen saturation and lower heart rate were linked to lower risk, while higher temperature and higher baseline heart rate were linked to higher risk.
What Is Still Unproven
All data came from one hospital, and the study was retrospective.
The model needed complete monitoring records, so it was built on longer, more complex operations in older patients with higher infection rates.
It did not account for what the anesthesia team did during surgery, such as medications or fluids that shape the vital signs.
It also predicts infection as a single group, so it cannot say whether the risk is pneumonia, a wound infection, or sepsis.
The model slightly overestimated risk in every procedure group, which means alert thresholds would need tuning to avoid alarm fatigue.
The authors position a high score as a trigger for closer monitoring rather than a direct treatment decision, and they call for multicenter and prospective testing.
For a deeper look at the tools reshaping the operating room, read The 2026 Surgical Technology Intelligence Report.
Sources
Blatter et al., End-of-surgery prediction of postoperative infectious complications from intraoperative vital-sign dynamics, npj Digital Medicine. Link
