AI Uses Operating Room Vital Signs to Predict Infection Risk Seconds After Surgery

Researchers at the University of Bern have developed an artificial intelligence model that predicts a patient's risk of postoperative infection immediately after surgery, using physiologic data collected during the operation rather than relying only on preoperative risk factors.

Why it matters

Postoperative infections remain one of the leading causes of surgical complications, contributing to longer hospital stays, higher healthcare costs, and increased mortality.

Traditional risk assessment is typically based on patient characteristics known before surgery, such as age, comorbidities, and procedure type, limiting clinicians' ability to detect evolving physiologic risk during the operation itself.

The study

The research team analyzed data from 10,719 surgical procedures performed at Inselspital Bern. Their AI model, called CARESCORE, combined routinely available clinical information with continuous intraoperative vital-sign data, including:

• Blood pressure

• Heart rate

• Oxygen saturation

• Temperature

• End-tidal CO₂

Rather than using raw monitoring signals, the system extracted interpretable trends and statistical features that reflected a patient's physiologic response throughout surgery.

Key findings

The model generated an individualized infection risk score within seconds after surgery ended.

Compared with models using only preoperative information, CARESCORE achieved substantially better predictive performance, reaching an AUROC of 0.88 (95% CI 0.85–0.91) for postoperative infectious complications.

The researchers also found that specific patterns in oxygen saturation and heart-rate dynamics were associated with a higher likelihood of later infection.

Why this is important

The work demonstrates that intraoperative physiologic data contain clinically meaningful information that is often overlooked.

Providing surgeons and perioperative teams with an immediate, explainable risk assessment at the end of surgery could enable earlier surveillance, targeted monitoring, and timely intervention before complications become clinically apparent.

Limitations

The model was developed and validated using data from a single institution (Inselspital Bern). External validation across multiple hospitals and diverse surgical settings is still required before routine clinical implementation.

---

Source: University of Bern

Study: npj Digital Medicine