An AI model trained on Greek hospital records predicted surgical complication risks with 94.3% accuracy, according to a new study that points to a possible path for earlier warnings in hospital surgery units.
The study was led by Constantinos Koutsojannis of the University of Patras in Greece and published in “Applied Sciences.” Researchers analyzed 19,965 anonymized patient records from a general surgery department at a Greek tertiary university hospital. The data covered a 10-year period, from 2013 to 2023.
The research focused on patients who showed signs of elevated risk after surgery. These included health care-associated infections, medication errors, equipment failures, and other adverse events. The dataset included 2,700 such cases, or 13.5% of all records.
AI model from Greek hospital data spots surgical risk
Researchers did not treat the system as a direct detector of proven medical mistakes. Instead, they built it to identify patterns linked to higher adverse-event risk. That distinction is important.
The hospital did not have a full set of formally reviewed medical error labels. So the team used proxy markers, including longer hospital stays and higher treatment costs. They checked those markers against ICD-10 complication codes to improve reliability.
The best-performing system was a Random Forest model. It reached 94.3% accuracy and an AUC-ROC score of 0.98, a measure of how well a model separates high-risk cases from lower-risk ones.
An ensemble voting model followed with 93.6% accuracy. Other tested systems included a J48 decision tree, a multilayer perceptron, and Naïve Bayes.
‘Hospital stay length’ emerges as a key warning sign
The AI model used Greek hospital data to find which factors mattered most. Hospitalization duration ranked as the strongest predictor. Initial diagnosis also played a major role.
Researchers said these signals could help clinicians spot patients who may need closer attention, such as those whose recovery extends beyond expected timelines after surgery.
The study suggests the tool could support electronic health record alerts. For example, a patient still in the hospital several days after surgery could be flagged for review.
That may help teams act earlier, especially when infections or other complications are developing. The system is meant to support doctors and nurses, not replace their judgment.
Greek surgical data fills a European research gap
The findings also address a gap in European and Greek surgical data. Many existing surgical risk tools rely on U.S. datasets. The authors compared their model with benchmarks such as ACS NSQIP, which they cited at about 90% accuracy. They said locally trained tools may better reflect regional hospital workflows and resource limits.
Still, the study has limits. It looked back at past records from one hospital. That can introduce bias. The use of hospitalization length as both a risk marker and a predictor also raises concerns about overlap. When researchers removed that variable, performance fell but remained clinically useful.
The authors said future work should test the model in more hospitals. They also called for explainable AI tools, stronger fairness checks, and human oversight. Before any real-world rollout, the model would need external validation and careful integration into clinical workflows.
See all the latest news from Greece and the world at Greekreporter.com. Contact our newsroom to report an update or send your story, photos and videos. Follow GR on Google News and subscribe here to our daily email!
