Development and validation of a predictive surgical risk model based on artificial intelligence algorithms to support personalized decision-making (SURGIRISK)

Gonçalo Carreiro da Silva

Key information

Authors:

Gonçalo Carreiro da Silva (Gonçalo Carreiro da Silva)

Supervisors:

Isaac Cano Franco; Mónica Duarte Correia de Oliveira (Mónica Duarte Correia de Oliveira)

Published in

November 20, 2025

Abstract

Postoperative complications remain a major challenge in healthcare, affecting up to 15% of patients undergoing major surgery and contributing significantly to morbidity, mortality, and healthcare costs. Accurate preoperative risk prediction is essential to improve outcomes and optimize perioperative care. This work corresponds to Phase I of the SURGIRISK project, which aimed to develop and retrospectively validate an artificial intelligence model to estimate individual surgical risk using real-world clinical and administrative data. The study was conducted at Hospital Clínic de Barcelona, using anonymized information from approximately 6000 major abdominal surgeries. The main outcome was the number of days patients remained alive and out of hospital within 90 days after surgery (DAOH90), integrating morbidity and mortality into a single recovery indicator. Data were preprocessed and harmonized, followed by model training and evaluation using standard performance metrics. Predictive accuracy improved across iterations, with the Area Under the Receiver Operating Characteristic rising from below 0.8 to 0.868 in the final configuration, and the model explained up to 43% of the variability in postoperative recovery. Evaluation with the APPRAISE-AI framework confirmed compliance with standards of data quality, transparency, and reproducibility. Overall, this first phase demonstrated the feasibility and methodological robustness of developing an AI-based surgical risk prediction model using routinely collected hospital data and established a solid foundation for future phases focused on external validation and clinical implementation.