Bridging the Artificial Intelligence Translation Gap in Perioperative Pulmonary and Respiratory Complications: A...

Perioperative pulmonary and respiratory complications (PPRCs) result in increased morbidity, mortality, and healthcare costs. Recent research has focused on employing artificial intelligence (AI) and machine learning to predict and analyze PPRCs using expansive perioperative datasets. Despite advances in feature engineering, deep learning architectures, and explainable AI, models are rarely implemented in routine clinical practice. In this narrative review, we synthesize evidence from 59 unique publications, corresponding to 63 phenotype-specific study-listings across six PPRC phenotypes (combined postoperative pulmonary complications, pneumonia/pulmonary infection, hypoxemia, respiratory failure, pulmonary embolism/venous thromboembolism, and pulmonary edema) and introduce an integrated analytical and translational framework to bridge translation gap. The proposed paradigm comprises a tiered framework for evaluating outcome ascertainment quality (TIER) (39 of 63 study-listings [61.9%] were classified as TIER A according to the proposed framework), an algorithm-cohort-phenotype fit model, a discrimination-calibration-utility triad for performance assessment, a translation gap framework highlighting interacting challenges across validation, deployment, and evidence generation, and a six-step framework for clinical translation. Most models lacked robust validation, with only 9 of the 59 unique publications (15.3%) reporting true cross-institutional external validation, thereby limiting their generalizability. Near-term clinical translation is more likely through institution-specific development and local adaptation than direct widespread deployment, supported by concrete recommendations for clinicians.