Unreasonable effectiveness of training AI models locally
Sepsis remains a leading cause of morbidity and mortality worldwide.1 The use of artificial intelligence (AI), and particularly machine-learning (ML) approaches, to predict which patients are at risk for sepsis in the hospital may improve patient-centred outcomes through early recognition and timely antibiotics. Yet, despite major interest in the use of ML applications in sepsis care, there are only a handful of successful examples of model implementation that save lives through early detection.2 3 The high cost and extensive system architecture required to test and implement novel ML applications have limited many institutions’ abilities to bring these models to the bedside. Unfortunately, this has resulted in a preponderance of studies on model development rather than implementation and reliance on proprietary models disseminated to health systems without validation or testing. One well-known sepsis predictive model developed by an electronic health record vendor (Epic Systems,...