Adapting surgical models to individual hospitals using transfer learning

Preoperative models to assess surgical mortality are important clinical tools in determining optimal patient care. The traditional approach to develop these models has been primarily centralized, i.e., it uses surgical case records aggregated across multiple hospitals. While this approach of pooling greatly increases the data size, the resulting models fail to reflect individual variations across hospitals in terms of patients and the delivery of care. We hypothesize that this process can be improved through adapting the multi-hospital data model to an individual hospital. This approach simultaneously leverages the large multi-hospital data and the patient-and-case mix at individual hospitals. We explore transfer learning to refine surgical models for individual hospitals in the framework of support vector machine by using data from both the National Surgical Quality Improvement Program and a single hospital. Our results show that transferring models trained on multi-hospital data to an individual hospital significantly improves discrimination for surgical mortality at the individual provider level.