Building and implementing a TBI prognostic model featuring real-time analysis of brain CT images

Scope of Work Duke will complete all work for the machine learning model building and implementation of the model into the Duke clinical workflow. For Aim 1 of the project, this work will include data extraction and cleaning, neural network architecture design, and model optimization and validation. For Aim 2, this work will include establishment of technical infrastructure for real-time image and access and processing, construction of a front-end dashboard in close collaboration with frontline clinicians, and deployment and prospective validation of the model. The latter step will also consist of education and training of hospital users. For Aim 3 of the project, Duke will guide staff at Jefferson through the model implementation and validation process, with the active integration and training performed by staff at Jefferson. In Aim 3 Duke will also run the experiments on multi-site model generalization, using retrospective data at both Duke and Jefferson. Data will be shared between Duke and Jefferson via secure ethernet transfer between Jefferson’s secure data warehouse and Duke’s Protected Analytics and Computing Environment. The end goal of the work will be to provide a sophisticated, high-accuracy, and seamlessly integrated tool for predicting the risk of actionable TBI complications over the course of a TBI patient’s hospital encounter. This method, which will augment decision-making for treating a complex neurological condition, will significantly improve overall TBI outcomes, reduce readmission rates, and minimize the extraordinary costs incurred by inefficient provision of healthcare resources.