Learning Data Driven Medical Checklists For Multi-Disciplinary Consult Scheduling

Checklists are simple tools that are often used to promote safety in clinical applications, because they can easily be integrated into a clinical workflow deployment without the need for extensive training or additional technology. Currently, machine learning in health focuses on large, complex deep neural networks.

In comparison with deep neural networks, checklists are far easier to use, understand, and scrutinize.

In practice, checklists are difficult to develop, and the vast majority of real-world checklists are hand-crafted by panels of experts.

Given the widespread integration of electronic health data over the past decade, our goal is to learn checklists from data for clinical tasks. The main application of our method, in partnership with Dr.

Leo Celi at Beth Israel Deaconess Medical Center (BIDMC) in Boston, is to predict mortality in patients eligible for Continuous Renal Replacement Therapy (CRRT).

A generated checklist for CRRT is clinically desirable in the intensive care unit as a mechanism to trigger a multidisciplinary discussion with the patient’s family, to consider the value of the treatment weighed against the physical burden and cost.

Thus we include operational considerations such as number of items in the checklist, the false negative/positive rate tradeoff, and fairness across protected groups.