Centre for Doctoral Training in AI-enabled Healthcare


Analysis of dynamic treatment strategies in electronic health records

The NIHR Health Informatics Collaborative for Critical Care (HIC-CC) contains rich longitudinal data for tens of thousands of patients across multiple hospitals since 2014. More than 250 different demographic, physiological and treatment variables are captured as often as hourly creating more than 250 million data points. Sepsis is the leading cause of morbidity and mortality within critical care. Appropriate antibiotics delivered in a timely fashion have clear efficacy in a laboratory setting, but such clinical evidence is typically flawed.

Problem: Clinical decisions to initiate antibiotics in sepsis are typically presented and analysed in an unrealistic binary scenario. The timing indicator is usually an administrative event (e.g. blood draw for culture). In reality, clinicians continually evaluate and update the decision to withhold or initiate antibiotics against an updating mental model of risk and benefit. Early initiation will lead to over treatment and drug resistance whereas delayed treatment will lead to harm from disease progression, and organ dysfunction. In other words, an adaptive treatment regime is employed so that when time-varying measures such as fever, inflammatory response and physiology cross a threshold, antibiotics are started.

Proposal: Electronic health records such as HIC-CC represent a complex, dynamic longitudinal description of health and treatment more akin to the clinical mental model. We propose to use the parametric g-formula to evaluate these clinical dynamic treatment regimes. These will permit development of pragmatic embedded Sequential Multiple Assignment Randomised Trials (SMART) that would be otherwise too costly to implement.


crtical care