OBJECTIVES: Early diagnosis of sepsis has been shown to reduce treatment delays, increase appropriate care, improve patient outcomes including reduced mortality. The machine learning prediction algorithm uses variables routinely collected at intensive care units (ICUs) to predict sepsis and has shown excellent predictive properties in clinical settings. This research aims to estimate the potential cost and cost-effectiveness impact of a machine learning algorithm forecasting the onset of sepsis, in an ICU setting in Sweden and the UK.

METHODS: The model is based on findings from a randomized, prospective clinical evaluation of the machine learning algorithm from literature sources and local price lists. Of particular interest is to model the relationship between time from sepsis onset to treatment and prevalence of septic shock and in-hospital mortality. The model base case assumes that the time to treatment coincides with the time to detection and that the algorithm predicts sepsis three hours prior to onset. Cost-effectiveness (CE) is evaluated versus clinical practice methods in Sweden (Sepsis-3 criteria) and in the UK (NEWS2).

RESULTS: Stochastic CE analyses showed that the machine learning prediction algorithm was a cost-effective treatment option in both Sweden and the UK, demonstrating ICERs well below an established threshold of €20,000 per QALY in most scenarios. In the model base case (BC) the machine learning prediction algorithm was dominant in 57% and 2% of the simulations in the Swedish and UK setting, respectively. The BC ICER was - €1,783 (cost-saving) and €4,832 in Sweden and UK, respectively. A three-hour faster detection implies a reduced in-hospital mortality, resulting in 356 and 1,469 lives saved per year in Sweden and UK, respectively.

CONCLUSIONS: A sepsis prediction algorithm will have a substantial cost and lifesaving impact for ICU departments and the health care systems in Sweden and the UK.