When Predictive Models Misallocate Resources

A story told by sepsis data sets.

This article is “fun” the way getting run out of the park by petulant geese is “fun”, but it shows two things – one of domain interest, and one of general interest.

In this study, the authors took hospitalization data sets and used typical machine learning methods to develop predictive models. First, they made a predictive model for patients to receive a final diagnosis of “sepsis”. Then, they used mortality data to predict which patients would be at highest risk were they to develop sepsis. Finally, they merged these predictive models to describe the heterogenous risk for poor outcomes in patients at varying risk for sepsis:

The key finding above – the one of “domain interest” – is that many patients identified at high-risk for sepsis are at very low risk for downstream mortality. Likewise, some patients at apparent lower-risk for sepsis are quite vulnerable for poor outcomes if their diagnosis were to be delayed or missed. This could, in theory, inform strategies of screening and prioritization for suspected infection. In practice, SEP-1 rules the day.

The key principle demonstrated by these observations – the “general interest” application – is developers of predictive models ought potentially consider cost and clinical effectiveness of actions resulting from the predictions. Where the prediction may expend resources – whether actual cost or a zero-sum game of clinical attention – there may be additional priorities to incorporate into tools meant to augment clinical decisions and maximize patient benefit.