Let's Keep the Conversation Going

  • Home
  • Blog
  • The Persistent Problem with Predictive Analytics

The Persistent Problem with Predictive Analytics


Predictive analytic solutions, which span the continuum from simple rubrics to complex neural net models, share a common boundary problem. At some point, the predictive model reaches a point where it is no longer accurate at identifying at-risk individuals. No matter how clean the data is or complex the math, all predictive models will reach this point.

Take the Braden scale for example. 

Braden is widely used to identify those at high-risk of a pressure ulcer. And Braden is relatively good at predicting high-risk individuals. The problem starts to appear when we look across the patient population. At some point, Braden starts to miss patients. Because Braden is based on a limited set of factors and because those factors do not change, there is no way to drive the flexibility needed that will enable accurate predictions when patients fall out of the high-risk stratification.

This is limitation is starkly illustrated when we compare the performance of Braden to the Cognitive Clinical Success Machine. Based on a longitudinal analysis of Jvion's solution compared to Braden, we found that:

  • Braden mischaracterized 12.5% of actual pressure ulcers as Low Risk vs. <2% with the Jvion tool
  • In patients with a Braden score in the "moderate" range, Jvion was 500% more efficient in detecting pressure ulcers
  • Jvion segmented 5.4% of patient population as high risk. This population yielded 74 more pressure ulcers than a Braden score applied to 100% of the population

The Cognitive Clinical Success Machine not only better identified who was at risk across the entire population, it helped improve the allocation of resources by more precisely identifying the number of patients that needed interventions and treatment.

The application of interventions brings up another predictive analytic shortcoming. Predictive analytic solutions can only tell who is at risk. They are one-dimensional. They focus on a discrete set of problems (e.g. pressure ulcers, infections, readmissions) and risk stratify a patient population based on a limited set of use cases. They can't account for the interrelated risks and biographical context that leads to the best intervention for an individual patient. In contrast, Cognitive Clinical Success Machines deliver multidimensional views that include why a person is at risk and what to do about it. Rather than pointing to accuracy, these machines focus on effectiveness. Effectiveness accounts for two broad categories of machine capability: 1. how well the machine identifies at-risk across the entire patient population and 2. the ability to enable effective interventions. This approach extends multidimensionality to deliver an "Eigen Biographical View" of the patient.

A patient's Eigen Biography accounts for all the factors—clinical, socioeconomic, social networks and support, occupational, inequalities, access—that contribute to a person's risk. The Eigen Biography delivers a complete, virtual view of a patient that is used to identify the best actions and interventions that will result in the best possible health outcome. Moreover, this view enables the identification of adjacent risk factors that can be influenced and mitigated often by the same intervention. This biography is constantly updated and exists in perpetuity within the machine. To put it another way, the cognitive machine is always "thinking" about every patient in much the same way as a clinician: as a complex, multidimensional individual.

Tags: ai braden scale