Within medicine, there are a large number of transactional interactions that generate data. Pattern recognition can be used to infer what might happen in the future, for example, what treatment might be effective given a particular set of circumstances (Foley and Fairmichael 2015). If it is possible to identify those who are at high risk of developing complications, then they can be more aggressively managed to aid prevention Foley and Fairmichael 2015).
The IHI have outlined the IHI Triple Aim, to:
- Improve the patient experience of care
- Improve the health of populations
- Reduce the per capita cost of healthcare
Models have been built, using routine data, that can identify “Triple Fail” events: those that simultaneously fail to meet each criterion of the Triple Aim (Lewis, Kirkham et al. 2013). Readmission models are commonly cited, but they could be extended to cover a wide range of other Triple Fail events such as:
- Over-medicalised death
- Starting haemodialysis prematurely
- Being offered over-invasive treatment when a preference sensitive decision aid would have nudged a patient towards a less invasive option
Impactibility models identify which of these high-risk patients are most likely to be amenable to intervention or which interventions would be most effective for which individuals. As such, impactability models represent a way of improving the cost effectiveness of the intervention since it is not “wasted” on high-risk individuals who are not going to benefit from the intervention (Lewis 2010).
These approaches fit well with the philosophy of proactively managing healthcare, which is particularly attractive to those health systems that combine both payer and provider (Foley and Fairmichael 2015). This was the case at Geisinger who employ a team of clinicians and industrial systems engineers to develop predictive models that use routinely collected data to identify situations where intervention could avoid costly events that adversely impact patient care (Foley and Fairmichael 2015).
As well as models to predict clinical events, Geisinger have expanded their modelling to predict system level events, such as spikes in hospital activity and to tackle logistical inefficiencies, such as patients who are likely not to attend appointments. In the later example, some clinics were experiencing no show rates of up to 47%. Patients were not receiving the care that they needed and clinic time was being wasted. A model was developed that drew 100 predictive variables from the EHR. This was refined until the 40 best predictors were identified. The model now stratifies all patients according to their risk of not attending. High-risk patients receive a phone call from the clinic prior to the appointment. This has resulted in a 24% reduction in no shows (Foley and Fairmichael 2015).
There is a belief that, in the future, data from electronic records can be integrated with genetic and social care data and perhaps even data harvested from social media and wearable technology to enhance predictive models further (Bates 2015, Dunbar-Rees 2015).
If health and social care data could be integrated, then it would become possible to build predictive models that estimate the future social care needs of patients currently moving through the healthcare system. This could reduce delayed discharges, improve outcomes and improve patient experience (Foley and Fairmichael 2015).
There was also a feeling that healthcare is behind some other industries in the way that it uses routine data for predictive modelling. For example, Amazon.com use historical data to suggest other purchases. The gas industry has also used routine data to develop a better predictive supply and demand models. This has allowed them to redesign their storage infrastructure, reducing the need for large community based storage facilities (Dunbar-Rees 2015).
Some approaches to impactibility modelling can reduce healthcare inequalities, others can worsen them (Lewis 2010).
- Impactibility models that seek to prioritise those people who are receiving sub-quality care will tend to reduce health care inequalities because low quality care tends to be commoner in more deprived areas (the inverse care law)
- Impactibility models that prioritise patients with particular diseases that tend to respond well to intervention, (the so-called ambulatory care sensitive diseases, like heart failure), are likely to reduce health inequalities because these diseases are often more prevalent in patients from lower socioeconomic groups.
- In contrast, impactibility models that attempt to de-prioritise those people who are least engaged (e.g., who have poor English language skills, have drug and alcohol problems, or cognitive impairment), might exclude some of the most vulnerable people in society.
Predictive modelling and impactibility modelling are effectively forms of screening because they generate true positives, true negatives, false positives and false negatives. Just as with other forms of screening, there are harms associated with false positives and false negatives. We should therefore use the Wilson and Jungner criteria (or some modification thereof) (Wilson and Jungner 1968), developed for the WHO, when considering the appropriateness of a predictive modelling programme (Lewis 2015).
Evidence