Trimming away hospital costs: A bi-product of saving lives with statistics

Healthcare service delivery in most systems can be described as fragmented at best. In many healthcare systems, there has been very little continuity of care and integration for services provided by General Practices (GP), Hospital and Health Services (HHS) and other healthcare providers. Integrated Care is a worldwide trend in healthcare reforms that focuses on co-ordinating these different services.

Gold Coast HHS is currently implementing Integrated Care though Gold Coast Integrated Care (GCIC), which is a cross-sector collaboration between teams of multiple disciplines to provide one shared care plan for patients. The overall aim of Integrated Care is to maximise efficiencies and increase quality of care for patients with complex and comorbid conditions (i.e. patients with more than one ailment).

Emergency hospital admissions account for a significant proportion of total hospital admissions and costs. Re-hospitalisations of patients with chronic and complex comorbid diseases make up a large proportion of Emergency Department (ED) admissions. One way to reduce hospitalisations of this subset of patients is to identify which patients are at high risk of hospitalisation, and prioritise them for case management in the community before they require hospitalisation. GCIC aims to utilise healthcare information gathered from an array of information sources at near real-time to stratify patients into disease/condition groupings, and then subsequent groupings associated with risk of hospitalisation. Information sources can include: HHS, GP, specialist medical, allied healthcare professionals, ambulance, aged care facilities or, remote home monitoring data.

Building a Risk Stratification model is a complicated process, requiring an appropriate risk metric, along with a predictive modelling algorithm applicable to the metric. Hospitals and GP’s have the capacity to record a large number of different fields in their patient and admission databases, and the cleanliness and integrity of these databases are paramount in determining a strong predictive model.

There are many ways to measure risk of Emergency Department admissions. For scenarios where large number of admissions are present and unavoidable, Integrated Care may only want to target serial offenders of the hospital system. In this case, it may be useful to model the number of times patients are admitted into the Emergency Department. On the other hand, it may be more useful to predict distinct events of admissions. This prediction is used in many US and UK Risk Stratification approaches. In this case, the metric to predict is the probability that a patient will have ED admission in a time period (e.g. 3/6/12/48 months).

Choice of predictive model is entirely dependent on the metric to be predicted. Risk Stratification is a great example of the use of the commercial application of Generalised Linear Modelling statistical techniques. For the case of a continuous response (e.g. length of stay), a multivariate linear regression may be useful. For the case of modelling probability of a distinct event, a logistic regression model is the popular method. Other methods also exist, such as a Regression Trees, which are useful in certain environments, but can have crippling limitations.

Most Australian HHSs use legacy enterprise systems for storing service and admissions records (HBCIS, ORMIS, etc). However, one benefit of these systems is that they encourage diligence in data capturing, boosting data cleanliness. The same cannot be said for GP practice data, where the health of the data can differ wildly between practices. This discrepancy is one of the big limiting factors in the scope of variables that a predictive algorithm can model with.

Variable selection is a time consuming process that must be guided visually, with medical specialist collaboration and Literature Review. Many readily available fields have been determined to be extremely strong predictors of admissions in healthcare systems, including Age, Condition/s, Comorbidities, Prior ED admission and presence of medication. When the choice of variable fields and data is clean, and usable, it is a good preliminary indicator of the strength and confidence that specialists can have in a predictive model.

Lastly, there is no single solution that can be used across all health systems. Risk Stratification models to be used on a target population must be generated from a historical sample that is analogous to the target population. For each health system requiring Risk Stratification, a bespoke model must be generated exclusive to prior models generated. Most health systems have different series of variables that are available to be modelled. Even when variables are the same, the underlying historical sample must be analogous to provide the best prediction.

Find out more about Biarri working across Hospital and Health Services >>