Latest predictive risk model tests demonstrate the value of incorporating GP electronic medical records

Incorporating additional datasets into the more commonly used risk models can boost their predictive power, potentially giving health care staff more time to identify and work with at risk patients to lower the chance of their conditions rapidly deteriorating according to a BMJ Open study published today.

However the authors warn that the success of a predictive risk model depends on many factors beyond statistical performance. The results come at a point when NHS commissioners are increasingly faced with the problem of choosing between predictive risk models and how to implement them with little information to guide their decisions.

Using logistic regression techniques researchers at New York University and the Nuffield Trust tested a range of data sources (hospital inpatient, A&E, outpatient and GP electronic medical records) across five primary care trust areas within England.

Although the choice of underlying data makes a big difference to the initial investment costs, no study until now has looked at the marginal value of different data types Dr Martin Bardsley, Director of Research at the Nuffield Trust and study co-author

Participants in the study were made up of 1,836,099 people aged between 18 and 95, making it one of the largest studies to date of how to predict hospital admissions in the UK.

The work looked at a number of elements that make up predictive models including the effect of: additional data sources, whether individual site predictive models were more accurate than models based on data pooled from multiple sites, and the effect of using different population sets (denominators).

Key findings include:

• The addition of more detailed data sets led to moderate improvement in the numbers of patients identified, with little or no loss of predictive value;

• The use of GP registry data for the denominator proved to be of significant importance. By including all patients in the area, not just those with prior hospital use, improved rates of case findings were observed;

• The models were less accurate at identifying patients with lower risk scores, meaning that the amount spent on their interventions would need to be reduced to achieve a financial break–even (the point where the cost of the intervention is off-set by cost savings from reductions in future admissions);

• Models calibrated to local data did not show a consistent improvement over models built on pooled data and then applied to an individual area.

Dr Martin Bardsley, Nuffield Trust Director of Research and study co-author commented:

‘Although the choice of underlying data makes a big difference to the initial investment costs, no study until now has looked at the marginal value of different data types.

“The research published today suggests there is a strong case for including A&E and outpatient records in the models. Both are readily available to commissioners and their inclusion not only improves case finding but is also helpful for developing profiles of high risk patients, information which is vital to designing effective interventions.”

Acknowledging that the use of GP electronic medical records presents significant challenges, Dr Bardsley continued:

‘Information from GP Practices can be a very powerful tool for improving care services including identification of people at high risk of emergency admission.

“However there are challenges in interpreting data as there are some differences in coding and recording between sites. In addition we also need to be mindful of the need to protect patient confidentiality but with the right safeguards this should not be an insurmountable barrier”