DREES – Predictive modelling for health insurance

No votes yet.
Please wait...

Version française ici.

In July 2017, the French DREES (Directorate of Research, Studies, Evaluation and Statistics) published a report « Studies and Results » entitled « The rise of predictive models in international health systems« . It is focused on those new methods of anticipating events related to the care system, combined with techniques of scoring individuals according to the probability of occurrence of particular health risks.

The quantity and the multiplication of the available data allow today to develop powerful statistical models of risk analysis. These are no longer based solely on simple data such as the age or sex of the individual, but can take into account many other parameters. This makes it possible to better anticipate health risks, to better size the health system (for care), but also to better regulate this environment. For example, preventive actions can be implemented in a targeted manner and innovative remuneration methods or incentives for actors can be defined. This has been developing over the last decade at international level

Definition of predictive models

The objective of these models is simple, to determine the probability that a given individual undergoesthe occurrence of a given risk within the next 30 days, 6 months or 1 year. Three cases are studied :

  • Hospitalization : The probability of experiencing unplanned hospitalization, often over a 1 year period.
  • Dependency : Likelihood of losing one’s autonomy and being transferred to EHPAD (French term for special homes for dependent people) or becoming a consumer of home help
  • Costs models : Likelihood of becoming a major consumer of care.

The data available and collected differs greatly between countries / models. They may include data :

  • Socio-economic: widely used for monitoring diseases such as cancer or diabetes or as risk factors for obesity or smoking.
  • Medical Diagnostic Elements (Hospital Codes)
  • Prescription and use of drugs

The difficulty is to be able to collect samples of structured data, important enough to determine rules. For example, establishing correlations between risk factors, diagnoses and drug use. It is then necessary to form groups of individuals with similar risks.

The performance of these models may be modest as for drug consumptions (20 to 25% of cases correctly anticipated) but may be more effective as for hospitalization (between 50% and 80% of Positive Predictive Values – VPP). Some special cases (cardiovascular diseases for example) show amazing reliability (more than 85% of cases correctly anticipated).

For this, the usual techniques of statistical analysis (linear regression in particular) are commonly used. Big-data techniques, decision trees and machine learning are also emerging for the development of models.

Scope of predictive models

The basic principle of all models is therefore to define groups of individuals with equivalent risks. This obviously makes it possible, from a macroeconomic point of view, to better organize the health system. In particular, this alters the management logic, from reactive to preventive functioning. However, application to a particular person presents confidentiality problems.

modeles predictifs

Allowances to insurers: German and Dutch models

Public health insurance bodies operate on the basis of a proportional contribution, not correlated to the state of health. In Germany, since 1996 the budgets allocated to each regional entity are therefore adjusted according to the actual risks of the population covered. This is done through the creation of 80 homogeneous risk groups.

In the Netherlands, the 2006 reform made health insurance compulsory for all. Public funds are redistributed to each insurer according to models using the same techniques, while avoiding medical selection at entry.

English and American Models: Networks of Care

Since 2013, the NHS has imposed the creation of local healthcare networks (the Clinical Commissioning Groups), which are responsible for planning and managing care in a given geographical area. This decentralization was carried out in parallel with a requirement for stratification of populations to better predict care pathways.

On the US side, the network system is also deployed. Healthcare providers are paid by an annual fixed fee from public or private funds (Medicaid, Medicare and private insurance companies). This is the principle of the Managed Care Organization, introduced in the 1980s. In the Obamacare, established in 2010, Accountable Care Organizations are set up. They are also networks, with a virtual budget defined in advance, but solidarity with overruns or savings. This makes it possible to incite everyone to moderate costs.
In both cases, taking into account a certain number of data makes it possible to refine these budgets.

Improve the care path with predictive models

Preventive measures are often too late and only implemented following a first visit to the hospital. The challenge is to anticipate risks and set up prevention on emerging risks.

The NHS has successively implemented a predictive model of re-hospitalization at one year (Patients at risk of re-hospitalization – PARR), followed by a one-year hospitalization (Combined Predictive Model (CPR). Then, combined with data from social services, a model of loss of autonomy at one year was also developed. The same programs have also been developed in Canada and Australia.

Benefits appear at several levels:

  • Limit unplanned hospitalizations
  • Preventing risks
  • Improve the management of patients at risk of re-hospitalization
  • Establishing home support programs
  • Communication of lists of patients at risk to general practitioners (by the Welsh Predictive Risk Service)
  • Deployment of postoperative control resources (telemonitoring, nursing, etc.)

On the insurers side, these models are particularly interesting to better anticipate risks and costs. More specifically, this would allow better management of patients with chronic diseases and those requiring the implementation of case management with personalized care.

Studies and evaluation

More generally, the use of risk profiles makes it possible to compare the performance of healthcare providers between different regions or entities, using similar indicators.

Risk groups are also used to assess the effectiveness of new care (telemedicine, specific prevention programs, etc.). This is done through monitoring of control groups and risk scores. The models thus determined show a VPP close to 50%.
In France, through the medico-administrative data collected, such predictive models could be developed in the near future in order to support actions to improve pathways or their evaluation.

DREES – Predictive modelling for health insurance

Vous pourrez aussi aimer

Laisser un commentaire

Votre adresse de messagerie ne sera pas publiée. Les champs obligatoires sont indiqués avec *