Predictive Modelling

No one likes to be profiled, whether using age, gender, or race as criteria, but is it different if it is for our health?

Predictive modelling is a hot new technology used to produce profiles and is of rising interest in healthcare. It uses analytical techniques associated with machine/computer intelligence (that is, neural networks and artificial intelligence) to support, enhance, or indeed, even replace clinical and managerial decision-making by people.

A recent pilot project in England explored using weather forecasts to anticipate health exacerbations in individuals with COPD using a model which linked clinical indicators to the weather.[i] Strategic health authorities in England have been looking at how to use predictive models to data-mine health databases to identify individuals at high risk.[ii] In other countries, predictive models are used by both insurance companies and healthcare providers to identify individuals at risk, manage their care and better target healthcare resources.

Progress in improving clinical decision-support technologies notwithstanding, it is absolutely critical to make sure that before modelling becomes more imbedded in routine use, that we assess the potential impact. In particular, patients need to be involved as their consent may be needed to legitimate health profiling.

The increased interest in the use of predictive modelling reflects advances on three fronts:

· The availability of health information from increasingly complex data bases and electronic health records;

· Better understanding of causal or statistical predictors of health, disease processes and multifactorial models of ill-health;

· Developments in computer models using machine intelligence (artificial intelligence, neural networks), which are exploiting developments from other industry areas such as climate modelling, environmental impact assessment, modelling consumer behaviour, and other areas.

The door is, therefore, opening to the use of statistical predictive models in health care because bringing these three factors together is highly incentivised in health policy through operational priorities to improve health system productivity, better allocate resources particularly toward high-need individuals, and more generally improve the social value derived from health care funding. Indeed, advances in the management of long-term conditions could be seen as depending on the use of predictive modelling of patient health information.

Currently, most people’s experience of modeling lies in population health and epidemiology telling us predictive things about groups of unidentified individuals. It was left up to clinician’s or patients to decide if this information was relevant to any particular identifiable person. In many cases, these types of models are relatively simple, running on spread sheet-type computer software using simple statistical relationships.

The new and predictive models identify individuals with particular health risks from databases of health information such as electronic health records. Once identified, these individuals, now associated with a predicted future health risk profile, can be assessed in terms of their likely future consumption of health resources, the predicted clinical course of disease, compliance with a course of medicine (such as patient behavioural response to a perceived adverse reaction), or other factors of interest. That is to say, these models identify high-risk, high-cost, high utilisation individuals and how their health status might change in the future in response to care. Therefore, models do not just find the individuals at risk, they can also determine the targeting of resources, and importantly, whether that resource is better managing this particular patient’s risk or used for some other purpose.

For some, this may sound too good to be true. Imagine how much easier it will be to budget and design care programmes, or identify patients who might be non-compliant with their medicines, or experience health exacerbations due to weather, or lifestyle choices. The golden age of customised healthcare beckons.

But we need to pause and reflect before we embrace this technocratic future:

1. Should individuals consent to be profiled in the first place since the results of all this technological activity is to produce risk profiles of named individuals?

2. Could principles of social justice be violated in virtue of how these predictive models are actually designed and work. [iii] In the case of artificial intelligence models, they must be “taught” how to predict, the selection of the patients it learns from predisposing the models to “think” in a particular way; it is possible that they learn to discriminate in ways that are unacceptable.

3. The evidence shows that predictive models out-perform unaided judgment of clinicians. [iv] If these tools are to support clinical decision-making, but are better at it, what is the clinical and regulatory significance of disagreement between a model and clinician – when the model has a better chance of being right?

4. Models are “black boxes”; their underlying logic is no more accessible to users than the computer code that runs your word-processor. How are decision-makers to know that these models are in fact fit-for-purpose?

The use of nonlinear predictive modeling and data mining of health databases is in its infancy as the NHS is still building technical infrastructure and capabilities through electronic health records. Primary Care Trusts do not yet model the health needs of their populations at this level of precision, and providers have difficulty anticipating resource requirements for next week, let alone next year. Patients have yet to learn how to become vested as the lynch pin for improving service quality, while current health policy in England favours the better targeting of resources for long-term health conditions.

What is to be done? We can set out basic and certainly preliminary requirements if predictive models are to enjoy public, managerial and clinical confidence and acceptability so that this comprehensive agenda of change can be better assessed:

1. The underlying logic and coding of predictive models needs to be independently assessed to quality-assure them as fit-for-purpose;

2. We need to understand the social justice implications of health profiling through data-mining and predictive modelling to ensure that principles of social justice are not violated – that is, we need to know if this is socially acceptable, and is agreed to by patients.

We are left with trying to balance the goal of improving health by using predictive modelling (with the best intensions, no doubt) with the potential invasion of individual privacy. In these days of vast databases, personal privacy can be challenged on all fronts, usually in the name of national security. The question for us in healthcare is whether there is a wider public interest that will permit the use of predictive modelling to trump personal privacy? Do we, in effect, own our health status, or not.

Discussion and debate is certainly timely.

[i]Met Office, Health Forecasting Project, www.metoffice.com/health/copd_forecasting.html, accessed 18 October 2005

[ii] The PARR project: “Patients at Risk of Re-Hospitalisation Case Finding Tool”, King’s Fund, London. www.kingsfund.org.uk/health_topics/patients_at_risk.html, accessed 18 October 2005.

[iii] RM Dawes. 2002. The ethical implications of Paul Meehl’s work on comparing clinical versus actuarial prediction methods, Symposium in Honor of Paul E. Meehl, Atlanta GA 31 May.

[iv] RM Dawes, A theory of rationality as a ‘reasonable’ response to an incomplete specification, Synthese, 1(2-2000)133-163.