Population Health Management (PHM) is a term that has become common in healthcare circles in the last few years. At a high level this describes an approach that uses data science to shape care for groups or individual people based on their clinical risks. It is seen as a solution to the combined challenges of ever-increasing complexity and so cost of healthcare and an ageing population who live with treatable disease for longer. Meeting this constantly changing need, means more capacity and higher costs in the current ‘reactive sickness service’ model or a change to a pro-active risk mitigation approach.

Segmentation and risk stratification are typical techniques used to understand and identify those at high risk. Then existing services are mapped to this need or new services designed and implemented. There is a slight disconnect in the language in that population health management is in fact always about delivering an identified intervention to one or more individuals, so though we talk about population health, in reality, as ever, healthcare is about care one person at a time.

PHM doesn’t necessarily require highly technical approaches, for example the annual NHS flu vaccination programme is rolled out to everyone over 50 years of age, as this group are at higher risk of serious negative outcomes from infection. This is a simple example of segmentation and intervention targeting based on that technique. Also, highly technical approaches are not always useful. The “Glasgow effect” is a well-researched example that demonstrates worse life expectancy for people who live in Glasgow. Despite complex mathematical approaches to remove confounders such as age, income, income inequality, health status and many more, the cause of this worse outcome is unclear and therefore it is not possible to design an intervention.

 

“A psycho-social problem that will not be fixed by targeting conventional risk behaviours” Harry Burns, Scotland’s Chief Medical Officer (2011)

This is not to say this is not a problem that cannot be addressed, but even the most advanced analytical approaches are unable to identify a specific risk factor to focus mitigation on, with complex multi-factorial issues requiring equally complex solutions.

PHM should be led by clinicians, not data scientists. Clinicians and Data Scientists need to work together to identify priority areas where risks can be identified, predictive factors isolated, and an intervention is available that will mitigate that risk. This comes to the heart of our model, identifying risk alone is not PHM. Targeting an effective intervention to those in the population who can benefit from it, are impactable, is key. Then designing a solution that uses data at an individual level across a population to calculate both risk and impactability and inform responsible clinicians that this individual is likely to benefit from that intervention is possible and will deliver value to patients and the system.

All of this talks about a specific risk, PHM approaches are always about ‘risk of’ a specific bad event, not generic ‘risk’. A range of data science approaches can be used such as:

  • Statistical models to identify predictive factors that can be modified
  • Markov chain analysis to identify factors that predict progression between categories of risk
  • Machine learning approaches such as anomaly detection algorithms to identify previously unknown risk factors
  • Person-centric model based on Bayesian evaluation of multiple risk fed into deep learning techniques to evaluate across a large population

All of this sounds very impressive and technically it can be. As a patient or a clinician, the technical approach used is irrelevant as long as it is proved to be accurate and improves outcomes for that individual. As with all things, it is always preferable to use the simplest possible approach that can do what is required.

Clinicians have been using data to inform their diagnostic and therapeutic decision making since time immemorial, experiential learning is a form of data gathering. So PHM is not a revolution but an evolution of existing ways of working, with new tools and approaches, enabled by the scale and detail of data and computational power now available. The outputs of analysis are presented to clinicians to make better decisions. As with all medicine the way a treatment works in the real world requires continuous evaluation, of the data models and interventions put in place. To date evidence suggests that only a small proportion of PHM programmes deliver on the promise and the reasons for this poor performance need to be understood.

Though modern computation techniques would allow algorithms to evaluate risks and propose responses in an automated fashion, PHM approaches do not advocate for computers to make or over-ride clinical decisions. PHM has more parallels with an expert systems approach. Medicine remains a human science, PHM continues the drive toward personalisation of care by providing clinicians with better information about the risks to and the best intervention for, the individual.