TOP
As a young psychiatrist, I went through a period of being excited about the brain. It seemed that emerging technologies in imaging and genetics would transform clinical psychiatry. This was encouraging news with mental health problems causing the most disability worldwide, being the leading cause of death among young adults and, in severe forms reducing an individual’s life span by a decade. And yet a recent British Journal of Psychiatry article rather gloomily concluded with the view that clinical practice in psychiatry hasn’t changed fundamentally for half a century.
Although the academic journals became dominated by biological investigations, like many of my colleagues honing their skills on the front line of clinical management while I endeavored to get the biology right, the psychosocial aspects of my patient’s condition became my principal focus. So what went wrong?
For a start, none of those technical developments became tools that psychiatrists could use routinely. In a recent study, Steele and Paulus identified problems in demonstrating that findings in small samples are true of the population as a whole; even when participant numbers are large (the researchers give the example of identifying brain changes in major depression), the findings do not tell you anything about the patient you have in the consulting room. Even very large genetic studies did not contribute to this purpose, as the effect sizes were small for the multiple abnormalities identified.1
Just as important is that psychiatric disorders are multivariant constructs. This means that taking account of a single measure is unlikely to reveal much about an individual patient (see my blog Can Suicide Risk Be Significantly Influenced by Underwriting?). Ultimately a group level analysis tells you only about the group and not the individual.
As a solution, Steele and Paulus propose “a multi-variant risk-prediction framework”.2 Effectively, this allows multiple aspects of a patient’s presentation to estimate the probability that a certain outcome is present (a diagnosis) or will occur within a given time (a prognosis). This process uses machine learning, a catch-all phrase for a variety of techniques that train an algorithm to make predictions about a given individual from many examples where the outcome is known.
Increasingly, computational methods like these are being used to predict an individual’s diagnosis, disorder severity and prognosis - for example, their response to antidepressants, or their risk for mild cognitive impairment progressing to dementia. Of course, these efforts rely on a large assumption that mapping between mental states and brain states is computable. While “risk calculator” tools are familiar to underwriters assessing certain disorders (e.g., cardiovascular risk), none of these models is yet available to psychiatrists to guide treatment let alone available to insurers for use in risk assessment.
This is important for Life and Health insurers because mental health issues are more widely acknowledged by individuals seeking treatment and have seen hugely increased media exposure. The insurance industry is responding by improving its handling of mental health issues and thereby providing more people with fair access to its products at reasonable cost. On the other hand, with good reason, the industry remains concerned at the numbers of mental health claims, their duration and cost.
Computational models have the potential to provide solutions to at least some of these issues. A wealth of data exists concerning mental health claims that could be used to develop multivariant models of risk for future claims. The factors that emerge from these models would allow the development of appropriate, targeted question sets for the applicant and the professionals involved in their care. For example, it could help insurers with particularly difficult problems, balancing the sensitivity of the issues involved, the range and complexity of the language used in an insurance context, and the need for efficient underwriting processes. It would also enable the use of data based on concrete industry experience rather than extrapolating data from historical clinical populations, which are often made up of the most severe cases and therefore unrepresentative of the insured population.
I sought to learn more about the processes involved in building computational models but retreated rapidly from a world where I understood nothing and immediately felt massively out of my depth. This kind of work clearly requires finding a common language from a diverse range of disciplines, where each entity brings expertise and imagination to the table. It has the potential to develop fairer and more affordable access to insurance products for those who experience mental health problems.
