Machine learning may produce “proxy measures” for brain health problems
A study published today by an interdisciplinary collaboration, led by Denis Engemann of Inria, demonstrates that machine learning from large population cohorts can produce “proxy measures” for brain-related health problems without the need an evaluation by a specialist. The researchers took advantage of the UK Biobank, one of the world’s largest and most comprehensive biomedical databases, which contains detailed and secure data on the health of the UK population. This work is published in the open access journal GigaScience.
Mental health problems have increased worldwide, with the WHO determining that there was a 13% increase in mental health problems and substance abuse disorders between 2007 and 2017. The burden of these diseases burden on society is considerable, having a negative impact on almost all areas of life. : school, work, family, friends and community involvement.
Among the many issues that hamper society’s ability to treat these disorders is the fact that diagnoses of these health issues require specialists; the availability of which varies considerably across the world. The development of a machine learning methodology to facilitate mental health assessments could provide a much needed additional means to help detect, prevent and treat these health problems.
To develop AI models sensitive to mental health, researchers from Inria (Saclay – Île-de-France) and their colleagues turned to UK Biobank to obtain the necessary data. UK Biobank stores not only biological and medical data, but also questionnaire data on personal circumstances and habits, such as age, education, tobacco and alcohol consumption, sleep duration and physical exercise. Specific for this study, these questionnaires also include socio-demographic and behavioral data, such as moods and feelings of individuals, and the biological data include magnetic resonance (MR) images of 10,000 brain scans of participants.
Inria scientists combined these two data sources to create models that approximate measures of brain age and scientifically defined traits of intelligence and neuroticism. These serve as “proxy measures”, which are proxy measures strongly correlated with specific diseases or outcomes that cannot be measured directly. Developing approximations in this way has been used successfully in the past to predict “brain age” from MRI images. This previous body of neuro-clinical work served as a starting point for Denis Engemann and his team.
In this work, we have generalized this methodology in two ways. First, we have shown that beyond biological aging, the same framework of indirect measures is applicable to constructs more directly linked to mental health. Second, we have shown that useful indirect measures can be derived from other data than brain images, such as socio-demographic and behavioral data. “
Denis Engemann, National Research Institute in Digital Sciences and Technologies
The researchers validated their indirect measurements by demonstrating the same results in a separate subset of UK Biobank data.
The results of the work here provide a glimpse into a future where psychologists and machine learning models could work hand-in-hand to produce increasingly refined and personalized mental assessments. For example, in the future, customers or patients can grant a machine learning model secure access to their social media accounts or mobile phone data, and then return useful proxy metrics to both the client and the mental health or education expert.
However, while AI can provide much needed assessment tools, human interaction will always be essential, as Engemann points out: -based cases and through social interaction, whether obtained using machine learning or classic testing. “
Dad, K., et al. (2021) Population modeling with machine learning may improve measures of mental health. GigaScience. doi.org/10.1093/gigascience/giab071.