Predicting mental health early and precisely has major implications for clinical management and practice, and ultimately life expectancy. This project will use machine learning techniques to produce robust modelling tools that aim to improve the precision of clinical practice in mental health. Machine learning will be used to predict and classify disease risk at an individual level (for dementia, anxiety, depression, and others) and to determine the interactive factors that influence mental health across people’s lifespans (e.g. genetics, cognition, demographics).
Explaining the science
Most research so far has focused on classifying patients versus healthy controls. In contrast, this work will utilise large-scale datasets from preclinical populations (i.e. individuals not showing symptoms of disease or affliction) and aim to predict mental health across repeated measurements in time, from the same individuals. This approach is critical for early diagnosis and intervention.
To address the need for training robust models with a comprehensive set of inputs as well as the need for future predictions to be based on clinic data alone, machine learning algorithms will be trained with 'privileged' data from bespoke, multi-modal MRI scans. These scans have high sensitivity to neurocognitive disruption but are not routinely collected at the clinic as they are costly, time-consuming, and often not tolerable by all patients.
Following the training phase, these robust models will be tested to see if they are able to classify new individual cases without the 'privileged' data; that is, from routine epidemiological and cognitive data collected at the clinic as part of medical records.
This approach has the potential to deliver a low-cost, robust, and sensitive means of detecting individualised risk of future disease progression. This work will provide the basis for tailoring clinical assessment to key factors for early diagnosis for mental health related disease, and developing personalised health tools for screening and intervention.
The aim is to develop predictive models based on machine learning approaches, to differentiate asymptomatic populations (those not presenting symptoms) at high versus low risk of mental health related disease. These models will be used to interrogate the neurocognitive factors that underlie cognitive health.
The proposed work has the following key objectives:
- Implement and validate machine learning models, capitalising on existing datasets, that can identify individualised profiles of cognitive health. These models should be biologically interpretable in order to determine the key factors that best determine mental health, and how they interplay.
- Extend these predictive models to compare metrics of interacting predictors across repeated measurements on the same individuals over time. Further to this, multi-modal deep learning architectures will be implemented that allow for cross-modal dataflow between feature extractors, extracting more interpretable features than uni-modal learning for the same amount of training data. In particular, Graph Attention Networks will be tested which can allow for very significant improvements on potentially sparse and small datasets from repeated measurements.
- Test the validity of this approach by training the algorithms produced with 'privileged' information from highly diagnostic data (i.e. brain imaging data) from a small sample, and predict mental health (i.e. high versus low risk) on an independent sample using only routine clinical data. Enhanced model predictions with privileged data will validate the proposed models as powerful diagnostic tools with potential applications in clinical practice.
The proposed work has the potential to deliver low-cost, robust, and sensitive tools for detecting individualised risk of mental health disease, which are scalable for widespread implementation in clinical practice.
The work aims to develop robust models that predict disease progression based on routine clinical testing (e.g. demographic questionnaires and cognitive testing) which is less costly and more applicable to most patients than MRI scans are. If implemented in nation-wide computing platforms that have the capacity to handle large-scale data across health centres, these models could provide low cost tools for precision diagnosis.
Further, the proposed project will provide the basis for developing a targeted data collection scheme based on critical factors for diagnosis, as revealed by machine learning modelling. These tools will allow for tailored clinical assessment of key factors for early diagnosis, making testing more efficient and cost effective, while less invasive and time consuming.
The longer-term goal is to integrate mobile technology and online data analytics methods with feedback to individuals about their personal health and changes across their lifespan. The project researchers are already working with industrial partners towards the development of such mobile applications for personalised diagnosis. Similar technology could be used to develop computerised decision-support systems that will inform clinicians about personalised health when making decisions about diagnosis and treatment.
March 2018: Project received seed funding from The Alan Turing Institute.