Artificial intelligence applications to precision mental health: artificial networks for predicting lifelong brain health

Predicting cognitive decline associated with older age is one of the top research priorities with major ramifications for clinical care and early interventions. Differentiating asymptomatic individuals at high vs. low risk of decline remains a challenge in clinical practice. In this project we will develop novel machine learning methods for predicting cognitive health based on multi-modal and heterogeneous data from healthy individuals. Generative Adversarial Networks (GANs) provide a particular promising route for this [1]. On the technical side, generative models use neural networks to find a parametrised random variable that can explicitly be sampled and that follows an unknown distribution. On the practical side, given N samples (e.g. images of human faces) from this unknown distribution, we aim to create more samples (e.g. new images of human faces) from the same underlying distribution using this parametrized random variable. Since their introduction [1] GANs have had a huge impact into applications, mainly in computer vision and photography where generators have been used to simulate new instances of photographs of faces/objects, being trained on samples from each of these classes. In the context of this project, we will investigate the potential of developing and training GANs that are able to generate -- from samples of an underlying unknown joint distribution of multi-modal data - complementing and diagnostic data (e.g. costly or invasive diagnostic PET scans) with clinical data that can be collected more routinely (e.g. cognitive tests, structural MRI). We will capitalise on existing datasets of healthy individuals that have both types of data (PET, MRI, cognitive testing): UC Berkeley (200 individuals); Alzheimer's disease neuroimaging initiative: ADNI (400 individuals). In order to not rely on a perfect and homogenously sampled dataset, we will start this by focusing on the recently introduced Wasserstein GANs [2] and our work in [3] that uses the Wasserstein GAN concept to learn bespoke (in our case joint) prior distributions trained on possible un-matched training samples (that means not all patients need to exhibit the same collected data). In this way, we are able to test the relevance of collected data (complementing and augmenting different subsets of the data), personalise screening (linking the GAN to a classifier that is trained to delineate individuals into different cohorts), and subsequently make screening and diagnosis more efficient and accurate. These objectives are in line with BBSRC priority areas on i) healthy ageing, ii) computational biology and the Industrial Strategy programme on 'Data to early diagnostics and precision medicine'. The project will provide the PhD candidate with training on interdisciplinary science combining artificial intelligence with brain imaging and clinical translation.