Abstract
Evidence of neuroanatomical abnormalities in subjects with a recent first episode of psychosis (FEP) has been heterogeneous, possibly due to the increased risk of false positives and heterogeneous findings associated with small samples that dominate the literature. In addition, the clinical impact of such findings has been limited. Machine learning promises to overcome this limitation, however, initial attempts to identify FEP have yielded inconsistent results. Within this movement, deep learning has recently emerged as a promising approach in areas such as visual and speech recognition, as well as other areas of medicine. Its ability to capture highly abstract and complex interactions may be useful to capture the characteristic subtle and widespread neuroanatomical changes of FEP.The overarching aim of this doctoral thesis was to investigate neuroanatomical abnormalities in FEP at the individual level in a mega-analytic approach. Structural Magnetic Resonance Imaging (sMRI) data was collated from five independent studies, totalling 1074 participants. FEP and healthy controls (HC) were first compared using voxel-based morphometry in a large-scale megaanalysis. This was followed by a thorough review of the current evidence for deep learning in psychiatric and neurologic neuroimaging. A deep neural network, along with other wellestablished methods for comparison, were then used to classify FEP and HC at each site separately to test for the reproducibility of findings. Finally, a deep neural network was used to classify the two groups in a large-scale mega-analysis.
Collectively, results revealed a pattern of fronto-temporal-insular changes identified both at group and individual level. Deep neural networks performed better than traditional machine learning approaches, albeit by a small margin. However, performances were lower than expected overall, ranging between 50 and 70%. Upon interpreting these results, I was able to show evidence for publication bias, suggesting that initial studies may have been over-optimistic. Consist with this, the large-scale deep learning analysis suggested that the reliable classification of FEP based on neuroanatomical data may be around 60%. In light of these results, future studies should continue the pursuit for larger samples combined with multimodal approaches to build more reliable and informative models.
Date of Award | 1 Sept 2020 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Andrea Mechelli (Supervisor), Stefania Tognin (Supervisor) & Isabel Valli (Supervisor) |