Abstract
Over the decades, neuroscientists have been actively seeking a neuroimaging-based biomarker capable of capturing cognitive abilities. Having the biomarker will provide us with an important tool to fight against psychopathology and neurological disorders. Nevertheless, these endeavours often meet with challenges. A robust neuroimaging-based biomarker should possess good psychometric properties. First, it should show a strong capability to predict the cognitive abilities of out-of-sample individuals not a part of the model-building process (known as predictability). Second, the neuroimaging-based biomarkers should also maintain consistent predictions over time (known as test-retest reliability). Third, the neuroimaging-based biomarkers should also be applicable beyond participants whose brains are scanned from the same protocol (known as cross-cohort generalisability). In this thesis, we used stacked machine learning to develop biomarkers for cognitive abilities from multimodal neuroimaging that possess these three psychometric properties. To demonstrate the benefits of our approach, we conducted three studies. In the first study, we created a stacked machine-learning pipeline that draws information across multiple types of brain MRI to predict cognitive abilities. We found that this stacking led to biomarkers that predicted cognitive abilities with high predictability and high test-retest reliability, especially when compared to a non-stacked machine learning (10.1016/j.neuroimage.2022.119588). In the second study, we examined a widely used biomarker called Brain Age in predicting cognitive abilities. The results demonstrated the statistical advantage of our approach over Brain Age (10.7554/eLife.87297.2). In the third study, we tested the applicability of the stacked machine-learning on three different large-scale datasets to assess its psychometric properties across diverse cohorts. The stacking led to consistently high predictability and high reliability across the three datasets with a moderate level of generalisability. Consequently, our study establishes that the stacked predictions could result in biomarkers that have all the required psychometric properties, solidifying its use for building a robust neuroimaging-based biomarker for cognitive abilities.