Democratising Machine Learning for researchers working in Alzheimer's space

研究总结

RESEARCH SUMMARY One of the key challenges in Alzheimer’s disease (AD) is the early detection of individuals who are at risk of developing the condition, and subsequently making predictions about how rapidly their condition will progress. To facilitate this, the parent grant of this proposal (R01-AG058676-01A1) brings together five leading well char- acterized Alzheimer’s cohorts to clarify risk and protective factors for Alzheimer’s dementia: the Adult Children Study (ACS), the Alzheimer’s Disease Neuroimaging Initiative (ADNI), the Australian Imaging, Biomarkers and Lifestyle Flagship Study of Ageing (AIBL), the Dominantly Inherited Alzheimer Network (DIAN) and the National Alzheimer’s Coordinating Center (NACC). The resulting cohort is critical for understanding the factors which precipitate or delay dementia diagnosis. The large size of this dataset motivates the application of machine learn- ing(ML) and artificial intelligence (AI) methodologies to address the critical questions of classifying individuals as at risk of AD and for making predictions about the future progression in terms of the disease, symptoms and function. However, there are stark contrasts between the AD space and those domains where ML/AI techniques have been used to great success, such as computer vision, text mining and audio analytics. In particular, the smaller sample size, even for the largest datasets collected to date, means that highly expressive models are just as likely to detect technical artefacts in the data as real biological signal, especially when looking at high-dimensional in- formation where AI/ML is most appropriate. Moreover, standard ML/AI approaches do not typically work with missing values, a common feature of clinical AD cohorts, where certain types of measurements are difficult to collect. Finally, the areas where AI/ML have fulfilled their promise are those where the barrier to entry for both ML and domain specialists has been reduced. This democratisation has been achieved largely by providing bench- mark datasets that have been processed to remove technical artefacts, have comprehensive documentation and adhere to under Findable, Accessible, Interoperable, and Reusable (FAIR) data principles. In this grant, we seek to conduct high-resolution, multi-modal harmonisation and cross-dataset imputation of the five leading AD cohorts aiming to improve their suitability for AI/ML approaches, proving these in a form that is FAIR, well-documented and easy to use with AI/ML frameworks. While the parent grant performs some harmoni- sation of summary statistics (e.g. average amyloid levels from PET imaging, cognitive test summary statistics) and imputation using classical approaches, this extension seeks to improve the level of granularity (e.g entire images, individual or small groups of questions in cognitive tests) at which harmonisation is performed, and incorporate biological prior knowledge into data imputation. This will be achieved by leveraging recent advancements in al- gorithmic bias-removal and matrix completion, and will incorporate our understanding of disease processes and the nature of the modalities being analysed. The latter will put additional constraints on the inferential, allowing it to produce more accurate and sensible estimations of measurements. We will provide FAIR-curated version of these datasets, along with software to enable researchers to adjust the level of harmonisation and imputation as needed, based on the strengths and weaknesses of the underlying algorithms developed and implemented in this proposal. The integration of the five largest longitudinal AD cohorts, as per the parent grant of this proposal, provides an invaluable opportunity to explore the power of AI/ML to improve our ability to detect AD early on and make accurate forecasts about individual’s change over time. This proposal seeks to bring modern AI/ML methods firmly into the AD community by producing curated de-biased datasets that can be seamlessly integrated into most existing ML pipelines. By improving the quality of the underlying data at a higher resolution than has been done before, predictive and prognostics models derived from this work are likely to be substantially more powerful than past approaches.