The view on Alzheimer’s Disease (AD) diagnosis has shifted towards a more dynamic process in which clinical and pathological markers evolve gradually before diagnostic criteria are met. Given the wide variability in data available per subject, inherent per-person differences, and the slowly changing nature of the disease, accurate prediction of AD progression is a significant, difficult challenge. The goal of this project is to devise novel Personalized Machine Learning Models that can accurately capture future changes in the key biomarkers and cognitive scores related to AD and other neurological conditions. As the basis for our framework, we use the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset–the largest publicly available dataset for AD research. These data are highly heterogeneous and multi-modal, and include imaging (MRI, PET), cognitive scores, CSF biomarkers, genetics, and demographics (e.g. age, gender, race). The developed models are the break-through in machine learning for health-care as they allow personalized forecasting of the diseases' progression - in contrast to the traditional "one-size-fits-all" approaches. This capability is of great importance to both clinicians and those at risk of AD since it is critical to early identification of at-risk subjects, construction of informative clinical trials, and timely detection of AD.
In the Affective Computing group, we have actively been working on the development of personalized machine learning models for future forecasting of AdasCog13 - a significant predictor of Alzheimer’s Disease(AD) in the cognitive domain – over the future 6, 12, 18, and 24 months, using the data of participants in the ADNI database. Specifically, in our latest work to be presented on August 09 in the premiere conference on Machine Learning for Healthcare (ML4HC), we introduced a modeling framework based on Gaussian Processes (GPs) that leverages the notion of "meta-learning" (learning how to learn). This approach learns automatically from previous participants' data what is the best forecasting model to apply to a new participant: the population-level or personalized model. This is important in cases when the participants' data are highly noisy or missing, in which case the population-level GP models are suboptimal. Conversely, when we have a good-quality past data of the target participants, these are used to effectively personalize the target model to a new participant, largely outperforming the population-level model on the future data of that participant. This has important implications for the design of clinical trials and also in gauging decisions of medical practitioners, allowing them to use smart and personalized AI when deciding what treatment to prescribe to their patients (by informing them of potential future outcomes for those patients, based on the medical history of the target patient but also large source of knowledge available from previous/other patients). For more details about this approach, check our paper ("Meta-Weighted Gaussian Process Experts for Personalized Forecasting of AD Cognitive Changes") that is provided below.