Machine-learning-based study for mobility measures as preliminary predictor for Alzheimer's disease risks and cognitive impairments
Approved Research ID: 96841
Approval date: January 12th 2023
80% of all recorded cases of dementia are due to AD, making it the most common cause of dementia and representing an annual global cost of $250 billion for its treatment and care. No curative treatment has been found to this day and those that slowdown the disease need to be administered as early as possible to maximize the benefits. It has been shown that structural brain changes due to AD occur 20 to 30 years before cognitive decline. The early diagnosis of this disease is therefore essential to allow the best patient support.
However, current screening-methods such as Cerebrospinal fluid analysis or Amyloid-PET scans are highly invasive. Additionally, imaging tests (i.e., PET, MRI) require expensive infrastructure (e.g., a single Amyloid-PET scan costs more than 1500$). Neuropsychological tests require a long time to conduct (up to 2 hours) and show practice effect, making it challenging to screen the population on a large scale. It would be of great value to identify low-cost, non-invasive and accessible biomarkers in order to better diagnose the disease.
Wearable technologies offer a suitable solution to this aim, allowing passive and continuous recording of data that can be investigated for AD biomarkers. Previous longitudinal studies have shown a significant reduction in mobility in mild NCD due to AD 12 years before clinical diagnosis. Neuroanatomically speaking, early AD is characterized by a deterioration in the medial temporal lobe first, which has been demonstrated to be important in spatial navigation.
So far, brain structures and cognitive functions have been found to be associated with physical activity on the one hand and polygenic risk on the other based on the UK Biobank data and on independent clinical datasets. Whether physical activity can be used as a preliminary predictor of AD markers in Machine Learning frameworks has not yet been demonstrated.
In this study we aim to investigate mobility as a biomarker for AD risks by using acceleration recordings together with clinical data to predict polygenic risks, brain structures parameters and cognitive performances using Machine learning algorithms. We will investigate through data-driven features extraction methods and unsupervised-feature extraction ones how this parameter can be used as preliminary predictor of AD risks independently or combined with clinical data. The models obtained during this study will then be transferred and generalized to other clinical datasets. Conclusive results would allow the implementation of new methods for pre-screening of AD.