Principal Investigator: Dr Stuart Reynolds
Physiosigns, Inc., 800 West California Avenue, Suite 200, Sunnyvale, California, 94086, United StatesTags: 25687, diabetes, heart disease, Hypertension, stroke
1a: We will use advanced machine learning techniques to build models that predict the likelihoods of a variety of chronic diseases and their co-morbidities (diabetes, heart disease, hypertension, stroke).
1b: The predictive models we are building can be used for early and accurate detection of the elevated risk of a variety of chronic diseases. Early detection of the elevated risk of such conditions occurring within a year or two of detection can trigger preventative measures in a given patient that may delay or even completely avoid the onset of said condition. The public health implications of this, if successful, should be significant. The UK Biobank Resource is nearly unique in its ability to support our research. We intend to publish our results.
1c: Physiosigns will use machine learning software to build models and perform tests to assess their reliability. Our team has been developing and applying machine learning technology to large data application areas for twenty years. Our core learning framework allows a variety of modelling techniques, such as logistic regression and neural network-based deep learning to be applied to thousands of unique partial data contexts (data subsets chosen automatically with online learning). The resulting individual models are combined with techniques drawn from mixtures of experts, typically resulting in predictions that are more accurate and robust than single predictor alternatives.
1d: We would like to use the full cohort – the maximum number available. Our methods can consider more risk factors and diseases with lower prevalence as the amount of data increases.
We would like to request a few additional data items:
Category 1010: Acceleration intensity distribution
All fields: 90092-90158
Category 1008: Physical activity measurement – Additional exposures
Bulk fields: 90001, 90004
Our reason: after reviewing the accelerometer data that we currently
have access to, we believe it to be insufficient to determine peak
intermittent activity, duration of moderate ‘bursty’ activity, and
other such metrics from the data. We would like to generate as many
activity-related features as possible, to give our machine learning
algorithms more to work with.
Last updated Jul 5, 2017