Abstract:
Electronic medical records are used to build adverse event prediction models for resource management. Many machine learning methods struggle with sparse, irregularly-sampled data. Carrying the last value forward or linear regression can interpolate missing values. Imputation and regular time-series resampling are done with Gaussian process (GP) regression. GP model structure, such as a covariance function, may require extensive, ad hoc investigation. Multivariate real-world clinical data with different time-series variables can make this difficult. We use Neural Process (NP) neural latent variable models to estimate missing values in clinical time-series data. The NP model uses a latent space conditional prior distribution to model local data variations to learn global uncertainty. This prior learns during training, unlike conventional generative modelling. NP model adapts to clinical data dynamics. We propose a variant of the NP framework for efficient modeling of latent-input mutual information and meaningful learned priors. The proposed method outperforms conventional methods in MIMIC III experiments.
Note: Please discuss with our team before submitting this abstract to the college. This Abstract or Synopsis varies based on student project requirements.
Did you like this final year project?
To download this project Code with thesis report and project training... Click Here
