Learning from incomplete data by combining physiological knowledge and machine learning

点击翻译按钮获取中文摘要

Data analytics is playing an increasingly critical role in our decision making. However, data-based prediction is often challenged by missing data. Machine learning is well suited for analyzing massive data, but its accuracy can be hampered by the quality of the data. This problem is particularly critical in health or environmental records, which are notoriously incomplete and erroneous. Errors in the data fed to a machine learning model may yield costly or even lethal mistakes in its prediction. Given this knowledge gap, this project seeks to develop a method that maximizes the amount of information that can be gleamed from a physiological dataset with missing data. To accomplish that goal, we will first develop an innovative data imputation method that is based on known physiology. Specifically, the method will be based on HumMod, a state-of-the-art computational model of human physiology. HumMod is presently formulated for a middle-age man only; hence, in Objective 1 we will extend HumMod to take into account sex and age, by developing instantiations of the model for a middle-age woman, an older man, and an older woman. In Objective 2, we will apply the sex- and age-specific HumMod models to clean data provided by our partners, and then apply machine learning analysis to predict individual health status. The accuracy of the prediction will be compared with analogous predictions made by the same machine learning model but applied to datasets cleaned by other data imputation methods. We expect our physiologically-based data imputation method to out-perform alternative methods, many of which can introduce bias or are sensitive to outliers. The impact of this project will be greatly strengthened by the participation of industry partner AstraZeneca Canada, who will provide cash contribution, and by non-profit Diabetes Action Canada, who will provide access to a large database.