Predictive analytics methods for rare outcomes prediction
Regression analysis and machine learning models are commonly used for predicting health outcomes in clinical and population-based studies. But these predictive models are typically less accurate in correctly predicting rare outcomes, often resulting in poor sensitivity and high specificity. Conventional approaches for handling this problem includes data sampling methods, penalized approaches, and use of alternative accuracy metrics. This study investigates the predictive accuracy of regression and machine learning models based on resampling and cost-sensitive methods for predicting rare outcomes.Using a combination of Monte Carlo methods and population-based data from stroke and epilepsy cohorts, we examine approaches for improving the accuracy of prediction models (wrt to sensitivity and specificity) compromising data quality.
Enhancing Treatment Decisions by Integrating Patient Data with Clinical Expert Knowledge
Prognostic risk scores such as TIMI and GRACE are commonly used for risk stratification in the management of coronary artery disease (CAD). However, these scores are known to over-estimate patient-specific risk of adverse outcomes because they fail to account for other important risk factors (e.g., fragility, commorbidities) that might be modified patient-specific risks. This inaccurate risk stratification has limited the uptake of standardized risk scores in the management of people with complex CAD (.e.g, multivessel CAD) clinical settingsIn this project, we examine how the integration of clinician expert knowledge with patient data might improve the accuracy of clinical risk prediction tools for prognostic purposes in chronic disease populations. Using a combination of elicitation methodology, data analysis, and bayesian prediction models, we develop more accurate prognostic tools for evaluating treatment options in the management of multivessel CAD, We are exploring the application of these methodologies to the diagnosis of dementia in individuals with mild cognitive impairment
Robust classification Models for Multivariate repeated measures design
Multivariate repeated measures (MRM) data, in which multiple outcomes are repeatedly measured at two or more occasions, are increasingly being collected in biological and ecological studies as researchers aim to understand long-term changes in multiple outcomes. These data are inherently complex, often characterized by non-Gaussian continuous distributions, are high-dimensional and have complex correlation structures. Repeated measures discriminant analysis (RMDA) that assume parsimonious covariance and/or mean structures have been recently developed for predicting group differences in repeated measures data. However, these RMDA procedures assume that the data follow a multivariate normal distribution, an assumption which may not be tenable in multivariate repeated measures data, which are usually characterized by non-normal distributions and/or high-dimensional data
This research develop novel robust classifiers based on multivariate generalized models and trimmed estimators for discriminating between population groups in MRM non-normal data
This research is supported by the NSERC Discovery Grant