Grant: Discovery Grants Program - Individual

A robust machine learning (ML) model is a model that can maintain its expected performance (e.g., accuracy, sensitivity, etc.) when tested against new data distributions. Designing robust ML models requires understanding challenges related to input distortions, where crafted changes (distortions) in input data can deteriorate the ML models' output. This research program aims to develop training and calibration methods to design decision trees (DTs) and convolutional neural networks (CNNs) that can adapt to input data distribution changes after model development. Other concerning challenges in developing robust ML models include 1) feature stability, 2) data quality, and 3) precision vs. recall, where there is a trade-off between precision and recall during model training. The long-term objective is to address these concerning ML models robustness challenges and discover polynomial time verification algorithms (i.e., ML model's correctness) to verify ML models' robustness and use them as configurable components to build a robust ML framework for critical applications. The short--term objectives of this research program are to develop ML models that are more resilient to input data distribution changes by developing: 1.A DT model training method to handle changes in input data distribution. 2.A verification method to guarantee upper bound error on the DT model. 3.A CNN model calibration method using real-time wearable sensor data. 4.A method to learn from a mix of labelled and unlabeled electronic health record (EHR) data. ML models' robustness enables the applicability of the ML models in critical applications such as healthcare. We will apply the proposed methods to leverage the potential of heterogeneous databases for user benefits, such as analyzing the EHR data to detect a critical health condition. For example, Sudden Cardiac Arrest (SCA) is an unexpected heart condition caused by loss of heart functionality in a person with a related or unknown cardiac disease. SCA presents a sophisticated challenge to the Canadian healthcare system because of its high probability of potential fatality. In Canada, it is estimated that 35,000 SCA occurs every year. Most SCA events happen in public places or at home, and few people survive. This research program will develop methods to address input data distribution challenge and uses these methods as configurable components to build robust ML frameworks for critical applications such as healthcare applications. More specifically, this research program will help identify patients at risk of SCA and reduce the healthcare system's financial burden, thus improving the Canadian economy and quality of life. The trained HQPs will acquire competitive skills transferable to many industrial sectors, increasing their chance of joining several industries or having their startup in healthcare that can help physicians diagnose and determine when patients need medical intervention before they need hospitalization.