Our Research

The vision of precision medicine is to diagnose patients more accurately and treat them more effectively, taking into account their individual genomic, lifestyle, and environmental factors. Machine learning will play a significant role in implementing this vision, using the capability of machine learning methods to learn from very complex training datasets and to produce consistent, repeatable predictions (diagnoses, prognoses, treatment recommendations) for test cases. My long-term objective is to improve the accuracy and trustworthiness of machine learning methods for precision medicine. Contributing toward this long-term objective, my short-term objectives are developing transfer learning methods and causal models for biomedical data. 

Transfer learning for biomedical data

Biomedical datasets tend to be small (e.g., a few hundred patients) and high-dimensional (e.g., attributes for 30,000 genes). On one hand, neural networks are promising for such applications due to their capacity to reduce high-dimensional inputs to lower-dimensional latent representations and to learn very complex functions. On the other hand, neural networks may overfit to small training datasets. Fortunately, in the biomedical domain, many large datasets are publicly available. We will investigate methods to transfer knowledge from large source domain datasets to small target domain datasets to improve the prediction accuracy on the target domain. 

Causal models for biomedical data

Most machine learning models exploit correlations between the input and output variables. Biomedical applications, however, require models that not only make accurate predictions but are based on causal relationships since scientists and physicians want to understand the underlying mechanisms. To address this need, we will explore causal models: (1) causal discovery (i.e., the discovery of structural causal models) and (2) causal inference. 

There are multiple connections between transfer learning and causal models for biomedical data. On one hand, causal relationships are likely to reflect underlying mechanisms of data generation, and causal models are therefore expected to transfer better between domains than correlational models. On the other hand, causal inference methods reduce selection bias from the domain of factual outcomes to the domain of counterfactual outcomes.

The primary research areas we are exploring are:

1. Universal domain adaptation and domain generalization

2. Calibrating machine learning models for out-of-distribution data 

3. Causal models for transfer learning

4. Causal inference

Real-world applications that have and continue to motivate our machine learning methodology research include precision medicine and precision agriculture.  We have long-term collaborations with the BC Children's Hospital, the Vancouver Prostate Centre (a leading cancer research centre), and Terramera Inc. (a successful startup company in the field of precision agriculture). 

There are different scenarios for transfer learning, i.e., domain adaptation (labelled data from one source domain, unlabelled target data) and domain generalization (no target data, but multiple source domains). The most general task of domain adaptation is universal domain adaptation, where both the source and target domains can have private classes that do not appear in the other domain. The main challenge in this setting is that only representations of the (apriori unknown) shared classes should be aligned. Universal domain adaptation methods can deal with new classes (not seen in the source domain) by classifying them as "unknown." However, there may be multiple new target classes. We investigate how to distinguish the different new target classes using clustering techniques. For domain generalization, our key question is how to learn lower-dimensional latent feature representations from the source domains such that the learned features can be transferred to unseen target domains. 

Calibrated machine learning models produce confidences that match the accuracy of their prediction. Calibration is especially important for safety-critical applications such as those in the biomedical domain. Most of the existing calibration methods address the in-distribution scenario, where the test data is from the same distribution as the training data. Calibrating models for out-of-distribution test data is a more challenging and relevant problem, as test data in the real world are likely from different distributions. We investigate how to calibrate models in the out-of-distribution setting. 

The goal of causal discovery is to discover causal relationships from observational data. Causal models are expected to transfer better between domains than models based on correlations, since they capture mechanisms of data generation and are more robust to spurious correlations. We will use Structural Causal Models, which use Directed Acyclic Graphs among the variables as the backbone, to represent causal models. We will use state-of-the-art causal discovery methods to learn the causal models from small, high-dimensional data. We investigate how to improve transfer learning using causal models in the case where some labelled target data is available and also in the more general case where the target data is unlabelled (i.e., domain adaptation). 

The goal of causal inference is to estimate treatment effects from observational data, and the potential outcomes framework is one of the most prominent approaches. A major challenge is the existence of confounders, affecting both the treatment and the outcome, which leads to selection bias and covariate shift between the different treatment groups. Some methods are grounded in the sufficiency of the propensity score for average treatment effect estimation. Other methods aim at learning a balanced representation which reduces the covariate shift while maintaining the capacity to predict the potential outcomes. Individual treatment effect estimation methods pursue this approach for binary treatments. Continuous treatments are more challenging since it is impossible to create one outcome prediction subnetwork for each treatment, as in the binary setting. 

We will investigate methods for causal inference that will also be employed in our causal models to estimate the strength of relationships between a variable and its parents. Compared to binary treatment effects for which learning unbiased representations has been widely studied, the more practical yet complicated continuous treatment setting has remained under-explored. We will adopt an information-theoretic approach to estimate the treatment effects in continuous single- and multi-treatment settings.