Merging equations, ecology, and medicine to inform on human and animal behaviours.
Our collaborative research team (CRT) works on a number of statistical methods to capture behavioural and physiological processes occurring at multiple scales that will help identifying a wider range of behaviors and medical conditions in human patients and animals. See the link of our team here.
You can read fun, non-technical summaries of some of our work: Research Summaries
Our current research projects include:
- Using labelled data to improve state predictions: While most biologging data available to researchers are unlabelled, many researchers have small additional labelled datasets that provide information on some of the true hidden states. Ecologists can identify some behaviours from in situ, video, or acoustic observations of animals, and patients can selfreport mood and be clinically assessed. Such labelled data has been used in combination with the unlabelled data in semisupervised machine learning approaches to help inform parameter estimation and improve predictions. We believe that semisupervised methods could be further developed to improve the analysis of biologging data.
- Improving realism through the incorporation of feedback mechanisms: The capacity of an individual to move in certain manners or their motivations to eat or rest, are driven in part by physiological processes (e.g., energy levels, hunger). In turn, these physiological processes are affected by the movement and behaviours of the individual, resulting in feedback mechanisms. While these physiological processes are rarely observed for animals, some processes, such as heart rate can be easily measured by wearable devices for patient data. However, most HMMs do not account for physiological processes nor their complex feedback with movement. Methodological development of HMMs that model both movement and physiology are necessary across fields, and our CRT we will focus on development and inclusion of physiological conditions and feedback mechanisms in ecology and medical science.
- Accounting for individuality in population inference: Biologging data are typically collected across multiple individuals, and thus provide a unique opportunity for researchers to understand individual differences, and account for them when making inference at the population level. When adequate covariates are available (e.g. age, sex, ambient temperature), HMMs can account for differences among individual time series by incorporating these variables into the state-dependent distributions (e.g. accounting for differences in travel speeds by size) or state switching dynamics (e.g. accounting for difference in the environment experienced). When covariates are insufficient to explain the variation (e.g. fundamental, and unmeasured, differences in phys iology or environmental conditions experienced), researchers have turned to adding random effects for the parameters of the state-dependent and state switching distribution using both a parametric and semi-parametric framework. However, their distributional assumption (e.g., normally distributed) is often overly restrictive and can oversmooth the distribution. Our CRT will explore and develop flexible random effects distributions to improve our capacity to account for individual variability and conduct inference.
- Inference, identifiability, and model validation: Basic HMMs are identifiable when the state-dependent distributions are distinct and the transition probability matrix is full rank and ergodic, and in practice are easy to fit to data when the state distributions are different enough across states and the temporal persistence in behaviour is high. However, HMMs often have complex likelihoods, and subsequently posterior distributions, that can lead to convergence and parameter identifiability problems. The types of model modifications currently applied to HMMs, as well as those we are suggesting, are likely to further complicate the fitting of these models. As such, a priority of our CRT will be to develop guidelines to help practitioners identify when they have sufficient data to estimate the parameters of complex models and the situations when the use of highly informative priors or outside information is required to induce identifiability.