Under Review
Function-Valued Topological Features of Brain Networks
Under review. Preprint link: [biorXiv]
Functional MRI can show us how brain regions "activate" in concert during tasks or at rest. Regions are then linked based on how often they co-activate, forming a network of brain regions. We know that adolescence and early development is an important time for the organization of these networks to form & shift.
These changes can be parsimoniously described using topological features of the brain network—but these features are sensitive to the thresholding of the network edge weights. We propose a functional data framework to relate network topology to person-level outcomes, while accounting for varying levels of thresholds & confounding that may arise from individual distributions of edge connectivity.
Accepted
Image Source: Wikimedia Commons
Dynamic Ensemble Prediction of Cognitive Performance in Spaceflight
In Scientific Reports (https://www.nature.com/articles/s41598-022-14456-8).
While in space, astronauts are exposed to
microgravity,
isolation and confinement,
environmental hazards (e.g., space radiation), and
occupational stressors (e.g., spacewalks).
These can affect their ability to get adequate sleep & remain alert, which is critical to mission success.
In this paper, we identify predictors of neurobehavioral alertness over the course of a 6-month spaceflight mission, using self-reported, cognitive, and environmental time series data collected from 24 astronauts on the International Space Station.
Given time-varying and discordantly-measured environmental, operational, and psychological measurements collected aboard the International Space Station, we propose an ensemble prediction model to accurately and dynamically predict neurobehavioral alertness at the individual level. The component models of the ensemble allow for flexible forecasting of both nonlinear associations and time-varying effects on predicted alertness.
Our method is broadly applicable to environmental studies where the main goal is accurate, individualized prediction of human behavior involving a mixture of person-level traits and irregularly measured time series.
Conditional Correlation Models with Association Size (CoCoA)
In Biostatistics (https://doi.org/10.1093/biostatistics/kxac032).
There is an inherent trade-off between speed and accuracy: you can either take the time to do things well, or rush and increase your chances of making errors. Speed and accuracy are therefore coupled, where the coupling strength is thought to change under different conditions (e.g., depending on incentives, motivation); in some cases, the two can become decoupled. We ask, what is the effect of attention on speed-accuracy coupling?
To quantify this effect, we conceptualize speed-accuracy coupling in terms of conditional correlations, where the attention and other confounders or variables of interest (i.e., age) are the conditioning variables. We identify and gauge the performance of three parametric and estimating equations-based estimators of conditional correlation, and then propose a novel measure of effect size, which we term the association size. We apply our framework (CoCoA: a Conditional Correlation Model with Association Size) to speed/accuracy data obtained from adolescents participating in a neurocognitive task.
Automated Analysis of Low-Field Brain MRI in Cerebral Malaria
In Biometrics (https://onlinelibrary.wiley.com/doi/10.1111/biom.13708).
Cerebral malaria is a serious complication of malaria infection, with a fatality rate of 15-20% in children, despite optimal treatment. Brain MRI has been useful in showing us how cerebral malaria works, as well as identifying children who have severe brain swelling and are therefore at greater risk of death.
But advanced MRI technology is not uniformly available across the globe—in low resource settings, low-field MRI scanners (which can produce lower-resolution images) are more common, and there is also a shortage of available radiologists to manually interpret MRI scans. These challenges motivate the development of fully automated methods that can assess patients' brain images, augmenting or replacing manual interpretation, while also accommodating reduced image quality.
We develop and validate a biologically and statistically principled method for the statistical image analysis of low resolution, noisy brain MRI. We leverage existing and publicly available, high-quality brain imaging data to identify brain tissue in the images in our low-resolution sample. The resulting volume-, intensity-, and curvature-based image features have excellent prediction performance and is validated on external data. We want the pipeline to be as accessible/reproducible as possible, so our code uses open-source software and publicly available resources, requires only the raw MRI scans as input, and is available on Github.