Research

Accelerometry, NHANES, neuroimaging, machine learning, and reproducible reporting.

A few threads run through my work: get messy biomedical data into a form people can analyze, write open-source software so others can do the same, and check that the methods hold up. Below are the projects I’m working on now, the papers I’d point someone to first, and the repositories I find most useful. Everything else is on the Software and Publications pages and on GitHub.

Accelerometry & wearables

Most of my current work is on wrist-worn accelerometers: turning raw, sub-second data into steps, sleep, non-wear, and activity measures that can be compared across devices and studies. With the JHU Wearables group I maintain a family of R packages, collected in activerse, that read data from many devices and run open-source algorithms, such as the Oxford step-count model, directly from R. A big part of this is benchmarking: we compare step-counting algorithms on publicly available data and in older adults in the ARIC study.

NHANES & population cohorts

NHANES is a nationally representative US sample with wrist accelerometry, and a lot of my time goes into making it usable. The 2011–2014 waves and the National Youth Fitness Survey released raw 80 Hz data, which is large and awkward to work with; nhanes_80hz is the pipeline we use to process it. We have used these data to compare step-counting algorithms across the US population and to study physical activity as a predictor of mortality, alongside similar work in the UK Biobank.

Neuroimaging

Much of my earlier work was on CT imaging of stroke: extracting the brain from head CT and segmenting intracerebral hemorrhage. Along the way I built R infrastructure for medical imaging, including Neuroconductor, a platform for imaging packages in R, and interfaces such as fslr and freesurfer that let you call the standard imaging tools from R. I still maintain these, and the CT work produced open methods and data, including a public high-resolution CT brain template.

Machine learning & AI

Machine learning runs through all of this. For hemorrhage on CT we went from random-forest classification of each voxel to 3D deep neural networks, and then to Bayesian deep learning that outperformed the volume estimators used in clinical trials. In wearables, many of the best step, activity, and sleep models are deep-learning models written in Python; I wrap them in R so they can be run on large cohorts and compared fairly. I also care about how models are evaluated: ROC and AUC can mislead when the predictor is binary. More recently I’ve written opendataloader to turn PDFs into structured text from R.

Reproducible reporting

I want analyses that can be rerun and teaching materials that rebuild themselves. ari, the Automated R Instructor, turns slides and a script into narrated lecture videos, so a course can be regenerated when the material changes, and didactr builds on it to automate whole courses. Most of my papers ship with public code and data, and this site is itself built with Quarto from plain data files.