I develop machine learning systems for healthcare workforce optimization and create statistical methods for experimental design. My research focuses on online learning, hierarchical modeling, and computational approaches to biomedical data analysis.
Statistical learning, computational biology, and data science
My old churn model measured how surprising a user’s silence was against their own history. That’s a real question with a real answer, but it isn’t churn risk: it flagged regulars who took a week off and ignored light users who’d quietly stopped. It passed every backtest for two years; a confused user in a focus group is what caught it. The replacement is the table banks use for late accounts – bin users by how long they’ve been gone and how much they normally use the product, count who came back – with the same empirical Bayes shrinkage as before, pointed at the right quantity this time.
Wen et al. show two forecasters with identical marginal predictions can make arbitrarily different decisions. A disjoint contextual bandit (one model per arm) is exactly their independence forecaster. In conjugate linear models the joint is exact and free, and partial pooling through a shared component gets it back.
Our most-hit endpoint took 200 milliseconds per shift, so someone capped it at one page of results. The cap ran for years and quietly made product decisions nobody agreed to: an eleven-item shelf, a two-second page, and features that never became ideas. Then a blocked MySQL upgrade forced us to finally read the function.
If you’re running a Bayesian model in a non-stationary environment, you need to forget old data. The obvious approach – scale the precision matrix by a constant – has a failure mode called covariance windup. This post works through three forgetting rules, ending with one borrowed from adaptive control that dominates the others.
My wife has a blog that I self-host. WordPress-based, so it’s MySQL, PHP-FPM, nginx. I’d like to think about its deployment/infrastructure approximately never.
Research in computational biology, statistical methods, and neuroscience
Neurons communicate with each other through electrochemical transmission at synapses. Microglia, the resident immune cells of the central nervous system, modulate this communication through a variety of contact-dependent and -independent means. Microglial secretion of active sialidase enzymes upon exposure to inflammatory stimuli is one unexplored mechanism of modulation. Recent work from our lab showed that treatment of neurons with bacterial sialidases disrupts neuronal network connectivity. Here, we find that activated microglia secrete neuraminidase-3 (Neu3) associated with fusogenic extracellular vesicles. Furthermore, we show that Neu3 mediates contact-independent disruption of neuronal network synchronicity through neuronal glycocalyx remodeling. We observe that NEU3 is transcriptionally upregulated upon exposure to inflammatory stimuli and that a genetic knockout of NEU3 abrogates the sialidase activity of inflammatory microglial secretions. Moreover, we demonstrate that Neu3 is associated with a subpopulation of extracellular vesicles, possibly exosomes, that are secreted by microglia upon inflammatory insult. Finally, we demonstrate that Neu3 is necessary and sufficient to both desialylate neurons and decrease neuronal network connectivity. These results implicate Neu3 in remodeling of the glycocalyx leading to aberrant network-level activity of neurons, with implications in neuroinflammatory diseases such as Parkinson's disease and Alzheimer's disease.
While hierarchical experimental designs are near-ubiquitous in neuroscience and biomedical research, researchers often do not take the structure of their datasets into account while performing statistical hypothesis tests. We present Hierarch, a Python package for analyzing nested experimental designs. Using a combination of permutation resampling and bootstrap aggregation, Hierarch can be used to perform hypothesis tests that maintain nominal Type I error rates and generate confidence intervals that maintain the nominal coverage probability without making distributional assumptions about the dataset of interest.
We have designed, synthesized, and applied a rhodol-based chromophore to a molecular wire-based platform for voltage sensing to achieve fast, sensitive, and bright voltage sensing using two-photon (2P) illumination. Rhodol VoltageFluor-5 (RVF5) is a voltage-sensitive dye with improved 2P cross-section for use in thick tissue or brain samples. RVF5 features a dichlororhodol core with pyrrolidyl substitution at the nitrogen center. In mammalian cells under one-photon (1P) illumination, RVF5 demonstrates high voltage sensitivity (28% ΔF/F per 100 mV) and improved photostability relative to first-generation voltage sensors. This photostability enables multisite optical recordings from neurons lacking tuberous sclerosis complex 1, Tsc1, in a mouse model of genetic epilepsy. Using RVF5, we show that Tsc1 KO neurons exhibit increased activity relative to wild-type neurons and additionally show that the proportion of active neurons in the network increases with the loss of Tsc1. The high photostability and voltage sensitivity of RVF5 is recapitulated under 2P illumination. Finally, the ability to chemically tune the 2P absorption profile through the use of rhodol scaffolds affords the unique opportunity to image neuronal voltage changes in acutely prepared mouse brain slices using 2P illumination. Stimulation of the mouse hippocampus evoked spiking activity that was readily discerned with bath-applied RVF5, demonstrating the utility of RVF5 and molecular wire-based voltage sensors with 2P-optimized fluorophores for imaging voltage in intact brain tissue.