MStat Seminar

In 2014, a particularly virulent strain of enterovirus, D68 (referred to as EV-D68) was responsible for an outbreak of severe respiratory illness in children, with 1,153 EV-D68 cases reported across 49 states. Some of these cases of severe illness even resulted in acute flaccid myelitis (or AFM). Despite this, there is no commercial assay for its detection in routine clinical care. Using the BioFire RP (which was not explicitly designed to differentiate EV-D68 from other picornaviruses), we formulate a model to distinguish EV-D68 from other human rhinoviruses/enteroviruses (HRV/EV) tested for in the panel. This model incorporates real-time PCR metrics from the RP HRV/EV assays. Six institutions in the United States and Europe contributed to the model creation, providing data from 1,619 samples spanning two years, confirmed by EV-D68 lab developed molecular methods, as inputs to the model. The final model demonstrated an overall sensitivity and specificity of 87.1% and 86.1%, respectively, among this dataset. Using the Trend data we survey for historical evidence of EV-D68 positivity and demonstrate a method for prospective real-time outbreak monitoring within the network. This model was used as a real-time monitoring tool during 2018, alerting the research network of EV-D68 emergence in July. One of the first sites to experience a significant increase, Nationwide Children's Hospital, confirmed the outbreak and implemented EV-D68 testing at the institution in response. Applying the model to historical Syndromic Trends RP data from 2014 to 2018, we find three potential outbreaks with predicted regional EV-D68 rates as high as 47% in 2014, 18% in 2016, and 29% in 2018. Future collaborations with these clinicians hope to determine if EV-D68 is uniquely responsible for AFM.

We'll explain what students are expected to do in the project and how the process usually unfolds. We plan to have a mock defense presented by a previous student, Chanel Roe, so you can see how the defense typically works, and of course there will be time to answer your questions.

Recommender Systems are a core part of many e-commerce platforms. To generate recommendations, practitioners must leverage both behavioral data and content data. This talk will present one method of incorporating flawed content data alongside behavioral data to increase diversity in recommendations. The model employed is a conditional variational autoencoder. The external label used as the condition is learned semi supervised by label propagation. This conditioning enables the use of style transfer, a technique common in computer vision, to be applied to recommendations to diversify the results.