Root Insurance Bidding Strategy Challenge: We want to propose a bidding strategy for online ad placements based on customer demographics to increase sales of our car insurance policies while minimizing cost and obtaining at least 400 policies sold per 10,000 customers.

Fingertip pulse oximeters are the current standard for estimating blood oxygen saturation without a blood draw, both at home and in healthcare settings. However, pulse oximeters overestimate oxygen saturation, often resulting in ‘hidden hypoxemia’: a patient has hypoxemia (dangerously low oxygen saturation), but the oximeter returns a healthy oxygen value. Unfortunately, oximeter overestimation of oxygen saturation is exacerbated for patients with darker skin tones due to light-based oximeter technology. This results in Black patients experiencing hidden hypoxemia at twice the rate of white patients. By combining pulse oximeter readings (SpO2) with additional patient data, we develop improved methods for estimating arterial blood oxygen saturation (SaO2) and identifying Hidden Hypoxemia. The predictions of our models are more accurate than pulse-oximeter readings alone, and remove the systematic racial inequity inherent in the current medical practice of using oximeter readings alone.

Acronym/Abbreviation (short form) disambiguation is one of main challenges when using NLP methods to uderstand medical records. While this topic has long been studied, it is still a work in progress. Current strategies often involve having manually curated datasets of abbreviations and train classifiers. The main problem of that approach is that curated datasets are sparse and don't include all the short forms. In Dec 2020, a paper came out where they created a large dataset of short forms as one of their steps in their pipeline to pre-train models. The goal of our project would be to build upon their short form disambiguation piece and create a tool to disambiguate a medical short form using its context. Example of the usage of our tool: original_sentence = "The patient states that she has had dizziness, nausea, some heartburn, and some change in her vision. She is gravida 6, para 4, AB 2. She has no history of adverse reaction to anesthesia." AB could stand for "abortion", "ankle-brachial", "blood group in ABO system", "A, B lines in Kerley lines". disambiguated_sentence = "The patient states that she has had dizziness, nausea, some heartburn, and some change in her vision. She is gravida 6, para 4, abortion 2. She has no history of adverse reaction to anesthesia."

We built a podcast recommendation engine that suggests episodes to a listener based on either a previous episode that they've heard or an episode description that they can input with freeform text entry.

We built a movie finder app that allows a user to enter some details they remember about a movie (along with some optional filter info on the genre and release year) and then predicts what movie the user is thinking of. To solve this NLP problem, our tool uses an embed-and-rerank model. We have precomputed vectorizations of movie plot information for the approximately 34,000 movies in our dataset.

Our model’s first step is to vectorize the user’s query and do a fast comparison to find the 100 closest plot vectors. Then it reranks these top 100 closest plots, performing a more thorough comparison using a neural network that semantically compares the plot fragments with the original query. Finally, we output the 10 movies which show up at the top of this new ranking.