Summer@Brown

"What course are you taking?"

That was probably the most commonly asked question at Brown this summer, usually spoken between two students who arrived from miles away, hoping to make new friends on campus for the duration of their two-week stay. Most people had easy answers: "Tropical Marine Biology. What about you?""I'm doing Cancer Informatics."

They'd smile, nod, and start asking questions about each other's class.

For me, it usually went a bit differently.

"What course are you taking?"

Deep breath. "I'm taking 'Building Biosensor Software for Public Health Research.'"

Then would surface the questioning smile or the "what?"

Before experiencing the class, I would have been just as confused as they were. I hardly knew what a biosensor was, let alone how to create software for it. I definitely wasn't expecting that by the end of my stay, I would have learned a completely new branch of Python programming and been able to use it smoothly enough to transform real raw data into something ready for interpretation.

For the first couple of days, we spent most of our 3-hour long class refreshing our knowledge of Python data types. We covered the basics––strings, integers, Booleans––but also the more complex ones like dictionaries and tuples. I ended up coding multiple advanced data analysis algorithms for homework, a task that many of my classmates teamed up on due to its length. Those homework assignments were no easy feat.

Later, we started diving into the medical background of the course. To clarify, a biosensor is a measuring device that takes in data at repeated intervals, usually to gather information regarding a health concern––some examples being a step or heart rate tracker. The ones we worked with specialized in alcohol. They measure a person's blood alcohol level in ug/L, tasked with sending out warning messages if levels get too high. Eventually, the goal for these biosensors is not only to help those struggling with drinking problems regulate themselves, but also determine if it is safe for them to participate in actions like driving.

For most of the second week, we learned how to take in real biosensor data and clean it of gaps, non-wear, and artifacts (unusual spikes) using pandas dataframes, a data type, similar to but more flexible than arrays, from the pandas Python library. Learning how to use the dataframes was without a doubt the most difficult part of the course––I had never yet worked with them and was initially unfamiliar with their functions and capabilities. However, by reviewing class code, watching informative videos, and jotting down notes, I was able to grasp the code enough to manipulate the data sets given to us in class.

Before we knew it, the learning phase of our course was over and it was time to complete our final project. The task was to choose between six real biosensor data sets, clean the data completely of messiness, and provide a final analysis of what the data showed us, whether it was reliable, and any warning messages that might be sent out to the user. I chose a data set that had a significant artifact in need of fixing.

My program code had three main functions: Check if there were any accuracy issues (ie. gaps, non-wear, or artifacts), fix such errors, and provide a final analysis of the data, including rise/fall durations, rates, averages, etc.. The first function tested false for gaps and non-wear, so all that remained was to tweak the artifact. Since the spike covered multiple time intervals, I decided to cut it out completely and treat it like a gap. First, I erased the data in the time interval encapsulating the artifact. Next, I located and calculated the slope between the points before and after the artifact. Finally, I filled in the interval with a line appropriate for the endpoints and slope. The pictures below display the unedited data, the refined version, and my final description, respectively.

*TAC stands for Transdermal Alcohol Concentration

Although the class was centered around programming and analysis, we also spent notable periods of class time discussing the more business and experimental-related aspects of the biosensor. In the testing stage, for example, researchers had users log the type and amount of alcohol they ingested, but a serious concern was whether the information was accurate. We debated the implementation of different systems like photo-trackers to account for such inaccuracy concerns. We also considered the design of the biosensors––usually taking the form of a wristband––against societal stigma. That is to say, how could the company make users feel comfortable wearing the trackers in a manner where it wouldn't attract unwanted attention?

After two weeks of answering questions like those, cleaning data, and describing results, I'd like to say that I am a proud expert in "Building Biosensor Software for Public Health Research."