My name is Natasha Rozario

Machine Learning Paper

In December 2020 I published a paper in the peer-reviewed Canadian Journal of Surgery entitled "Can machine learning optimize the efficiency of the operating room in the era of COVID-19?" I had read in the newspaper about the backlog of elective surgery cases in Ontario as a result of the pandemic. Upon discussing the issue with my father, a surgeon, I learned that the average procedure time of the last 10 surgical procedures is used to schedule surgery. As a machine learning enthusiast I knew there was a better way to leverage historical data to optimize OR efficiency.

Using OR-tools, an open-source software suite developed by Google AI, I created a python script that inputs large volumes of historical operative booking data from a spreadsheet and calculates new booking times to reduce overtime and undertime frequencies. With 36 months of data from 15 surgeons from multiple divisions, we demonstrated a theoretical 21% reduction in overtime, 19% reduction in undertime, and a cost-savings of $469,000 over 3 years. We cowrote a paper discussing our findings and the potential for more machine applications in medicine.

This project was a passion of mine as I truly believe that my calling as a computer scientist is to create solutions in medicine and innovate in this intersection of fields. Solutions that enable us to better utilize our limited and precious resources are crucial in a future where we face system-level challenges. I believe there is tremendous potential for impact when it comes to technology applications in health care.

Related articles:
• Waterloo Stories
• Waterloo News
• Waterloo Computer Science
• Waterloo Math
• The Record
• Education News Сanada
• Technology.org

The code for the analysis can be viewed here

Speech Analysis for Heart Failure

While pursuing my Master in Health Informatics in 2023, I’ve been working at the Centre of Digital Therapeutics at the University Health Network in Toronto as a research analyst. I’ve been very fortunate to find a role where I can apply both my computer science background and interest in health applications of technology. 

During my time here, I have been investigating if it is possible to analyse speech data to predict decompensation in people with chronic heart failure. One of the symptoms of worsening heart failure is oedema, where fluid builds up in the tissues. This is usually detected by observing a change in the patient’s weight, however this is not a very sensitive measure and is only effective at detecting decompensation not predicting it.

However edema would also result in vocal cords becoming heavier and stiffer, resulting in changes to the patient’s voice. These changes are more sensitive to small changes in heart failure condition and preliminary research suggests analysing voice markers in short speech samples as short as 30 seconds can be effective at predicting hospitalisations before they occur.

My work with this project involved developing an iOS app for daily collection of data, and developing a methodology for preprocessing and enhancing the speech data. This step is especially important since recordings will be done at the patients’ homes and there is limited control over the environmental conditions. 

I used a pretrained neural network in PyTorch to remove the background noise from the speech data (architecture is based on this paper) and the difference is so clear!

Before:

After:

The paper with our results is still in progress, but we won second place at the Transform HF Ideathon; I’m so excited to see where the work goes from here!

  The final product can be viewed here.

Space Comp

Another term goes by and I have completed yet another project with the Coffee 'N Code club. This time, I created a bot using Javascript that competes with other bots in a physics-based war game.

I was really excited to code more in javascript, as my experience with the language at the time consisted of the few js functions I had implemented on this website and the VS Code extension I built with Node.js during my last co-op.

All the instructions and resources to build the game is available on Github here.

Gooseshooter

As Director of Technology of the UW Virtual Reality Club, we accomplished a lot in terms of growth with our new website and many new members. I particularly enjoyed the workshop that I designed and hosted at the end of the term.

It was the first project-building workshop that we had hosted and as a strong proponent of learning by doing I truly believed that this would be a good introduction to game building.

Together we built a markerless AR goose shooter game using Unity that can be run on any iOS or android device (get Mr. Goose!).    

All the lesson materials and resources can be found here.

Reddit Bot

 So the Coffee and Code club held a 7 part workshop on how to build a Reddit bot that searches through comments and looks for cyberbullying which was super cool since it was a mind opener to social responsibility.

We went through a few examples such as how a spam detector might detect spam from keywords in emails such as "buy", "free", or "sale".

We then went through an example of getting the program to interact with Reddit. We got 100 comments from the UWaterloo and the UofT subreddits and used these comments to train the program into recognising if a comment was from UWaterloo or UofT. After training the program with 5 comments from each thread, the model accurately sorted new comments a little more than 50% of the time.

Finally, we set out to actually build the Reddit bot to detect cyberbullying. We learned how to process the incoming comments by removing non-alphanumeric characters and stemming each word (for example, the stem of 'going' is 'go'). Using a naive Bayes classifier and a Hate Speech and Offensive Language database the model classifies comments as cyberbullying or not. The end result always performed with an accuracy of more than 70%. 

Pictures and screenshots can be found here.

Catalyst 2017: Business of Science Pitch

 So I just came back from the Catalyst Engineering summer programme where my group won 1st place in the Business of Science Pitch Competition! In the competition, we had to create an innovative product and pitch it in front of four judges at the Communitech building. For more information about the exact nature of the competition, click here.

So my group created a product called the Personal Alert Monitor, or PAM for short. It's a wearable tech that is used to contact people if the user's safety is somehow compromised. It can be applied to many different situations and used by a variety of people. For example, an elderly person could contact immediate family if they have fallen down the stairs, or a young adult can contact their friends or family if they are walking alone in the city at night and feel unsafe.

PAM would be connected to the user's phone via bluetooth. When the user wants to activate it, he/she would press a button on the face of the device which connect to the PAM app already downloaded on the user's phone. The user would have already set up a message to send as well as contacts to send it to beforehand.

The device would connect to the user's phone using a Arduino Pro Mini and an Arduino Bluetooth Module. I have built a prototype using the Arduino Uno and Arduino Bluetooth Module.

To show the message being sent from the PAM to my phone, I downloaded Blueterm from the Google Play Store, an open source terminal emulator that can connect to any serial device.