Booth Id:
CBIO022
Category:
Computational Biology and Bioinformatics
Year:
2023
Finalist Names:
Lukacsy, Bence (School: Barrington High School)
Abstract:
Although colonoscopies are among the most effective methods for detecting colorectal cancer, the CDC reports that approximately 1 in 3 adults forgo their recommended screening. Using machine learning deployed as both a website and API, I created a non-invasive, accurate, and accessible method for assessing colorectal cancer risk.
Collected data included basic health information and cancer diagnosis and was used as input features and labels for the machine learning models, respectively. These models were trained until the change in loss between predicted and expected values stopped approaching zero over ten iterations. The completed models were then subject to various metric analyses and the model with the strongest raw scores was selected as the deployment model. Subsequently, a website and API were created to allow the input of basic health information and the receival of a prediction.
The general accuracy of the model is above 98%, comparable to that of a colonoscopy. The number of false negatives approached 2.5% of the testing data, and for false positives, the number was close to 1%, both of which outperform most screening tests for colorectal cancer, such as Cologuard. All other performance metrics were above 97%, and the website and API are publicly available, accessible, and completely free.
This research demonstrated the possibility of creating an accessible risk-assessment tool for colorectal cancer. The use of machine learning and its deployment allows for increased access to medical recommendations without the need for invasive or expensive screening tests, and could potentially encourage millions of Americans to get their recommended in-person assessment.