Inspiration
As engineers with multidisciplinary interests, we are excited to leverage data science to support equitable, accessible, and timely care for all, irrespective of medical and socioeconomic context.
What it does
We have built a classifier that is accurately able to predict the survival of AIDS patients given demographic, medical and treatment information.
How we built it
Our baseline includes a logistic regression classifier, utilizing the sigmoid function to predict probabilities and classifies based on a 50-50 decision boundary. We improved on this baseline via a random forest ensemble model that can resist overfitting and generalizes well to new data. We tuned hyperparameters such as class weight, maximum depth, and maximum features to increase viability.
Challenges we ran into
We experimented with a variety of simple and machine-learning-based models to select for highest performance. This included experimentation with simple linear regression, LOWESS smoothing, Emax pharmacological modeling, K-nearest neighbors, logistic regression, decision trees, and random forests. In addition, hyperparameters required extensive fine-tuning for optimal results.
Accomplishments that we're proud of
We are excited to report a 93% AUC score via the random forest ensemble model!
What we learned
We learned how to effectively tune hyperparameters to maximize model performance.
What's next for SurvivAll: An AIDS Early-Detection Support Tool
At Team SurvivAll, we hope to expand our training dataset to include more diverse demographic data, especially nuanced stratifications as opposed to binary groupings; this would be fundamental to improving representation across demographics and supporting equitable AIDS treatment. We also plan to utilize modeling approaches such as XGBoost and survival modeling techniques to better account for time-to-event data.
In practice, our model acts as a data-driven decision-support tool for clinical professionals, helping to identify high-risk patients early.
Log in or sign up for Devpost to join the conversation.