CMKL Students Use AI to Predict Breast Cancer Recurrence with Bayesian Models
At CMKL University, a recent student-led project is breaking ground in the world of healthcare and machine learning, using Bayesian survival models to predict when breast cancer might return for patients.
Machine Learning Meets Medical Insight
The team — Settawut Chaithurdthum, Sivakorn Phromsiri, and Ittiphat Kijpaisansak — used publicly available clinical data to analyze recurrence timelines in breast cancer patients. Their goal: to compare various survival analysis models and determine which approach provides the most accurate and interpretable results for predicting patient outcomes.
“This isn’t just data crunching,” said team member Sivakorn. “We’re trying to create something useful — a tool that could eventually help doctors and hospitals prepare better.”
The students explored both traditional and modern techniques for analyzing survival data. They studied Weibull AFT, a model that predicts how long something lasts based on specific factors; and the Cox Proportional Hazards model, which looks at how different variables affect the chance of an event happening over time. On the modern side, they examined DeepSurv, a deep learning model that personalizes risk predictions, and Random Survival Forests, an ensemble method that builds multiple decision trees to estimate survival outcomes.
Their standout method? Bayesian frailty models, which allow predictions to evolve over time as new data becomes available — ideal for long-term conditions like cancer.
“The power of the Bayesian method is that it continuously evolves,” explained Dr. Irving, the professor who advised the project. “The more data you feed into it, the more accurate it becomes. Initial assumptions gradually give way to real, data-driven insights.”
What the Model Can Do
The team used data from over 700 breast cancer patients to build models that could predict two things: how soon the cancer might come back and how likely it is to return at different times.
Even with a small dataset, the results were promising. The best performer was DeepSurv, a deep learning model, with a Concordance Index (C-index) of 0.642.
The C-index is a score that tells how well a model predicts outcomes — 0.5 means random guessing, while 1.0 means perfect prediction. So, a score above 0.5 is considered good enough to use, and the higher the number, the better.
The Bayesian Weibull model with Gamma frailty came in close behind with a C-index of 0.628. While it was slightly less accurate than DeepSurv, it's easier for doctors to understand and apply in real-world settings — making it a strong option for clinical use.
“This is especially important in medicine,” the team emphasized. “Doctors don’t just need a prediction — they need to know why it was made and how confident the model is.”
Real-World Potential
While the project is still in its early stages, the team hopes to grow it further. Their next steps include expanding the dataset, refining the Bayesian model, and ultimately sharing their results with hospitals in Thailand.
“If we can scale this up, we’d love to see it implemented at various hospitals,” said Sivakorn. “We’re serious about building something that works.”
The students also noted how similar models have been applied to predict COVID patient outcomes, suggesting a wide range of future applications in hospital management, ICU planning, and personalized care.
