Work in Progress seminar by Dongxu Han, discussing Controlling Sensitivity of Gaussian Bayes Predictions based on Eigenvalue Thresholding
21 March 2019
Dongxu Han begin his seminar on Friday the 15th of March, with background information on his research which he studied during his PhD. Referring to how designing a clinical decision support system for doctors could aid patient’s health and speed up the diagnosing process by using new and existing images and data to learn from. The system would be able to provide a greater accuracy of prediction on future illnesses and diseases, this would aid Doctors, cut waiting times down in hospitals and reassure patients on their diagnosis. Dongxu moves on to show examples of the decision score system and how It would calculate the data it was give.
The system would rely on a sensitivity input/output for the function, from this, the system will be able to predict the pattern outcome in health. He combined the well-known PCA (Principle Component Analysis) approach with Bayesian classification approach. Dongxu explains how reducing the dimension can help increase sensitivity. In particular, he showed that small Eigen vectors have less effect on sensitivity therefore one can reduce the number of small eigen vectors. He provided a threshold to be used to cut off the number of small eigen vectors one can reduce in practise. With this, the sensitivity factors for the model relays dimensionality and the eigenvalues/vectors and by raising the dimensionality, this will therefore increase the sensitivity.
He reflects the data he had collected and by removing either small or large added values, this would increase or decrease in sensitivity. The results can be shown below. As For the experimental data set he provided us with, he used 2620 mammography images of breast cancer, both benign and malignant.
He then explains the graphs by referring to the coalition of decreasing large values (blue points) and small values (orange points), adjusting the accuracy and sensitivity of the experimental data set accordingly.
Our work in progress speaker, Dongxu Han, gave us a better understanding of his research and progress in designing a clinical decision support system by controlling sensitivity. This seminar also gave students and staff the opportunity to benefit from the discussion of ‘Controlling Sensitivity of Gaussian Bayes Predictions based on Eigenvalue Thresholding’, possibly incorporating the new knowledge into their own research. For individuals who may be considering a course in computing or for those who may have a general interest, these discussions provide an insight into current students’ research and where it will lead them.
Find out more about our courses in computing.