|
I am interested in improving clinical decision making for cancer patients and individuals at high risk for cancer. By matching classes of genomic signatures with classes of clinical outcomes -- times to events such as primary carcinogenesis, metastases, and recurrence -- we can get a better idea of which patients should get what interventions at what time. This problem of linking genomic classification with survival analysis is the goal of my research.
Genomic data is extremely high dimensional, providing tens of thousands of correlated noisy signals per patient sample. Yet patient samples are not nearly as numerous. Classification algorithms trained on such relatively small training sets tend to be noisy. This problem is further compounded if the training sample actually consists of a heterogenous mixture of sub-populations.
I am exploring a combination of approaches to address these challenges. To handle heterogeneity of the training sample, I am minimizing integrated square error, as opposed to maximizing a likelihood, in logistic models. Minimizing L2 distance, while less efficient than likelihood maximization, enjoys robustness to data heterogeneity. To deal with the imbalance between samples and covariate dimension and covariate correlation, I am applying L1 and L2 penalties (the elastic net) to regularize solutions to the integrated square error optimization problem. To deal with the non-convexity of the L2 minimization problem I solve a series of increasingly more accurate convex relaxations.
Proton auroral intensifications and injections at synchronous altitude: E.C. Chi, S.B. Mende, M.C. Fok, and G.D. Reeves,Geophysical Research Letters Vol 33, March 2006.
Different Algorithms for Normal and Protection Paths: R. Gupta, E. Chi, and J. Walrand, Journal of Network and Systems Management 13(1) pp13-33, May 2005.
Proactive Resource Provisioning: E. Chi, M. Fu, and J. Walrand, Computer Communications 27 (12) pp.1174-1182, July 2004.
Sharing Normal Bandwidth During a Failure: R. Gupta, E. Chi, and J. Walrand, Proceedings Seventh INFORMS Telecommunications Conference, Boca Raton, Florida, March 2004.
Different Algorithms for Normal and Protection Paths: R. Gupta, E. Chi, and J. Walrand Proceedings DRCN 2003, Banff, Canada, October 2003.
Proactive Resource Provisioning for Voice over IP: E. Chi, M. Fu, and J. Walrand Proceedings SPECTS 2003, Montreal, Canada, July 2003.
Histopathology of laser skin resurfacing: S. Thomsen, B. Baldwin, E. Chi, J. Ellard, J.A. Schwartz, Proceedings of SPIE Vol 2970:May 1997
|
