Statistical Learning Techniques on a High-dimensional, Richly Structured Feature Space To Predict Orthodontic Treatment Outcomes and Optimize Treatment Parameters

• Activity Number: 522
• Type: Contributed
• Date/Time: Thursday, August 11, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Christopher Overton*+ and Michael Zakharevich and Xiaorong Chen
• Companies: Align Technology and Align Technology and Align Technology
• Keywords: clinical outcomes ; computational geometry ; dimension reduction ; clustering ; predictive modeling ; statistical learning

Abstract

We survey Align Technology's program for predictive modeling of clinical outcomes, necessitated by our nontraditional approach of specifying entire sequences of orthodontic treatment steps at once. We benefit from 10^6's of consistent patient records, each with 10^5's of variables. Challenges include unstructured and missing data (e.g., outcomes and varying supplementary techniques.) We compare benefits of statistical learning techniques and discuss the tight optimization problem of balancing optimal expected treatment outcome with length/inconvenience of treatment. Our research paradigm combines iterative rounds of several components: detection of features from 3-D patient scans, most of sufficient statistical power only cumulatively; parsing of clinical text; building of low-variance clusters by case type and clinician; deriving from low-level features a critical threshold of predictors to test both specialist dogma and new hypotheses; dimensionality reduction---an emerging technique to locate more densely populated regions of high-dimensional feature spaces; and current machine learning techniques to model outcome probabilities within each cluster.