Program Information
Cross-Decomposition Methods for Learning Clinically Relevant Measures of Patient Similarity in Oncology
C Berlind1*, C Ahern1 , W Lindsay1 , P Gabriel2 , C Simone3 , (1) Oncora Medical, Inc., Philadelphia, PA, (2) University of Pennsylvania, Philadelphia, PA, (3) University of Maryland, Baltimore, MD
Presentations
MO-RPM-GePD-JT-3 (Monday, July 31, 2017) 3:45 PM - 4:15 PM Room: Joint Imaging-Therapy ePoster Theater
Purpose: We study the quality, defined as alignment with radiotherapy outcomes, of distance metrics in transformed patient feature spaces and investigate whether cross-decomposition methods applied to patient and outcome variables can improve metric quality over unsupervised dimensionality reduction.
Methods: In this IRB-approved analysis, a dataset of 16,689 radiotherapy courses performed at one institution from 2008-2015 was created from EMRs and treatment planning systems using automated extraction software. Over 230 variables spanning patient demographics, medical history, and tumor characteristics were extracted for each patient, as well as 68 outcomes including survival, recurrence, and adverse events scored per CTCAEv4.0. Five Euclidean distance metrics were compared: four in 20-dimensional spaces transformed by partial least squares (PLS), canonical correlation analysis (CCA), principal component analysis (PCA), and independent component analysis (ICA), and one in the 732-dimensional (one-hot encoded) original feature space. A 50-50 train-test split was used. Global quality was measured by rank correlation (Spearman's ρ) between metric and outcome distances. Local quality (MNND) was measured using mean outcome distance between a patient and its 10 nearest (according to the test metric) neighbors, averaged over all test patients.
Results: The cross-decomposition methods, PLS (Spearman's ρ = 0.31) and CCA (ρ = 0.29), significantly outperformed ICA (ρ = 0.23), PCA (ρ = 0.20), and the original space (ρ = 0.14) in global quality. PLS (MNND = 1.954) and CCA (MNND = 1.955) also outperformed ICA (MNND = 1.986), PCA (MNND = 1.987), and the original feature space (MNND = 1.973) in local quality.
Conclusion: Our study, the first to investigate aligning patient similarity with radiotherapy outcomes, suggests that while unsupervised dimensionality reduction can improve outcome alignment over metrics in the high-dimensional patient feature space, cross-decomposition such as PLS and CCA provide additional global and local outcome alignment, likely due to their direct use of outcomes data.
Funding Support, Disclosures, and Conflict of Interest: Christopher Berlind, Christopher Ahern, and William Lindsay are employed by and hold securities in Oncora Medical, Inc., a software company in Philadelphia, PA.
Contact Email: