All Times ET
Keywords: Radiomics, Multilevel Modeling, Horseshoe prior, Missing Data
An important objective in oncology is the creation of a standardized set of criteria to predict and monitor tumor response to treatment and for outcome prognosis based on objectively measured biomarkers. In addition to the traditional role of imaging for staging and post treatment follow-up of HNSCC, there is increasing interest in the use of quantitative image extracted or radiomic features for characterization of HNSCC. Image analysis algorithms extract mathematically defined features of the tumor's appearance giving rise to high-dimensional matrix covariates.
Many challenges arise from the structure of radiomic data. Namely, an efficient and reliable variable selection technique is required to select a reasonable number of radiomic features to be used in prediction of key outcomes such as lymph node metastasis. Variable selection and prediction should reflect the heterogenity among tumor sites, however, site-stratified inference can result in low statistical power. In addition, there is a considerable amount of missing data among important predictors such as the presence/absence of human papilloma virus (HPV).
We propose a Bayesian hierarchical model that can address radiomic feature selection and prediction in a unified framework while dealing with complexities such as missing values in predictors. The hierarchical nature of the model enables information borrowing across tumor sites while allowing site-specific variable selection and parameter estimation. Integrating variable selection and missing data handling together with inference, results in predictions with adequate representation of uncertainty associated with each of these procedures. We present the results of the analysis as Bayesian feature selection outcomes across sites and the accuracy for predictions of lymph node metastasis.