A Statistical Explainable Learning Model Optimizing Co-localization of Multidimensional Positivity Thresholds in Immunotherapy Decision-Supporting

Tumor Mutation Burden (TMB) serves as a recognized stratified biomarker for immunotherapy. However, its one-dimensional representation of non-synonymous genetic alterations has been contentious. Specifically, the uniform quantification of mutations by TMB, coupled with measurement inaccuracies, complicates the accurate determination of a positive threshold for classifying patients. Parallel to this, assessing immunotherapy benefits requires the joint analysis of multiscale endpoints, namely discrete tumor response and sequential time-to-event, presenting a pressing challenge for clinical computation. Recognizing the intertwined nature of these challenges, we address the inter-sample bias inherent in multidimensional mutation biomarkers within the framework of multiscale endpoint fusion analysis, aiming for a more robust and comprehensive patient stratification. By combining the concept of corrected-score with a soft-threshold strategy, and utilizing the attention mechanism alongside the multiple instance learning, we propose a statistically explainable learning model optimizing co-localization of multidimensional positivity thresholds for immunotherapy categorical decision-supporting.