Year: 2026
Type: dissertation
Abstract: Data-driven machine learning connects characteristics of data to implementation decisions. Previous approaches tended to rely on field-specific experience such as previously successful classification models on ``similar'' data. How can the similarity of data be defined if they are collected from independent sources? One way is commensurability. In this dissertation we called datasets commensurable if variables have equivalent meaning across datasets. We focused on high-dimensional medical images of tissue acquired using mass spectrometry imaging (MSI). Technical and biological variation make comparison across images challenging. Despite this challenge, it is common practiced for spectra to be combined across images to find generalizable biomarkers of tumors. We showed that for the datasets used, MSIs are incommensurable and thus that aggregation of spectra across images should be performed with caution. Another way to compare datasets is data complexity. Data complexity measures are quantified characteristics of data that relate to the potential classification accuracy of a dataset. We used data complexity to predict the classification accuracy of over 20 classification models for 17 datasets; this is called meta-learning. The meta-models were applied to MSIs and showed that not only do data complexity of images differ, but aggregated images have the complexity of challenging natural-language datasets. This gave us a foundation to tackle the question: can we learn stable patterns from incommensurable data? We used a stacked ensemble model to find biomarkers of tumor tissue from MSIs. We showed biomarkers learned from aggregated images missed clinically relevant biomarkers that were captured when training on individual images. We discussed our findings in the context of personalized medicine.
Author: Theriault, Rachel (Lynn)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7209896573