Work

Human and the Machine: Vulnerabilities and Biases in Automated Data Analysis Pipelines

Year: 2020
Type: dissertation
Abstract: Accurate prediction of a cancer patient’s outcome is imperative to treatment selection and decision support. Recently, automated data analysis pipelines have been explored to model and predict complex patient outcome relationships using information extracted from existing data sources; for example, radiomics features that are designed to quantify patient images. However, consideration of biases imposed by human knowledge and interactions with these systems, and their resulting models, is required to obtain robust and informative results. In this thesis, automated data analysis pipelines are developed for exploration and prevention of unintended biases. A comprehensive radiomic pipeline was established and validated during an in-depth analysis of a highly cited radiomic signature; concluding that the signature was a surrogate for tumour volume and that intensity and texture values were not pertinent for prognostication. Safeguards were subsequently suggested to refine methodologies and ensure productive radiomic developments. Based upon the recommended safeguards, automated image quality classification methods were developed. In particular, convolutional neural networks (CNN) were trained to automatically classify head and neck (H) computed tomography images based upon the presence of dental artifacts. The CNNs were additionally validated using six external datasets and demonstrated increased classification performance with transfer learning. The suggested safeguards were also enforced in the development of operator-controlled data analysis pipelines with different levels of human interaction requirements. The pipelines predicted loco-regional failure in H patients using combinations of clinical, imaging (i.e. radiomic) and radiation therapy (i.e. RTx-omic) information. These pipelines provided insights into the importance of data curation and quality, as well as operator, data and methodology bias awareness as it pertains to result interpretation in operator-controlled pipelines. Overall, this work demonstrates the utility of automated data analysis pipelines for cancer patient outcome prediction using combinations of traditional and emerging features. Future work standardizing data governance processes will be required to actualize clinically impactful implementations of automated data analysis pipelines.
Author: Mattea Welch
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W3161668203
Cites: 181
Cited by:
Related to: 10