Georgia Institute of Technology
Towards Interpretable Machine Learning for Healthcare using Domain Knowledge
Abstract
dc:description.abstractMachine learning holds the potential to drive transformative advancements in drug discovery, healthcare decision support, and personalized medicine. Nevertheless, the complexity and lack of transparency in deep learning models create obstacles in comprehending their internal mechanics and building confidence in their predictions, thus impeding their widespread use in healthcare. Interpretable machine learning surfaces as a resolution to these obstacles, guaranteeing the reliability and credibility of models in critical scenarios by offering understandable explanations to human users. This dissertation aims to augment the interpretability of machine learning within healthcare contexts through the integration of domain-specific knowledge, thus ensuring that the interpretations are grounded firmly in clinical relevance. The specific aims are: Aim 1: Streamline domain knowledge synthesis and interpretation. In an effort to streamline the extraction and comprehension of domain-specific information from the multitude of randomized clinical trials on PubMed, we have constructed a specialized dataset for their classification and filtration, known as BioSift. Leveraging the filtered articles obtained through BioSift, we have crafted a unique biomedical entity dataset, TrialSieve, tailored to optimize the meta-analysis process in clinical studies about drug repurposing. Moreover, we have also implemented benchmarks for both BioSift and TrialSieve with the help of supervised, weakly-supervised, semi-supervised, and zero-shot algorithms. Furthermore, we have devised a two-tiered methodology to generate abstractive lay summaries from entire biomedical documents, ensuring they are readable and easily understood, thereby supplementing the entities. Finally, we have created CCS Explorer, a comprehensive end-to-end system equipped with a user-friendly graphical interface that facilitates summarization and the detection of participant, outcome, and intervention entities. Aim 2: Enhance interpretable model features through domain knowledge. Despite the prevalent use of deep learning superseding feature-based techniques, we advocate for an interpretable feature-based approach named SeizFt for the detection of seizures from physiological signals, which proves superior to deep learning alternatives. Furthermore, we introduce interpretable feature-based methodologies aimed at assessing the risk of infection and relapse in pediatric acute leukemia patients by utilizing electronic health records from a limited cohort. This culminates in a model that is in harmony with clinical guidelines. The strategies for both seizure detection and risk stratification highlight the crucial role played by steps such as preprocessing, augmentation schemes, feature selection, and glass-box model selection. Aim 3: Develop interpretable domain-aware representations from deep learning embeddings. In an attempt to reconcile feature-based and deep learning approaches, we put forward algorithms that combine the interpretability of features, derived from domain knowledge, with the superior accuracy of black-box deep learning models. Utilizing sleep staging as a representative case, we present two methods, namely SERF and NormIntSleep. These techniques map deep neural network embeddings onto an interpretable feature space, dictated by domain knowledge, allowing for clinical interpretability without compromising the accuracy associated with black-box models.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Al-Hussaini, Irfan
- Advisors dc:contributor.advisor
-
- Mitchell, Cassie S.
- Romberg, Justin
- Committee members dc:contributor.committeemember
-
- Chau, Duen Horng
- Heck, Larry
- Prakash, B. Aditya
- Landsness, Eric C.
Subjects
dc:subject × 12Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/75607