Michigan Technological University
DEEP LEARNING, MULTI-STAGED MACHINE LEARNING, AND REINFORCEMENT LEARNING TO IMPROVE CARDIAC RESYNCHRONIZATION THERAPY
Abstract
dc:description.abstractThis thesis includes four chapters. In Chapter one, we discussed the clinical background of cardiac resynchronization therapy (CRT) and the current issues towards patient response. Moreover, we considered previous machine learning (ML) and deep learning (DL) methods to improve the decision-making process, both standard supervised learning and multi-stage learning. Finally, we introduced the dataset which will be used in the following Chapters. In Chapter two, we developed a multi-input fusion DL model which utilizes SPECT MPI polar maps images and tabular data in the form of clinical features, ECG parameters, and derived SPECT MPI parameters. Using transfer learning (TL) to train the image component model via VGG16, while using a standard multilayer perceptron for the tabular component trended towards improve over current guidelines and conventional statistical ML. Additionally, Grad-CAM is employed to provide explainability in the DL model’s decision making with respect to the polar map inputs. In Chapter three, we constructed a multi-stage ML model which splits the CRT decision making process into two stages: 1) Clinical and ECG parameter stage 2) SPECT MPI stage. Using the staged data, the model attempts to form a decision using only the first stage; however, if the uncertainty is too high, the model will then consider the addition of the second stage of data. A multi-stage framework provides more clinical interpretability and a more accurate modeling process for clinicians where multiple sequential decisions are assessed weighing associated costs. In Chapter four, we created a multi-stage DL model using similar data staging. The model is constructed in two parts: 1) A deep autoencoder which processes the staged data into fixed length data embedding. The second stage data is treated as missing-not-at-random data when at the first initial stage. 2) A deep reinforcement learning (RL) agent works upon the processed embeddings to perform sequential decision making. The RL agent was trained with volumetric differences in blood ejection as a reward to learn the optimal policy to recommend CRT on a per-patient level. The multi-stage DL model is flexible for unstructured inputs, such as images, and can easily introduce domain specific knowledge in the form of rewards.
Degree
thesis:*- Name thesis:degree_name
- Master of Science in Statistics (MS)
- Level thesis:degree_level
- Campus Access Master's Thesis
- Discipline thesis:degree_discipline
- Department of Mathematical Sciences
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Larsen, Kristoffer A.
- Committee member dc:contributor.committeemember
-
- Kui Zhang
- Contributors dc:contributor
-
- Qiuying Sha
- Weihua Zhou
Subjects
dc:subject × 7Identifiers
dc:identifier.*- Identifier
- https://digitalcommons.mtu.edu/etdr/1754