Thesis
Predicting acute and late symptoms of head and neck cancer treatment using deep learning and longitudinal patient reported outcomes
University of Iowa
Master of Science (MS), University of Iowa
Spring 2021
DOI: 10.17077/etd.005874
Abstract
In clinical treatments for cancer patients, Patient-Reported Outcome (PRO) surveys are mainly used to monitor patients' symptoms during and after treatment. In the course of their treatment, patients experience a number of symptoms with different severity and PRO surveys are used to reflect the patients' conditions during and after treatment. Acute symptoms, or symptoms experienced by the patients shortly after treatment, are sometimes a limiting factor that prevents completion of the initial treatment and requires changes in treatment's dose or duration. Late symptoms or symptoms experience some time after treatment, are used as a proxy of the patient's quality of life (QoL) after treatment. Researchers use PRO surveys to investigate the relationships among symptoms and predict symptoms to guide treatment selection and improve the long-term QoL of the patients. In the case of cancer patients, PRO surveys are recorded every week during the patient's visit to the clinic and at different follow-up times after the treatment has concluded. In this sense, PRO data can be modeled as time-series data.
In this thesis, we apply Long-Short Term Memory (LSTM) neural networks to PRO survey data in order to predict symptom ratings in both acute and late stages. An LSTM model is a type of recurrent neural network proven effective at predicting time-series data. In this work, the acute stage refers to week six during treatment and the late-stage up to 18 months after treatment. The data used in this project corresponds to MDASI questionnaires collected from head and neck cancer patients treated at the MD Anderson Cancer Center. We apply several methods for missing data imputation and evaluate the performance of the LSTM model for each method. We show that the LSTM model is effective in predicting symptom ratings under the RMSE and NRMSE metrics. In our experiments, the LSTM model also outperforms other machine learning models. We also show that the late-stage symptom rating has a smaller range than the acute-stage symptom rating.
Details
- Title: Subtitle
- Predicting acute and late symptoms of head and neck cancer treatment using deep learning and longitudinal patient reported outcomes
- Creators
- Yaohua Wang
- Contributors
- Guadalupe Canahuate (Advisor)Mathews Jacob (Committee Member)Tyler Bell (Committee Member)
- Resource Type
- Thesis
- Degree Awarded
- Master of Science (MS), University of Iowa
- Degree in
- Electrical and Computer Engineering
- Date degree season
- Spring 2021
- DOI
- 10.17077/etd.005874
- Publisher
- University of Iowa
- Number of pages
- viii, 34 pages
- Copyright
- Copyright 2021 Yaohua Wang
- Language
- English
- Description illustrations
- color illustrations
- Description bibliographic
- Includes bibliographical references (pages 30-34)
- Public Abstract (ETD)
When cancer patients start receiving treatment, different symptoms with different seriousness may occur during and after treatment. For head and neck cancer patients, Patient-Reported Outcome (PRO) surveys are mainly used to monitor patients’ symptoms overtime and treatments may be adjusted based on the seriousness of the symptoms. Understanding symptom severity from modeling PRO data to better serve patients, would allow to identify potential severe side-effect symptoms and personalized patient’s treatment accordingly.
In this paper, we approach the symptom prediction problem for both acute and late stages. The idea is simple. Since PRO surveys take patients’ response from the questionnaires over time, if there is a way to learn from patients’ response patterns over time, we can predict patients’ response in the future. As a result, Long-Short Term memory (LSTM) neural network offers a way to build a prediction model. Due to the fact that LSTM model can memorize data in the past few time stamps, it’s an ideal model to be used on the PRO time-series data. After comparing the performance among LSTM and other machine learning models, it turns out that LSTM model has an overall better performance. Such results are exciting since it means LSTM can be a good candidate to predict patients’ symptoms and later on help investigating the relations among symptoms.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9984097276102771
Metrics
5 File views/ downloads
50 Record Views