Longitudinal deep learning models for integrating patient-reported outcomes in toxicity and survival prediction of head and neck cancer

Yaohua Wang

doi:10.25820/etd.007868

Back

Dissertation

Longitudinal deep learning models for integrating patient-reported outcomes in toxicity and survival prediction of head and neck cancer

Yaohua Wang

University of Iowa

Doctor of Philosophy (PhD), University of Iowa

Spring 2025

DOI: 10.25820/etd.007868

Files and links (1)

pdf

PhD_Thesis_Dissertation_of_Studying_the_Impact_of_Long_Short_Term_Memory__LSTM__on_Patient_reported_Outcome__PRO_Camera_Ready6.43 MB

Embargoed Access, Embargo ends: 06/26/2027

Abstract

In clinical practice, head and neck cancer (HNC) patients usually have severe side effects during the cancer treatment that subsides over time. However, some patients have persistent late toxicity that negatively affects the patient's quality of life (QoL) after the treatment. To monitor patients' symptom burden, clinicians often collect symptom ratings directly from patients using symptom questionnaires known as patient-reported outcome (PRO) measures. These PRO can be collected repeatedly over time and could be used to better inform treatment decisions. However, their incorporation into clinical practice presents numerous challenges. First, missing data are common if patients skip certain questions or entire questionnaires. Second, PROs are subjective: a symptom rating of 5 for one patient could be a 10 for another. Finally, only a subset of patients experience late toxicity, while most see their symptoms diminish over time. To address these challenges, we implement and evaluate a family of Long Short-Term Memory (LSTM) recurrent neural networks used for imputing missing values and predicting patients' late toxicity. To normalize the PRO ratings, we subtract the baseline from subsequent ratings, and the use of symptom clusters gives us a more balanced toxicity distribution. Our experiments show that LSTM-based methods effectively handle missing-value imputation and late-toxicity prediction, outperforming more standard machine learning and other statistical approaches. Building on these predicted symptom ratings and sequential treatment variables, we developed a Symptom Burden Model (SBM) to generate a binary (low/high) long-term indicator of symptom burden, i.e. 12 months after treatment. This indicator was then incorporated into a Cox regression model, alongside other clinical features, to predict overall survival. Our results show that including the long-term symptom burden indicator in the Cox model increased the test c-index from 0.804 to 0.849, demonstrating a notable improvement in survival prediction.

Bidirectional Long Short-Term Memory (Bi-LSTM)

Late Toxicity

Patient Reported Outcomes (PROs)

Survival Prediction

Symptom Burden Prediction

Symptom Severity Prediction

Details

Title: Subtitle: Longitudinal deep learning models for integrating patient-reported outcomes in toxicity and survival prediction of head and neck cancer
Creators: Yaohua Wang
Contributors: Guadalupe M Canahuate (Advisor)
Tyler Bell (Committee Member)
John Buatti (Committee Member)
Mathews Jacob (Committee Member)
Kishlay Jha (Committee Member)
Resource Type: Dissertation
Degree Awarded: Doctor of Philosophy (PhD), University of Iowa
Degree in: Electrical and Computer Engineering
Date degree season: Spring 2025
DOI: 10.25820/etd.007868
Publisher: University of Iowa
Number of pages: xii, 86 pages
Language: English
Date submitted: 03/28/2025
Description illustrations: illustrations, graphs, tables
Description bibliographic: Includes bibliographical references (pages 72-86).
Public Abstract (ETD): Head and neck cancer patients often experience severe side effects during treatment, which for many subside over time. However, some continue to suffer from persistent late toxicity, sig- nificantly diminishing their quality of life. To better anticipate these outcomes, we analyzed data from a weekly symptom-reporting questionnaire called the MDASI-HN, collected both during and after treatment. Using advanced deep learning methods, specifically Long Short-Term Memory (LSTM) networks, we addressed the challenges of missing responses and subjective symptom rat- ings to predict late toxicity 12 months post-treatment. We then combined these predictions with other clinical information to build a “Symptom Burden Model”, ultimately improving our ability to predict overall survival outcomes. Our findings show that incorporating patient-reported symptom data into survival models substantially enhances prediction accuracy, highlighting the potential of patient-centered data to improve treatment decisions and long-term care for head and neck cancer patients.
Academic Unit: Electrical and Computer Engineering
Record Identifier: 9984831231502771

Metrics

5 Record Views