Files and links (1)
Eric Anyimadu Manuscript 212.50 MB
Embargoed Access, Embargo ends: 06/29/2027
Dissertation
A multimodal deep learning approach to oncologic outcome prediction in head and neck cancer
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2026
DOI: 10.25820/etd.008393
Abstract
Head and neck cancer (HNC) patients often experience complex, long-term treatment- and disease-related side effects that substantially impair their quality of life and, in some cases, survival. A large volume of heterogeneous data is routinely collected for these patients, including patient-reported outcomes (PROs), organ-specific dose–volume histograms (DVHs), treatment parameters, and detailed clinical and disease characteristics. These data provide complementary perspectives on patient risk and have the potential to enhance clinical decision making. However, effective integration of such multimodal data into predictive models remains challenging due to missing data, high dimensionality, intricate interdependencies, and longitudinal variability.
This dissertation presents a comprehensive multimodal machine learning framework designed to address these challenges and fully leverage available data for toxicity and survival modeling in HNC to support data-driven decision making.
The framework first employs collaborative filtering–based imputation to fill missing PRO data while preserving symptom inter-relationships. To incorporate high-dimensional PRO data into survival models, dimensionality reduction techniques, including principal component analysis and autoencoders, are applied to derive compact and informative representations. Symptom cooccurrence clustering is also used to identify clinically meaningful patient subgroups.
Temporal modeling using bidirectional long short-term memory networks are used to capture dynamic symptom trajectories and improve long-term outcome forecasting. In parallel, a spatially informed deep learning normal tissue complication probability (NTCP) model integrates DVH-derived features with contrastive representation learning to enhance the prediction of radiation-induced toxicities, such as mandibular osteoradionecrosis and xerostomia.
Finally, these components are unified within an end-to-end multimodal deep learning pipeline that integrates clinical, dosimetric, treatment, and patient-reported data. By combining imputation, representation learning, and temporal modeling, the proposed framework improves predictive performance, enhances interpretability, and supports proactive, patient-centered decision making in head and neck oncology.
Details
- Title: Subtitle
- A multimodal deep learning approach to oncologic outcome prediction in head and neck cancer
- Creators
- Eric Ababio Anyimadu
- Contributors
- Guadalupe Canahuate (Advisor)Thomas Casavant (Committee Member)G. Elisabeta Marai (Committee Member)Yang Liu (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 2026
- DOI
- 10.25820/etd.008393
- Publisher
- University of Iowa
- Number of pages
- xii, 142 pages
- Copyright
- Copyright 2026 Eric Ababio Anyimadu
- Language
- English
- Date submitted
- 04/22/2026
- Description illustrations
- illustrations, tables, graphs
- Description bibliographic
- Includes bibliographical references (pages 122-142).
- Public Abstract (ETD)
Head and neck cancer patients often face lasting side effects from both the disease and its treatment, which can significantly affect daily life and, in some cases, survival. At the same time, hospitals routinely collect a wide range of information about these patients, including symptoms reported by patients themselves, details about radiation dose to specific organs, treatment plans, and clinical characteristics. While each of these data sources offers valuable insight, combining them effectively to guide care remains difficult due to missing information, the complexity of the data, and how patients’ conditions change over time.
This work presents a new approach that uses advanced machine learning to bring these different types of data together in a meaningful way. First, it fills in missing symptom information by learning patterns from similar patients. It then simplifies complex symptom data into more manageable forms while still preserving important details. The approach also identifies groups of patients who share similar symptom patterns, which may help clinicians better understand different experiences of the disease.
To better track how symptoms evolve, the model analyzes patient data over time, allowing it to make more accurate predictions about future outcomes. It also uses detailed radiation data to improve predictions of treatment-related side effects, such as jaw bone damage and chronic dry mouth.
By integrating all of this information into a single framework, this work aims to improve the ability to predict patient outcomes, support more personalized treatment decisions, and ultimately enhance quality of life for individuals with head and neck cancer.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9985176974902771
Metrics
1 Record Views