Graduate student Sumaya Hossain will be defending her thesis titled “A Unified Multimodal Framework for Alzheimer’s Disease Staging with Graph Learning, Explainability, and Clinical Narrative Generation.”

Sumaya Hossain thesis defense

• Date: Monday, April 27
• Time: 8:30-9:30 a.m.
• Location: BARC 1142
• Current major: M.S. of data science
• Thesis committee chair: Dr. Bayazit Karaman
• Committee members:  Dr. Parisa Hajibabaee (co-advisor), Dr. Abdulaziz Alhamadani and Dr. Susan LeFrancois.

Abstract

Alzheimer’s disease (AD) staging requires integrating multimodal clinical, imaging, and biomarker data, yet classification remains challenging at the mild cognitive impairment (MCI) stage due to feature overlap with adjacent classes. Graph neural networks (GNNs) model inter-patient relationships, but graph construction is often treated as fixed, and the link between prediction, explanation, and clinical interpretation remains limited.

This work proposes an interpretable framework for three-class AD staging (NC, MCI, AD) using a hybrid SpectralGCN-based population-graph model trained on the ADNI dataset. Graph construction and temporal modelling are treated as design variables and evaluated through systematic ablation. A multi-component explainability framework combines feature attribution, graph neighborhood influence, counterfactual analysis, and uncertainty estimation to provide patient-level insights into model behavior. A grounded natural language generation (NLG) pipeline converts these outputs into structured clinical summaries, with a companion OASIS-3 framework that applies the same grounded generation strategy to structured multimodal biomarkers without a GNN prediction model.

The model achieves a balanced accuracy of 0.87 and a macro-averaged AUC of 0.97. Demographic and genetic edge construction yields the strongest performance, while temporal modelling provides the largest improvement in MCI classification. MCI remains the most challenging class, with distributed feature contributions and predominantly aleatoric uncertainty. The NLG pipeline produces zero false positives and high factual accuracy across both datasets, reaching 100% field-level accuracy on ADNI and 96.93% field-level accuracy on OASIS-3. These results highlight the role of graph construction and graph-based explanation in AD staging and show that combining explainability with controlled generation enables clinically interpretable and factually grounded outputs for patient-level summaries.