Biomedical NLP for Alzheimer's Disease

This project was completed as an MSc research dissertation at University College London, using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Clinical records contain rich signal about disease progression, but it is locked in unstructured text. The goal was to surface latent patterns from that text and combine them with structured biomarkers, rather than to build a diagnostic classifier.

The corpus was built from unstructured ADNI medical-record text. Each patient document was enriched with structured signals available in ADNI: APOE4 genotype (a genetic risk factor), MMSE cognitive score, and hippocampal volume normalized by intracranial volume — a standard adjustment for individual head-size variation.

Raw clinical text was extracted, tokenized, and lemmatized, then passed through bigram detection to capture multi-word clinical phrases, synonym normalization to collapse clinically equivalent terms, and frequency filtering with deduplication to remove noise before corpus construction.

Documents were modeled with Latent Dirichlet Allocation (LDA): each document is treated as a mixture of latent topics, and each topic as a distribution over vocabulary. LDA is a bag-of-words model — it does not capture word order or deep contextual meaning, which shaped how results were interpreted.

Clinical text, APOE4 genotype, MMSE score, and normalized hippocampal volume were fused into a single enriched patient document, giving the topic model structured biomarker context alongside free text.

Topic models were evaluated primarily through coherence scoring across different numbers of topics, passes, and iterations, alongside qualitative inspection of topic-keyword distributions and corpus visualizations.

• This was an exploratory research prototype, not a clinically validated diagnostic system.
• No diagnostic accuracy is claimed anywhere in this write-up.
• LDA is a bag-of-words model and does not capture word order or deep contextual meaning.
• Any illustrative data shown on this page (example tokens, topic labels, chart values) is clearly marked as illustrative, not a reported result.

Working with real clinical text surfaced how much preprocessing quality drives topic-model quality — normalization and filtering choices mattered more than model tuning. It also reinforced the limits of bag-of-words methods for clinical narrative, motivating later work with transformer-based representations.

• Python
• Gensim
• Natural Language Processing
• Latent Dirichlet Allocation
• Topic Modeling
• Text Preprocessing
• Coherence Evaluation
• ADNI
• Data Visualization