Alzheimer's Disease Classification Using Language Models

Team Members

Catherine Chen, cchen207
Nate Gillman, ngillman
Michal Golovanevsky, mgolovan

Introduction

Alzheimer’s disease (AD) is one of the most common neurodegenerative diseases. There is no known cure for AD, but early detection of the disease is crucial to delay progression. In recent years, deep learning models have greatly contributed to AD diagnosis research. The Alzheimer’s disease neuroimaging initiative (ADNI) dataset contains an adverse events log that details (in natural language) events for which the study participants were hospitalized. We propose to feed this data into a transformer model [1] to predict whether a patient gets Alzheimer’s. This is a binary classification problem. Time allowing, we would like to expand this to a 3-class classification problem--Alzheimer’s, mild cognitive disorders, and control.

Related Work

For our classification task, we plan to utilize word embeddings from Clinical Bert. These have already been pre-trained on clinical corpora, and we plan to fine-tune them for our particular data set.

[1] Emily Alsentzer, John Murphy, William Boag, WeiHung Weng, Di Jindi, Tristan Naumann, and Matthew McDermott. 2019. Publicly available clinical BERT embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop.

Original repo: https://github.com/EmilyAlsentzer/clinicalBERT
More up to date repo: huggingface transformers (https://github.com/huggingface/transformers)

Data

We will use data from the Alzheimer’s disease neuroimaging initiative (ADNI). The initiative keeps a log of adverse hospital visits each patient from the study experienced. We will use the visit notes as our text corpus. Every patient has a label - AD, MCI, and CN. Based on the notes of each visit, we will predict whether or not a patient has the disease.

There are a total of around 1000 study participants that have hospital notes. Each note section is summarized in a few sentences. Thus, we do not anticipate compute power to be an issue. The only preprocessing we may consider is redacting the word “Alzheimer’s” from the corpus. We will also need to merge two documents to combine the labels with each patient ID and their notes associated with the ID.

Methodology

We propose to feed this data into a transformer model (Clinical BERT) to predict whether a patient gets Alzheimer's. We will do fine-tuning specific to our classification task, using the Clinical Bert pre-trained word embeddings. Past researchers have had success with using clinical BERT to accomplish large-scale natural language tasks that are relevant to patient outcomes. We propose doing something similar--using clinical BERT to aid in the diagnosis of Alzheimer's.

Metrics

We will implement several different models to compare performance of our model against other baselines. We will compare Clinical BERT to other language models we previously implemented in class (RNNs and vanilla Transformers).

The notion of “accuracy” indeed applies to our model. Since we are trying to solve a classification task, we can measure our progress by examining the accuracy for each class, and the accuracy overall.

Base: random chance (33.33%)
Target: 80% and above (as it would be comparable to other papers on the dataset)
Stretch: 100% ;)

Ethics

What broader societal issues are relevant to your chosen problem space?
The methods that we develop will generalize far beyond the classification of Alzheimer’s. Our work will result in a pipeline that can solve the general classification problem of mapping clinical notes to patient diagnoses.

Who are the major “stakeholders” in this problem, and what are the consequences of mistakes made by your algorithm?
In the biomedical problem space, the consequences of mistakes made by any ML algorithm can be detrimental since the major “stakeholders” problem are patients. For this problem in particular, if a patient is misdiagnosed with mild cognitive disorder instead of Alzheimer's, then they may not get the proper care they need and their condition may actually worsen more rapidly.