Using explainable artificial intelligence to identify linguistic biomarkers of amyloid pathology in primary...

Recent success has been achieved in Alzheimer's disease (AD) clinical trials targeting amyloid beta (β), demonstrating a reduction in the rate of cognitive decline. However, testing methods for amyloid-β positivity are currently costly or invasive, motivating the development of accessible screening approaches to steer patients toward appropriate diagnostic tests. Here, we employ a pre-trained language model (Distil-RoBERTa) to identify amyloid-β positivity from a short, connected speech sample. We further use explainable AI (XAI) methods to extract interpretable linguistic features that can be employed in clinical practice.

We obtained language samples from 74 patients with primary progressive aphasia (PPA) across its three variants. Amyloid-β positivity was established through the analysis of cerebrospinal fluid, amyloid PET, or autopsy. 51% of the sample was amyloid-positive. We trained Distil-RoBERTa for 16 epochs with a batch size of 6 and a learning rate of 5e-5, and used the LIME algorithm to train interpretation models to interpret the trained classifier's inference conditions.

Over ten runs of 10-fold cross-validation, the classifier achieved a mean accuracy of 92%, SD = 0.01. Interpretation models were able to capture the classifier's behavior well, achieving an accuracy of 97% against classifier predictions, and uncovering several novel speech patterns that may characterize amyloid-β positivity.

Our work improves previous research which indicates connected speech is a useful diagnostic input for prediction of the presence of amyloid-β in patients with PPA. Further, we leverage XAI techniques to reveal novel linguistic features that can be tested in clinical practice in the appropriate subspecialty setting. Computational linguistic analysis of connected speech shows great promise as a novel assessment method in patients with AD and related disorders.