In a recent study published in Alzheimer’s & Dementia, researchers developed a method for predicting the progression of Alzheimer’s disease (AD).
Study: Prediction of Alzheimer’s disease progression within 6 years using speech: A novel approach leveraging language models. Image Credit: fizkes/Shutterstock.com
Background
Individuals with mild cognitive impairment (MCI) have a heightened risk of AD. Therefore, accurate prediction of the progression from MCI to AD can help in treatment-related decisions, selection for trials of new drugs, and participation in rehabilitation programs. AD pathology has been conventionally assessed using neuroimaging techniques or biomarkers.
Various studies have evaluated these (conventional) methods for predicting MCI to AD progression. However, they are expensive and invasive, limiting their applicability.
By contrast, neuropsychological tests (NPTs) are the most accessible for cognitive decline assessment. Computer-based approaches have been tested for predicting MCI-to-AD conversion using NPTs. Speech in NPTs can be used to predict cognitive decline.
Artificial intelligence-based diagnostic models using acoustic and linguistic features from NPTs have been developed.
The Framingham Heart Study (FHS) has been recording NPTs since 2005, and the recordings have been used to build diagnostic tools. Previously, the study’s authors applied natural language processing (NLP) techniques on recordings to place individuals across the dementia spectrum.
About the study
In the present study, using speech data, researchers developed a method to predict AD progression within six years. The FHS monitored a cohort of 166 people with cognitive complaints. Each individual underwent an hour-long NPT that was digitally recorded and stored. Education information, health risk factors, and apolipoprotein E (APOE) alleles were available.
The study focused on MCI-to-AD progression and not on normal cognition to MCI or AD because of the limited utility of NPTs in predicting cognitive decline without (signs of) cognitive deterioration.
The team developed a tool to transcribe voice recordings in their previous work automatically. This tool was used to transcribe subjects’ audio files. Each sentence was labeled according to the specific sub-test.
Different vector embeddings were obtained for NPTs based on specific segments of each transcript. The Universal Sentence Encoder, a deep learning-based model, generated vector embeddings.
Training data were increased by randomly sampling from transcripts to generate abbreviated versions, which were subsequently encoded. Besides, sub-test content was separately encoded, creating eight specific embeddings.
A logistic regression model was trained on the quantitative data associated with each sub-test content. Embeddings from abbreviated versions were used as independent input, generating multiple scores for each transcript.
A transcript average score (TAS) was generated from these multiple scores. An ensemble logistic regression model was generated using sub-test scores and TAS to predict the likelihood of MCI to AD progression within six years.
Model performance was evaluated using a stratified group k-fold cross-validation approach. Besides, an internal cross-validation was performed for feature selection and dimensionality reduction.
Performance metrics included the area under the receiver operating characteristic curve (AUC), accuracy, precision, sensitivity, and specificity.
Findings
Of the 166 subjects with MCI, 90 progressed to AD dementia within six years. AD dementia included mixed dementia and AD with/without stroke. The mean time to AD was 2.7 years.
Older females with lower education and those carrying the APOE ε4 allele were more likely to progress to AD. Besides, females who progressed to AD were 1.4 years older, on average, than males.
The model incorporating demographics, APOE carrier status, health factors, and text features (viz., NLP model) achieved an F1 score of 79.9% and an AUC of 78.5%.
The corresponding figures for the model with only text features were 79.4% and 77.8%, respectively. The model with text and demographic features had an AUC of 77.5% and an F1 score of 79.6%.
The model with only NPT scores had an F1 score of 75.5% and an AUC of 71.3%. The AUC and F1 scores of the model with only demographic features were 68.8% and 71.1%, respectively.
A model based on a mini-mental state examination had an AUC of 60.7%. The model with only health factors achieved an AUC of 66.2%.
Conclusions
In sum, the researchers illustrated the potential of automated speech recognition and NLP in predicting progression to AD among people with MCI. The proposed model predicted AD progression with a sensitivity of 81.1, specificity of 75%, and accuracy of 78.2%.
This approach allows for an accessible and non-invasive AI-based prediction without involving genetic or laboratory tests or imaging, making it ideal for remote assessments.
Further large-scale studies are required to corroborate these findings and validate their generalizability, given that the cohort was predominately White.