AI Speech Analysis Predicts Alzheimer’s with 78.5% Accuracy

Researchers created a technique for forecasting the course of Alzheimer’s disease (AD) in a recent study that was published in Alzheimer’s & Dementia.

Context

AD risk is higher in people with mild cognitive impairment (MCI). Thus, precise forecasting of the progression from MCI to AD can be beneficial for treatment choices, drug trial selection, and involvement in rehabilitation initiatives. The pathophysiology of AD has traditionally been evaluated utilizing biomarkers or neuroimaging methods.

These (traditional) techniques for forecasting the transition of MCI to AD have been assessed in numerous research. Nevertheless, their usefulness is limited due to their high cost and invasiveness.

In contrast, the easiest assessments of cognitive loss to do are neuropsychological tests (NPTs). NPTs have been used to test computer-based methods for MCI-to-AD conversion prediction. It is possible to forecast cognitive impairment in NPTs using speech.

Using language and auditory characteristics from NPTs, artificial intelligence-based diagnosis models have been created.

Since 2005, NPTs have been recorded as part of the Framingham Heart Study (FHS), and the recordings have been utilized to create diagnostic tools. The authors of the study previously used natural language processing (NLP) methods on audio recordings to categorize people according to different stages of dementia.

Concerning the study
In this study, a strategy to predict the progression of AD within six years was established by researchers utilizing voice data. A group of 166 individuals with cognitive problems was observed using the FHS. Every person had an NPT for an hour, which was digitally recorded and saved. There was access to health risk variables, education data, and apolipoprotein E (APOE) alleles.

Due to the limited usefulness of NPTs in predicting cognitive decline without (evidence of) cognitive deterioration, the study concentrated on the transition of MCI to AD rather than normal cognition to MCI or AD.

In order to automatically transcribe voice recordings from their earlier work, the team created a tool. The audio files of the subjects were transcribed using this tool. The particular sub-test was used to label each sentence.

For NPTs, several vector embeddings were acquired according to particular transcript segments. Vector embeddings were produced by a deep learning model called the Universal Sentence Encoder.

By taking random samples from transcripts to create condensed versions, which were then encoded, training data were increased. Additionally, eight distinct embeddings were created by individually encoding the sub-test information.

Each subtest’s quantitative data set served as the training set for a logistic regression model. Embeddings from condensed versions were utilized as separate input, resulting in several transcript ratings.

These numerous scores were combined to get a transcript average score (TAS). Sub-test results and TAS were used to create an ensemble logistic regression model that predicted the probability of MCI to AD progression within six years.

An strategy known as stratified group k-fold cross-validation was used to assess the performance of the model. In addition, feature selection and dimensionality reduction were carried out by internal cross-validation.

The area under the receiver operating characteristic curve (AUC), accuracy, precision, sensitivity, and specificity were among the performance measurements.

Results

Ninety of the 166 MCI patients developed AD dementia within six years. AD dementia encompassed AD with or without stroke as well as mixed dementia. There was a 2.7-year mean time to AD.

AD progression was more common in older, less educated females and those with the APOE ε4 genotype. In addition, the average age of females who advanced to AD was 1.4 years older than that of males.

An F1 score of 79.9% and an AUC of 78.5% were attained by the model that combined text features, health factors, demographics, and APOE carrier status (i.e., the NLP model).

Comparable numbers for the model with just text features were 77.8% and 79.4%, in that order. The AUC and F1 score of the text and demographic feature-rich model were 77.5% and 79.6%, respectively.

AUC was 71.3% and F1 score was 75.5% for the model that solely included NPT scores. The model using only demographic data has an AUC of 68.8% and an F1 score of 71.1%.

The AUC of a model that was based on a mini-mental state exam was 60.7%. With just health factors included, the model’s AUC was 66.2%.

In conclusion
The researchers’ findings demonstrate the utility of NLP and automated speech recognition in forecasting the development of AD in individuals with mild cognitive impairment. With an accuracy of 78.2%, specificity of 75%, and sensitivity of 81.1, the suggested model predicted the course of AD.

This method is perfect for remote evaluations since it enables an accessible and non-invasive AI-based prediction without requiring imaging, laboratory testing, or genetic testing.

Since the cohort was primarily White, more extensive research is needed to confirm these results and ensure their generalizability.

For more information: Prediction of Alzheimer’s disease progression within 6 years using speech: A novel approach leveraging language models, Alzheimer’s & Dementia, https://alz-journals.onlinelibrary.wiley.com/doi/10.1002/alz.13886