Premium
Early dementia diagnosis from spoken language using a transformer approach
Author(s) -
Wahlforss Alfred,
Jonasson Alexander Aslaksen
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.043445
Subject(s) - transformer , test set , computer science , dementia , artificial intelligence , set (abstract data type) , natural language processing , spoken language , speech recognition , training set , machine learning , engineering , medicine , disease , voltage , electrical engineering , pathology , programming language
Background Diagnosing dementia early is a critical factor in order to apply ameliorating factors. Currently, it is both difficult and costly to make an accurate diagnosis of early dementia. We present an approach where an accurate diagnosis is achieved using a machine learning model that analyses the spoken language of patients describing the cookie‐theft picture. Method The dataset used is Dementia Bank, which is the largest public dataset of spoken word dementia diagnostics. The experiments were carried out on the cookie‐theft section of the DementiaBank Pitt Corpus, which is a dataset containing audio and transcripts of patient interviews in which the patients perform the cookie‐theft test, a commonly used tool for diagnostics. The dataset consists of 243 control samples and 309 samples from patients with dementia. Although the dataset includes transcripts along with parts‐of‐speech tags produced by linguists, the goal of this study was to automate the process of diagnosis. As such, automatic speech recognition (ASR) was used to generate text transcripts from the audio files. The generated transcripts were split into two sets; one training set consisting of 70% of the original transcripts, and a test set containing the remaining 30%. The training set transcripts were then used as input to various transformer models, including BERT and ROBERTA. These models were then trained on the training set to fine‐tune their parameters. During evaluation, the transcripts from the test set are fed into the transformer model, which then classify the transcript as either belonging to the control group or the dementia group. Accuracy is used as an evaluation metric. Result We achieve an accuracy of 90% on the Dementia Bank dataset using the transformer model BERT, which according to a review study by Lyu, Gang (2018) is state‐of‐the‐art when using the entire dataset. Conclusion These results give the possibility to introduce a cost‐effective screening method for signs of early dementia. These results give the possibility to introduce a cost‐effective screening method for signs of early dementia. Hopefully, our approach could put an end to the current subjective diagnostics which is expensive and has varied results.