Comparative study of Deep Classifiers for Early Dementia Detection using Speech Transcripts

The applications of artificial intelligence are increasing day by day. The world is transformed by AI technologies like automated systematic processes and virtual assistance. Dementia is a disorder that causes disintegration in the mental capacity of an individual. This is not a disease but a group of symptoms caused by various other conditions and is common in older people. Detecting dementia in the early stages will reduce complications and provide access to medical attention and medications. Dementia is detected in patients by examining their ability to think, communicate and their physical movements. This is done by performing numerous physical and mental status exams. This paper aims to detect early signs of dementia in patients using English speech transcript files. The proposed models employ considerable deep learning and natural language processing techniques like GloVe, Word2Vec, Doc2Vec word embeddings and LSTM, BiLSTM, GRU, BERT, RoBERTa and ALBERT. The models were trained using the Pitt Corpus from the DementiaBank dataset. The best accuracy obtained was 0.812 using the BERT+BilSTM model, and the best F1 score obtained was 0.81 by the ALBERT+BiLSTM model.