Search

In this article, we developed an interview framework and natural language processing model for estimating cognitive function, based on an intake interview with psychologists in a hospital setting. The questionnaire consisted of 30 questions in five categories. To evaluate the developed interview items and the accuracy of the natural language processing model, we recruited participants with the approval of the University of Tokyo Hospital and obtained the cooperation of 29 participants (7 men and 22 women) aged 72–91 years. Based on the MMSE results, a multilevel classification model was created to classify the three groups, and a binary classification model to sort the two groups. For each of these models, we tested whether the accuracy would improve when text augmentation was performed. The accuracy in the multi-level classification results for the test data was 0.405 without augmentation and 0.991 with augmentation. The accuracy of the test data in the results of the binary classification without augmentation was 0.488 for the moderate dementia and mild dementia groups, 0.767 for the moderate dementia and MCI groups, and 0.700 for the mild dementia and MCI groups. In contrast, the accuracy of the test data in the augmented binary classification results was 0.972 for moderate dementia and mild dementia groups, 0.996 for moderate dementia and MCI groups, and 0.985 for mild dementia and MCI groups.

The number of people with dementia is increasing each year, and early detection allows for early intervention and treatment. Since conventional screening methods are time-consuming and expensive, a simple and inexpensive screening is expected. We created a standardized intake questionnaire with thirty questions in five categories and used machine learning to categorize older adults with moderate and mild dementia and mild cognitive impairment, based on speech patterns. To evaluate the feasibility of the developed interview items and the accuracy of the classification model based on acoustic features, 29 participants (7 males and 22 females) aged 72 to 91 years were recruited with the approval of the University of Tokyo Hospital. The MMSE results showed that 12 participants had moderate dementia with MMSE scores of 20 or less, 8 participants had mild dementia with MMSE scores between 21 and 23, and 9 participants had MCI with MMSE scores between 24 and 27. As a result, Mel-spectrogram generally outperformed MFCC in terms of accuracy, precision, recall, and F1-score in all classification tasks. The multi-classification using Mel-spectrogram achieved the highest accuracy of 0.932, while the binary classification of moderate dementia and MCI group using MFCC achieved the lowest accuracy of 0.502. The FDR was generally low for all classification tasks, indicating a low rate of false positives. However, the FNR was relatively high in some cases, indicating a higher rate of false negatives.