Machine Learning‐Based Brief Screening Instrument for Dementia: “Six‐Questions Based Dementia Screen Test”
Kuan‐Ying Li, Tang‐Wei Huang, Chen‐Wen Yen, Ching‐ Fang Chien, Ling‐Chun Huang, Yuan‐Han Yang- Psychiatry and Mental health
- Cellular and Molecular Neuroscience
- Geriatrics and Gerontology
- Neurology (clinical)
- Developmental Neuroscience
- Health Policy
- Epidemiology
Abstract
Background
Machine learning (ML) applied to healthcare is increasingly and it provides innovative solutions to solve complex real‐world problems. In dementia, ML has been applied to dementia detection, discriminating different types of dementia, and prediction progression of diseases. The aim of the study was to develop a new, brief clinical tool based on machine learning for dementia screening.
Method
The training data was from 533 participants (311 dementia and 202 non‐demented participants) screened with “Six‐Question Dementia Screening Test” (6Q‐DS). The 6Q‐DS contained six simple questions to evaluate mood, memory, orientation, and concentration. We used e‐Xtreme Gradient Boosting (XGBoost) algorithm, an ensemble ML based on decision trees, to establish the prediction model for dementia.
Result
A total 533 participants, 186 men and 347 women with mean age 77.6 ± 8.2 (mean± SD) years, were recruited. The dementia group was significantly older and had lower education years compared to non‐dementia group (p < 0.01 for both). The receiver‐operator characteristic (ROC) analysis showed that the area under the curve (AUC) was 0.91 for discriminating non‐dementia vs. dementia, with the sensitivity as 0.86, specificity 0.81 and accuracy 0.84. We found that the top 3 features are the second calculations for serial 3’s test, feeling of depression and the third calculations for serial 3’s test. We hypothesized that the second and third calculations required more cognitive demands because of holding information longer and depression is frequent in dementia. In our sixth question “Please count down 100 by 3 for 5 times (Serial 3’s)”, we found that the second and third calculations for serial 3’s test were important features to discriminating dementia from non‐dementia. We also found that the feeling of depression is the second important feature to distinguish dementia in our data set.
Conclusion
We proposed the new, ease‐to‐use instrument with good classification of screening dementia based on ML. The 6Q‐DS showed an acceptable accuracy with a relatively high sensitivity and specificity. Further works are warranted to validate the model in different clinical settings and disease severity.