Preprint · Dec 27, 2024
Digitizing Cognitive Function Assessment through Smartphone Integration with a Deep Learning Model for Scoring the Intersecting Pentagon (Preprint)
By Yu CY, Ma W, Liao YZ et al.
BACKGROUND Cognitive function assessment is pivotal for evaluating neurodegenerative disorders. Traditional assessments like the Mini-Mental State Examination (MMSE) are clinician-administered and can be time-consuming. Machine learning offers potential advantages in accuracy, efficiency, and accessibility for these assessments. OBJECTIVE To develop and validate a deep learning-based approach for digitizing and automating the scoring of the intersecting pentagon test, and to integrate the model into a smartphone application for efficient and accessible cognitive assessments of neurodegenerative disorders. METHODS The 926 scanned MMSE records for this study were obtained from the geriatric psychiatry and neurology outpatient clinic at Taipei Veterans General Hospital, Taiwan. The MMSE paper test data was processed, particularly focusing on the intersecting pentagon item, which was then converted into a standardized image format. Various deep learning models based on convolutional neural networks were trained on this data. Additionally, the model was integrated into a smartphone application, allowing real-time assessment. RESULTS The analysis identified a significant correlation between the intersecting pentagon score and key variables including age, years of education, and overall MMSE score, alongside most MMSE sub-items. The MobileNet model emerged as the most effective for scoring the intersecting pentagon, achieving a F1 score of 0.92 with recall and precision rates of 0.96 and 0.89, respectively. Furthermore, it showed a mean absolute error of 2.22 and a R-squared of 0.87 when predicting MMSE scores based on the pentagon drawing, age, and years of education. CONCLUSIONS This study highlights the benefits of applying deep learning to digitize the MMSE, particularly through automated scoring of the intersecting pentagon test. By integrating demographic information, we've shown a deep learning approach to predict MMSE scores solely based on the pentagon drawing. The model's integration into a smartphone app underlines the practicality of digital tools in cognitive assessments, enhancing accessibility and efficiency in evaluating neurodegenerative disorders.