TensorFlow

TensorFlow implements machine learning methods for building and training neural networks to analyze biological and clinical data, including prediction of cognitive impairment from neuropsychological test data.

Key Features:

• Artificial neural network training: Training of artificial neural network models was used to classify cognitive states (NC, MCI, ADD).
• Multi-center neuropsychological data support: Models were trained on multi-center neuropsychological test data from the Seoul Neuropsychological Screening Battery comprising 14,926 formal assessments.
• High-dimensional input handling: Models were trained using a dataset comprising 46 variables derived from neuropsychological assessments.
• Model evaluation metrics: Performance was quantified using accuracy, sensitivity, and specificity across evaluated datasets.
• Cross-dataset evaluation: Model performance was assessed across ten randomly selected datasets, including balanced and clinic-based cohorts.
• Feature selection with Recursive Feature Elimination (RFE): RFE identified 'time orientation' and '3-word recall' from the Mini-Mental State Examination (MMSE) as highly predictive features and enabled reduction to a 12-variable subset including age and education.
• Reported predictive performance: Reported accuracies included 96.66 ± 0.52% (balanced dataset) and 97.23 ± 0.32% (clinic-based dataset) for detecting cognitive impairment, and 95.49 ± 0.53% and 96.34 ± 1.03% for classifying NC, MCI, or ADD.

Scientific Applications:

• Cognitive impairment prediction: Predicting presence of cognitive impairment (MCI and Alzheimer's disease dementia) from neuropsychological test scores.
• State classification: Distinguishing normal cognition (NC), mild cognitive impairment (MCI), and Alzheimer's disease dementia (ADD).
• Diagnostic feature identification: Identifying MMSE items (time orientation and 3-word recall) and a reduced feature subset as predictive markers for cognitive impairment.
• Clinical performance assessment: Evaluating sensitivity and specificity of automated classifiers in balanced and clinic-based datasets.

Methodology:

An artificial neural network was trained in TensorFlow on 46 variables from 14,926 Seoul Neuropsychological Screening Battery assessments, evaluated across ten randomly selected datasets with accuracy, sensitivity, and specificity reported, and feature selection via Recursive Feature Elimination reduced the inputs to a 12-variable subset including age and education.