Comparison of machine learning algorithms for predicting cognitive impairment using neuropsychological tests

Area of Science:

• Computational neuroscience
• Medical informatics
• Cognitive psychology

Background:

• Neuropsychological tests (NPTs) are crucial for assessing cognitive function but can be time-consuming and costly to interpret.
• Developing efficient methods for analyzing NPTs is essential for timely diagnosis of cognitive decline.
• Machine learning offers a promising approach to optimize NPT interpretation and classification.

Purpose of the Study:

• To optimize systematic NPTs using machine learning.
• To develop classification models for differentiating healthy controls (HC), mild cognitive impairment (MCI), and Alzheimer's disease dementia (ADD).
• To evaluate the performance of various machine learning algorithms for cognitive impairment classification.

Main Methods:

• A dataset of 14,926 subjects from 46 NPTs (Seoul Neuropsychological Screening Battery - SNSB) was utilized.
• Machine learning classification (two- and three-way) was performed using scikit-learn.
• Seven algorithms (NB, RF, DT, KNN, SVM, AdaBoost, LDA) were compared based on accuracy, sensitivity, specificity, PPV, NPV, and AUC.

Main Results:

• The random forest (RF) algorithm demonstrated superior performance across all classification tasks.
• RF achieved high accuracy, sensitivity, specificity, PPV, and NPV, with AUCs up to 99% for differentiating HC, MCI, and ADD.
• The models, trained on 29 selected NPT features, showed excellent predictive capabilities.

Conclusions:

• The RF algorithm is the optimal classification model for NPT data, outperforming other algorithms.
• These machine learning models can effectively diagnose MCI and ADD, even with normal NPT results.
• The developed models enhance diagnostic accuracy, optimize cognitive evaluation, and minimize missed diagnoses.