Utility of machine learning algorithms in classification of progressive cognitive impairment in Alzheimer's disease: A retrospective cohort based on China

Abstract

Background

Distinct risk factors influence Alzheimer's disease (AD) stage stratification, yet effective tools for early diagnosis and prognosis remain limited, especially in middle-aged populations.

Objective

To develop machine learning models for predicting cognitive decline and identifying early markers of stage stratification in a middle-aged Chinese cohort.

Methods

We conducted a retrospective study on 451 patients from 2017 to 2021 (aged 45–90 years, 47.7% male). All participants were classified into normal, mild cognitive impairment (MCI), AD. Neuropsychological scale, epidemiological and laboratory parameters were collected. Four machine learning algorithms, the Least Absolute Shrinkage and Selection Operator (LASSO) regression, Random Forest, Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were employed with 10-fold cross-validation. Model performance was measured using area under the receiver operating characteristics curve (ROC-AUC) and area under precision and recall curves (PR-AUC), classification confusion matrices, sensitivity, accuracy, precision, recall, F1 Score.

Results

Models demonstrated high ROC-AUC and satisfactory PR-AUC, with LASSO and SVM excelling in the MCI group (recall: 85.3% and 93.1%; F1 score: 78.4% and 78.3%, respectively). Mini-Mental State Examination (MMSE) scores differed significantly across stages, except for advanced-stage items such as naming, language repetition, and language understanding.

Conclusions

These multi-dimensional machine learning models show promise as effective tools for predicting AD stage stratification, enabling targeted monitoring and early intervention for at-risk patients.