Predicting conversion in cognitively normal and mild cognitive impairment individuals with machine learning: Is the...

Area of Science:

• Neurology
• Artificial Intelligence
• Biomarker Research

Background:

• Machine learning (ML) aids in diagnosing the spectrum of mild cognitive impairment-Alzheimer's disease (MCI-AD).
• Current ML models often require data not routinely available in clinical settings.
• This study introduces a novel ML approach using standard clinical data to predict cognitive decline.

Purpose of the Study:

• To develop and validate a multi-step ML model for predicting cognitive worsening in individuals at risk for Alzheimer's disease.
• To assess the model's predictive accuracy using standard clinical data (SCD).
• To investigate the impact of cerebrospinal fluid (CSF) biomarkers on prediction accuracy and understand the model's decision-making process.

Main Methods:

• Participants with normal cognition and MCI were selected from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
• Participants were categorized based on total tau/amyloid beta 1-42 ratios.
• A multi-step ML approach was employed to predict 3-year conversion, utilizing SCD and CSF biomarker information. Shapley Additive Explanations (SHAP) analysis was performed.

Main Results:

• The ML model achieved an overall accuracy of 84% in predicting cognitive worsening.
• Accuracy was higher in subgroups: 86% for patients with negative CSF and 88% for those with AD-like CSF.
• SHAP analysis revealed distinct predictors and cut-offs for conversion between CSF-positive and CSF-negative groups.

Conclusions:

• The developed ML approach demonstrates good predictive accuracy for cognitive decline using standard clinical data.
• Categorization based on CSF biomarkers significantly enhances the predictive accuracy of the model.
• Optimizing cut-offs for neuropsychological tests could further improve the prediction of cognitive conversion.