Machine learning prediction of dementia conversion in mild cognitive impairment: A two- to six-year follow-up study

OBJECTIVES: Mild cognitive impairment (MCI) involves measurable cognitive decline that does not yet significantly disrupt daily functioning but may signal increased risk of dementia. Reliable prediction of dementia conversion in MCI is essential for early intervention and optimized clinical trial design. This study aimed to evaluate the predictive performance of various machine learning (ML) classification algorithms using clinical and neuropsychological data. METHODS: Data were drawn from the Gothenburg MCI Study and included 347 patients from a memory clinic, of whom 84 (24%) converted to dementia within two to six years. We applied 11 ML classification algorithms (logistic regression, linear discriminant analysis, naïve Bayes, k-nearest neighbors, LASSO, ridge regression, elastic net, decision tree, random forest, gradient boosting, and support vector machine (SVM)) to predict dementia conversion based on 54 clinical predictors (e.g., cerebrospinal fluid biomarkers, neuropsychological test scores, comorbidities, and demographics). In a second step, we included delta scores reflecting change in neuropsychological test performance from baseline to follow-up. RESULTS: Without delta scores, LASSO, ridge, elastic net, random forest, and SVM performed best, achieving accuracy ≥0.87, kappa = 0.64, and AUC-ROC ≥0.90. These models demonstrated high specificity (0.94) but moderate sensitivity (0.68). Including delta scores improved performance, with ridge and elastic net achieving accuracy = 0.90, kappa = 0.73 and 0.72, AUC-ROC = 0.94, specificity = 0.96, and sensitivity = 0.73. The elastic net model yielded a positive predictive value of 0.85 and a negative predictive value of 0.92. CONCLUSIONS: ML models incorporating clinical and cognitive change data can accurately predict dementia conversion in MCI, supporting their utility in clinical decision-making.