Machine learning for predicting dementia and mild cognitive impairment in the heliad cohort: Contrasting self-reports and clinical signs

In this study we evaluated the ability of Random Forest and XGBoost classifiers to discriminate among dementia, mild cognitive impairment (MCI), and any cognitive diagnosis (dementia or MCI) using data from 2,062 community‐dwelling adults aged 65 and older in the HELIAD cohort. Models were trained on four feature subsets: (A) self-and caregiver-reported measures, (B) combined medical and neuropsychological assessments, (C) neuropsychological assessments alone, and (D) the union of A and B. Three binary classification tasks were defined: dementia versus no dementia, MCI versus no MCI, and any diagnosis versus no diagnosis. In order to address class imbalance, data preprocessing techniques included SMOTE oversampling, stratified 80:20 train–test splits, median imputation, and recoding missing values. The final models were assessed on hold-out data after hyperparameters were adjusted using randomized search optimization of the positive-class F-score. Both algorithms achieved near-perfect discrimination using self-reported and informant-reported features (test F₁ ≥ 0.82; AUC‐ROC ≥ 0.995). The most significant predictors, according to feature importance analyses, were Clinical Dementia Rating memory, personal care, community affairs, and judgment scores, as well as reported symptom onset. On the other hand, models trained on medical and neuropsychological variables performed worse (F₁ 0.28–0.68; AUC‐ROC 0.73–0.91), with the greatest contribution coming from neuropsychological measures like semantic fluency, list-learning, and MMSE. These results highlight the importance of structured subjective measures of functioning for effective cognitive impairment screening at the population level.