Prediction of the conversion from subjective cognitive decline to mild cognitive impairment or Alzheimer’s disease by integrating imputation techniques and machine learning based on multimodal clinical data

Introduction: Subjective Cognitive Decline (SCD) is an early indicator of Alzheimer's disease (AD), and accurate identification of high-risk individuals is essential for timely intervention. Predicting the progression of SCD remains challenging, particularly in resource-limited clinical settings where advanced neuroimaging is not routinely available. Methods: We developed a predictive framework to forecast 10-year conversion using multimodal clinical data, including fluid biomarkers, cognitive assessments, apolipoprotein E4 (APOE4), and demographics, utilizing 2423 visit-level samples from 635 participants in the Alzheimer's Disease Neuroimaging Initiative (ADNI). After participant-level splitting and preprocessing for missingness and redundancy, 7 machine learning models and 2 heterogeneous ensembles were constructed and evaluated using standard performance metrics. Finally, the main model was further interpreted using calibration analysis, SHapley Additive exPlanations (SHAP)-based interpretation, and robustness analysis. Results: Support vector machine trained on plasma-based data with MICE imputation showed the best balanced performance (balanced accuracy=0.70, sensitivity=0.72, specificity=0.68, F1-score = 0.50, MCC=0.33, AUC=0.73). SHAP revealed executive function, episodic memory, plasma A β 40 and p-tau, and APOE4 as key predictors. Discussion: Our framework identifies individuals at high risk of SCD progression using cost-effective, nonimaging data. This makes it scalable to primary care settings and underserved populations.