Development and External Validation of a Machine Learning Model for Classification of Mild Cognitive Impairment and...

Background and Objectives: As society ages, the number of patients with cognitive impairment is increasing. Machine learning methods that use structured clinical and cognitive-assessment data during routine diagnostic work-up may support and monitor structured classifications of cognitive status. This kind of approach can improve early screening, reduce physicians' workload and develop greater support for personalized treatment. To develop an XGBoost-based machine learning model using the National Alzheimer's Coordinating Center (NACC) dataset and to evaluate the model's precision with clinician-assigned diagnosis in a Latvian retrospective cohort study. Materials and Methods: The research was designed as a retrospective external validation cohort study that used two data sources. Firstly, the National Alzheimer's Coordination Center (NACC) longitudinal dataset was used to train the ML model. Secondly, medical records gathered from Pauls Stradins Clinical University Hospital dating from 2020 to May 2025 were used to evaluate the algorithm's precision. Results: In the NACC study, the weighted four-class model achieved an overall accuracy of 84.0% and a balanced accuracy of 70.9%, but the SCD class remained poorly classified. After reframing the model to a three-class model the performance grew stronger for normal cognition, mild cognitive impairment (MCI) and dementia. Class distribution in the Latvian cohort consisted of dementia (n = 138); MCI (n = 13); and subjective cognitive decline (SCD) (n = 2). Dementia was identified most strongly-124/138 (sensitivity-89.9%). MCI was correct in 9/13 cases (sensitivity-69.2%). SCD cases were excluded. Overall, the model agreed with the neurologist-assigned diagnoses in 88.1% of the cases (133/151). Conclusions: The ML classification model has high precision when comparing with neurologist-assigned diagnoses, but it struggles to separate adjacent early-stage diagnoses, meaning that it did not reliably identify SCD. These findings support further methodological development and the implementation of prospective research. Nevertheless, this technology has high potential for being integrated in the future to aid triage and early screening, especially when advanced diagnostics are limited.