Machine learning model for predicting the conversion to dementia using the Cube Copying Test

Background Early detection of dementia requires highly accurate and efficient screening tests that minimize patient burden. Objective To develop a machine learning model predicting dementia conversion within 3–5 years using Cube Copying Test (CCT) drawings at baseline. Methods This retrospective study analyzed CCT drawing data from 767 patients at the Center for Comprehensive Care and Research on Memory Disorders (2011–2020). Of the 2303 patients who met the inclusion criteria, 534 were excluded due to mild cognitive impairment (MCI) persistence, pending diagnoses, or new neurovascular diseases, while 1002 were lost to follow-up. Eligibility criteria included a baseline Mini-Mental State Examination (MMSE) score ≥24, absence of dementia diagnosis or anti-dementia medication intake, and completion of a 3–5-year follow-up without meeting exclusion criteria. Results Of 767 patients, 457 converted to dementia (318 with Alzheimer's disease, 116 with dementia with Lewy bodies, and 23 with frontotemporal dementia) within 3–5 years, while 310 did not. The model achieved an area under the curve of 0.85 for predicting dementia conversion. Shapley Additive exPlanations analysis identified PatchCore-derived features as the strongest predictors, distinguishing drawing patterns of converters and non-converters. Conclusions In patients who convert to Alzheimer's disease, dementia with Lewy bodies, or frontotemporal dementia, the very early stages of constructional apraxia-like symptoms already exist at the preclinical stage or MCI stage. Applying deep learning-based anomaly-detection models can detect these early drawing distortions that differ from normal aging and contribute to improving the performance of dementia-conversion prediction.