We developed and validated an adaptive machine-learning–based algorithmic framework that enables personalized differential diagnosis of neurodegenerative diseases according to specific clinical questions. Differentiating neurodegenerative disorders presenting with cognitive decline or movement disorders, particularly in their early stages, remains a major clinical challenge. While 18F-FDG PET imaging substantially improves diagnostic accuracy, it requires highly trained specialists for accurate image interpretation. In contrast to previous approaches that address predefined diagnostic comparisons (e.g., Parkinson’s disease versus multiple system atrophy), the proposed algorithmic framework dynamically adapts to the specific clinical question and the composition of the available data, allowing for a more individualized diagnostic approach for each patient. For model development, we used a database of 18F-FDG PET brain images from patients with different neurodegenerative disorders examined at the University Medical Centre Ljubljana, including individuals with Alzheimer’s disease (AD), the behavioral variant of frontotemporal dementia (bvFTD), dementia with Lewy bodies (DLB), Creutzfeldt–Jakob disease (CJD), Parkinson’s disease (PD), multiple system atrophy (MSA), progressive supranuclear palsy (PSP), corticobasal degeneration (CBD), as well as healthy controls. Each case was characterized using the expression of different disease-specific metabolic patterns (e.g., ADRP, FTDRP, PDRP, MSARP, PSPRP). The algorithmic framework incorporates multiple machine-learning classifiers, including logistic regression, support vector machines, and naïve Bayes. Model performance was evaluated using leave-one-out cross-validation and assessed with standard metrics such as the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. The primary goal of this study was not to achieve maximal diagnostic accuracy of individual models, but rather to develop a context sensitive diagnostic framework that also accounts for uncertainty and prediction stability across different clinical scenarios.
|
