Context: Alzheimer's disease (AD) and dementia with Lewy bodies (DLB) are the most common neurodegenerative dementia types. It is important to differentiate between them because of the differences in prognosis and treatment approaches.
Objective: Investigate if sparse partial least squares (SPLS) classification of cortical thickness measurements could differentiate between AD and DLB.
Methods: Two independent cohorts without MR-protocol alignment in Norway and Slovenia with 97 AD and DLB subjects were enrolled. Cortical thickness measurements acquired with Freesurfer were used in subsequent SPLS classification runs. The cohorts were analyzed separately and afterwards combined. The models were trained with leave-one-out cross validation and test datasets where used when available. To study the impact of MR-protocol alignment, the classifiers were additionally tested on sets drawn exclusively from the independent cohorts.
Results: The obtained sensitivity/specificity/AUC values were 94.4/88.89/0.978 and 88.2/94.1/0.969 in the Norwegian and Slovenian cohorts, respectively. Both cohorts showed AD-associated pattern of thinning in mid-anterior temporal, occipital and subgenual cingulate cortex, whereas the pattern supportive for DLB included thinning in dorsal cingulate, posterior temporal and lateral orbitofrontal regions. When combining the cohorts, sensitivity/specificity/AUC were 82.1/85.7/0.948 for the training and 77.8/75/0.731 for the testing datasets with the same pattern-of-difference. The models tested on datasets drawn exclusively from the independent cohorts did not produce adequate accuracy.
Conclusion: SPLS classification of cortical thickness is a good method for differentiating between AD and DLB, relatively stable even for mixed data, but not when tested on completely independent data drawn from different cohorts (without MR-protocol alignment).