A Machine Learning–Based Approach to Discrimination of Tauopathies Using [¹⁸F]PM-PBB3 PET Images

Background: We recently developed a positron emission tomography (PET) probe, [¹⁸F]PM-PBB3, to detect tau lesions in diverse tauopathies, including mixed three-repeat and four-repeat (3R + 4R) tau fibrils in Alzheimer's disease (AD) and 4R tau aggregates in progressive supranuclear palsy (PSP). For wider availability of this technology for clinical settings, bias-free quantitative evaluation of tau images without a priori disease information is needed. Objective: We aimed to establish tau PET pathology indices to characterize PSP and AD using a machine learning approach and test their validity and tracer capabilities. Methods: Data were obtained from 50 healthy control subjects, 46 patients with PSP Richardson syndrome, and 37 patients on the AD continuum. Tau PET data from 114 regions of interest were subjected to Elastic Net cross-validation linear classification analysis with a one-versus-the-rest multiclass strategy to obtain a linear function that discriminates diseases by maximizing the area under the receiver operating characteristic curve. We defined PSP- and AD-tau scores for each participant as values of the functions optimized for differentiating PSP (4R) and AD (3R + 4R), respectively, from others. Results: The discriminatory ability of PSP- and AD-tau scores assessed as the area under the receiver operating characteristic curve was 0.98 and 1.00, respectively. PSP-tau scores correlated with the PSP rating scale in patients with PSP, and AD-tau scores correlated with Mini-Mental State Examination scores in healthy control–AD continuum patients. The globus pallidus and amygdala were highlighted as regions with high weight coefficients for determining PSP- and AD-tau scores, respectively. Conclusions: These findings highlight our technology's unbiased capability to identify topologies of 3R + 4R versus 4R tau deposits.