Semiquantitative [¹²³I]FP-CIT SPECT metrics combined with machine learning improve clinical differentiation of Parkinson's disease and atypical parkinsonian syndrome.
Source: PubMed, NCBI / U.S. National Library of Medicine
To evaluate whether semiquantitative striatal [¹²³I]FP-CIT SPECT-derived metrics improve clinical differentiation of degenerative parkinsonism using an integrated machine learning approach. This cross-sectional study included 487 patients with Parkinson's disease (PD) and 219 with atypical parkinsonisms (APS), classified as progressive supranuclear palsy (PSP, n = 127), multiple system atrophy parkinsonian type (MSA-P, n = 37), multiple system atrophy cerebellar type (MSA-C, n = 12), and corticobasal degeneration (CBD, n = 43). All participants underwent a [¹²³I]FP-CIT SPECT. Striatal [¹²³I]FP-CIT uptake was quantified using anatomical (caudate, putamen, ventral striatum) and functional (limbic, executive, sensorimotor) parcellations to calculate specific binding ratios, asymmetry indices, and inter-regional ratios. Discriminative performance of each metric was evaluated using receiver operating characteristic (ROC) curves analyses. A random forest classifier integrating all semiquantitative metrics was trained and validated, enabling data-driven identification diagnostic pathways. Caudate-to-putamen and sensorimotor-to-limbic inter-regional ratios showed the strongest discriminative performance (AUC up to 0.95) for differentiating PD from APS. The random forest achieved a 64% overall accuracy with high per-class specificities (> 84%) and revealed two diagnostic pathway
Abstract
To evaluate whether semiquantitative striatal [¹²³I]FP-CIT SPECT-derived metrics improve clinical differentiation of degenerative parkinsonism using an integrated machine learning approach. This cross-sectional study included 487 patients with Parkinson's disease (PD) and 219 with atypical parkinsonisms (APS), classified as progressive supranuclear palsy (PSP, n = 127), multiple system atrophy parkinsonian type (MSA-P, n = 37), multiple system atrophy cerebellar type (MSA-C, n = 12), and corticobasal degeneration (CBD, n = 43). All participants underwent a [¹²³I]FP-CIT SPECT. Striatal [¹²³I]FP-CIT uptake was quantified using anatomical (caudate, putamen, ventral striatum) and functional (limbic, executive, sensorimotor) parcellations to calculate specific binding ratios, asymmetry indices, and inter-regional ratios. Discriminative performance of each metric was evaluated using receiver operating characteristic (ROC) curves analyses. A random forest classifier integrating all semiquantitative metrics was trained and validated, enabling data-driven identification diagnostic pathways. Caudate-to-putamen and sensorimotor-to-limbic inter-regional ratios showed the strongest discriminative performance (AUC up to 0.95) for differentiating PD from APS. The random forest achieved a 64% overall accuracy with high per-class specificities (> 84%) and revealed two diagnostic pathways. A lower caudate-to-posterior putamen ratio, primarily grouped PSP, CBD and MSA-C, where lower contralateral sensorimotor uptake pointing to PSP while higher values and ipsilateral caudate uptake subregion further distinguishing CBD from MSA-C. A higher caudate-to-posterior putamen ratio, included PD and MSA-P, where lower ipsilateral caudate uptake together with a higher sensorimotor-to-cognitive ratio mainly differentiate both groups. Integrating anatomical and functional [¹²³I]FP-CIT SPECT metrics within a machine learning framework enhances the clinical differentiation of degenerative parkinsonisms and supports [¹²³I]FP-CIT SPECT as a robust in vivo disease biomarker.