Development and validation of the automated imaging differentiation in parkinsonism (AID-P): a multicentre machine learning study

Abstract

Background Development of valid, non-invasive biomarkers for parkinsonian syndromes is crucially needed. We aimed to assess whether non-invasive diffusion-weighted MRI can distinguish between parkinsonian syndromes using an automated imaging approach. MethodsWe did an international study at 17 MRI centres in Austria, Germany, and the USA. We used diffusion-weighted MRI from 1002 patients and the Movement Disorders Society Unified Parkinson's Disease Rating Scale part III (MDS-UPDRS III) to develop and validate disease-specific machine learning comparisons using 60 template regions and tracts of interest in Montreal Neurological Institute space between Parkinson's disease and atypical parkinsonism (multiple system atrophy and progressive supranuclear palsy) and between multiple system atrophy and progressive supranuclear palsy. For each comparison, models were developed on a training and validation cohort and evaluated in an independent test cohort by quantifying the area under the curve (AUC) of receiving operating characteristic curves. The primary outcomes were free water and free-water-corrected fractional anisotropy across 60 different template regions. Findings In the test cohort for disease-specific comparisons, the diffusion-weighted MRI plus MDS-UPDRS III model (Parkinson's disease vs atypical parkinsonism had an AUC 0•962; multiple system atrophy vs progressive supranuclear palsy AUC 0•897) and diffusion-weighted MRI only model had high AUCs (Parkinson's disease vs atypical parkinsonism AUC 0•955; multiple system atrophy vs progressive supranuclear palsy AUC 0•926), whereas the MDS-UPDRS III only models had significantly lower AUCs (Parkinson's disease vs atypical parkinsonism 0•775; multiple system atrophy vs progressive supranuclear palsy 0•582). These results indicate that a non-invasive imaging approach is capable of differentiating forms of parkinsonism comparable to current gold standard methods.Interpretations This study provides an objective, validated, and generalisable imaging approach to distinguish different forms of parkinsonian syndromes using multisite diffusion-weighted MRI cohorts. The diffusion-weighted MRI method does not involve radioactive tracers, is completely automated, and can be collected in less than 12 min across 3T scanners worldwide. The use of this test could positively affect the clinical care of patients with Parkinson's disease and parkinsonism and reduce the number of misdiagnosed cases in clinical trials.

Automated classification of neurodegenerative parkinsonian syndromes using multimodal magnetic resonance imaging in a clinical setting

“…Improvement in DTI acquisition parameters, especially standardization of the resolution and slice gap, may improve classification performances as suggested previously. 25 In line with previous pathological and imaging studies, the best features for the differentiation between PD and PSP were midbrain and third ventricle volumes 6,7,13,17,28,29 and FA in the SCP. 17,[29][30][31] The putamen volume was highly discriminant between MSA-P and PD, PSP and MSA-C patients in our study, MSA-P being characterized by a prominent putamen atrophy.…”

“…Future work could involve calculating free-water and free-water-corrected measurements in clinical protocols for improved classification accuracy. 25 In conclusion, our study showed that an automated categorization of parkinsonian syndromes was applicable to patients with early to moderately advanced parkinsonism recruited in a clinical environment, despite the variability in scanners and acquisition parameters, volumetry being a robust discriminative biomarker. Medical centers could benefit from such an approach in order to increase diagnostic accuracy and patient management.…”

“…[13][14][15][16][17] Recently, several studies have demonstrated that machine learning algorithms trained with MRI data could differentiate between parkinsonian syndromes with high accuracy. [18][19][20][21][22][23][24][25] Nevertheless, most of these studies were based on one single type of MRI acquisition protocol, either T1-weighted volumetry [18][19][20] or diffusion-weighted data, 25 with three studies using a multimodal approach combining volumetry and diffusion 21 or including R2* relaxometry 22 or spectroscopy. 23 Two studies have relied on large cohorts, including 1002 25 or 464 subjects, 19 whereas others studies have investigated smaller samples.…”

“…[18][19][20][21][22][23][24][25] Nevertheless, most of these studies were based on one single type of MRI acquisition protocol, either T1-weighted volumetry [18][19][20] or diffusion-weighted data, 25 with three studies using a multimodal approach combining volumetry and diffusion 21 or including R2* relaxometry 22 or spectroscopy. 23 Two studies have relied on large cohorts, including 1002 25 or 464 subjects, 19 whereas others studies have investigated smaller samples. 18,[20][21][22][23] Only two studies have included at the same time PD, PSP, MSA-P and MSA-C subjects, 19,23 while other studies have not differentiated between MSA-P and MSA-C, 18 have only included MSA-P patients 25 or MSA-C patients 24 or have not included MSA 20,21 or PSP patients.…”

“…23 Two studies have relied on large cohorts, including 1002 25 or 464 subjects, 19 whereas others studies have investigated smaller samples. 18,[20][21][22][23] Only two studies have included at the same time PD, PSP, MSA-P and MSA-C subjects, 19,23 while other studies have not differentiated between MSA-P and MSA-C, 18 have only included MSA-P patients 25 or MSA-C patients 24 or have not included MSA 20,21 or PSP patients. 22 Moreover, these studies have mainly been designed in a research environment and tested without an independent replication cohort.…”