Objective: To assess the feasibility of automated classification of different parkinsonian subtypes using clinical and imaging variables in a single center study.

Background: The clinical differential diagnosis of parkinsonian syndromes can be challenging. Imaging markers, both qualitative and quantitative, have provided assistance. For the latter, machine-learning approaches have been employed already (1,2). The combination of clinical signs and imaging markers, however, has not been assessed to date.

Method: Data on 253 patients was available with grossly different numbers of patients for each disease entity. 20 clinical traits of parkinsonian syndromes dichotomized to „present: yes/no“, and 2190 ROI-based imaging markers derived from T1- and DTI-imaging were fed into 10 different machine-learning-models capable of dealing with multinomial outcomes and missing values. The algorithms were compared using three-fold nested cross-validation.

Results: The winning models were random forest classifiers and boosted regression trees yielding a mean AUC of 0.83. AUCs ranged between 0.84 and 0.89 for the classification of IPS vs. all, 0.83 and 0.88 for PSP vs all, 0.84 and 0.94 for MSA vs. all and 0.61 and 0.91 for CBS vs. all. Model calibration showed high concordance of predicted probabilities and observed frequencies of entities. The variables most important for classification were responsiveness to L-dopa, volume of the middle cerebellar peduncle, pontine volume and presence of vertical gaze paresis.

Conclusion: A number of machine-learning algorithms were simultaneously assessed using a dataset composed of clinical and imaging variables. Support in the diagnosis of one parkinsonian syndrome vs. all other syndromes is possible if the goal is the classification of a patient into one particular syndrome vs. all other entities.

References: (1) Huppertz HJ et al., „Differentiation of neurodegenerative parkinsonian syndromes by volumetric magnetic resonance imaging analysis and support vector machine classification“, Mov Disord. 2016 Oct;31(10) (2) Scherfler C et al., „Diagnostic potential of automated subcortical volume segmentation in atypical parkinsonism“, Neurology. 2016 Mar 29

« Back to 2019 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/differential-diagnosis-of-parkinsonian-syndromes-combining-clinical-and-imaging-data-in-a-machine-learning-approach/