Assisted Parkinsonism Diagnosis Using Multimodal MRI—The Role of Clinical Insights

Background: While automated methods for differential diagnosis of parkinsonian syndromes based on MRI imaging have been introduced, their implementation in clinical practice still underlies considerable challenges. Objective: To assess whether the performance of classifiers based on imaging derived biomarkers is improved with the addition of basic clinical information and to provide a practical solution to address the insecurity of classification results due to the uncertain clinical diagnosis they are based on. Methods: Retro- and prospectively collected data from multimodal MRI and standardized clinical datasets of 229 patients with PD (n = 167), PSP (n = 44), or MSA (n = 18) underwent multinomial classification in a benchmark study comparing the performance of nine machine learning methods. A predictor space of imaging variables, either with or without clinical information, was investigated. Classification results were assessed using multiclass AUCs. Individual predicted probabilities were visualized to address diagnostic uncertainty. Results: Clinical diagnosis was accurately confirmed using machine learning models with only small differences when using imaging and clinical signs versus imaging variables only (expected multiclass AUC of 0.95 vs. 0.92). Still, multinomial classification is hampered by imbalanced class frequencies. The most discriminatory variables were responsiveness to levodopa, vertical gaze palsy, and the volumes of subcortical structures, including the red nucleus. Conclusion: Machine-learning-assisted classification of MR-imaging biomarkers gathered in routine care can assist in the diagnosis of parkinsonian syndromes as part of the diagnostic workup. We provide a visual method that aids the interpretation of neuroimaging-based classification results of the three main parkinsonian syndromes, improving clinical interpretability.