Objective: To determine how multimodal neuroimaging data can contribute to the classification of Parkinson’s disease (PD) patients with mild cognitive impairment (MCI) from PD without MCI using multivariate pattern analysis, a machine learning (ML) technique.

Background: Cognitive decline is one of common non-motor symptoms in PD, one quarter of patients are experiencing this during the course of their disease. Most of the studies use group-level analysis exploring neurobiological markers of PD with MCI, which have limitation when planning clinical application of findings in individual patient level. Recent studies suggested multivariate pattern recognition technique, a subfield of ML, can be applied to analyze neuroimaging data overcoming the limitations of group-level analysis.

Method: In total, 23 PD patients (11 PD non-MCI and 12 PD with MCI) and 12 age-matched healthy controls (HC) participated and completed [¹¹C]DTBZ PET scan and structural MRI scan (3.0 T) followed by the neuropsychological test. MCI were identified based on neuropsychological test results. Availability of vesicular monoamine transporter 2 (VMAT2) that measured by [¹¹C]DTBZ PET were obtained from three striatal subregion (motor, associative and ventral striatum) in each hemisphere and grey matter volume measure from Broadman map that is obtained from T1 MRI were used as input. The group classification was tested using support vector machine, a supervised ML algorithm.

Results: With VMAT 2 level input, 100% of accuracy was obtained for separating PD from HC and the motor striatum is the most contributed neuroanatomical biomarker, while only 69.3% of accuracy were obtained for classifying PD with MCI from PD non-MCI. With volumetric data input, the accuracy for classification of PD from HC was only 47.28 %, however PD with MCI from PD non-MCI can be classified with 78.4% accuracy which is superior than the performance of VMAT 2 input. The right anterior entorhinal cortex (BA 34) contributed most for this classification.

Conclusion: Our study showed that measure of dopamine level is superior input for separation of PD from HC while cortical volume data may be superior for separating PD with non-motor symptoms from idiopathic PD. Our findings suggest that use of multimodal neuroimaging data in ML approach possibly improve classification and diagnosis power when the pathological markers of disease subtype are heterogeneous like PD and its non-motor symptoms.

« Back to 2019 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/classify-parkinsons-disease-with-mild-cognitive-impairment-machine-learning-approach/