Hybrid Machine Learning Methods for Robust Identification of Parkinson’s Disease Subtypes

Objectives: It is important to subdivide Parkinson’s disease (PD) into specific subtypes, since homogeneous groups of patients are more likely to share genetic and pathological features, enabling potentially earlier disease recognition and more tailored treatment strategies.

Methods: We aimed to identify PD subtypes by advanced hybrid machine learning (ML) methods that are robust to variations in the number of subjects (n) and features (p). We applied multiple feature reduction and cluster analysis methods to cross-sectional and timeless data, derived from longitudinal datasets (years 0, 1, 2 & 4; Parkinson’s Progressive Marker Initiative). Segmentation of dorsal striatum (DS) on DaT SPECT images was performed: (i) via MRI, and (ii) directly on SPECT. We considered 15 datasets: 5 with solely non-imaging clinical information (1 timeless; 4 cross-sectional), and also including SPECT images segmented using MRI (5 sets) or SPECT itself (5 sets). The timeless or collective cross-sectional data consisted of ~900 studies. Radiomic features of DS were extracted using our standardized SERA software. Hybrid ML systems were constructed invoking: 16 feature reduction algorithms (FRAs), 8 clustering algorithms (CAs) and 19 classifiers (Cs), optimized via automated ML hyperparameter tuning and 5-fold cross-validation. Homogeneity & separation clustering evaluation method (CEA) was used on each trajectory (hybrid system) for a range of 2-10 clusters. We selected optimal subtypes through our modified information criterion (MIC): likelihood consisting of clustering performance (distance of each subject to center of cluster) coupled with either Average of Classifier Performance (AOCP) or Average of Correlation Factor (AOCF). In cross-sectional analysis, we cross-linked subgroups obtained from each year via multiple classifiers based on training and testing results (details below). Our findings were further confirmed by a statistical approach (High Dimensional Hotelling’s T² Test).

Results: Initially selected disease subtypes via CEA were not consistent across hybrid ML methods. Subsequently, MIC enabled us to select more consistent clusters across different hybrid methods. When using non-imaging clinical information only, the clusters were not robust to variations in features, whereas utilizing SPECT information enabled consistent generation of clusters (rest of this work). We arrived at 3 clusters in years 0, 1, 2, and 4, and on timeless datasets. To identify similarity between subtypes, first we utilized training and testing process of K-means. After algorithm training using a specific year, we associated subjects of another year with the resulting clusters (testing process). After comparing the associated subjects with the original test subgroups, we were able to identify similar clusters. We were able to verify 3 distinct subtypes, across years in cross-sectional data and also in timeless data. When subsequently repeating training and testing processes using multiple classifiers (for 72 hybrid systems consisting of 4 FRAs, 3 CAs, and 6 Cs, experimentally selected based on improved performance), similar results were obtained. Moreover, Hotelling’s T² test re-confirmed our findings. Based on heat map analyses, the 3 identified PD subtypes were: i) motor-dominant, ii) non-motor-dominant, and ii) mixed motor/non-motor. Furthermor, relative to MRI-based segmentation of SPECT images, the clusters generated using SPECT-based segmentation remained less consistent when the number of subjects changed.

Conclusions: Appropriate hybrid ML framework and independent statistical tests enabled robust identification of 3 subtypes in PD subjects. This was achieved by combining clinical information with SPECT images segmented using MRI. The resulting PD subtypes were robust to the number of subjects and number of features.

• Download figure
• Open in new tab
• Download powerpoint