Parkinson's disease (PD) is a chronic and progressive neurodegenerative disease that most commonly affects people over 65. Neuronal loss is the primary cause of the primary motor symptoms of Parkinson's disease.

The presence of motor and non-motor symptoms significantly impacts the quality of life of PD patients. According to a recent study, cognitive and psychiatric changes are directly linked to PD progression.

Although there is no permanent treatment for PD, temporary symptomatic treatments with levodopa and dopaminergic agonists can significantly improve quality of life by alleviating early symptoms.

Materials and methods

We outline our various methods for data selection, image processing, machine learning, and analysis below. The data and code are publicly available.

Multi-dimensional datasets (timeless and cross-sectional) were generated, segmentation and feature extraction were performed, and hybrid machine learning was utilized.

To compare clusters derived from different datasets, we employed two methods. Finally, we evaluated our results by dividing datasets into small groups.

First-stage analysis, including optimal cluster selection

In the first step, we applied our 35 datasets to the first category of HMLS. Initial samples of disease subtypes via ICEM were inconsistent across trajectories and datasets.

Afterward, we used all datasets for the second and third HMLSs, enabling ensemble and collective selection of an optimal cluster number.

Objectives

Subdividing Parkinson's disease (PD) into subtypes can lead to earlier recognition and custom-tailored treatment. The objective was to identify reproducible PD subtypes robust to patient numbers and characteristics variations.

Methods

We applied multiple feature-reduction and cluster-analysis methods to cross-sectional and timeless data extracted from longitudinal data (years 0, 1, 2 & 4; Parkinson's Progressive Marker Initiative).

885 PD/163 healthy-control visits; 35 datasets using non-imaging, conventional-imaging, and radionics features from DAT-SPECT images.

A hybrid machine-learning system was constructed using 16 feature-reduction algorithms, eight clustering algorithms, and 16 classifiers (C-index clustering evaluation was used on each trajectory).

Following this, we performed the following:

1. Identification of optimal subtypes.
2. Multiple independent tests to assess reproducibility.
3. Further confirmation by a statistical approach.
4. Test of reproducibility by sample size.

Results

Clusters generated without radionics information were not robust to changes in features, whereas collections generated using radionics information were consistent.

As a result of k-means training and testing, as well as Hotelling's T2 test, we found three distinct subtypes.

In terms of dopaminergic deficit (imaging), with some escalation of motor and non-motor symptoms, mild, intermediate, and severe are the three categories of PD.

Conclusion

We aimed to identify robust PD subtypes by incorporating clinical and imaging data. For optimal cluster selection, we utilized hybrid machine learning systems with a thorough analysis of trajectories and extensive robustness analysis.

Using radionics features enabled the identification of clusters that were more robust to changes in components and samples.

Source: https://jnm.snmjournals.org/content/62/supplement_1/107