Exploring the Potential Imaging Biomarkers for Parkinson's Disease Using Machine Learning Approach

Abstract

Parkinson's disease (PD) is a neurodegenerative disorder characterized by motor and neuropsychiatric symptoms resulting from the loss of dopamine-producing neurons in the substantia nigra pars compacta (SNc). Dopamine transporter scan (DATSCAN), based on single-photon emission computed tomography (SPECT), is commonly used to evaluate the loss of dopaminergic neurons in the striatum. This study aims to identify a biomarker from DATSCAN images and develop a machine learning (ML) algorithm for PD diagnosis. Using 13 DATSCAN-derived parameters and patient handedness from 1309 individuals in the Parkinson's Progression Markers Initiative (PPMI) database, we trained an AdaBoost classifier, achieving an accuracy of 98.88% and an area under the receiver operating characteristic (ROC) curve of 99.81%. To ensure interpretability, we applied the local interpretable model-agnostic explainer (LIME), identifying contralateral putamen SBR as the most predictive feature for distinguishing PD from healthy controls. By focusing on a single biomarker, our approach simplifies PD diagnosis, integrates seamlessly into clinical workflows, and provides interpretable, actionable insights. Although DATSCAN has limitations in detecting early-stage PD, our study demonstrates the potential of ML to enhance diagnostic precision, contributing to improved clinical decision-making and patient outcomes.

Keywords: AdaBoost; DATSCAN; Parkinson’s disease; basal ganglia; classification; machine learning.

Figure 1

Direct pathway (red), indirect pathway (yellow), and nigrostriatal pathway (black).

Figure 2

The distribution of key demographic and clinical characteristics of the participants: (a) cohort at enrollment; (b) sex at birth; (c) handedness; (d) side most affected at PD symptom onset.

Figure 2

The distribution of key demographic and clinical characteristics of the participants: (a) cohort at enrollment; (b) sex at birth; (c) handedness; (d) side most affected at PD symptom onset.

Figure 3

The participants’ symptom profiles and treatment statuses distribution: (a) initial symptom (at diagnosis)—resting tremor; (b) initial symptom (at diagnosis)—rigidity; (c) initial symptom (at diagnosis)—bradykinesia; (d) initial symptom (at diagnosis)—postural instability.

Figure 3

The participants’ symptom profiles and treatment statuses distribution: (a) initial symptom (at diagnosis)—resting tremor; (b) initial symptom (at diagnosis)—rigidity; (c) initial symptom (at diagnosis)—bradykinesia; (d) initial symptom (at diagnosis)—postural instability.

Figure 4

The participants’ age at enrollment in the PPMI project.

Figure 5

The duration from PD diagnosis to enrollment in the PPMI project.

Figure 6

The ML process.

Figure 7

Histograms of features. HANDED 1—right, 2—left, 3—mixed. The rest of the values are unitless SBRs.

Figure 8

The Pearson correlation coefficients. HANDED 1—right, 2—left, 3—mixed. COHORT 1—healthy controls, 0—PD. The rest of the values are unitless SBRs.

Figure 9

Confusion matrices for classifiers trained on all features and computed on a test set.

Figure 10

Feature importances using the LIME method.

Figure 11

The distribution of feature importances.

Figure 12

Confusion matrices for classifiers trained on con_putamen feature and computed on test set.

Figure 13

ROC curves for the models along with AUC scores.

Figure 14

AdaBoost classifier decision function values: (a) decision function values for PD and healthy controls observations for single con_putamen feature; (b) histograms of decision function values for PD and healthy controls observations for single con_putamen feature.