Machine Learning Shows Promise in Assessing Parkinson's

Liam Davenport

September 08, 2023

COPENHAGEN, Denmark — Machine learning techniques could soon be used for the differential diagnosis of parkinsonian syndromes and to predict gait dysfunction for patients with Parkinson's disease (PD) if validated in further studies, two new studies suggest.

In the first, 18F-fluorodeoxyglucose (FDG) positron-emission tomography (PET) scans from more than 260 individuals with parkinsonian syndromes across sites in Slovenia and the US were analyzed using three machine learning models.

All the models were able to differentially diagnose patients with either PD, multiple system atrophy (MSA), or progressive supranuclear palsy (PSP) with an accuracy approaching 90%, study presenter Matej Perovnik, MD, PhD, Department of Neurology, Ljubljana University Medical Centre, Ljubljana, Slovenia, reported.

"The diagnostic performance here, although not perfect, is nonetheless quite encouraging and equals that of many other biomarker/imaging based diagnostic tools," Ronald B. Postuma, MD, neurologist, McGill University Health Centre Research, Montreal, Quebec, Canada, who was not involved in the study, said in a press release.

In the second study, a team from Thomas Jefferson University, in Philadelphia, Pennsylvania, showed that a machine learning technique that was based on combined brain scans and clinical features from more than 40 patients with PD could identify gait dysfunction with an accuracy of nearly 80%.

"This supports the exploration of machine learning methods for predicting gait dysfunction in Parkinson's disease" using region-of-interest (ROI) brain scans, the team writes in the abstract to the study.

Both reports were presented here at the International Congress of Parkinson's Disease and Movement Disorders (MDS) 2023.

Differentiating Between Parkinsonian Syndromes

FDG-PET brain imaging and network analysis can be used to identify specific disease-related metabolic brain patterns related to PD, PSP, and MSA, Perovnik told the audience.

By quantifying the expression of these patterns in an individual, a subject score can be obtained that can be used in the differential diagnosis and the tracking of disease progression, he said.

To automate the process, the researchers sought to develop a differential diagnostic algorithm that combined FDG-PET scan features with machine learning to differentiate between parkinsonian syndromes.

They collated information on specific disease-related patterns in two populations of patients with parkinsonian symptoms, one from Slovenia and the other from the US. Patients' final clinical diagnoses were unknown at the time of imaging.

The team derived three datasets for the machine learning models: one based on previously validated clinical features in the Slovenian cohort, another on analogous feature patterns in the US cohort, and the third using a support vector machine based on 95 brain ROIs.

Included in the study were 265 FDG-PET scans from across the two sites. There were 161 patients with PD, 57 with MSA, and 47 with PSP.

The overall accuracy of the three models was comparable ― 86% for the Slovenian-based model, 85% for the US-based model, and 89% for the ROI model.

Perovnik said this shows that "metabolic brain networks identified in one institution can be used at different sites." Visualization of the ROI-based model indicated that "similar metabolic topographies can be identified using different methodologies."

"It's become quite clear that there are reproducible differences in overall patterns of FDG-PET between parkinsonian syndromes," Postuma commented, "and that these changes can be seen very early in the disease process."

One of the difficulties, he noted, is that often, the patterns partially overlap. This suggests machine learning tools "may be very useful, as they do not necessarily need to prespecify a specific pattern."

Predicting Gait Dysfunction

Gait function is normally "assessed with the naked eye," explained Tsao-Wei Liang, MD, medical director at the Jefferson Comprehensive Parkinson's Disease and Movement Disorders Center, Thomas Jefferson University, who is co-author of the second study.

Moreover, this takes place in the clinic, "which is often a very heterogeneous clinical environment, and there's no one standard way to do it," he told Medscape Medical News.

Consequently, the results are highly variable, "as we all may view things slightly differently in the office."

While rating scales can supplement gait analysis, "we rely on a single digit score" derived from them, Liang explained, "and unfortunately, there's a lot lacking in the description."

The team therefore set out to "understand gait dysfunction in a deeper way" by comparing the performance of eight machine learning techniques in predicting gait dysfunction in 43 patients who underwent diffusion tensor imaging (DTI) on MRI.

The images were initially assessed using different techniques to generate seven brain maps. The results were combined with age, sex, and MDS-Unified Parkinson's Disease Rating Scale–III scores and were fed into the machine learning models.

Compared to standardized office assessments, a multilayer perceptron model with five hidden layers and rectified linear activation performed the best on the binary classification of gait dysfunction vs no gait dysfunction, with an accuracy of 77.8%.

"This study suggests that a machine learning approach of DTI analysis may have potential in predicting gait dysfunction in PD patients if confirmed by larger confirmative studies," commented Klaus Seppi, MD, professor of neurology at the Medical University Innsbruck, Innsbruck, Austria, in a press release.

Future studies should "explore if this pattern changes with disease progression," he added.

No funding for the study was declared. The authors have disclosed no relevant financial relationships.

International Congress of Parkinson's Disease and Movement Disorders (MDS) 2023: Abstracts 297 and 1599. Presented August 28.

For more news, follow Medscape on Facebook, X, Instagram, and YouTube.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.