AI can help differentiate Parkinson’s disease from essential tremor

Researchers in China have developed and evaluated machine learning algorithms to better identify and classify parkinsonian tremors and those essential for diagnosis. The results, published in Frontiers in neurosciencethey offer a reference for the intelligent diagnosis of Parkinson’s disease (PD) and “show promise for use in wearable devices for tremor suppression,” the authors wrote.

Because the symptoms of PD are complex and become more severe in the later stages, early diagnosis of the condition and introducing effective treatment are key to managing the disease.

However, the overlapping features between PD and essential tremor (ET), such as upper limb tremors, can make it difficult to differentiate between the conditions, and the task usually depends on clinicians’ clinical experience. According to the authors, about a quarter of PD patients are misdiagnosed as having ET.

Furthermore, “some efficient and accessible non-invasive biomarkers such as tremor signals including tremor acceleration and surface electromyogram (sEMG) have been studied for differentiation between PD and ET,” they wrote.

A total of 7 predictive models created to help clinicians distinguish between PD and ET (random forest [RF]extreme enhancement of the gradient [XGBoost]vector machine support [SVM]backpropagation neural network [BP]ridge classification [Ridge]logistic regression [LR]and convolution neural network [CNN]).

Researchers compared the relative importance of various upper limb postures, tremor characteristics, and demographics in diagnosing each disease.

Nearly 400 patients were recruited between June and November 2020, each with ET or PD confirmed with tremors in the upper limbs; information on tremor was collected through a system of medical devices and evaluated in 4 different postures for each patient.

As each patient rested, stretched, flew, and flew vertically, the devices recorded acceleration and sEMG measurements, for a total of 40 tremor variables collected from each patient.

80% of the data collected was used as a training set in each model, while the remaining 20% ​​was used as a validation set; the proportion of patients with PD or ET was consistent across each set.

“The average value of AU-ROC [area under receiver operating curve]which was calculated 10 times, was used as an indicator to evaluate the model to determine the different combinations of parameters for each model, “added the researchers.

Of the 398 patients evaluated, 257 had PD and 141 had ET.

Analyzes revealed:

• The ensemble learning models, including RF and XGBoost, showed the best overall predictive ability with accuracy greater than 0.84 and AUC greater than 0.90
• The other 5 models lacked significant predictive ability
• The dominant frequency of flexor sEMG, mean flexor sEMG amplitude, resting posture, and wing posture had a greater impact on the diagnosis of PD, while gender and age were less important.

The relatively small sample size marks a limit to the current analysis, while more patients with ET will be needed in future studies.

“With the further acquisition of ET subject data in future work, the performance of the models will be further improved and more valuable results will be obtained,” concluded the authors.

Reference

Xing X, Luo N, Li S, Zhou L, Song C, Liu J. Identification and classification of parkinsonian tremors and essential for diagnosis using machine learning algorithms. Anterior neurosciences. Published online March 21, 2022. doi: 10.3389 / fnins.2022.701632