Identification and Classification of Parkinsonian and Essential Tremors for Diagnosis Using Machine Learning Algorithms

Abstract

Due to overlapping tremor features, the medical diagnosis of Parkinson’s disease (PD) and essential tremor (ET) mainly relies on the clinical experience of doctors, which often leads to misdiagnosis. Seven predictive models using machine learning algorithms including random forest (RF), eXtreme Gradient Boosting (XGBoost), support vector machine (SVM), logistic regression (LR), ridge classification (Ridge), backpropagation neural network (BP), and convolutional neural network (CNN) were evaluated and compared aiming to better differentiate between PD and ET by using accessible demographics and tremor information of the upper limbs. The tremor information including tremor acceleration and surface electromyogram (sEMG) signals were collected from 398 patients (PD = 257, ET = 141) and then were used to train the established models to separate PD and ET. The performance of the models was evaluated by indices of accuracy and area under the curve (AUC), which indicated the ensemble learning models including RF and XGBoost showed the best overall predictive ability with accuracy above 0.84 and AUC above 0.90. Furthermore, the relative importance of sex, age, four postures, and five tremor features was analyzed and ranked showing that the dominant frequency of sEMG of flexors, the average amplitude of sEMG of flexors, resting posture, and winging posture had a greater impact on the diagnosis of PD, whereas sex and age were less important. These results provide a reference for the intelligent diagnosis of PD and show promise for use in wearable tremor suppression devices.

Machine Learning in Tremor Analysis: Critique and Directions

“…This is to make each feature comparatively similar in magnitude, allowing ML algorithms based on gradient descent to iterate/converge more smoothly. Some authors used standardization to scale the features to ensure comparability between ML model performance, 32 whereas others normalized, minimum to maximum, their power‐spectrum density (PSD)–based features before implementing a two‐stage algorithm 46 …”

“…As tremor intensity physiologically fluctuates with time, the recording length governs which temporal aspects of the signal are included in the analysis. For most studies, the length per tremor signal recorded ranged from 20 38,82 to 30 32 to 60 seconds, 51 representing common clinical examination durations. Whereas some studies suggested that differentiation accuracy plateaus with recording lengths from 5 seconds upward, 58 other studies for tremor detection “in‐the‐wild” analyzed data collected over up to 26 hours per patient 62 …”

“…80,81 Our search identified several disadvantages in the tremor ML literature so far, as only 12 studies (of 36, representing 455 participants) stated the diagnostic criteria applied, and only 22 stated the summary clinical/demographic details. The patient treatment state was reported in some papers 33,43,47,[53][54][55]58,64,67 : whereas some patients were recorded in their medication on state, 37,62 other reports included patients on various medication states, 32,49 or throughout a medication cycle. 64,67 These aspects greatly limit the generalizability of results, as the clinical context is paramount for interpretation.…”

“…74 Cross-correlation and autocorrelation features are generally accepted as good representatives of tremor characteristics. 89 Most of the reviewed accelerometer publications engineered 1 to 10 features from both time and frequency domains, 34,39,42,47 representing amplitude and regularity, 32,51 spectral power, 51 fast Fourier transform coefficients, 32 and spectrograms. 59 Transitioning from a hypothesis-driven toward a data-driven approach, a selection of studies increased the number of features that they engineered (Fig.…”

“…The majority of publications (29, n = 1059 patients) related to the detection and quantification of PD tremor, five (n = 187 patients) focused on ET, four on the differentiation between PD and ET, and one each on the quantification of physiological and fatigue‐induced enhanced physiological tremor, respectively. The sample size ranged between n = 398 32 and n = 1 33 . The majority of studies were conducted under laboratory conditions based on “scripted protocols” with fixed recording length and positions, whereas nine studies were “unscripted,” that is, recording patients during activities of daily living (ADL; see Table 1).…”