Table 6.

Comparison with a selection of prior work.

Work	Target and result	Data	Feature	Technology
Bandini et al [10]	Found PDa patients have lower average facial expression movement distance; facial expression recognition for PD	17 PD patients, 17 healthy control subjects	Average distance of 49 facial key points in the facial expression movement	Face tracing, SVMb
Rajnoha et al [11]	Identified PD hypomimia by analyzing static facial images; less accurate compared with video-recording processing method.	50 PD patients, 50 healthy control subjects	128 facial measures (embedding) by CNNc	Face detector-based (HOGd), CNN, traditional classifiers (eg, random forests, XGBoost)
PARKe framework by Langevin et al [12]	PARK instructs and guides users through 6 motor tasks and 1 audio task selected from MDS-UPDRSf and records their performance by videos	127 PD patients, 127 healthy control subjects	Facial features: facial action units (AUs); motion features: motion magnitude metric of fingers and hands based on FFTg	OpenFace tool version 2, FFT
Our method	Proposed facial landmark features from videos to diagnose PD using facial expressions and achieved outstanding performance	33 PD patients, 31 healthy control subjects, 176 records	848 facial expression amplitude features and tremor features of facial key points; 65 features were left after feature compression	Face ++, traditional classifiers (LRh, SVM, DTi, RFj), LSTMk, LASSOl

^aPD: Parkinson disease.

^bSVM: support vector machine.

^cCNN: convolutional neural network.

^dHOG: histogram of oriented gradients.

^ePARK: Parkinson's Analysis with Remote Kinetic-tasks.

^fMDS-UPDRS: Movement Disorder Society Unified Parkinson Disease Rating Scale.

^gFFT: fast fourier transform.

^hLR: logistic regression.

ⁱDT: decision tree.

^jRF: random forest.

^kLSTM: long short-term memory.

^lLASSO: least absolute shrinkage and selection operator.