Table 1.

Summary of highlighted works regarding freezing of gait (FOG) detection with wearable sensors.

Publication	Analysis Methods	Sensors and Location	Participants	Main Results
Moore et al. [14]	Threshold method to analyze the power of specific frequency bands	Accelerometers located in the shank	11 PD	Identification of the increasing of power in specific frequency bands when FOG appears. Detection of 78% of FOG events.
Bächlin et al. [60]	Threshold analysis from three different frequency bands	9 accelerometer signals from Daphnet [68]. sensors located in the ankle, knee, and lower back	10 PD (8 with FOG)	Reduction of false detections with the addition of a Total-band threshold. A sensitivity of 73.1% and specificity of 81.6%.
Mazilu et al. [44]	ML techniques with Bächlin et al. [60] and statistical features	9 accelerometer signals from Daphnet [68]	10 PD (8 with FOG)	Sensitivity and specificity over 95% with 10fold cross-validation. A sensitivity of 66.25% and specificity of 95.38% with LOSO cross-validation.
Moore [61]	Threshold analysis	7 sensors located at the lumbar back, thighs, shanks, and feet	25 PD	Identification of the shank and back as the most convenient places to the sensors. Sensitivity 84.3% specificity 78.4%.
Tripoliti et al. [62].	ML techniques in a four steps method	6 accelerometers and 2 gyroscopes attached to different parts of the body	16 People (5 healthy, 6 PD with no FOG, and 5 with FOG)	Sensitivity of 89.3% and specificity of 79.15% with LOSO evaluation considering only patients with FOG symptoms.
Zach et al. [63]	Threshold detection with Moore et al. [14] features	A single triaxial accelerometer placed at the waist	23 PD patients with FOG	A lumbar sensor is identified as the best place for FOG detection. A sensitivity of 75% and specificity of 76%.
Ahlrichs et al. [15]	SVM classifier with frequency and statistical features	Single waist-worn sensor with a triaxial accelerometer	20 PD (8 with FOG and 12 with no FOG)	Frequency-based features could be reliably used to detect FOG. A sensitivity of 0.923, and specificity of 1 using data from 5 patients for testing.
Rodríguez-Martín et al. [59]	SVM classifier with statistical and spectral features validated with R-10fold and LOSO	Single waist-worn sensor with a triaxial accelerometer	21 PD	A sensitivity 88.09% and specificity 80.09% with R-10fold cross-validation, and a sensitivity of 79.03% and specificity of 74.67% for LOSO evaluation.
Samà et al. [64]	ML algorithms with a reduced version of the features proposed by Rodríguez-Martín et al. [59]	Single waist-worn sensor with a triaxial accelerometer	15 PD	Systematical reduction of the number of features. A sensitivity of 91.81% and specificity 87.45% for R-10-fold, and sensitivity of 84.49% and specificity 85.83% in LOSO evaluation.
Camps et al. [65]	DL and ML techniques. A novel spectral data representation	9-channel waist-worn IMU with accelerometer, gyroscope, and magnetometer	21 PD	The use of CNN with novel spectral data representation. AUC of 0.88, a sensitivity of 91.9 and a sensibility of 89.5 when testing with data of 4 patients.
Mohammadian et al. [66]	Novelty detection with CNN denoising autoencoders	9 accelerometer signals from Daphnet [68]	10 PD (8 with FOG)	Validation of a method to detect abnormal movement without the need for labeled data for training. Average AUC of 0.77.
San-Segundo et al. [67]	DL and ML algorithms validated in four different data representations	9 accelerometer signals from Daphnet [68]	10 PD (8 with FOG)	Validation of DL-based systems with CNN with a novel MFCC data representation. The analysis of the use of previous and posterior windows. AUC of 0.931 and an EER of 12.5% with LOSO cross-validation.