Table 1.

Algorithm descriptions including overview over default and tuned algorithm parameters. GSDA^a, GSDB, and GSDC are algorithm versions described and validated for the lower back [11].

Domain	Algorithm name (reference)	Description	Original sensor position	Algorithm parameters (default)	Algorithm parameters (optimized)
Machine learningb	Brand (2022) [17]	Use of deep convolutional neural network to discriminate gait and nongait segments based on accelerometer data.	Wrist	N/Ac	CNNd trained on Mobilise-D optimization data set (see Methods)
Time domaine	Gu (2017) [41,42]	This method finds peaks in the summed and squared (RMSf) acceleration signal. It uses multiple thresholds to determine if each peak belongs to a step or artifact.	Wrist	verisense_k=3 sim_thres=–0.5 cont_thres=4 mag_thres=1.2	verisense_k=2 sim_thres=–0.8 cont_thres=4 mag_thres=1.2
Time domaing	Hickey (2017) [43]	Window-based threshold comparison of combined SD of 3D acceleration signal and vertical acceleration.	Lower back	ThresholdStill=0.2 ThresholdUpright=–0.5	ThresholdStill=0.2 ThresholdUpright=–0.5
Template basedg	Iluz (2014) (GSDA) [44]	Convolution of input signal with a gait cycle template (sine wave). Detection of local maxima in convolution result to define regions of gait.	Lower back	Vertical and anteroposterior acceleration used (lower back) activity_thres=0.01 min_bout_length=5 template_len=0.5 cm_norm_thres=0.4	Vertical and anteroposterior acceleration replaced by acceleration norm activity_thres=0.04 min_bout_length=10 template_len=1 cm_norm_thres=2.5
Template basede	Karas (2019) [31]	Template-based method (considering covariance between a scaled and translated pattern function) for stride detection based on adaptive empirical pattern transformation.	Wrist	sim_MIN=0.85 dur_MIN=0.8 dur_MAX=1.4 ptp_r_MIN=0.2 ptp_r_MAX=2.0 mean_abs_diff_med_p_MAX=0.5 mean_abs_diff_med_t_MAX=0.2 mean_abs_diff_dur_MAX=0.2	sim_MIN=0.3 dur_MIN=0.2 dur_MAX=3.0 ptp_r_MIN=0.2 ptp_r_MAX=3.0 mean_abs_diff_med_p_MAX=0.5 mean_abs_diff_med_t_MAX=0.5 mean_abs_diff_dur_MAX=0.5
Time domainb	Kheirkhahan (2017) [45]	Based on ActiGraph activity counts using sliding windows and adaptive thresholds.	Lower back	Walking threshold=0.75	Walking threshold=0.6
Time domaing	Paraschiv-Ionescu (2019) (GSDB and GSDC) [46]	Locomotion period detection based on detected steps from the Euclidean norm of the accelerometer signal. Consecutive steps are associated to gait sequences.	Lower back	GSDB: th=0.1 GSDC: th=0.15	Wrist: th=0.35
Time domaing	Paraschiv-Ionescu (2020) [47]	Extension of Paraschiv-Ionescu (2019). It applies an improved preprocessing strategy for the acceleration norm including an iterative succession of smoothing and enhancement stages. Furthermore, a data-adaptive threshold was introduced.	Lower back	N/A	N/A
Frequency domaing	Wavelets (Proprietary, Center for the Study of Movement, Cognition, and Mobility. Tel Aviv Sourasky Medical Center, Tel Aviv, Israel)	Time-frequency analysis using wavelets.	Lower back	Vertical and anterio-posterior acceleration used (lower back)	Vertical acceleration replaced by acceleration norm
Machine learningb	Willetts (2018) [20]	Activity detection using random forests and hidden Markov models to detect various activity modes. Only the output for “walking” activity was considered.	Wrist	Epoch length: 30 seconds	Epoch length: 1 second

^aGSD: gait sequence detection.

^bProgramming language is Python (Python Software Foundation).

^cN/A: not applicable.

^dCNN: convolutional neural network.

^eProgramming language is R (R Foundation for Statistical Computing).

^fRMS: root-mean-square.

^gProgramming language is Matlab (MathWorks).