Table 1.

Main characteristics of the included studies. LLA (lower limb amputees); ULLA (unilateral lower limb amputees); BiLLA (bilateral lower limb amputees) (TFA (transfemoral amputees); TT (transtibial); TTA (transtibial amputees); BiTTA (bilateralt transtibial amputatees) TKA (through the knee amputees); LLP (lower-limb prosthesis); TAA (total ankle amputees); TRA (transradial amputees); HlULA (High-level upper limb amputee).

Article	Study Design	Participants	Aim	Procedure	Outcome
Beausoleil S. et al., 2019 [34]	Case Study	15 LLA Patients	To assess kinematic gait parameters during 6MWT and clinical applicability of IMU.	Post-Rehab assessment of gait, during 6MWT, with IMU (3D Acc + 3D Gyro) on both feet.	High stance and cadence variability on both limbs. High and variable minimal Toe clearance on AL. Gait kinematic parameters variability are correlated with future falls. Relevant IMU applicability in clinical context.
Maqbool H.F. et al., 2017 [35]	Controlled Clinical Trial	8 Healthy Control, 1 TFA Patients, 1 TTA Patient	To evaluate the reliability of RT Gait event detection algorithm in both flat and inclined surfaces for TFA patients.	Patients walked for 10 min at self-selected walking speed over flat surfaces and walk up and down a ramp (5° inclination) with a 6-DOF IMU (Acc + Gyro) fixed on the shank and insole with footswitches. The algorithm was written in MAT-lab.	100% detection accuracy for Initial Contact and Toe Off with different prostheses. Reliable algorithm for gait event detection.
Seel T. et al., 2014 [36]	Case Study	1 TFA Patients	To asses Joint Angle using IMUs and validate it’s measure with a 3D optoelectronic movement detection system (MDS).	IMUs (Xsense MTw) mounted on proximal and distal legs as well as foots. Body marker for 3D MDS (Vicon V612) mounted on both legs. Participant is requested to walk 10 m at self-selected speed. Data are gathered and confronted with 3D MDS	Joint angle calculation using accelerometer and gyroscope showed high precision and correlation with 3D MDS (1° RMSE for prosthetic leg and 3° for contralateral).
Wentinik E.C. et al., 2014 [37]	Case Study	3 TFA, 3 TKA	To detect onset of gait with IMUs.	Footswitches positioned in the heel center and under the first metatarsal head. Two inertial sensors (Xsens, Acc + Gyro) placed on anterior side of the proximal and distal leg. sEMG electrodes placed on residual muscles of the amputated leg. Each patient is asked to walk.	IMUs demonstrated reliable in detecting onset of gait in both healthy (Gyro) and prosthetic limb (Acc).
Simonetti E. et al., 2021 [38]	Case Study	1 TFA	To use a framework of MIMUs to evaluate BCoM acceleration and instantaneous velocity. Validate the measure versus a 3D MDS.	Full body marker set + 7 MIMUs on feet, shanks, thighs and trunk. The participant is requested to walk at self-selected speed through an 8 m path with 3 force platform in the middle.	Moderate to Strong correlation between MIMUs and Force Platform for SCoMs and BCoM acceleration and velocity.
Paradisi F. et al., 2019 [39]	Case-Control Study	20 TTA,20 Healthy Control	To investigate upper body acceleration and how these propagate from pelvis to head.	3 MIMUs located at head, sternum, and lumbo-pelvic segment. Participants were asked to walk thorough a 10 m pathway at self-selected speed.	Amputees have a larger coefficient of attenuation of acceleration from pelvis to sternum, greater medio-lateral and head acceleration. Attenuation coefficient may be a useful index for mobility assessment in LLA.
Dauriac B. et al., 2019 [40]	Case study	9 TFA	To evaluate the walking speed by estimating COM speed during gait cycle using a single IMU integrated in a microprocessor-controlled knee ankle prosthesis.	Several sped and slop conditions were tested at treadmill	This method estimates the walking speed with a 9% of RMSE in patients walking on a treadmill with 0° slope. The RMSE slightly increased when the slope is taken to 5% (but still acceptable).
Major M.J. et al., 2016 [41]	Controlled Trial	20 LLP (8 TFA, 9 TTA, 2 TT/TTA, 1 TT/TFA), 5 Healthy Control	To asses step length (SL) in patients with LLP.	A three-axis accelerometer was fixed at lumbar level in 20 LLP. The patients were asked to walk in a 20-m pathway from a standing position to a complete stop.	SL was correlated positively with previous literature study. Method was validated only on Healthy subject but not for LLP users.
Howcroft J. et al., 2014 [42]	Case study	11 TTA	To investigate if accelerometer derivate measures can differentiate between dynamic states and how those data correlate with clinical measures scores.	Community Balance and Mobility scales, Balance Berg scales, Prosthesis Evaluation Questionnaire were administered to the participants. An inertial sensor was affixed to the pelvis and then the participants walked in two scenarios: a 10-metre path on level ground (LG) and an 8-metre path covered by foam mattresses (uneven ground—UG).	Statistically significant differences were found between LG and UG walking in TTA participants. Stride time, vertical and AP acceleration FFT first quartile and ML Harmonic ratio were greater in UG than LG. Vertical acceleration and cadence were greater in LG than UG. ML acceleration range, AP acceleration standard deviation and stride time were correlated with change in clinical outcome measures scales.
Lamoth C.J. et al., 2010 [43]	Controlled Trial	8 TFA, 8 Healthy Control	To asses variability and stability of gait in LLA patients and healthy subjects.	All participants were equipped with a tri-axial accelerometer and walked for 6 min in various context: (i) indoor walking, (ii) indoor walking with cognitive dual tasks, (iii) outside walking (even terrain) in a square circuit (260 m long), (iv) outside walking (uneven terrain) in a square circuit (260 m long).	There was significant statistically differences in trunk acceleration (variability on ML acceleration) and walking speed (LLA patients are slower than healthy subject) in amputees’ group. Those two parameters are directly correlated with stability of the gait.
Tura A. et al., 2010 [44]	Controlled Trial	10 TFA, 10 Healthy Control	To evaluate a method for assessing gait regularity and symmetry of LLP users using a single accelerometer.	All participants were equipped with a single tri-axial accelerometer mounted at thoracic level and foot insoles. Patients are asked to walk a straight path 70 m long at natural, lower and faster speed. Step and stride regularity and duration were used to determine symmetry and regularity of the gait.	Step and stride regularity and step and stride duration are good index of regularity and symmetry of gait. A single accelerometer is capable to determine these parameters with good sensibility and specificity.
Clemens S. et al., 2020 [45]	Cohort Study	65 TTA, 63 TFA	To evaluate test-retest reliability of IMU based measures of segmental symmetry between lower limbs and differences between TTA and TFA in segmental symmetry score (SSS) and segmental repeatability score (SRS).	Participants wore knee sleeves equipped with 4 IMUs. They were asked to undergo a 10MWT on a Zeno Electronic Walkway system. Using sagittal angular velocities of thigh and shank SSS and SRS where calculated.	Good test-retest reliability, can differentiate between healthy and AL. Cannot differentiate between TFA and TTA.
Daines K.J.F. et al., 2021 [46]	Cohort Study	89 LLA (4 BiTTA, 1 TT/TFA, 63 TTA, 18 TFA, 2 TKA, 1 TAA)	To evaluate if the use of a random forest classificator is able to classify risk of fall in LLA.	An android smartphone was placed in posterior pelvis. All patients performed a 6MWT in a 20 m pathway. Data were collected in a custom-made application installed on to the smartphone.	Random forest classificator applied to data collected with a smartphone showed a good specificity (near 95%), good accuracy (81.3%) in classifying risk of fall in LLA patients.
Shawen N. et al., 2017 [47]	Controlled Trial	7 TFA, 10 Healthy Controls	To develop a classifier that integrates data from healthy participants to detect falls in individual with LLA.	All participants carried a Galaxy S4 Smartphone (Acc + Gyro) in natural position (pocket, hand, waist) during activities of daily living. 3 LLA participants took the phone with them for three days for quantifying false alarms.	Using a machine learning approach, data recorded from smartphones regarding angular and linear accelerations of healthy subjects can be used to classify falls risk in LLA subjects more specifically than a threshold approach (2 false alarms vs. 122).
Hordacre B. et al., 2015 [48]	Cohort Study	47 TFA	To asses activity and participation at home and various settings both for fallers and non- fallers LLA.	All participants were equipped with a stepwatch 3 activity monitor sensor and a GPS linked to the prosthesis. The community activity was defined as counting steps outside the house in various settings, and the home activity as counting steps inside the house. Participation was defined as an event in which participants had to leave home.	A statistically significant difference was demonstrated between LLA fallers versus non-fallers participants for commercial activity, recreational activity and total community activity. In addition, a statistically significant difference was found in recreational and total community participation.
Kapti et al., 2013 [49]	Case study	1 Healthy Subject	To investigate the use of accelerometric data recorded from TTA for trajectory control of an experimental active ankle joint prosthesis	Two acceleration sensors were used to register AP, ML and Vertical acceleration on the sound leg.	Data acquired from registration from the sound leg of a TTA may be used for controlling the trajectory of an LLP with active ankle joint users.
Chang M. et al., 2019 [50]	Case Study	4 TTA	To use a fuzzy logic system for terrain detection and automatic prosthetic ankle angle correction.	All participants wore a prosthesis with smart ankle system (equipped with an IMU sensor and a load cell for GRF detection) and walked on five different terrain condition (flat, upslope, downslope, upstairs and downstair) for at least 20 steps.	This fuzzy logic system had a 97.5% accuracy in terrain detection.
Su B.Y. et al., 2019 [51]	Case Study	1 TFA, 10 Healthy Participants	To evaluate a new method for training and intent recognition system using Convolutional Neural Network (CNN) algorithm.	Three IMUs were positioned on thigh, shank and ankle of the healthy leg. All participants were instructed to walk at a comfortable speed and walked among different motion states as well as steady state.	CNN can be used effectively for intent recognition with a system of 3 IMUs, and potentially to control a powered prosthesis for allowing natural transition trough motion states.
Keri MI et al., 2021 [52]	Case Study	1 TFA	To develop a low cost IMU based vibratory feedback system and use it to trigger prosthesis motion illusion (kinesthetics illusion -KI).	Vibratory feedback system (VFS) was composed with: an Arduino microcontroller, two 3-DOF Gyroscope, a lithium battery, a vibratory actuator. The accuracy of the VFS is quantified using an MDS and commercial IMU. Vibratory actuator was fixed on thigh and IMUs to a Robotic arm.	Participant in this study experienced KI for 16 degrees in knee flexion. Illusion of motion may improve gait parameters and reduce risk of falling.
Krasoulis A. et al., 2019 [53]	Controlled Trial	12 Healthy controls, 2 TRA	To develop a multi-grip classification system for prosthesis control in TRP users.	For HC 16 EMG-IMU sensors were placed in two 8 Sensor row on the forearm. For TRA 12 and 13 sensor were placed on the stump. Participants were asked to execute different grasp for calibration (power grasp, lateral grasp, tripod grasp, index pointer and hand opening). Consequentially they were asked to pick an object that stimulate a specific grasp.	Authors developed a multi-grip classification system using only two EMG-IMU sensors that can be used for real time prosthesis control during grip tasks.
Sharba G.K. et al., 2019 [54]	Controlled Trial	4 Healthy Control, 1 HlULA	To develop a RT shoulder girdle movement classifier to help high level ULA to control a prosthetic hand.	EMG and 3DOF Acc. were fixed on shoulder girdle of all participants. A set of five motions were chosen for classification: (i) elevation, (ii) depression, (iii) protraction, (iv) retraction and rest. The above classification was the used to control elbow, wrist and fingers of a 3D printed prosthesis.	Results showed a 92.8% accuracy in classifying shoulder girdle movement of the ULA participants. Classification was used to control a 3D printed UL prosthesis.
Ladlow P. et al., 2019 [55]	Controlled Trial	19 LLA (9 ULLA and 10 BiLLA), 9 Healthy Control	To asses validity of an algorithm combining data from accelerometer and HR (GT3X+ + Polar T31) monitor to assess energy expenditure (EE) during Physical Activity versus Actiheart Monitor (AHR)	All participants wore a Metamax 3B mask for Indirect calorimetry and were equipped with an AHR and a Polar T31 HR monitor. An Actigraph GTX3+ (3-DOF Acc) mounted on the waist near the shortest residual limb. All participants, then, are asked to walk on a treadmill at 5 progressive velocities and two slope (2% and 5%). Physiological Cost Index is then calculated (ΔHR/Walking speed).	The use of integrated Acc. data and HR data provided the most valid estimation of EE in ambulatorial setting for both amputation group. Level amputation impacts on accuracy of predicting EE.
Ladlow P. et al., 2017 [56]	Controlled trial	10 ULLA,10 BiLLA, 10 Healthy Control	To assess the impact of anatomical positioning of GT3X+ activity monitor in LLA participants and to develop algorithm on predicting EE.	All participants wore a Metamax 3B mask for Indirect calorimetry and were equipped with a GT3X+ activity monitor on either side of the waist above the hip and at L2 level. Participants were asked to walk on a treadmill at 5 progressive speeds. Moreover, all participants performed a sitting-based arm crank ergometry.	The anatomical positioning of accelerometers impacts the ability to predict EE in LLA. The positioning that better correlates with EE is on the amputated side of the waist, just above the hip.
Smith J.D. et al., 2021 [57]	Cross-sectional Study	23 TTA 9 TFA 3 BiTTA	To determine step count and step count accuracy with different activity monitor and O2 consumption during a 2MWT.	All participants were equipped with an Actigraph GT9X+ and a Garmin Vivofit ® 3 both on wrist and ankle of the non-dominant side. A modus Stepwatch 4 is placed on the non-dominant ankle in addition to the above sensors. All Participants are fitted with a Polar HR sensor and a Cosmed 5 portable metabolic analyzer. After three minutes sitting, participants performed a 2MWT as fast and safetly possible.	There were no differences in distance walked, VO2, HR and RPE between different amputation level. Step count and cadence were greater in TTA vs. TFA. Stepwatch on the ankle and Vivofit on the wrist provided the most accurate step count.
Desveaux L. et al., 2016 [58]	Cross-sectional	15 TTA	To asses if TTA patients with diabetes meet recommended level of physical activity and daily steps count. To investigate if physical functioning measures are correlated with objective measures of physical activity.	Participants were provided with a Stepwatch activity monitor (SAM) fixed around the ankle of intact limb. Participants were asked to wear SAM for 9 consecutive days. Physical activity was measured by steps count and number of minutes engaging activity involving >90 steps/min. Each participants underwent a 2MWT and performed an L test. Activities-specific Balance Confidence Scale (ABC) and WHO QoL-Brief Questionnaire were administered.	Despite improvement in functional mobility (L test) over 6-month follow-up, step count were below 6500/day and participants spent <150 min/week for vigorous physical activity (>90 steps/min). These results indicate the needs of post-rehabilitation intervention to promote active lifestyles.
Kim J. et al., 2021 [59]	Randomized Cross-over Trial	10 TTA	To quantifying metabolic cost, step count, walking, perception of mobility and quality of life between powered and non-powered prostheses users.	Participants were randomly assigned to perform testing with a powered prosthesis or with an unpowered prosthesis. All participants were equipped with two ActiGraph GT9X Link (one mounted on the prosthetic foot and one mounted on the prosthetic pylon) and a GPS enabled system on their phone active for two weeks. At the end of the two weeks, data were collected and participants underwent a metabolic measurement with Kosmed K4b2.	Authors did not find any differences in metabolic cost between powered prosthesis.