7.2 Slip-Related Fall Prevention Strategy
7.2.2 Detection
7.2.1.5 Requirements
Previous research proposed and attempted to meet several conditions that the fall prevention system should accomplish. The most often used fall prevention system identified needs were: i) flexible customi-sation amongst various users [40]; ii) assisted-as-needed behaviour [40, 354]; iii) no (or very limited) disturbance to the subjects, as demonstrated by: a) lightweight and comfortable to wear while walking [37,40,81]; b) compact design [81]; c) mechanical compliance between the subject and the exoskeleton [354]; and d) positioning heavy parts of the device away from the actuation joint [37, 81, 354]; iv) high torque development in a short time [37,81]; and v) adapt to the mechanical needs of the subject’s mobil-ity and intents on a constant basis. The fulfilment of these requirements enables the system to ensure a normal gait in the absence of aid and quick assistive torque supply to counteractLOBoccurrences in the presence of assistance. Furthermore, in the case of the proposed fall prevention strategy, the actuation of the assistive system must be accomplished within the actuation time period (100 ms).
7.2.2.1 Monitoring Variables Selection
The heel acceleration, shank-to-ground angle, and hip angle were identified as relevant kinematic variables.
Other variables were also incorporated to increase the number of possibly important kinematic character-istics and undertake a more thorough selection. As a result, the knee angle and shank angular velocity were also taken into account [45, 112, 340, 351]. The following criteria were used to try an objective selection of the most important monitoring variables. As a result, the majority of the criteria were aimed at describing technical aspects of the variables, such as: i) the simplicity of data processing required to extract the variable in real time from sensor data (Criterion 1); ii) based on the scientific literature, the capacity of the variable’s signal to efficiently identify gait events (Criterion 2); iii) the number of sensors required to compute the variable, the ease of sensor positioning, and if further instrumentation beyond the assistive device is required (Criterion 3); iv) bibliographical proof of the variable or its corresponding body segment (e.g. knee) being used to research and/or detect human biomechanical responses to slips (Criterion 4). In addition, certain movies from the slip-like perturbation protocol’s perturbation trials were visually analysed in order to detect visual modifications in human body segments caused by a slip-like perturbation. This allowed to collect visual signals regarding the changes caused by gait perturbations in the temporal evolution of the variable signal (Criterion 5). Furthermore, an innovation criteria was included in order to account for whether the variable has previously been addressed in a fall prevention strategy in the scientific literature, to the best of the author’s knowledge (Criterion 6).
As an outcome, the knee angle and shank angular velocity appear to be the most appropriate variables for detecting slip-inducedLOB. Despite the relevance of heel acceleration in slip biomechanics [350], it was overlooked for two reasons: i) the challenge of appropriately positioning and using wearable sensors attached to the heel; and ii) the impact of the foot on the floor puts significant noise into the heel accel-eration signal, limiting its usability and capacity to identify gait events. Beschorner et al. [350] gathered the heel acceleration data using reflexive markers, which alleviated the noise problem. The hip angle and shank-to-ground angle variables were also discarded, owing to the necessity to integrate the angular ve-locity signal in real time to produce the angle signal, which represents computational costs. The encoder from thePKOprovides the knee angle, and its extraction would not require any integration and so would not cause any drift issues. Subsequently, the identification of slip-inducedLOB was carried out while the knee angle and shank angular velocity variables were monitored.
7.2.2.2 Central Pattern Generators Controllers
Human movement and vital vegetative functions are recognised to be recurrent and cyclic processes.
Despite the importance of neuromuscular dynamics and sensory feedback in modulating these rhythmic functions, the functional activity of the neuronal circuits located in the spinal cord, i.e.,CPG, is attributed to the foundation of cyclic activity pattern generation [355,356]. The word ”central”means that the peripheral
nervous system and its sensory feedback are not used to generate rhythm [47]. Thereby, the use of biomimetic or biologically inspired CPG controller systems to monitor and regulate human locomotion variables becomes appealing, because such motion is almost certainly controlled by spinal oscillators, i.e., biologicalCPG[357,358]. The artificial CPG is thus expected to synchronise with the biological one, which plays a critical role in rhythmic movement support [357,359]. In terms of rehabilitation, the artificialCPG would provide assistive torque to the controlled joint whenever appropriate, allowing for the compensation of biologicalCPGdeficiencies, such as those caused by a brain lesion, towards healthy locomotion [357].
CPGcontrollers based on nonlinearAdaptive Frequency Oscillators (AFO)emerge as a dependable method to aid in the detection of abrupt and unexpected gait perturbations [360]. An AFO is a mathematical instrument that can synchronise its output to a frequency component of a periodic or quasi-periodic input signal while learning its key properties like amplitude and phase. A network ofAFO, i.e., aCPGcontroller, may then constantly synchronise with and estimate a periodic or quasi-periodic input signal [359–361].
An unanticipated perturbation during steady walking would cause irregular changes in the input signal, prompting theAFOto search for new signal patterns associated with various frequencies. This would soon divert the input signal from theCPG’s expected trajectory, allowing for the early and effective detection of an unexpected gait perturbation [360]. When a perturbation is identified, the artificial CPGactivates a robotic assistive system to deliver timely assistive torque at the controlled joints to counterbalance the LOBand encourage an efficient balance recovery [40].
Because human biped locomotion consists of a periodic or quasi-periodic motor activity, it may be decomposed into the sum of periodic or quasi-periodic signals [362]. As such, prior knowledge of the periodicity of human locomotion can be performed by taking advantage of nonlinear oscillators’ ability to generate stable rhythmic patterns, i.e., limit cycle behaviour, which is useful for decomposing the respective signals into a sum of sinusoidal waves that can be learned by an oscillator network [359,361]. TheCPG controller must have the same number ofAFOas the number of major frequency components required to correctly characterise the input signal, i.e., the learning signal. If the number of oscillators is inadequate to account for all of the essential frequency components of the input signal, the oscillator network will only learn and adapt to the higher-power frequency components [363]. Thereby, the learnt signal generated by theCPGwill be a rather approximate approximation of the input signal. Righetti et al. [363] state that if the number of oscillators is more than the number of frequency components to learn from the input signal, one of two things might happen: i) some oscillators will not converge towards any frequency and so contribute nothing to the learned signal; or ii) several oscillators will code the same frequency component and the sum of their corresponding amplitudes will equal the amplitude of the relevant frequency component.
In comparison to other alternative approaches, Ijspeert et al. [364], Tropea et al. [360], and Santos et al. [47] identified some intriguing aspects that makeCPGcontrollers suited for monitoring human locomo-tion: i) can create stable limit cycles that are resistant to disturbance. If the rhythmic pattern is disrupted, the controller quickly returns to its prior cyclic behaviour; ii) can be userd to control distinct segments or
modules within the same system. The variousCPGcan be linked together via a phase connection. As a consequence, theCPGmodel design is ideally suited for distributed implementation [47]; iii) feature a few control parameters that allow them to modulate locomotion based on changes in direction and speed.
This characteristic enablesCPGto generate online trajectories with smooth modulations even when the control parameters are abruptly changed; iv) allow for reciprocal entrainment between the mechanical system and them; v) do not require any training before implementation, unlike other algorithms, because the algorithm’s learning process is integrated into network dynamics. vi) do not have large computing cost since no demanding signal or algorithmic processing is required; vi) offer low computing costs since no intensive signal or algorithmic processing is required. and vii) Once the frequency bandwidth of the reg-ulated signals is known, they may be configured to monitor only these signals (all the higher frequency components can be associated toLOBreactions). Hence, unlike training-based algorithms, theCPGtuning does not need the use of signals captured during sophisticated unexpected gait disruption procedures, as only steady-state walking parameters are employed to tune the system.
7.2.2.3 Threshold-based algorithms
A simple threshold-based approach has been shown to be successful and generalizable for detecting slip-like perturbations based on the error caused between the actual kinematics and the kinematics predicted by an oscillator network [360]. In the presence of a perturbation, the error signal rapidly grows and exceeds the specified threshold values, allowing for early and effective detection of postural changes. This timely identification would allow a powered orthosis worn by the individual to give mechanical help in order to reduce the danger of a fall [40,360]. In this regard, the capacity of a simple and an adaptive threshold algorithm to identify perturbations was investigated. The simple threshold algorithm began by determining whether the current sample,𝑖, of the error signal was between the set fixed threshold levels. If the error value exceeded one of the thresholds indicating an abnormal condition, a warning was issued and a counter variable,𝑐, was increased.𝑐 was otherwise reset. Second, the number of consecutive warnings was checked to see if it exceeded the number of acceptable warnings,𝑟. The variable𝑟 was used to enable a more consistent perturbation detection by reducing the amount of false alarms generated by individual samples that exceeded a threshold but were not perturbations. If𝑐is less than𝑟, the algorithm moves on to the assessment of the next sample,𝑖+1. The algorithm, on the other hand, recognised a perturbation if𝑐was equal to or larger than𝑟. The detection time was determined to ensure that the perturbation was correctly detected. To obtain this metric, it is necessary to obtain the time difference between the actual beginning of the perturbation, as determined by pert sample, and the perturbation onset recognised by the algorithm. The detection was considered a false alarm if it was discovered before the perturbation began or later than 1 second after it occurred. The perturbation was declared effectively identified if it was noticed within the 1 second interval following its beginning. This time interval was used as a reference since earlier studies indicated that falls can occur in as little as 1 second [40,365].
In comparison to the fixed threshold, the adaptive threshold technique follows a similar strategy. Unlike the fixed threshold strategy, this approach allowed for the use of contextual information from prior samples in the setting of threshold values. The mean,𝜇, and standard deviation,𝜎, of the m-sized window before the present sample𝑖 were acquired in order to compute the dynamic thresholds tailored to each sample.
To improve the effectiveness of the subject-specific perturbation detection, the thresholds were additionally determined based on the coefficients𝑎and𝑏 assigned to each subject. The coefficients𝑎and𝑏 show how the standard deviation,𝜎, affects the computation of the upper and lower thresholds, respectively.
The concept of perturbation detection was identical to the fixed threshold approach once the thresholds were calculated. These algorithms’ threshold and window size parameters were set for each participant and are explained further. The variable𝑟was set to 3 in order to identify a perturbation only if three or more consecutive samples exceeded one of the threshold values. Because both threshold-based algorithms do not need knowledge from future samples, they are thought to conduct online perturbation detection in real-world contexts.
7.2.2.4 Requirements
Despite the fact that few research [40] explain the detection of slip-inducedLOB, several conditions for the detection stage of the fall prevention strategy were identified in order to validate the perturbation detection performance: i) The detectionACCof actual perturbations must be more than 75%; ii) the real perturbations’ mean detection time must be less than the detection time further provided for the fall prevention strategy (360 ms); and iii) the number of false perturbations detected must be smaller than the number of right perturbations detected. Given that this chapter is still in its early stages, the present purpose of this work is to determine whether the perturbation detection algorithm can reach acceptable rather than ideal performance. Hence, satisfying the above-mentioned parameters would demonstrate acceptable performance of the perturbation detection method and open the path for future optimisation of the detection process.