A Systematic Review of Gait Analysis in the Context of Multimodal Sensing Fusion and AI

Background: Neurological diseases are a leading cause of disability and mortality. Gait, or human walking, is a significant predictor of quality of life, morbidity, and mortality. Gait patterns and other kinematic, kinetic, and balance gait features are accurate and powerful diagnostic and prognostic tools. Objective: This review article focuses on the applicability of gait analysis using fusion techniques and artificial intelligence (AI) models. The aim is to examine the significance of mixing several types of wearable and non-wearable sensor data and the impact of this combination on the performance of AI models. Method: In this systematic review, 66 studies using more than two modalities to record and analyze gait were identified. 40 studies incorporated multiple gait analysis modalities without the use of artificial intelligence to extract gait features such as kinematic, kinetic, margin of stability, temporal, and spatial gait parameters, as well as cerebral activity. Similarly, 26 studies analyzed gait data using multimodal fusion sensors and AI algorithms. Results: The research summarized here demonstrates that the quality of gait analysis and the effectiveness of AI models can both benefit from the integration of data from many sensors. Meanwhile, the utilization of EMG signals in fusion data is especially advantageous. Conclusion: The findings of this review suggest that a smart, portable, wearable-based gait and balance assessment system can be developed using multimodal sensing of the most cutting-edge, clinically relevant tools and technology available. The information presented in this article may serve as a vital springboard for such development.


I. INTRODUCTION
H UMAN walking or gait involves intricate coordination and interplay between cortical and subcortical areas, the cardiorespiratory system, and the musculoskeletal system resulting in reflex or controlled movement.While normal gait can be defined as a series of rhythmic, systematic, and coordinated movements of the limbs and trunk, which results in the change in position of the body's center of mass, The authors are with the Biomedical Engineering Department, Khalifa University, Abu Dhabi, United Arab Emirates (e-mail: 100062079@ku.ac.ae; aamna.alshehhi@ku.ac.ae; herbert.jelinek@ku.ac.ae; 100052964@ku.ac.ae; kinda.khalaf@ku.ac.ae).
individual gait patterns are unique and influenced by age, personality, mood and sociocultural factors [1].The period between one heel strike and the following heel strike when measuring forwards motion is defined as a gait cycle [2] (as shown in Fig. 1, [3]).
In addition to providing the key means for mobility, gait is a sensitive indicator of overall health status that can reveal the status and progression of underlying health challenges including neurological conditions (e.g., sensory or motor impairments and cognitive impairments) and musculoskeletal/orthopedic problems (e.g., osteoarthritis and skeletal deformities), to cardiovascular and metabolic conditions (e.g., heart failure, respiratory insufficiency, peripheral arterial occlusive disease, and obesity), and to ageing-associated ambulatory dysfunction and trauma [4], [5].In the elderly, gait disorders are also linked to impaired proprioceptive function associated with polyneuropathy, frontal gait disorder associated with vascular encephalopathy, poor vision, as well as osteoarthritis [6].As precursors of falls, gait deficits are considered as the most common cause of severe injuries in the aging population, while slow gait correlates with mild cognitive impairment (MCI) and future occurrence of dementia [7].
Gait analysis has emerged as a quantitative method for investigating a wide range of walking challenges and gait irregularities [4].Gait analysis, supported by modern sensing technology and computational algorithms, can be used in clinical and research laboratory investigations, security, and everyday living contexts, allowing for the identification, monitoring, and intervention.
Instrumented gait analysis (IGA) is considered the gold standard for gait assessment [8].IGA measures and analyzes different aspects of human gait applying spatiotemporal, kinematic, and kinetic measures.Traditional IGA systems consist of motion capture cameras, force plates, instrumented walk-ways, and treadmills, whereas contemporary IGA systems consist of miniaturized peripheral sensing systems, computational platforms, and modalities [8].IGA-based quantitative assessment can improve the diagnosis, outcome prediction, and rehabilitation of different gait impairments [9], [10], [11], [12].Smart wearable technologies, multi-modal physiological network sensors, sensor fusion techniques, and AI-driven computational platforms are becoming the center of interest when it comes to evaluating human mobility and gait.Such cutting-edge movement analysis technologies allow for gait and balance assessment to be more objective, accurate, quantifiable, and sensitive to alterations brought on by age, illness, or trauma [13].These technologies provide synchronous monitoring of spatiotemporal gait parameters (swing, stance, cadence, step width, step length, etc.), as well as upper/lower limb dynamics, including kinematics (angular positions, velocities, and accelerations), kinetics (forces, moments, plantar pressure and center of pressure), in conjunction with associated musculoskeletal, cardiorespiratory and neurological changes [14].Currently, there are two main types of modalities used for gait acquisition: sensor-based (SB) and vision-based (VB) [15].Although SB gait capture demands more complicated sensing technology, it has the capability of providing precise quantitative data using either floor sensors or wearable sensors in addition to utilizing either marker free or marker-based recording [15], [16].
Although the importance of the role that cortical areas play in normal gait is well known, decoding the precise relationship between cortical activity and gait characteristics remains elusive.Numerous studies indicate that cortical involvement is essential in human gait [17], [18].In the last two decades, evidence for cortical involvement in human locomotion has been provided by neuroimaging studies based on position emission tomography [19], electroencephalography (EEG) [20] and functional near-infrared spectroscopy (fNIRS) [21].EEG and fNIRS are increasingly gaining acceptance in the scientific community owing to their non-invasiveness, mobility, and ease of use [22], [23].EEG is one of the first methods developed for capturing cortical activity and has several applications in gait analysis [24].fNIRS is a relatively new technique which effectively captures brain hemodynamic [25].A large amount of multimodal data is, however, difficult to interpret and requires a computational system to describe and connect the many variables.Proprietary algorithms or models of artificial intelligence (AI) are computational systems that are currently being intensely investigated for prediction, as well as the diagnosis or prognosis of different clinical pathology [26].
Researchers in the field of gait analysis are utilizing AI techniques including machine learning (ML), support vector machine (SVM), and neural network (NN).These techniques have the potential to provide ways for obtaining, storing, and evaluating multifactorial complicated gait data, capturing its non-linear dynamic variability, and giving the crucial advantages of predictive analytics.ML models are a subset of AI models, comprising supervised and unsupervised learning.Both have advantages and disadvantages with supervised learning models being time-consuming, and labels for input and output variables need specialized knowledge.In contrast, unsupervised learning algorithms might provide radically erroneous outcomes [27].Sensor data collection, preprocessing, feature extraction, feature selection, classification, and result interpretation are all becoming standard procedures for gait classification, prediction, and analysis based on AI [28], [29], [30].
Several systematic reviews have evaluated the viability of utilizing wearable and non-wearable sensors individually as unimodal detectors for gait features associated with various neurological disorders [15], [16].Furthermore, while studies like [31] and [32] conducted in-depth examinations of deep gait recognition techniques and related privacy and security issues, they predominantly concentrated on the analysis of gait patterns using vision sensors.This analysis, however, omitted the exploration of the potential benefits stemming from multimodal data integration, including the fusion of cerebral activity and kinematic data.Moreover, the clinical implications of their findings remained less apparent, lacking investigations into the gait characteristics of individuals afflicted by neurological conditions like Parkinson's disease, as well as the impact of exoskeleton-assisted rehabilitation on gait dynamics.In contrast, our review paper specifically addresses these unexplored aspects, shedding light on the synergistic potential of sensor fusion, AI models, and clinical relevance in the realm of gait analysis.To the best of our knowledge, there has been no comprehensive evaluation of the efficacy of multimodal sensor fusion and AI prediction models for gait analysis.In this context, the overarching research question guiding this systematic review is: How does the integration of multiple gait analysis modalities through fusion techniques, coupled with AI models, impact the quality of gait analysis and the effectiveness of AI models for diagnosing and prognosing neurological diseases, with a focus on wearable and non-wearable sensor data?
In particular, this review highlights: 1. the need to integrate several gait analysis methods in order to extract kinematic, kinetic, margin of stability, temporal, and spatial gait parameters, as well as brain activity, in order to assess the pace and intensity of movement.
2. the utilization of multimodality AI analysis of the patient's walking pattern to compensate for the one-sidedness of single modality gait recognition systems that only learn gait alterations in a single measurement parameter.

II. REVIEW METHODOLOGY
Finding studies, vetting them for inclusion, and extracting data for this review were all completed in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement recommendations (Fig. 2) [33].

A. Search Strategy
A total of 660 articles were collected from Google Scholar, PubMed, Scopus, and IEEE Xplore databases.The term "gait analysis" was combined with at least two of the following search words/terms: Electroencephalography (OR EEG); Electromyography (OR EMG); Inertial Measurement Unit (OR IMU); Functional near-infrared spectroscopy (OR FNIRS); Hear rate (OR HR); Electrodermal activity (OR EDA); and Ground reaction force (OR GRF).In addition to scanning databases, the reference lists of all chosen papers were examined to identify any relevant research that may have been missed during the first search.

B. Inclusion and Exclusion Criteria
Articles published between January 2000 and July 2023 were considered.In the first stage of screening, duplicated articles were excluded, leaving English journal papers.In subsequent steps, abstracts only were excluded and full texts were evaluated in order to include only publications that utilized multi-modal sensors for analyzing gait and to exclude articles with animal experimentation and those pertaining to upper limb prostheses.To prevent the inclusion of unreliable gray literature, we implemented stringent selection criteria.We prioritized publications from reputable journals that had undergone peer review and evaluated author credibility, methodological rigor, and consistency of findings.These measures were implemented to ensure that only credible, highquality sources were included in our study (Fig. 2).While a comprehensive quality assessment for all 66 articles was not feasible, this stringent selection process was instrumental in upholding study quality, contributing to the overall integrity of our review's findings.

III. GAIT ANALYSIS USING MULTIMODAL SENSING FUSION
Fusion of multimodal data is a potential advancement for human movement research, such as enhanced activity detection and more accurate gait assessment [34].By quantifying the spatiotemporal, kinematic, kinetic, muscular, and physiological features of people, a fusion method provides a more complete understanding of gait impairment [16] and rehabilitation opportunities [35].The current review identified one study that recorded cortical and kinematics data from EEG and IMU, respectively, while 19 studies examined gait by integrating EEG and EMG sensor data.Eight research papers used EMG data with HR data, seven studies combined EMG with IMU data, and two studies combined EMG with fNIRS data.Fig. 3 presents a clustered bar chart showing the several combined gait analysis approaches.Several widely used fusion approaches are also described in this overview (Table I).

A. EMG & EEG
Twenty research studies employed EEG and EMG together, six of these focused on Parkinson's disease (PD).Timefrequency analysis of electrophysiological data reported by Roeder et al. [23] indicated substantially reduced corticomuscular coherence (CMC) and EMG power at low beta frequencies in older and PD individuals, whereas shorter swing time was noted for PD patients compared to healthy ones.Günther et al. [36] explored muscle and EEG activity during the freezing of gait associated with Parkinson's disease and how it varies from both walking and deliberate halting.At the outset of halt and freeze of gait events, they observed an increase in EEG-EMG coupling [37], [38].Venuto et al. [39] evaluated a non-invasive wearable embedded cyber-physical system for PD monitoring and discovered that the system can extract changes in walking pattern between PD and healthy patients.Alterations in central common drive to ankle muscles in response to visually directed foot positioning was reported by Jensen et al. [40], who noted that the corticospinal tract is involved in the modification of gait when visually directed foot placement is necessary.
Furthermore, several researchers investigated hybrid EMG-EEG data collected during walking of healthy individuals in order to explore the contribution of the motor cortex to muscle activation.In study [41], coherence and directionality analyses were utilized to determine if the motor cortex contributes to plantar flexor muscle activity during the stance phase and push-off phase of gait.There was substantial EEG-EMG and EMG-EMG coherence in the beta   and gamma frequency ranges.Similarly, the beta band activity was shown to originate mostly in the bilateral pre-motor cortices when a beamforming analysis was performed [42].Examining changes in cortical power and CMC during the gait cycle, Roeder et al. [43] reported evoked responses at spinal and cortical populations as well as positive time lag between EEG and EMG during human walking.

B. EMG & HR
The association of HR and EMG with gait has been reported by eight studies.The objective of these was to examine muscle activation, heart rate, and oxygen consumption in a variety of gait scenarios including walking at different walking speeds and the impact of age, and gender.The magnitude of the vertical ground response force during the toe-off phase varied significantly as a function of gender and gait speed [44].However, only HR was affected in the study by Groppo et al. [45] when applying negative pressure to the lower body and recreating a hypergravity condition.Similarly, Masumoto et al. [46] investigated the effect of walking in water on muscle activity, stride frequency, and HR and EMG responses with respect to age.While walking in water, older participants had greater hip musculature activity and reduced ankle plantar flexor activation.Nordic walking was also employed to the relationship between pole force and physiological reactions [47].Furthermore, Simpson et al. [48] developed backpack load position recommendations for hikers following the observation higher load positions resulted in weaker EMG activity in the gastrocnemius medialis muscles.

C. EMG & IMU
Integrating wearable EMG and IMU sensors provides granular gait metrics including kinematics and muscle activation.Seo and Kim [49] tested a novel phenotype for gait rehabilitation (machine of gait training and rehabilitation -MGTR).The MGTR was found to a strong association with knee angle during normal gait.In addition, and EMG data demonstrated ability to discriminate between limb kinematics and muscle co-contraction strategies to appropriately diagnose gait deficiencies by segmenting EMG signals via IMU data [50], [51], [52].In addition to these features, Krausz et al. [53] recommended the incorporation of an eye movement tracker that might improve the performance of the EMG and HR models.

D. fNIRS & EMG, fNIRS & EEG
Quantitative information of gait has the potential to significantly advance our understanding of cortical control of movement in both normal and abnormal gait patterns.In four publications, the assessment of the hemodynamic response using fNIRS was paired with EMG and EEG separately to detect cortical hemodynamic body movements.Peters et al. [54] utilized fNIRS and EMG to distinguish between active and passive overground locomotion when using a robotic exoskeleton.Here passive walking was associated with higher levels of oxyhemoglobin in the right frontal cortex than active gait.Two studies combined fNIRS and EEG and focused on gait analysis for PD patients [25], [55].Orcioli-Silva et al. [55], looked into how dopaminergic drugs affected cortical activity in PD patients as they walked freely and avoided obstacles.They found a promising effect on the β and γ power in the EEG CPz channel with dopaminergic medication and an increase in step length and step velocity.The same group investigated how tremor dominant (TD) and postural instability gait disorder (PIGD) motor subtypes of PD affected cortical activity during free walking and obstacle avoidance indicated that individuals with postural instability gait condition have more pre-frontal brain activity than tremor dominating patients [25].

E. EEG & IMU, EDA & HR
To reduce movement noise in EEG recordings during gait, Kilicarslan and Vidal [56] collected EEG and IMU data.IMU sensors were used to analyze the total head movement and segment the gait, while an adaptive de-noising framework was built to describe and manage the motion artifact contamination in EEG readings.Using EDA HR data, Anwer et al. [57] the impact of load bearing strategies on gait metrics, which indicated that carrying burdens with both hands increased gait symmetry, dynamic balance, and decreased the chance of falling.

IV. GAIT ANALYSIS USING MULTIMODAL SENSING
FUSION AND AI Statistical characteristics and frequency domain features are typically generated for use with AI models.Recent gait analysis research employing AI and multimodal fusion sensors are outlined in Table II.Twenty-six publications have utilized AI models to analyze multimodal fusion sensing data for gait analysis.The majority of these focused on combining EMG data either with IMU (12 studies) or EEG (7 studies).

A. Feature Extraction
Several studies retrieved statistical characteristics from processed raw time-series data using a sliding window approach, which was particularly prominent for EMG data, such as mean, median, entropy, standard deviation, root mean square variance, etc., [58] and [59].The frequency-domain characteristics represent yet another set of feature extraction.Hasan et al. [60] employed wavelet synchro squeezed transform (WSST) to recover EEG alpha and beta event-related desynchronization from the data epochs.Mezzina and De Venuto [61] analyzed the brain signal patterns using fast Fourier transformation (FFT).The EMG power spectrum along with waveform length, median and mean frequency, and modified mean frequency were the most frequently employed techniques for transforming original EMG temporal sequences into the frequency domain [58], [59].
With its ability to standardize the process of automated extraction of features, deep learning algorithms such as long short term memory (LSTM) and convolutional neural network (CNN) have become more prominent in the field of gait analysis.While EMG data are closely correlated with muscle force, they are collected with a substantial phase delay due to filtering, which makes accurate prediction difficult.To overcome this issue, artificial neural network algorithms were suggested to automatically extract signal characteristics in order to circumvent a possible bottleneck.LSTM was differentiated from other artificial neural networks by its capacity to ignore irrelevant information [62].To efficiently extract the inhibition and excitation mechanisms among EMG data and to emphasize the association between combined efforts of muscle activations and movements, LSTM with its power of forgetting unnecessary information, jointly analyses the signals of all the channels [63], [64].Similarly, Duan et al. [65] used CNN-based deep learning to efficiently extract walking-related characteristics from multimodal inputs after converting them into a 2D matrix.CNN is an end-to-end algorithm that offers superior learning capabilities since it bypasses the need for feature extraction.Local features on both temporal and spectral scales can be extracted via convolutional layers, which is useful for signal classification and motion detection.Moreover, to efficiently extract characteristics from disparate data sources, Jun et al. [66] supplied sequential skeleton and average foot pressure data into recurrent neural network (RNN)-based encoding layers and CNN-based encoding layers, respectively.Likewise, Zhao et al. [35] proposed a novel hybrid model for learning the locomotion disparities between PD patients.A spatial feature extractor (SFE) was built to understand the spatial information of multiframe time-series data and to output spatial features following dimensionality reduction.In the meantime, a novel correlative memory neural network (CorrMNN) architecture was created to measure the correlation in bimodal gait data and extract the dynamic gait variations to generate temporal features [35].Similar neural network was developed to capture the changes in multimodal PD handwriting data using a novel Spatio-temporal Siamese neural network [67].Meanwhile, the Multiview gait identification problem was addressed by proposing a spiderweb graph neural network (SpiderNet) to deal with visual gait data appropriately.By constructing a multi-view active graph convolutional neural network, it outperformed other methods in representing the Spatio-temporal and structural knowledge underlying the gait data [68].As well, the associated Spatio-temporal capsule network (ASTCapsNet) utilized by Zhao et al. [69] demonstrated significant results when combining multiple independent datasets that rely on unimodal data, such as vision data, force-sensitive data, and ground reaction force data.The common spatial pattern (CSP) neural network was also used to extract characteristics from EEG brain signals.CSP technique relies on the computation of a collec-tion of spatial filters that increases the variance of one class of EEG signals and decreases the variance of another class [70].

B. EMG & IMU
Assistive devices, such as exoskeletons, play a significant role in rehabilitation, leading Su et al. [71] to suggest a framework for machine learning that uses two multilayer perceptrons (MLP) to forecast locomotor modes and recognize gait events.Likewise, a deep RNN was employed as an intention prediction model for powered prosthesis based on knee joint motion prediction [72].Wang et al. [73] also made accurate predictions for the continuous joint angle using a hierarchical planner, however, according to research conducted by [74], bidirectional long short term memory (BiLSTM) generates the most accurate predictions of lower limb joint angles throughout the entire locomotion cycle.Donahue's study confirmed these findings by showing that kinetic waveforms can be estimated using machine learning from real-world running data without the need for feature engineering using the BiL-STM architecture [75].Exoskeleton control and user activity monitoring might be improved with more precise estimates of joint moment [76].In addition, the end-to-end training made it possible for LSTM predictor to be used for continuous kinematics prediction through the regression process, when transmission delay is introduced by exoskeletons [63].Also, multiple linear regression yielded an F1 score of 0.9 for approximating IMU angular velocity profiles and subsequently locomotion events using EMG data [77].

C. EMG & EEG
In the categorization of walking patterns [65], the combination of EEG and EMG data offers better performance compared to employing a single modality of EEG or EMG.Tortora et al. [64] designed and tested a hybrid human-machine interface (hHMI) for deciphering leg walking phases using a Bayesian fusion of EEG and EMG inputs.hHMI performed considerably better than its single-signal inputs.In addition, during the time-frequency domain analysis of the brain signal changes, logic networks primarily handled differentiating an unexpected loss of balance event from typical movements [61], [78].MLP was employed to assess brain and muscle health in the earliest stages of PD.This aided prescribing the correct medication to limit the progression of the disease, as well as identifying the various phases of PD and distinguishing PD from non-PD Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE II RECENT STUDIES THAT EMPLOYED AI AND MULTIMODAL FUSION SENSORS FOR GAIT ANALYSIS
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE II (Continued.) RECENT STUDIES THAT EMPLOYED AI AND MULTIMODAL FUSION SENSORS FOR GAIT ANALYSIS
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.[59].Similarly, Hou et al. [79] obtained high sensitivity and specificity for real-time FoG detection utilizing a CNN deep learning model and an adaptable, wireless sensor network.

D. IMU & EEG
Using AI models while merging IMU and EEG was reported in several research such as cortical changes associated with the intention of acceleration during self-paced walking, multimodal biometric authentication system, and FoG prediction for PD patients.Specifically, utilizing processed pre-acceleration EEG, a SVM classifier with radial basis kernel was developed to discriminate between the constant speed and accelerated speed conditions.This might be used to detect the intention to accelerate stride and then operate an associated assistive device adaptively [60].In addition, Bajpai et al. [80] an ensemble model comprised of two neural Networks for clinical and personal applications.Transfer learning was used to learn user-specific FoG-related characteristics for personal use.Besides, Zhang et al. [81] designed a multimodal biometric authentication system based on RNN to protect against biometric authentication-related threats.

E. Other Multimodals
Multimodal gait analysis enhanced by AI also involved data from various sensors, like motion capture cameras, accelerometers, and pressure sensors, to comprehensively assess an individual's walking pattern.Alharthi and Ozanyan [82] investigated the integration of reflective marker trajectories, force-plates, and EMG sensors.They discovered that multimodality fusion produced more accurate predictions than single modality methods, such as single stream CNN, Vision Transformer, and statistical classifiers.Likewise, using data from pressure sensors, accelerometers, and gyroscopes, Moon et al. [83] demonstrated the same conclusion, showing the highest accuracy for gait recognition using a combination of CNN, RNN, and Self-Attention models.Meanwhile, the multivariate squeeze-and-excitation network proposed in [97] demonstrated 91.31% accuracy in the recognition of human locomotion.Fig. 4 depicts the best performance of the various gait analysis models that have utilized EMG-IMU, EMG-EEG, IMU-EEG, and other multimodal sensing fusion data.
V. DISCUSSION Gait is not only necessary for human mobility, independence and everyday life functioning, but it is also a key predictor of quality of life, health status and mortality, as well as the progression of underlying pathophysiology [98].Examining gait patterns, particularly spatiotemporal, kinematic, kinetic, and balance gait features, can shed light on the quality of gait in association with overall health status and functionality.In gait research, smart wearable technologies and artificial intelligence, such as machine learning and deep learning techniques, are gaining growing interest.Despite their limited use in clinical settings, these methods hold great potential for changing how gait is quantified by collecting, storing, and evaluating multifactorial complex gait data while also capturing its non-linear dynamic characteristics and variability.While neural networks have been used in a small number of research studies, the findings are encouraging and warrant more investigation.Several studies employed multimodal sensing fusion to design and improve lower limb prosthesis [39], [49], [53].The results showed that the optimum separability, repeatability, clustering, and desirability across subjects and activities were achieved by integrating characteristics from vision, EMG, IMU, and goniometer sensors [53].Hence, future applications of this sensing fusion in a forward predictor for powered lower-limb prostheses and exoskeletons might benefit from the incorporation of vision-based ambient data.
Promising results were also revealed by the EEG and EMG hybrid modality [24], [42], [55], [85], [87].Chung and Wang [44] focused on the effects of age and gender.They discovered that females exhibit higher GRF during the heel-strike and toeoff phases, as well as more tibialis anterior muscle activation.On the other hand, Short et al. [87] investigated unilateral CP and found that children with CP had more cortical activity when walking.Research using the LDA classifier trained with fused IMU and EMG data achieved the highest accuracy compared to other types of classifiers trained with the same fusion of signals.However, this conclusion is not generalizable because of the wide variation of data collection settings and feature extraction strategies across investigations.LDA was mainly used for motion intention prediction and joint trajectory generation.In contrast, the single study that used the SVM classifier with EEG and EMG data had the lowest accuracy.Specifically, SVM was able to predict 9 out of 12 acceleration events with a mean delay of −741ms.According to Huang et al. [72], RNN performs better than both support vector regression (SVR) and conventional artificial neural networks (ANN).Similarly, the LSTM predictor performed better than the SVM predictor [62].On the other hand, gait events such as foot contact and toe-off were recognized accurately with MLP due to the meaningful IMU data that depicts these events clearly [71].
Higher number of electrodes with EEG also showed improved results in EEG-EMG multimodal classification [65].Meanwhile, cortical and subcortical changes lead to PD patients poor coordination and abnormal gait [59].Other factors that impact the performance of the models and lead to some variation in the findings include offline prediction that demonstrated superior model performance compared to online prediction.In addition, a number of models reported in the literature lacked modern computational methodologies and feature selection procedures or inadequately defined reference data such as knee joint angle.More research into the connectivity between cortical areas and classification accuracy is needed when EEG data is included into multimodal fusion AI.
Subject variability and label deficiency are intrinsic issues not just in gait assessment and staging, but in all healthcare settings.Therefore, future research might concentrate on loss design, data augmentation, or prototype learning methodologies to address this difficult but realistic topic.Similarly, creating unsupervised or semi-supervised approaches may lessen the need for time-consuming annotations and pave the way for more robust model creation.For this paradigm shift to be aligned with personalized gait abnormality assessment and rehabilitation, the loop must be closed with artificial intelligence models that include both static and dynamic features, as well as sophisticated data reduction and individualized feature selection of the most important gait characteristics.
The field of neurological disease assessment is undergoing notable trends that could reshape diagnostic approaches.A prominent trend involves integrating gait analysis methods, encompassing kinematic, kinetic, and cerebral activity measurements.This holistic approach provides a more nuanced understanding of movement patterns, enhancing the diagnostic potential of gait analysis.Additionally, the rise of multimodal sensor fusion, which combines wearable and non-wearable sensor data through advanced techniques, is becoming increasingly prevalent.This trend significantly improves diagnostic accuracy by capturing data from multiple sources, contributing to more precise assessments.Moreover, artificial intelligence models are gaining traction, particularly when coupled with multimodal fusion.This combined approach holds the potential to revolutionize neurological disease diagnosis, offering advanced predictive capabilities by processing complex datasets and identifying patterns that might not be discernible through individual modalities alone.
Despite the promising advancements, several open issues warrant careful consideration.Firstly, determining effective strategies to fuse various sensor data types optimally remains a challenge.This necessitates exploring fusion techniques that maximize the benefits of multiple data sources while ensuring robustness.Real-world validation of wearable systems across diverse settings and patient groups is imperative to establish their reliability and generalizability.Moreover, developing AI models that provide interpretable results is an ongoing concern, as comprehensible outcomes are vital for gaining clinicians' and patients' trust.Creating practical systems for long-term patient monitoring and disease tracking presents a technological challenge that requires attention to ensure usability and accuracy.Lastly, integrating gait analysis systems into clinical workflows and addressing adoption challenges remain key issues, emphasizing the importance of standardized practices and seamless incorporation into medical routines.

VI. LIMITATIONS
While this review did record the locations of the sensors and highlighted the most desired site by researchers, it was unable to draw any firm conclusions about the optimal number and placement of sensors.Even though all studies assessed gait, not all used the same experimental design, methodology, or conditions.This impacts both the obtained data and AI-based quantification of gait.
Different AI research groups also train their AI models with distinct datasets making it challenging to evaluate the performance of two AI models.Furthermore, software variation and its potential impact on algorithm efficacy were not considered due to lack of information in the included research studies.Likewise, there is also some uncertainty as to whether or not adding more sensors reduces or improves the reliability of the AI findings.
Certain machine learning models are known to be overfitting, and do not perform well on new datasets, making the scarcity of benchmark data all the more worrisome.The usefulness of an artificial intelligence model drops significantly if it lacks the ability to consistently generalize to novel, previously encountered situations.

VII. CONCLUSION AND FUTURE WORK
This systematic review primarily discussed the relevance of gait analysis utilizing fusion approaches and AI models that have been developed for this purpose.Among the 66 research articles, 44 utilized EMG signals as part of the multimodal fusion data.The relevance of combining multiple forms of wearable and non-wearable sensor data and the influence of this combination on the performance of AI models might thus be investigated further.The significance of cortical activity in evaluating gait might also be investigated in future research utilizing EEG and fNIRS.This allows researchers to measure several forms of mental stress and cognitive processes along with a variety of aberrant gait patterns.While deep learning is only employed in a small number of gait analysis studies, the findings are encouraging and warrant more investigation.Consequently, this review article serves as a starting point for the design and validation of a smart portable wearable-based gait and balance assessment system employing current tools and technologies that have been specifically developed for clinical use.

INSTITUTIONAL REVIEW BOARD STATEMENT Not Applicable.
INFORMED CONSENT STATEMENT Not Applicable.

Fig. 3 .
Fig. 3. Graph depicting the results of 40 studies utilizing multimodal sensing fusion for gait analysis.

Fig. 4 .
Fig. 4. Performance of several AI models in gait analysis using sensing fusion data.Figure shorthand is listed in Table II's footnotes.
Manuscript received 23 May 2023; revised 31 August 2023; accepted 12 October 2023.Date of publication 17 October 2023; date of current version 26 October 2023.This work was supported by the Healthcare Engineering Innovation Center, Khalifa University, United Arab Emirates, under Grant 8474000132.(Corresponding author: Rateb Katmah.)

TABLE I RECENT
RESEARCH ON GAIT ANALYSIS WITH MULTIMODAL FUSION SENSORSAuthorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE I (
Continued.)RECENTRESEARCH ON GAIT ANALYSIS WITH MULTIMODAL FUSION SENSORSAuthorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE I (
Continued.)RECENTRESEARCH ON GAIT ANALYSIS WITH MULTIMODAL FUSION SENSORSAuthorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.