This systematic review was registered on PROSPERO (International Prospective Register of Systematic Reviews) under registration number CRD420261285256. The review was conducted and reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines (Checklist 1). Minor methodological changes were made during full-text screening to align the risk-of-bias assessment tools with the final included study designs. These refinements did not affect the overall objectives, eligibility criteria, or review conclusions.

A comprehensive literature search was conducted across 4 electronic databases: PubMed, Scopus, Web of Science, and IEEE Xplore. The search covered studies published between January 2015 and March 2025, with the final search performed in March 2025. These databases were selected to ensure coverage of biomedical, rehabilitation, engineering, robotics, and wearable technology literature.

To improve sensitivity and minimize the risk of missing relevant studies, the search strategy combined keywords and synonyms related to exoskeleton technologies and the knee joint. These included terms such as “exoskeleton,” “robotic exoskeleton,” “wearable robot,” “powered orthosis,” and “robotic orthosis,” combined with “knee” or “knee joint,” and terms related to human or clinical testing (eg, “human,” “patient,” “clinical,” “trial,” and “experiment”). Studies that focus solely on simulation, musculoskeletal modeling, finite element analysis, or in silico approaches were excluded at the search level where possible.

Database-specific Boolean search strings were developed and adapted for each platform. The full reproducible search strategies for all databases are provided in Multimedia Appendix 1.

The search was limited to English-language publications due to feasibility constraints in screening and data extraction. Gray literature, trial registries, and industry- or manufacturer-reported studies were excluded because this review focused on peer-reviewed scientific evidence. These decisions may introduce language and publication bias, and this is acknowledged in the limitations.

Following the initial database search, 71 duplicate records were removed, resulting in 2418 unique articles. Study selection was performed in multiple stages.

First, an AI-assisted abstract screening was applied to articles retrieved from Web of Science, IEEE Xplore, PubMed, and Scopus. Screening was performed using the GPT-4 large language model accessed through the OpenAI API. The model was not additionally trained on domain-specific datasets and relied on its pretrained language understanding capabilities. The filtering prompt and classification code were developed by reviewer AZ and applied prior to manual screening (provided in Multimedia Appendix 2). The AI-assisted screening used a prompt-based binary classification approach in which each abstract was evaluated to determine whether the study focused exclusively on knee exoskeletons. The model was instructed to retain studies by default and to exclude articles only when the abstract explicitly involved joints other than the knee (eg, hip, ankle, foot, or multijoint lower-limb systems). In ambiguous cases, studies were retained for further manual evaluation. This conservative filtering strategy was intentionally designed to prioritize sensitivity and minimize false exclusions.

Of the 2418 unique records, 1089 were excluded solely through AI-assisted abstract screening. To validate the performance of the AI-assisted filtering process, a random subset of AI-excluded records was manually reviewed by the authors. This validation confirmed that the majority of excluded studies involved multijoint or non–knee-focused exoskeleton systems. After AI-assisted screening, 71 articles from Web of Science, 21 from IEEE Xplore, 29 from PubMed, and 997 from Scopus remained eligible for subsequent manual screening.

Second, articles retained after AI-assisted screening underwent manual review to confirm eligibility and ensure that the studies focused exclusively on knee exoskeletons and represented original research rather than review articles. Manual screening was performed independently by 4 reviewers (DK, Zeeshan, AN, and ARK). Following AI-assisted and manual screening, 1166 articles were retained for full-text assessment. Full-text screening further evaluated eligibility based on the predefined inclusion and exclusion criteria, resulting in 112 studies selected for data extraction.

Data extraction was performed for all eligible studies following full-text screening. For each study, the following information was systematically extracted and organized into structured tables: author name and DOI, article title, database source, device name, weight, mechanism type, control system, actuators, range of motion, clinical testing details, and patient type. These data formed the basis for qualitative and comparative analysis of knee exoskeleton technologies and applications.

The risk of bias was assessed based on the design and methodological characteristics of the included studies. This approach was adopted because the reviewed literature included randomized clinical trials, nonrandomized clinical studies, experimental feasibility studies involving healthy participants, and technical validation studies. Risk of bias was assessed according to the study design: randomized controlled trials were evaluated using the RoB-2 tool. Nonrandomized clinical studies involving patient populations were assessed using the ROBINS-I (Risk of Bias in Nonrandomized Studies of Interventions) tool, which evaluates bias due to confounding, participant selection, intervention classification, deviations from intended interventions, missing data, outcome measurement, and selective reporting. Experimental and feasibility studies conducted in healthy participants were assessed using a modified Downs and Black checklist, as these studies typically focus on technical feasibility, biomechanical performance, usability, or short-term functional outcomes rather than clinical efficacy. Risk-of-bias assessment was independently performed by 3 reviewers (DK, PKJ, and Zeeshan), and disagreements were resolved through discussion. Nonhuman validation studies were not assessed using clinical risk-of-bias tools, as they do not evaluate health outcomes, and were analyzed descriptively with emphasis on device architecture, actuation strategy, sensing configuration, control method, and technical validation outcomes.

Risk-of-bias assessments were incorporated into the narrative synthesis, with greater emphasis placed on findings from studies demonstrating lower risk of bias or higher methodological quality.

A total of 32 articles met all inclusion criteria and were included in the final review.

Figure 1 illustrates the overall study selection process, detailing the reasons for excluding certain articles and their corresponding numbers.

To reflect the multidisciplinary nature of this field and to ensure clarity in synthesizing the evidence, the Results section was divided into two complementary parts. The first part focuses on the engineering perspective, summarizing design concepts, actuation strategies, and control approaches that enable the functionality of knee exoskeletons (Table 1). The second part addresses the clinical perspective, highlighting patient populations, rehabilitation goals, and functional outcomes associated with the use of knee exostheses (Table 1). This structured approach highlights the importance of integrating engineering advances with clinically relevant outcomes to ensure effective rehabilitation.

Clinical and biomechanical outcomes were frequently associated with specific engineering design characteristics across the included studies, as shown in Table 3. Motor-driven exoskeletons with gait-phase-based control were among the most commonly investigated systems in neurological rehabilitation and were associated with improvements in gait coordination, swing-phase assistance, and walking support in individuals with stroke and spinal cord injury [29,42,44,57]. Similarly, assist-as-needed and adaptive control strategies were developed to modulate assistance in response to user performance and task demands, thereby supporting individualized rehabilitation approaches [48,57].

Passive and compliant exoskeleton systems have been commonly investigated for repetitive movement assistance and the reduction of muscular effort in healthy individuals [40,49,52]. However, these systems generally provided lower torque output and limited adaptability for pathological gait correction. In contrast, higher-torque motorized systems (10‐40 Nm) were more frequently used for stair climbing, perturbation recovery, and poststroke rehabilitation tasks that require greater movement support and stabilization [21,33,43].

Several studies also emphasized the importance of wearable comfort, joint alignment, and interface compliance for user acceptance and stability during assisted gait [45,54]. Hybrid rigid-soft architectures, self-aligning mechanisms, and hinge-free exosuit designs were specifically developed to reduce joint misalignment and preserve more physiological gait kinematics while maintaining assistive functionality [31,40,54,56]. Overall, the reviewed studies demonstrated that rehabilitation outcomes were influenced not only by torque generation capacity but also by sensing modalities, control strategies, biomechanical alignment, and human-robot interaction design.

Over the last 10 years, significant progress has been made in knee exoskeleton research, particularly in actuator design, sensor integration, and control system development [38,39,48,54,56]. From an engineering perspective, the domain has evolved significantly in complexity. Nevertheless, a closer look at the literature reveals that clinical validation has not kept up with this advancement. The risk-of-bias assessment indicates that current evidence supporting knee exoskeletons is methodologically heterogeneous. Although engineering reporting and technical characterization are generally robust, clinical evidence remains constrained by small sample sizes and the moderate risk of bias. Therefore, the strength and generalizability of the current evidence regarding long-term clinical efficacy should be interpreted with caution.

Many published studies continue to be conducted in laboratory settings, primarily involving healthy subjects rather than those with neurological or orthopedic conditions [29,32,45,48-50,55]. Although these investigations are essential for confirming feasibility and safety, they fail to fully address the intricacies of rehabilitation in practice.

Motor-driven exoskeletons remain the leading technology in this area due to their ability to accurately adjust assistive torque and their effective integration with control frameworks that use torque and impedance principles [31,34,42,54,56]. In structured walking activities, these systems can consistently provide assistance that varies with the gait cycle phase. Nevertheless, achieving greater torque capability often requires larger actuators and transmission mechanisms, thereby increasing distal mass and inertia [31,40,42,54]. These mechanical additions may influence natural gait mechanics and metabolic demand, yet these effects are not consistently quantified. In rehabilitation settings, the trade-off between torque capacity and wearability becomes particularly relevant, as long-term use depends on comfort and tolerance rather than peak output.

Passive and hybrid systems have emerged as attempts to address these limitations by emphasizing transparency and reduced mechanical burden [32,40,49,52]. Passive elastic devices have demonstrated measurable biomechanical effects during controlled tasks such as squatting and jumping [49,52]. However, their assistance profiles are inherently task-specific and less adaptable to the variability observed in pathological gait. Hybrid rigid-soft architectures seek to preserve torque capability while improving interface compliance and load distribution [31,54,56]. Although conceptually promising, comparative data demonstrating superior long-term outcomes are still limited. Taken together, the current evidence suggests that future progress may depend less on maximizing mechanical output and more on refining the balance between assistance magnitude, comfort, and adaptability to individual impairment.

The review demonstrated that rehabilitation outcomes were strongly influenced by engineering characteristics such as actuation type, control architecture, and interface compliance, highlighting the importance of biomechanically aligned and task-specific exoskeleton design.

The observed results in clinical and biomechanical parameters can be directly interpreted using fundamental knee biomechanics. When walking on a flat surface, knee biomechanics are characterized by coordinated changes in joint kinematics and kinetics, particularly during the initial stance phase, in which knee flexion and extension moments regulate shock absorption and stability [66,67]. During daily activities, tibiofemoral forces can reach 2‐4 times body weight during level walking and up to 5 times body weight during stair descent, highlighting the substantial mechanical demands placed on the joint [68,69]. These biomechanical requirements emphasize the need to develop an exoskeleton adaptable to various tasks.

Figure 8 shows a biomechanical model of the human lower limb, representing the femur and tibia as rigid bodies interconnected by ligamentous and contact elements, used to estimate knee joint kinematics and loading. This framework provides a physiological basis for designing knee exoskeletons that deliver assistive torques aligned with natural joint motion and internal force distributions.

Clinical data show that altered knee mechanics lead to pain and functional disorders. Although the peak moment of external knee adduction shows a weak overall correlation with knee pain in osteoarthritis, dynamic anomalies, such as varus deformity, are associated with pain, underscoring the importance of frontal-plane stability [70]. After total knee replacement, knee extension moments decrease in patients, and joint load increases during downhill walking, indicating continued biomechanical vulnerability in the early stages of rehabilitation [71]. Similarly, patients after anterior cruciate ligament reconstruction exhibit altered kinematics during rehabilitation [72]. Overall, knee exoskeletons must integrate accurate anatomical sensing, adaptive control, and compliant actuation to modulate joint moments without disrupting natural movement, underscoring the critical need to align engineering design with clinical biomechanics.

Across studies, the quality of the human-robot interface consistently influences system performance, even when not explicitly analyzed. Mechanical misalignment between the anatomical knee axis and the exoskeleton joint can introduce parasitic torques and increase soft tissue stress [40,42,43,51,54]. Several groups have proposed self-aligning mechanisms, polycentric linkages, and improved suspension strategies to address this issue [33,42,54]. These refinements have been associated with improved user comfort and perceived naturalness of movement.

Despite these advances, many clinical investigations do not systematically report alignment errors, device migration, or long-duration wear tolerance [31,44,52,57]. This omission is clinically meaningful. Rehabilitation depends on repetition and sustained engagement; even minor discomfort or instability may alter neuromuscular activation patterns or reduce adherence over time. Emerging work examining embodiment and user perception suggests that when assistance feels mechanically coherent with voluntary movement, users integrate the device more effectively [44,57]. For this reason, interface optimization should be considered central to rehabilitation efficacy rather than an accessory design feature.

Control architecture remains a defining feature of knee exoskeleton systems. FSM-based controllers remain widely used in clinical contexts for their deterministic phase segmentation and predictable assistance timing [21,31,33,38,39,41,45-47,51]. In patient populations where safety is paramount, this reliability provides practical advantages.

More advanced strategies, including impedance control [38,54], model predictive control [39], human-in-the-loop optimization [30], EMG-driven assistance [36,50], and neuromusculoskeletal model-informed machine learning [48], have demonstrated promising personalization capabilities in controlled environments. For example, human-in-the-loop optimization has reduced metabolic costs in healthy participants [30], and EMG-based systems enable modulation of assistance based on muscle activation signals [36,50]. However, these approaches are rarely evaluated in larger or clinically heterogeneous cohorts [44,57].

Furthermore, few studies report performance under fatigue, uneven terrain, prolonged daily use, or sensor drift [30,39,48]. Without such validation, it is difficult to determine whether algorithmic complexity yields meaningful clinical benefit. At present, clinical implementation appears to favor robustness and predictability over adaptive sophistication.

Across the reviewed literature, phase-specific assistance consistently contributes to functional improvement. Systems that target swing-phase knee flexion or stance-phase extension tend to improve toe clearance, knee extension, and gait symmetry while preserving voluntary muscle contribution [46,50,51,53]. This observation suggests that the timing of assistance may be more critical than continuous torque magnitude.

In contrast, trajectory-enforced systems may risk reducing active neuromuscular participation [42,44]. Since rehabilitation aims to facilitate motor relearning rather than to replicate passive movements, this distinction is important. Studies in pediatric cerebral palsy populations demonstrate that appropriately timed extension assistance can reduce crouch gait severity without suppressing voluntary engagement [50,51], supporting the concept of deficit-targeted assistance.

However, most outcome assessments are performed during active device use [31,44,57]. Postintervention follow-up is uncommon, making it difficult to determine whether improvements represent true motor adaptation or temporary mechanical augmentation. Longitudinal retention studies remain necessary to clarify this distinction.

Short-term improvements in knee kinematics, gait symmetry, and metabolic efficiency were consistently reported across studies. However, these findings were predominantly derived from studies conducted in healthy participants under controlled laboratory conditions [30,31,40,49].

Evidence from clinical populations, including individuals with stroke, spinal cord injury, and cerebral palsy, remains comparatively limited and heterogeneous [29,42,44,51,57]. While some studies reported improvements in gait parameters and walking efficiency, these findings are based on small sample sizes, short-duration interventions, and variable study designs. Therefore, the strength of evidence supporting clinical effectiveness is substantially lower than that observed in healthy populations.

Importantly, most reported improvements occurred during active exoskeleton use, with limited assessment of retention or transfer of gains after device removal [31,44,57]. This raises uncertainty about whether the observed improvement represents true motor rehabilitation or temporary performance augmentation driven by mechanical assistance.

Short-term improvements in knee extension, step symmetry, and metabolic efficiency have been reported in stroke, spinal cord injury, pediatric cerebral palsy, and postsurgical populations [28,31,36,38,42,44,46,50,51]. In spinal cord injury, powered assistance reduced ambulatory effort compared to orthotic support [42]. Early postoperative implementation after total knee arthroplasty has been reported to be feasible and safe [28].

Nevertheless, most studies involve small cohorts and brief intervention periods [29,44,57]. Outcome measures frequently emphasize kinematic and biomechanical metrics, whereas participation-level outcomes and quality-of-life indicators are less consistently reported [35,37,53]. While biomechanical improvements confirm device functionality, sustained functional independence and integration into daily activities remain the ultimate rehabilitation goals.

The available evidence therefore supports short-term performance enhancement, but durable rehabilitation benefit has not yet been conclusively demonstrated. Larger randomized trials with extended follow-up are required to determine whether knee exoskeleton–assisted training yields meaningful, lasting functional gains.

Collectively, the literature indicates that knee exoskeleton technology has achieved a high degree of mechanical and algorithmic sophistication [39,48,54,56]. However, clinical validation remains comparatively early in development [31,44,57]. The field appears technologically advanced yet still consolidating in terms of rehabilitation evidence.

Future progress will likely depend on closer integration between biomechanical analysis, interface refinement, control robustness, and longitudinal clinical evaluation. Demonstrating sustainable functional outcomes in real-world contexts will ultimately determine whether knee exoskeletons transition from promising prototypes to established rehabilitation tools.

This review is limited by substantial heterogeneity in device design, control strategies, populations, and outcome measures, which precluded quantitative meta-analysis. Many included studies were exploratory or feasibility-focused and frequently involved healthy participants, limiting generalizability to clinical populations. Reporting of mechanical specifications, adverse events, and long-term follow-up was inconsistent, constraining cross-study comparison. Additionally, most studies assessed outcomes only during device use, providing limited insight into retention or neuroplastic effects.

Furthermore, a substantial proportion of included studies were conducted in healthy participants or involved early-stage feasibility or prototype validation, limiting the direct clinical applicability of the findings.

Future research should more explicitly align mechanical design and control strategies with specific rehabilitation objectives. Mechanical parameters such as torque capacity, range of motion, and assistance timing should be justified based on task demands and population needs rather than maximized by default. Adaptive and uncertainty-aware control approaches that show promise in healthy users should be systematically evaluated in clinical populations with attention to safety, usability, and interindividual variability.

Most importantly, future studies should move beyond short-term feasibility testing and controlled laboratory demonstrations. Well-designed, multicenter randomized controlled trials are needed, with larger, clinically representative samples, predefined primary outcomes, and standardized intervention protocols. Longer follow-up periods should be included to determine whether improvements persist after device removal and whether observed gains reflect true rehabilitation effects rather than temporary device-assisted performance.

Future trials should include patient-important outcomes, not only biomechanical parameters. These should include functional independence, walking, endurance, fall risk, participation in daily activities, quality of life, and user satisfaction. Greater standardization of outcome measures that integrate biomechanical, functional, and participation-level domains would strengthen cross-study comparisons and improve clinical interpretation. This would help determine whether knee exoskeletons can provide meaningful rehabilitation benefits beyond short-term improvements observed during assisted use.

In summary, knee exoskeleton devices demonstrated consistent short-term improvements in biomechanical and performance-related outcomes, particularly under controlled laboratory conditions [30,31,40,49]. However, the current evidence is largely derived from small-scale studies, frequently involving healthy participants or short-duration feasibility trials [29,44,57].

Evidence from clinical populations remains limited and heterogeneous, with insufficient high-quality data to support routine clinical implementation in physical therapy and rehabilitation practice. While preliminary findings suggest potential benefits for gait rehabilitation, these results should be interpreted cautiously due to methodological limitations, including small sample sizes, lack of long-term follow-up, and variability in outcome measures [29,42,44,57].

Future research should prioritize well-designed, multicenter randomized controlled trials with clearly defined primary outcomes, longer intervention and follow-up periods, and inclusion of patient-centered outcomes such as functional independence, fall risk, and participation. These factors are critical for establishing the true clinical value and translational potential of knee exoskeleton-assisted rehabilitation.