The Search for the Holy Grail in Running Biomechanics: Is There an Ideal Movement Profile for Minimizing Mechanical Overload?

Abstract

Background:

Running biomechanics can influence injury risk, but whether the combined effect of different biomechanical factors can be identified by individual running profiles remains unclear. Here, we identified distinct biomechanical profiles among healthy runners, examined lower limb mechanical load characteristics, and evaluated potential implications for injury risk.

Hypothesis:

Multiple factors would serve as a common denominator allowing identification of specific patterns.

Study Design:

Cross-sectional.

Level of Evidence:

Level 2.

Methods:

Step cadence, stance time, vertical oscillation, duty factor, vertical stiffness, peak ground reaction force (GRF), and anteroposterior, lateral, and vertical smoothness were determined from 3-dimensional kinematic data from 79 healthy runners using a treadmill at 2.92 m/s. Principal component analysis, self-organizing maps, and K-means clustering techniques delineated distinct biomechanical running profiles. Mutual information analysis, Kruskal-Wallis, and Pearson’s Chi-squared tests were conducted.

Results:

Five biomechanical profiles (P1-P5) demonstrated different running mechanical characteristics: P1 exhibited low cumulative and peak mechanical load due to a combination of high duty factor, low step cadence, and longer stance time; P2 showed characteristics associated with the lowest peak mechanical load due to reduced peak GRF and greater smoothness; P3 and P5 showed contrasting running patterns, but maintained moderate smoothness and peak GRF; and P4 exhibited the highest peak mechanical load, driven by high GRF, low duty factor, and high vertical oscillation.

Conclusion:

The 5 profiles appear to be associated with different lower limb load patterns, highlighting previously unrecognized connections between biomechanical variables during running. Some variables contribute to increased peak and cumulative load, whereas others help reduce it, underscoring the complex interplay of biomechanical factors in running.

Clinical Relevance:

Identifying distinct running profiles can help clinicians better understand individual variations in mechanical load and injury risk, thus informing targeted interventions, such as personalized training adjustments or rehabilitation programs, to prevent injuries and enhance performance in runners.

Biomechanical analyses are employed extensively as an objective approach to understanding and preventing running-related injuries. However, the literature on biomechanical risk factors for running injuries presents a contradictory landscape. A recent systematic review reported that 11 out of 21 running-related biomechanical risk factors are inconsistent among studies, reflecting the complex and multifactorial nature of these injuries. ³³ There is a clear need to transcend the traditional method of mapping individual risk factors, moving towards a comprehensive concept of an injury risk profile. ⁵ Identifying different movement profiles could be crucial in exploring innovative injury prevention strategies. ³⁵

The search for distinct movement profiles has garnered considerable attention in the literature. Building upon previous studies that have explored this concept in different contexts, such as knee osteoarthritis patients, ²³ recent investigations have delved into the characterization of running patterns. van Oeveren et al ³² introduced a classification identifying various running profiles, including hop, bounce, stick, push, and sit, based on the dual-axis framework theory. However, this classification still lacks empirical validation, as many of the proposed variables can interact. For example, in clinical practice, it is common to identify runners with a running technique that combines low cadence (push runners) and low duty factor (bounce runners), indicating the intricate nature of movement patterns and their interplay in the running biomechanics domain.

In medicine, using machine learning techniques to recognize patterns has become a powerful tool across various domains. A study by Droppelmann et al ¹³ demonstrated the high accuracy of deep learning algorithms in detecting tendon anomalies across multiple imaging modalities, suggesting their use as a valuable complementary tool in detecting musculoskeletal conditions. ¹³ These results highlight the potential of machine learning in facilitating precise diagnosis and informing clinical decision-making. This trend of utilizing artificial intelligence has also extended to the domain of running biomechanics in studies aiming to classify runners based on sex, speed, and injury status.^1,2,18 Applying machine learning techniques to biomechanical parameters makes it possible to identify commonalities in movement profiles across different people and describe distinct strategies each subgroup employs to handle loading response and propulsion, ultimately shedding light on the mechanical overload profile. By elucidating the complex interactions and strategies employed by injury-free runners, it could be possible to have some insights regarding the existing running profiles. This, in turn, could serve as a basis for understanding how different biomechanical parameters interact during running, allowing researchers and clinicians to structure new studies on biomechanical profiles and interventions for each subgroup.

This study aimed to apply machine learning techniques to biomechanical running parameters to identify movement profiles in injury-free runners. Furthermore, we aimed to describe each group's different demographic characteristics, loading, and propulsion strategies, providing insights into profiles that may lead to possible mechanical overload. We hypothesized that multiple factors would serve as a common denominator that would allow for the identification of specific patterns.

Methods

Participants

In this cross-sectional study, 79 healthy runners were included by convenience to participate in the study (Table 1). They were enrolled from 2019 to 2022. To be included, participants had to be recreational runners with ≥1 year of experience in treadmill running. They should have reported no running-related musculoskeletal pain in the lower limbs during the year before the assessment that caused a restriction or stoppage of running (distance, speed, duration, or training) for ≥7 days or 3 consecutive scheduled training sessions. In addition, participants had to be between 20 and 40 years of age and have a body mass index <30 kg/m². The study received approval from the institutional Ethical Committee for Human Experiments of the Instituto Nacional de Traumatologia e Ortopedia (approval no. 2.984.498). All participants provided informed consent before taking part in the study.

Table 1.

Demographic and biomechanical variables for the running profiles

	P1	P2	P3	P4	P5	P value differences
Sample size, n	12	16	15	14	22	n/a
Sex	10M (83.3%), 2F (16.7%)	10M (62.5%), 6F (37.5%)	7M (46.7%), 8F (53.3%)	7M (50%), 7F (50%)	9M (40.9%), 13F (59.1%)	.16
Age, y	28.0 (27.0-29.0)	30.0 (28.0-31.8)	29.0 (28.3-33.0)	29.5 (27.0-32.0)	29.0 (28.0-33.0)	.41
Body mass, kg	82.9 (73.2-86.7)	68.6 (60.3-77.2)	70.1 (55.5-77.9)	71.4 (59.6-81.6)	68.6 (64.0-75.1)	.04* P1 > P2, P4.
Height, cm	181.8 (176.7-183.4)	168.8 (165.5-174.3)	166.0 (163.8-177.4)	169.4 (163.2-177.1)	167.5 (162.0-170.0)	.003* P1 > P2, P4.
BMI, kg/m2	25.4 (23.9-25.9)	24.9 (21.0-26.5)	22.9 (20.7-27.3)	24.5 (23.0-25.9)	24.8 (22.4-26.3)	.78
Foot strike	10 RF (83.3%), 1 MF (8.3%), 1 FF (8.3%)	9 RF (56.3%), 5 MF (31.3%), 2 FF (12.5%)	11 RF (73.3%), 3 MF (20.0%), 1 FF (6.7%)	9 RF (64.3%), 2 MF (14.3%), 3 FF (21.4%)	16 RF (72.7%), 6 MF (27.3%), 0 FF (0.0%)	.43
Cadence, steps/min	49.3 (47.9-50.6)	53.6 (52.9-55.1)	50.0 (48.4-51.8)	51.7 (49.2-52.3)	54.5 (53.3-55.3)	<.001* P1, P3, P4 < P2, P5.
Stance time, seconds	0.84 (0.81-0.88)	0.81 (0.80-0.83)	0.83 (0.81-0.84)	0.78 (0.75-0.81)	0.77 (0.74-0.79)	<.001* P4 < P1. P5 < P1, P2, P3.
Vertical stiffness, kNm	25.5 (23.4-28.3)	24.6 (23.4-29.2)	23.4 (20.6-24.8)	25.4 (23.5-28.8)	31.0 (28.1-32.8)	<.001* P5 > P1, P2, P3, P4.
Vertical oscillation, m	0.10 (0.09-0.10)	0.09 (0.08-0.10)	0.10 (0.10-0.11)	0.11 (0.11-0.12)	0.09 (0.08-0.09)	<.001* P4 > P1, P2, P5. P3 > P2, P5.
Duty factor a.u.	0.35 (0.34-0.36)	0.35 (0.34-0.35)	0.34 (0.33-0.34)	0.33 (0.32-0.33)	0.34 (0.33-0.34)	<.001* P4 < P1, P2. P5 < P1.
Peak GRF, %BW	2.23 (2.22-2.28)	2.24 (2.21-2.29)	2.32 (2.30-2.39)	2.43 (2.36-2.48)	2.31 (2.29-2.36)	<.001* P4 > P1, P2. P3 > P1.
Anteroposterior smoothness	–3.71 (–3.77 to −3.63)	–3.64 (–3.73 to −3.60)	–3.74 (–3.78 to −3.70)	–3.91 (–3.93 to −3.80)	–3.75 (–3.80 to −3.71)	<.001* P1, P2 > P4.
Lateral smoothness	–3.37 (–3.65 to −3.25)	–2.87 (–3.04 to −2.48)	–3.08 (–3.22 to −2.87)	–3.61 (–3.66 to −3.41)	–3.37 (–3.46 to −3.24)	<.001* P2, P3 > P1, P4, P5.
Vertical smoothness	–2.53 (–2.56 to −2.52)	–2.50 (–2.54 to −2.48)	–2.52 (–2.54 to −2.46)	-2.57 (-2.63 to −2.54)	–2.54 (–2.56 to −2.51)	.001* P2, P3 > P4.
Biomechanical characterization	Low load (cumulative and per step)	Low load per step	Moderate load per step	High load per step	High cumulative load

Continuous variables are expressed as median (quartile 25% to 75%) and categorical variables are expressed as frequencies (% proportion). Cadence, stance time, vertical oscillation, duty factor, and smoothness are dimensionless variables. BMI, body mass index; BW, bodyweight; F, female; FF, forefoot strike; GRF, ground reaction force; M, male; MF, midfoot strike; P, profile; RF, rearfoot strike.

^*P < .05.

Running Data Collection

Kinematic data were sampled at 250 Hz using an 8 high-speed camera motion analysis system (Bonita 10, Nexus 2, Vicon). Reflective markers were attached to anatomical landmarks in the lower limbs and trunk. ³⁷ Four rigid plastic shells, each containing 4 markers, were placed bilaterally on the thighs and shanks. ³⁷ The anatomical markers on the thighs and shanks were removed after an initial static trial and reconstructed using the cluster markers during dynamic trials. A functional calibration trial was conducted to calculate joint centers. ¹⁵

Before testing, all participants performed a warm-up by walking on a motorized treadmill (model Loopband Run 600 XT Pro, Technogym, Gambettola FC) at 1.40 m/s (5 km/h) and running at 2.22 m/s (8 km/h) for 4 minutes, wearing their regular running shoes. Subsequently, they ran for 4 minutes at 2.92 m/s (10.5 km/h) while kinematic data were collected during the first 30 seconds of the third minute of running. The average of the 10 central stride cycles of the right lower limb was used for data analysis.

Biomechanical Data Processing

A least-square pose estimator was employed to minimize the effects of passive and active soft tissue movement artifacts on cluster markers. ⁸ The data were filtered using a fourth-order zero-lag low-pass Butterworth filter with a cutoff frequency of 12 Hz. Segment reference systems were defined according to the International Society of Biomechanics recommendations, ³⁴ and the 3-dimensional (3-D) lower limb angles were calculated using Nexus 2 software (Vicon Motion Systems). The geometric center of the 4 pelvic markers was used as a surrogate for the center of mass (CoM). ³⁶ The foot velocity algorithm was utilized to determine the stance and swing phases of each cycle. ²⁹

Step cadence was calculated as the number of foot contacts during the 30 seconds of data collection and multiplied by 2 to obtain the number of steps per minute. Stance time was determined as the time between foot contact and foot off. Vertical oscillation was calculated as the range of vertical displacement of the CoM during a stride cycle. ³² Duty factor was calculated as the ratio of stance time over stride time. ³² Vertical stiffness and peak active GRFs (peak GRF) were calculated based on the equations validated by Morin et al ²⁷ and applied by Giovanelli et al. ¹⁶ Anteroposterior, lateral, and vertical smoothness were calculated using the SPARC function (spectral arc length). ⁴ The trajectory of the heel marker was used to calculate smoothness. Smoothness reflects both practice-related skill improvements and the underpinning functional health of the neuromuscular system. ²¹

Cadence, stance time, and vertical oscillation were normalized according to Hof’s proposal to scale gait data to body size. ¹⁷ Peak GRF was normalized by individual body weight. Runners were classified as rearfoot, midfoot, or forefoot strikers, according to Altman and Davis, ³ based on the angle of the foot at the instant of ground contact. This study defined “load” as the forces exerted on tissues or biological structures. ²⁰ GRF was employed as a proxy for peak active load, owing to its strong correlation with joint moments. ²⁸ In addition, “cumulative load” was characterized as the relationship between the magnitude of the load and the number of loading cycles (such as step cadence in running). ²⁰

Statistical Analyses

The study used Leporace’s 4-step process using unsupervised machine learning to identify running profiles (see Appendix). ²³ The first step involved employing a principal component analysis (PCA) to select key biomechanical variables from a matrix of runners and variables. The variables were standardized, and a covariance matrix was used to compute PCA components and scores, retaining those explaining 90% of the variance.

Next, a self-organizing map (SOM) analyzed the PCA scores to find running patterns. The SOM used unsupervised learning to cluster data into a 2-dimensional grid, preserving the original data’s topology. Neurons in the SOM competed based on a “winner takes all” rule, with a 3-phase training process involving competition, cooperation, and weight updates. The optimal SOM size was determined by testing maps with varying neuron counts and analyzing errors and inactive neuron percentages. This method employs unsupervised competitive learning, mirroring an artificial neural network and identifying a novel data configuration. It effectively reduces the dimensionality of the multivariate dataset by clustering them into neurons, thus augmenting the capacity for subgroup analysis and unveiling previously unnoticed relationships.

K-means clustering was then applied to divide the dataset into homogenous groups, with multiple iterations to ensure optimal partitioning. Runners’ profiles were defined based on these clusters, using descriptive analyses and spider charts to visualize key characteristics, standardized with Z scores.

Kruskal-Wallis tests were used for numeric variable comparisons among profiles, and Pearson’s Chi-squared tests compared categorical variables such as sex and foot strike patterns. A significance level of 0.05 was set for all analyses conducted using Matlab software (2015, The Mathworks) and the somtoolbox package (http://www.cis.hut.fi/projects/somtoolbox/). All analyses are consistent with the CHAMP statement. ²⁶

We utilized mutual information (MI) to quantify the shared information between continuous biomechanical variables and peak GRF, ¹⁰ aiming to understand the relationships between these variables across different biomechanical profiles. MI was computed using a MATLAB-based approach, where the continuous data were first discretized according to Sturges’ rule to determine the optimal number of bins for each variable. The number of bins (numBins) was calculated using the equation numBins = ceil(log2[n] + 1), where n is the number of samples for each variable. The MATLAB discretize function was then used to assign datapoints to the appropriate bins. This discretization allowed us to capture both linear and nonlinear dependencies between variables, which might be missed by traditional correlation methods like Pearson’s correlation. Normalizing the MI values ensured comparability across variables with different entropies. The MI analysis in this study helped identify which variables shared the most information with GRF. While MI normalized ranged from 0 to 1, values closer to 1 indicate higher sharing of information between the variables.

Results

The results of the machine learning algorithms suggested that the most optimal division of runners was into 5 profiles (P1 to P5). Table 1 describes the demographic and biomechanical characteristics of each profile.

P1 demonstrated a higher body mass (P = .04) and height (P = .003) compared with P2 and P4, yet without yielding differences in BMI (P > .05). There was no difference in the proportion of women and men among the groups or in the initial contact pattern (P > .05). There were no statistically significant differences in demographic variables among P2, P3, P4, and P5 (P > .05).

The spider chart in Figure 1 indicates the predominance of P1 for lower step cadence, longer stance time, and higher duty factor. The combination of these variables, as supported by the mutual information results (see Appendix), presented a higher association with lower peak GRF (Figure 1), suggesting a low cumulative and active peak load.

Figure 1.

Running P1 is characterized by a lower step cadence, a longer stance time, and a higher duty factor, which leads to lower peak GRFs. The spider chart illustrates the distribution of the 9 biomechanical variables utilized in the machine learning algorithm (standardized by Z score). GRF, ground reaction force; P, profile.

P2 was characterized by higher cadence, lower vertical oscillation, and a higher duty factor and smoothness. The variables with the highest mutual information with peak GRF were duty factor, vertical oscillation, and anteroposterior and lateral smoothness (see Appendix). This combination was associated with the lowest peak GRF (Figure 2), indicating a low peak load biomechanics profile.

Figure 2.

Running P2 is characterized by higher cadence, lower vertical oscillation, and an higher duty factor, leading to the lowest peak GRFs and improved smoothness. The spider chart illustrates the distribution of the 9 biomechanical variables utilized in the machine learning algorithm (standardized by Z score). GRF, ground reaction force; P, profile.

P3 showed no clear predominance, featuring a lower step cadence, higher vertical oscillation, and a longer stance time. The combination of these variables resulted in lower vertical stiffness and moderate peak GRF, suggesting an intermediate profile related to load biomechanics (Figure 3). The variables with the highest mutual information with peak GRF were duty factor, stance time, and vertical oscillation (see Appendix).

Figure 3.

Running P3 is characterized by a lower step cadence, higher vertical oscillation, and a longer stance time, resulting in lower vertical stiffness and moderate peak GRFs. The spider chart illustrates the distribution of the 9 biomechanical variables utilized in the machine learning algorithm (standardized by Z score). GRF, ground reaction force; P, profile.

Profile 4 (P4) exhibited higher vertical oscillation, shorter stance time, and a lower duty factor. These 3 variables had the highest mutual information with peak GRF (see Appendix). This combination was associated with an increased peak GRF and reduced smoothness (Figure 4), suggesting a high peak load.

Figure 4.

Running Profile 4 is characterized by a higher vertical oscillation, shorter stance time, and lower duty factor, leading to increased peak GRFs and reduced smoothness. The spider chart illustrates the distribution of the 9 biomechanical variables utilized in the machine learning algorithm (standardized by Z score). GRF, ground reaction force; P, profile.

P5 demonstrated a higher cadence, lower vertical oscillation, moderate duty factor, and shorter stance time. These 4 variables had the highest mutual information with peak GRF (see Appendix). This combination was associated with an increased vertical stiffness and moderate peak GRF (Figure 5), suggesting a high cumulative load biomechanics profile.

Figure 5.

Running P5 is characterized by a highercadence, lower vertical oscillation, and a shorter stance time, resulting in highervertical stiffness and moderate peak GRFs. The spider chart illustrates the distribution of the 9 biomechanical variables utilized in the machine learning algorithm (standardized by Z score). GRF, ground reaction force; P, profile.

In the Appendix, we present a figure and an educational video featuring 3-D reconstructions from running P1 to P5 to summarize and illustrate the results.

Discussion

The main finding of this study was the identification of 5 distinct biomechanical running profiles among runners. Each profile elicited a unique lower limb mechanical load profile, potentially leading to varying injury risks. Furthermore, we found that certain combinations contribute to increased load whereas others appeared to reduce it. For instance, participants in P1 exhibit a lower step cadence, which the literature suggests may be a risk factor for injuries due to the potential for increased mechanical loads. ²² However, they also demonstrate low peak GRF, suggesting reduced peak mechanical loads. In this profile, the higher duty factor appears to play a more significant role in reducing mechanical load compared with step cadence, contributing to enhanced forward propulsion and a lower cumulative load due to fewer steps per specific distance.^20,31 This highlights the complex interaction between variables, where one factor may mitigate the impact of another rather than one resulting directly from the other. Although our findings indicate no significant difference in sex or foot strike patterns across the groups, these results should be interpreted cautiously due to the low number of female participants and midfoot/forefoot strikers in some profiles, which may limit the generalizability of the biomechanical loading patterns.

The dual-axis model describes running styles at a given speed, focusing on 2 primary variables: step cadence and duty factor. ³² This model suggests that these 2 parameters alone could adequately delineate a runner’s biomechanical profile. However, emerging evidence implies that this perspective might be simplified. For instance, research by Patoz et al ³⁰ and Lussiana et al ²⁵ indicated that step cadence and duty factor alone do not adequately characterize running economy. This limitation may be attributable to the complex interplay among biomechanical variables, emphasizing the intricacy of running biomechanics and the shortcomings of models focusing on isolated factors. ³³ Our findings support and expand upon this emerging perspective, revealing distinct biomechanical profiles and intricate interactions between variables. These insights advocate for a more comprehensive understanding of running biomechanics, with significant implications for acute and chronic long-term training effects and injury prevention.^9,16

Among the identified profiles, Profile 2 exhibits characteristics linked with the lowest peak lower limb load: reduced peak GRFs and enhanced smoothness across the 3 planes of movement. Despite a high step cadence and moderate stance time, runners in this group appear to maintain a high duty factor, likely attributable to a shortened swing phase. Consequently, they exhibit less vertical oscillation, resulting in diminished GRFs and mitigated peak lower limb load during the absorption of external mechanical loads.

On the other hand, P4 exhibits biomechanical characteristics indicative of the highest peak lower limb load. Predominant features of this profile include high GRFs, low duty factor, and increased vertical oscillation. Such an increase in GRF is associated with an increased injury risk in runners.^19,33 This results in a running style with reduced smoothness. This group may benefit from a retraining program to reduce peak forces, increase contact time, and enhance running smoothness. ²⁴ Derie et al ¹¹ demonstrated that a music-based biofeedback gait retraining program effectively reduces peak forces by decreasing vertical oscillation of the CoM, which could be ideal for runners in this group. In addition, this group should be cautious when transitioning to a forefoot strike pattern, as it could further reduce contact time, ¹⁴ thereby decreasing the duty factor even more.

P1 demonstrated a combination of low step cadence and a longer stance phase, which may reduce peak load due to the high duty factor leading to lower peak GRFs. ⁶ As our results indicated, the low cadence observed in this profile suggests an increased step length, inferred from the inverse relationship between cadence and step length at a constant running velocity. Interestingly, this longer step length did not correspond to an increase in vertical oscillation, highlighting the complex and nonlinear interplay of biomechanical variables. This mechanism may reduce cumulative load by decreasing the total number of steps required to cover a given distance. However, it remains uncertain whether these running mechanics are influenced by the group’s higher body mass and height, despite all variables input into the model being normalized to individual body dimensions. ¹⁷ Despite the low cadence, this group demonstrated low mechanical load, indicating that there may not be a need for running retraining by increasing cadence. These findings are significant because increasing cadence is a validated method for reducing GRFs, ¹² but our study suggests it may not be necessary for all runners with low cadence, as P1 compensates with other biomechanical variables to maintain low load.

P3 and P5 represent divergent running patterns. Whereas P5 is marked by a high step cadence and increased vertical stiffness, P3 is defined by a low step cadence and reduced vertical stiffness. Despite this, both profiles maintain moderate smoothness and peak GRF. Given the higher cadence, P5 may incur an increased cumulative load compared with P3. Although stiffness may be considered beneficial to athletic performance, high stiffness levels have also been associated with an increased likelihood of lower extremity injury. ⁷ Thus, P5 could benefit from neuromuscular training aimed at modulating lower limb vertical stiffness to aid in the absorption of mechanical loads and increase the smoothness of movement. Conversely, P3 may be the most eligible group for cadence retraining to reduce mechanical overload.

Clinical Implications

This study offers new insights into running biomechanics by highlighting the nonlinear interactions among biomechanical variables, challenging the traditional approach of examining isolated risk factors. Our findings suggest 5 distinct biomechanical profiles, each associated with different loading and propulsion strategies that may affect injury risk. Rather than a “one-size-fits-all” approach, these profiles indicate the potential for more personalized injury prevention strategies. Clinical practitioners could use these insights to tailor training regimens based on individual biomechanical profiles, possibly improving the balance between an athlete’s capacity and demands and potentially reducing injury prevalence. Future studies are needed to evaluate the effects of neuromuscular training and running technique modifications, such as changes in cadence, foot strike pattern, and impact reduction, on the load distribution and biomechanical profiles identified here. Understanding the effects of these modifications could inform more personalized training and rehabilitation strategies.

Limitations

Our study has inherent limitations. First, the sample consisted exclusively of healthy runners, which aligns with our study’s objectives but limits the generalization to runners with pre-existing injuries, chronic health conditions, or novice runners. Second, while we analyzed various biomechanical variables, our analysis did not include other factors that may contribute to joint overload and injury risk, such as joint moments, stride asymmetry, and pelvic motion. Furthermore, the cross-sectional design of the study offers only a single snapshot of running biomechanics. Longitudinal studies would be necessary to provide a more comprehensive understanding of how these biomechanical profiles relate to injury risk over time.

We also acknowledge that the sample size, while sufficient for the initial development of our model, was relatively small. Future studies should aim to validate these findings in larger populations and include both male and female runners, with a higher representation of each foot strike pattern. We also did not control for injury history beyond the past year, which could have influenced the observed biomechanical patterns. Fatigue, known to alter running mechanics, was also not considered a variable in this study. Finally, while we proposed potential clinical applications based on our findings, these should be interpreted as hypotheses that require validation in larger cohorts and longer-term studies to ensure their efficacy in injury prevention.

Conclusion

We identified 5 distinct biomechanical running profiles among healthy runners based on 3-D kinematic data. These profiles appear to be associated with different lower limb load patterns, highlighting previously unrecognized connections between various biomechanical variables during running. Some variables contribute to increased peak and cumulative load, while others help reduce it, underscoring the complex interplay of biomechanical factors in running. Given the cross-sectional nature of the study and the sample size, our findings should be interpreted with caution. Nonetheless, this insight contributes to a more comprehensive understanding of running biomechanics and provides a foundation for future research to develop strategies for injury prevention.