October 22, 2025
A growing body of research in resistance training suggests that exercising at longer muscle lengths may confer distinct advantages for muscular adaptation. Specifically, training muscles in a lengthened position has been shown to promote hypertrophy [1], [2] and induce larger improvements in dynamic strength as compared to training in shortened positions [3]–[5]. In fact, recent research suggests that performing exercises through a partial range of motion (pROM) in the lengthened portion of the movement can elicit adaptations that are comparable to those achieved using a full range of motion (fROM), which has been considered the gold standard for optimizing strength and hypertrophy [6]–[10]. These findings have been amplified by the science-based lifting community, particularly through online forums and social media content [11], [12].
Despite the robust evidence, translating these insights into practical training recommendations remains nontrivial. Most studies on range of motion (ROM) in resistance training either measure performance across multiple predefined joint-angle intervals [2] or use qualitative labels such as “full” and “partial” repetitions [13]. This introduces ambiguity about which joint angles constitute a partial repetition, whether those definitions are consistent across exercises with different strength curves, and whether the achieved ROM during training matches the intended prescription. Furthermore, most studies also concentrate on a single exercise, potentially limiting generalizability even when precise joint-angle definitions are defined. Moreover, while many studies use randomized controlled trials to compare outcomes – such as muscle thickness or strength – between groups [4], [14], detailed aspects of movement execution are overlooked, such as repetition tempo, the durations of concentric and eccentric phases of an exercise, and the within-set variability under different ROM conditions. As a result, sources of movement variation and their effects on performance remain under-explored [15]. Consequently, it is unclear whether performance differences arise solely from a reduced ROM and whether exercise-specific effects play a role. Finally, the absence of motion analyses, due to the cost and logistical complexity of employing laboratory motion-capture systems, further limits understanding of how kinematic and temporal factors may contribute to the performance and adaptation differences reported in the extant literature.
To address these gaps, this study develops an AI-based pose estimation pipeline to analyze upper-body resistance exercises performed under lengthened-partial (pROM) and full-range (fROM) conditions. The dataset consists of publicly available videos from [13], which examined the effects of pROM and fROM on muscular adaptation. Each video is processed to extract a joint-angle time series for every exercise (8) and ROM condition (2), from which several metrics are derived, including the starting and ending angles of each repetition, the per-repetition ROM (maximum minus minimum angle), average duration of repetitions, durations of the concentric and eccentric phases, and active body side (left or right). We use this data to construct a crossed, random-effects, meta-analytic regression model that captures baseline participant- and exercise-level variation, as well as changes in response to the performance of lengthened partials. Analysis of this model, together with several post-hoc statistical tests, enables us to: (i) identify differences in repetition tempo and phase durations between pROM and fROM; (ii) compare performance variability across participants and exercises; and (iii) establish an exercise-independent criterion for defining pROM assessing its consistency across exercises. In doing so, this study presents a novel application of AI-based pose landmark detection and tracking to generate insights in exercise science, demonstrating its potential to enhance the precision and scope of experimental research.
Analysis of 280 resistance-training videos processed by the AI pose estimation pipeline revealed that pROM repetitions corresponded to approximately 56% of the full range of motion (fROM) and were consistent across most exercises (except the incline press), indicating that lengthened partials represent a relatively standardized fraction of total movement range. Repetitions performed under pROM were slightly shorter in duration, primarily due to faster eccentric phases, suggesting that pROM alters both the mechanical and temporal characteristics of movement execution. Variance decomposition showed that pROM increased both between-participant and between-exercise variability as compared to fROM, reflecting greater heterogeneity in how the movement was performed. However, the relative contribution of each source of variability remained stable across ROM conditions, implying that lengthened partials amplify existing differences without altering their underlying structure. Together, these findings demonstrate that lengthened partials exhibit distinct kinematic characteristics beyond the prescribed ROM and, despite introducing additional variability, a consistent and quantifiable definition can still be established across exercises.
In this section, we describe the videos that were collected, the AI pipeline used for video processing, and the process used to construct the final tabular dataset to derive quantitative insights.
Videos were obtained from the dataset accompanying [13], which investigated the effect of lengthened partial (pROM) versus full ROM resistance training across eight upper-body exercises in 26 participants using a within-subject design. In the study, each participant performed a series of upper-body exercises under both ROM conditions during an 8-week intervention period. Training was conducted twice weekly, comprising four exercises per session and four sets per exercise, under the supervision of at least one research assistant. Study participants performed all sets to momentary muscular failure, defined as the inability to complete another repetition despite maximal effort, while research assistants provided verbal encouragement and monitored adherence to the prescribed ROM. The order of ROM training was randomized across sessions, loads were adjusted to maintain the target repetition range and intensity, and multiple repetition ranges were used across exercises. Finally, all training sessions were separated by at least 48 hours.
| Exercise | Full ROM | Partial ROM | ||
| Videos | Repetitions | Videos | Repetitions | |
| Bayesian Curl | 19 | 136 | 22 | 200 |
| Cable Pushdown | 20 | 154 | 20 | 157 |
| Dumbbell Curl | 18 | 175 | 17 | 162 |
| Dumbbell Overhead Extension | 13 | 151 | 11 | 106 |
| Dumbbell Row | 17 | 163 | 17 | 188 |
| Flatpress | 17 | 149 | 15 | 160 |
| Incline Press | 17 | 197 | 17 | 142 |
| Lat Pulldown | 21 | 184 | 19 | 158 |
| Total | 142 | 1309 | 138 | 1273 |
Not all combinations of exercise types and ROM classifications were represented in the video repository. Of the 3091 available videos, 6 were identified as duplicates2, and 23 were excluded due to excessive occlusion from identity masking or poor camera quality. The high number of excluded videos reflects a limitation of AI pose analysis evaluation. Although the approach can estimate human body landmarks during brief occlusions or out-of-frame movements, prolonged obstructions substantially reduce the accuracy of ROM metrics, resulting in unreliable assessments. Furthermore, most excluded videos contained few usable repetitions, justifying their removal.
The remaining 280 videos were manually verified to assign exercise labels and ROM classifications (partial vs. full); a detailed breakdown is provided in Table 1. To assess the accuracy of the pose estimation pipeline, roughly 28% of the videos (78 videos from 9 participants) were manually annotated to confirm both the body side that was used to perform the exercise (left or right) and the number of exercise repetitions. These manual observations were compared with the algorithm’s outputs, and discrepancies greater than two repetitions per set were flagged for review and used to help refine the pipeline. In this test set, the final pose estimation pipeline correctly identified the exercising body side in 73 videos (93.6%), achieved an average deviation of 1.55 repetitions from the ground truth, and agreement within \(\pm\)2 repetitions in 85.9% of clips (68 videos).
For each video, several summary statistics were computed, including per-repetition ROM (mean and standard deviation in brackets), average repetition duration, and mean concentric and eccentric phase durations. When participants had multiple recordings for the same exercise and ROM condition, their data was aggregated into a single record. To reduce bias, the first and last repetitions of each set, often atypical due to initial setup errors or fatigue, were excluded from subsequent analyses. Table 2 summarizes the data obtained from the pose estimation pipeline and suggests that there are clear and expected differences in range-of-motion between fROM and pROM across exercises. While temporal metrics indicate that fROM repetitions seem to be, on average, slightly longer in duration than pROM repetitions, concentric durations were generally comparable. Furthermore, the number of repetitions per exercise and ROM condition varied, suggesting task- or movement-specific strategies rather than uniform volume adjustments. Finally, body-side proportions indicate a modest overall left-side bias that is more pronounced under pROM than fROM.
Each video was processed using MediaPipe Pose [16], a deep-learning–based computer vision framework for real-time human pose estimation. The system performs a multi-stage sequence of visual inference steps on every frame, including image preprocessing, person detection, landmark identification, and temporal motion tracking. These stages are powered by pre-trained convolutional neural networks that are optimized on large-scale human motion datasets.
| Average | Rep | Eccentric | Concentric | Exercise | Body | |
| Exercise | ROM (°) | Duration (s) | Duration (s) | Duration (s) | Repetitions | Side |
| Bayesian Curl | 100.92 (38.72) | 4.52 (0.88) | 2.39 (0.49) | 2.13 (0.62) | 7.84 (1.93) | 0.58 |
| Cable Pushdown | 112.61 (12.18) | 4.32 (0.94) | 1.89 (0.75) | 2.43 (0.41) | 8.35 (1.96) | 0.60 |
| Dumbbell Curl | 127.65 (16.45) | 4.32 (1.05) | 2.22 (0.57) | 2.09 (0.79) | 11.83 (4.87) | 0.50 |
| Dumbbell Overhead Extension | 127.86 (44.61) | 4.01 (0.54) | 2.29 (0.53) | 1.72 (0.43) | 13.37 (3.95) | 0.54 |
| Dumbbell Row | 96.53 (23.71) | 3.70 (0.58) | 2.11 (0.52) | 1.59 (0.44) | 10.58 (2.97) | 0.53 |
| Flatpress | 125.16 (28.09) | 3.58 (0.35) | 1.86 (0.40) | 1.72 (0.44) | 9.85 (2.90) | 0.47 |
| Incline Press | 115.51 (55.32) | 3.44 (0.50) | 1.71 (0.36) | 1.73 (0.52) | 12.12 (2.18) | 0.47 |
| Lat Pulldown | 118.80 (30.31) | 3.80 (0.69) | 2.16 (0.66) | 1.64 (0.44) | 11.59 (5.65) | 0.62 |
| Bayesian Curl | 88.28 (19.06) | 3.81 (0.68) | 2.04 (0.38) | 1.77 (0.69) | 10.86 (4.01) | 0.36 |
| Cable Pushdown | 78.91 (25.61) | 3.92 (0.67) | 1.52 (0.54) | 2.39 (0.43) | 8.74 (2.35) | 0.40 |
| Dumbbell Curl | 91.93 (19.30) | 4.31 (0.93) | 2.34 (0.35) | 1.96 (0.72) | 10.06 (1.91) | 0.65 |
| Dumbbell Overhead Extension | 110.07 (52.87) | 4.03 (1.52) | 2.22 (0.83) | 1.81 (1.17) | 14.26 (8.54) | 0.82 |
| Dumbbell Row | 74.89 (12.11) | 3.44 (0.41) | 1.94 (0.32) | 1.50 (0.30) | 12.42 (4.27) | 0.65 |
| Flatpress | 86.78 (37.04) | 3.24 (0.39) | 1.53 (0.43) | 1.71 (0.43) | 12.62 (4.25) | 0.73 |
| Incline Press | 57.72 (40.01) | 3.29 (0.50) | 1.57 (0.44) | 1.72 (0.47) | 9.69 (2.47) | 0.65 |
| Lat Pulldown | 92.38 (30.45) | 3.92 (0.82) | 2.21 (0.56) | 1.71 (0.59) | 9.68 (2.89) | 0.53 |
In the first stage, frames are normalized and transformed into a format suitable for 2D human landmark recognition. A person detection network then identifies the bounding region containing the human subject. Within this region, a pose estimation network predicts the two-dimensional (2D) coordinates and confidence scores of 33 anatomical landmarks, which represent key joints and reference points across the human body. Each landmark is then assigned a visibility probability indicating the confidence that it has been correctly localized. To stabilize the trajectories over time, a temporal tracking module maintains landmark identity across consecutive frames, reducing jitter, and ensuring motion continuity despite the possibility of partial occlusions or rapid movement.
In each frame, the joint angle \(\theta\) can be computed using the coordinates of three adjacent landmarks: a proximal point (a); a joint center (b); and a distal point (c). That is, \[\theta := \cos^{-1}\left( \frac{(a - b) \cdot (c - b)}{\|a - b\| \; \|c - b\|} \right)\] where the vectors \(a, b\) and \(c\) represent the two-dimensional pixel coordinates of the relevant anatomical landmarks, \(\| \cdot \|\) is the \(L^2\) norm, and \(\cos^{-1}(\cdot)\) is the inverse cosine. All coordinates are scaled to the native pixel dimensions of the video to preserve proportionality across participants and mitigate different camera setups. Only frames where all contributing landmarks exceeded a visibility threshold of \(0.5\) were included in the analysis [16]. However, to ensure the analyzed signal reflected the intended movement, the pipeline first attempted to record an exercise-mapped angle (e.g., elbow flexion for curls; shoulder flexion for lat pulldown). If the mapped signals were incomplete, a dominant angle was selected using a coverage score. Because all exercises were performed unilaterally, body side (left vs. right) was also inferred from the selected angle.
As a consequence of this process, a continuous time series of joint angles was obtained for each video, which represented the frame-by-frame kinematic profile of the exercise being performed. To reduce the impact of outliers and tracking errors, the ROM for each exercise was defined as the difference between the \(5^{\textrm{th}}\) and \(95^{\textrm{th}}\) percentiles of the joint-angle distribution across all frames. To reduce frame-to-frame fluctuations and produce smoother motion trajectories, the joint-angle time series were processed using a Savitzky–Golay filter [17] with a window length of eleven frames and polynomial order of two. This technique attenuates high-frequency noise while preserving the natural curvature and peaks of the original signal. Such smoothing is important in pose estimation because even minor pixel-level jitters can introduce artificial variability in angle measurements and distort the derived metrics [18], [19]. Furthermore, small gaps in landmark visibility (\(< 2\) seconds) were linearly interpolated to reduce excessive signal loss, while longer occlusions were left unfilled to ensure ROM measurements were accurate.
The smoothed joint-angle time series were analyzed with a trough-based method that detects flexion-extension cycles from angular minima and maxima. We automatically adjusted the detection thresholds to each clip’s motion and cadence (using robust amplitude and an autocorrelation-based cadence estimator) to avoid double-counting and missing repetitions. These adaptive rules were bounded by conservative floor and ceiling parameters to avoid over- or under-segmentation. When the primary joint angle value was unreliable (e.g., brief occlusions or moving out of frame), we used two fallbacks, selecting the option that produced the most plausible cadence: (i) data from a nearby joint on the same side was scaled to match the focal joint’s ROM; and (ii) blended the information from both joints and weighted each according to how visible they were in the video.
The thresholds for the time between repetitions, ROM, and prominence were empirically tuned to the dataset (minimum inter-repetition distance between troughs: 2.00 seconds; minimum repetition duration from peak to trough: 0.30 seconds; minimum ROM: 10°; minimum prominence: 5-10°) based on the manual review of 78 videos. For each detected repetition, several metrics were extracted, including the starting and ending angle of the repetition, total exercise duration (seconds), concentric and eccentric phase durations (seconds), body side (left or right), and per-repetition ROM (maximum minus minimum angle). An internal extremum within each repetition was used to divide the movement into concentric (shortening) and eccentric (lengthening) phases.
Given that video exclusions are plausibly related to observable factors such as exercise type and setup, but participants recording the videos were unaware the footage would later be analyzed, we treat missingness as missing at random [20]. Under this assumption, and given the within-subject design of the original experiment, we leveraged the joint-angle time series for statistical analysis. However, not all participants completed every exercise under both pROM and fROM, and the number of successfully segmented repetitions varied by exercise and ROM condition, resulting in an unbalanced and crossed dataset. Thus, we fit a crossed random-effects meta-regression [21] with participants and exercises as random factors.
In this framework, participant \(i\) contributes performance score \(Y_{ieg}\), an estimated standard deviation \(s_{ieg}\), and number of repetitions
\(k_{ieg}\) for exercise \(e\) and ROM condition \(g \in \{\textrm{pROM},\textrm{fROM}\}\). Because participants and exercises are random samples from larger
populations, and each participant completed multiple exercises under both ROM conditions, we analyze the following model: \[\label{meta95regress}
Y_{ieg} = \beta_0 + \beta_1\,\mathbf{1}\{g=\mathrm{pROM}\} + \beta_2\,\mathbf{1}\{\mathrm{sex}_i=\mathrm{F}\}
+ p_i + u_e + (q_i + v_e)\,\mathbf{1}\{g=\mathrm{pROM}\} + \varepsilon_{ieg},\tag{1}\] where \(\beta_0\) denotes the common intercept, \(\beta_1\) represents the average fixed
effect of the pROM condition relative to fROM, and \(\beta_2\) is the fixed effect of sex noting that only four participants were female. The random terms \(p_i\) and \(q_i\) correspond to participant-specific intercept and slope effects, respectively, capturing between-participant heterogeneity in both baseline performance and in their response to the pROM condition. Similarly, \(u_e\) and \(v_e\) represent exercise-specific random intercept and slope effects, which account for systematic differences across exercises and their differential sensitivity to pROM. The random
effects are assumed to follow bivariate normal distributions: \[\begin{bmatrix} p_i\\ q_i \end{bmatrix}
\sim \mathcal{N}\!\left(\mathbf{0},
\begin{bmatrix}
\tau_p^2 & \xi\,\tau_p\tau_q\\
\xi\,\tau_p\tau_q & \tau_q^2
\end{bmatrix}\right),
\qquad
\begin{bmatrix} u_e\\ v_e \end{bmatrix}
\sim \mathcal{N}\!\left(\mathbf{0},
\begin{bmatrix}
\tau_u^2 & \rho\,\tau_u\tau_v\\
\rho\,\tau_u\tau_v & \tau_v^2
\end{bmatrix}\right),\] where \(\tau_p^2\) (\(\tau_u^2\)) and \(\tau_q^2\) (\(\tau_v^2\)) denote the variance
components for the intercept and slope terms at the participant (exercise) level, and \(\xi\) (\(\rho\)) represents the corresponding correlation between the intercept and slope effects. The
residual term \(\varepsilon_{ieg}\sim\mathcal{N}(0,\sigma^2_{ieg})\) represents observation-level error, with known sampling variance \(\sigma^2_{ieg}=s^2_{ieg}/k_{ieg}\) which is used to
define inverse-variance weights \(w_{ieg}=1/\sigma^2_{ieg}\). Model parameters were estimated using restricted maximum likelihood (REML) in the metafor package [22] within the R statistical computing environment [23]. We report the model-based REML standard errors provided by the metafor package, noting that the significance of all fixed effect terms were unchanged when using cluster-robust standard errors computed by
participant or exercise [24].
Conceptually, the fixed-effect terms quantify the overall difference between ROM conditions after adjusting for sex, which has been previously shown to influence performance outcomes [25], [26]. The participant-level random intercepts and slopes capture individual variability in both baseline performance and responsiveness to the pROM condition, reflecting that participants may differ not only in their innate abilities, but also in how their performance changes across ROM conditions. Similarly, exercise-level random intercepts and slopes allow the magnitude of the pROM effect to vary across exercises. This crossed random-effects specification appropriately partitions the total variance into within-participant, between-participant, and between-exercise components, thereby accounting for the repeated-measures nature and crossed design of the study. The inclusion of inverse-variance weights further ensures that exercises or participants with more precise performance estimates (e.g., based on a greater number of repetitions or lower measurement error) exert greater influence on the estimation of fixed and random effects.
To confirm the robustness of our findings, we systematically compared alternative plausible formulations of the meta-regression model. Specifically, we assessed models that varied in their random-effects structure, including those that (i) only incorporated participant-level random intercepts; (ii) only included exercise-level random intercepts; and (iii) allowed random slopes for the pROM condition at either the exercise or participant levels only. We found that, in addition to being theoretically justifiable, 1 was the best-fitting model as it achieved the lowest Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) values [27].
We proceed by conducting four sets of analyses. In Section 3.1, we examine the mechanical and temporal features of the different movements using Model 1 to both validate the AI-derived data and identify additional defining features of lengthened partials in addition to ROM. Then, in Section 3.2, we analyze the covariance structure of the random-effects model to identify sources of variation and how they are influenced by the treatment (pROM). In Section 3.3, we evaluate the relative contributions of between-participant and between-exercise variation to total variance. Finally, in Section 3.4, we propose a standardized definition of pROM by expressing ROM on a logarithmic scale and using this definition to estimate a data-driven, exercise-agnostic threshold.
Let \(Y_{ieg}\) denote the repetition tempo or phase duration of each exercise and ROM condition. For benchmarking and validation, we also assess average range-of-motion (in degrees). By estimating Model 1 , the sign and magnitude of \(\beta_1\) indicates whether pROM is associated with systematic changes in these metrics. A significant effect would suggest that performing lengthened partials influences not only ROM (by definition), but also the overall execution of the different exercises.
6pt
| Rep Duration (s) | Eccentric Phase (s) | Concentric Phase (s) | Range of Motion (°) | |||||
|---|---|---|---|---|---|---|---|---|
| 2-3 (lr)4-5 (lr)6-7 (lr)8-9 | Estimate | SE | Estimate | SE | Estimate | SE | Estimate | SE |
| Intercept | 3.831\(^{***}\) | 0.187 | 2.080\(^{***}\) | 0.149 | 1.804\(^{***}\) | 0.133 | 120.390\(^{***}\) | 8.228 |
| pROM | -0.303\(^{**}\) | 0.114 | -0.201\(^{*}\) | 0.098 | -0.119 | 0.101 | -59.054\(^{***}\) | 12.043 |
| Female | 0.220 | 0.231 | 0.071 | 0.161 | 0.041 | 0.152 | 13.824 | 8.851 |
| \(\tau_{p}^2\) | 0.169 | 0.126 | 0.094 | 683.211 | ||||
| \(\tau_{q}^2\) | 0.254 | 0.097 | 0.099 | 268.491 | ||||
| \(\xi\) | 0.759 | 0.517 | 0.499 | 0.501 | ||||
| \(\tau_{u}^2\) | 0.209 | 0.131 | 0.104 | 315.928 | ||||
| \(\tau_{v}^2\) | 0.172 | 0.229 | 0.074 | 480.158 | ||||
| \(\rho\) | 0.846 | 0.931 | 0.766 | -0.261 | ||||
| AIC | 1090.041 | 1106.184 | 840.542 | 32406.720 | ||||
| \(Q_E\) | 2974.69\(^{***}\) | 4764.71\(^{***}\) | 4329.97\(^{***}\) | 184693.07\(^{***}\) | ||||
| \(Q_M\) | 4.00\(^{*}\) | 2.20 | 0.74 | 13.24\(^{***}\) | ||||
Table 3 reports parameter estimates from the crossed meta-regression models corresponding to Model 1 . Across all models and after adjusting for participants’ sex, the pROM condition was associated with significantly shorter repetition durations (\(\beta_1=-0.30\), SE \(=0.11\), \(p<0.01\)) and a markedly reduced range of motion (\(\beta_1=-59.05\), SE \(=12.04\), \(p<0.001\)). Effects for the eccentric (\(\beta_1=-0.20\), SE \(=0.10\), \(p<0.05\)) and concentric (\(\beta_1=-0.12\), SE \(=0.10\), n.s.) phases were smaller in magnitude and not as statistically robust as the other outcomes. Nevertheless, the sum of the phase estimates (\(\sim-0.32\) s) is consistent with the overall reduction in repetition duration. Notably, the sex coefficient (female vs.male) was small and nonsignificant across all outcomes, indicating that the inclusion of sex as a covariate did not materially alter the estimated pROM effects.
The estimated variance components indicated substantial heterogeneity among repeated observations at both the exercise and participant levels. The correlation between exercise-specific random intercepts and slopes (\(\rho\)) was strong and positive for repetition duration (\(\rho=0.85\)), eccentric phase (\(\rho=0.93\)), and concentric phase (\(\rho=0.77\)), suggesting that exercises with longer baseline durations tended to exhibit smaller pROM-induced reductions. In contrast, the negative correlation for range of motion (\(\rho=-0.26\)) implies that exercises characterized by greater baseline ROM displayed somewhat larger decrements under pROM conditions. At the participant level, the corresponding correlations (\(\xi\)) were moderate to strong, and positive across all outcomes (ranging from \(\xi=0.50\) to \(\xi=0.76\)), indicating that individuals who generally performed slower or with larger ranges of motion tended to maintain those relative tendencies across both ROM conditions.
Moderator tests (\(Q_M\)) showed that the pROM condition accounted for a significant proportion of variance in repetition duration (\(Q_M=4.00\), \(p<0.05\)) and range of motion (\(Q_M=13.24\), \(p<0.001\)), whereas the effects for the eccentric and concentric phases did not reach statistical significance (\(Q_M=2.20\) and \(0.74\), respectively). Nonetheless, substantial residual heterogeneity (\(Q_E\), all \(p<0.001\)) persisted across all models, indicating that unmeasured factors contributed to variability.
The results confirm that although pROM was associated with a substantially reduced ROM (as expected), the corresponding decrease in repetition duration – while statistically significant – was relatively modest (\(\sim\)8%). Because participants in [13] were prescribed load adjustments to maintain target repetition ranges and effort intensity, both ROM conditions likely produced similar time under concentric tension. Given the relatively greater importance of contraction-phase duration [28], [29], this may help explain why prior studies have often reported comparable muscle hypertrophy and strength adaptations between full and partial ROM training conditions [14], [30], [31].
We next characterize how performance variability differs between ROM conditions at both the participant and exercise levels. To do so, we conducted two sets of hypothesis tests. The first evaluated whether baseline performance (intercept) and condition-specific responsiveness to pROM (slope) were statistically uncorrelated; a significant result would imply that pROM has a heterogeneous effect on performance outcomes. The second tested whether the variance components associated with fROM and pROM were equal at a given level; a significant result implies that the magnitude of between-participant or between-exercise variation differed across ROM conditions.
More specifically, to assess exercise-level sensitivity to pROM, we tested: \[H_{0}^{(\rho)} : \rho = 0.\] This hypothesis was evaluated using a likelihood-ratio test (LRT) by comparing a model with an unstructured (UN) exercise-level covariance matrix (\(\rho \neq 0\)) to an identical model with a diagonal (DIAG) covariance structure where \(\rho=0\). The test statistic \(\Lambda_{\rho} = 2(\ell_{\text{UN}} - \ell_{\text{DIAG}})\) is asymptotically \(\chi^2_1\) under \(H_{0}^{(\rho)}\), where \(\ell_{\text{UN}}\) and \(\ell_{\text{DIAG}}\) denote the maximized log-likelihoods of the respective models. A significant result indicates a nonzero correlation between intercept and slope terms, implying that exercises with higher baseline performance exhibit systematically different pROM effects.
We next tested whether between-exercise variability differed between fROM and pROM. \[H_{0}^{(uv)} : \tau_{u}^2 = \tau_{v}^2.\] This was evaluated by comparing the UN covariance model to a compound-symmetry (CS) model that constrains \(\tau_{u}^2 = \tau_{v}^2\) while allowing \(\rho\) to vary freely. The LRT statistic, \(\Lambda_{uv} = 2(\ell_{\text{UN}} - \ell_{\text{CS}})\) is again asymptotically \(\chi^2_1\) under \(H_{0}^{(uv)}\). A significant result implies that the degree of between-exercise variation in pROM responsiveness differs from that of baseline performance of fROM.
| Rep Duration (s) | Range of Motion (°) | |
|---|---|---|
| \(H_{0}^{(\rho)}\) | 8.920\(^{**}\) | 0.482 |
| \(H_{0}^{(uv)}\) | 0.199 | 0.312 |
| \(H_{0}^{(\xi)}\) | 15.966\(^{***}\) | 6.723\(^{**}\) |
| \(H_{0}^{(pq)}\) | 1.484 | 6.464\(^{*}\) |
Analogous tests were conducted at the participant level. Specifically, we examined the extent of individual heterogeneity in response to pROM: \[H_{0}^{(\xi)} : \xi= 0,\] as well as differences in the magnitude of participant-level variation between ROM conditions: \[H_{0}^{(pq)} : \tau_{p}^2 = \tau_{q}^2.\]
We report results for Rep Duration (s) and Range of Motion (°) in Table 4, as these outcomes showed significant fixed effects in Table 3. At the exercise level, the correlation test, \(H_{0}^{(\rho)}\), was significant for repetition duration (\(\Lambda = 8.92\), \(p < 0.01\)) but not for ROM (\(\Lambda = 0.48\), \(p = 0.49\)). This indicates that exercises with higher baseline repetition duration (fROM) exhibited systematically different responses to pROM, whereas no such relationship was observed for range of motion. In contrast, the variance test, \(H_{0}^{(uv)}\), was not significant for both outcomes, suggesting that the overall magnitude of between-exercise heterogeneity did not differ between fROM and pROM conditions, i.e., \(\tau_{u}^2 = \tau_{v}^2\). At the participant level, both correlation tests, \(H_{0}^{(\xi)}\), were significant for repetition duration (\(\Lambda = 16.00\), \(p < 0.001\)) and for ROM (\(\Lambda = 6.72\), \(p < 0.01\)), indicating that individuals with higher baseline performance responded systematically differently to pROM. The variance test, \(H_{0}^{(pq)}\), was moderately significant for ROM (\(\Lambda = 6.46\), \(p < 0.05\)) but not for repetition duration, providing weak evidence that the magnitude of between-participant heterogeneity differed between fROM and pROM conditions. Given this finding, assuming \(\tau_{p}^2 = \tau_{q}^2\) may not be warranted.
Overall, the results suggest that variability in response to pROM may be driven primarily by participant-level differences rather than exercise-specific factors. This pattern is particularly evident in the ROM outcomes, where both the intercept–slope correlation and the magnitude of participant-level heterogeneity differ between fROM and pROM. In contrast, exercise-level variability remains relatively stable across conditions. We further explore this finding in the next section.
In the previous section, we examined the structure of heterogeneity in Model 1 by testing whether components of the covariance matrices differed between ROM conditions at both the participant and exercise levels. In this section, we investigate how differences at the parameter level translate into the total observed variability in the outcomes. Specifically, two complementary analyses were conducted to: (i) quantify whether pROM increases the overall magnitude of between-participant or between-exercise variability relative to fROM; and (ii) determine whether the relative contributions of participants and exercises to total variability differ between ROM conditions.
For the first analysis, we tested whether the absolute magnitude of between-participant and between-exercise heterogeneity was greater under pROM than fROM. To this end, define \[V_{p,\mathrm{fROM}} = \tau_p^2, \quad V_{p,\mathrm{pROM}} = \tau_p^2 + \tau_q^2 + 2\xi\,\tau_p\tau_q, \qquad V_{e,\mathrm{fROM}} = \tau_u^2, \quad V_{e,\mathrm{pROM}} = \tau_u^2 + \tau_v^2 + 2\rho\,\tau_u\tau_v,\] to be the total between-participant and between-exercise variability under each ROM condition, respectively. Thus, we define the following variance contrasts \[D_p = V_{p,\mathrm{pROM}} - V_{p,\mathrm{fROM}} = \tau_q^2 + 2\xi\,\tau_p\tau_q, \qquad D_e = V_{e,\mathrm{pROM}} - V_{e,\mathrm{fROM}} = \tau_v^2 + 2\rho\,\tau_u\tau_v,\] to quantify the change in variability induced by performing pROM as opposed to fROM. One-sided parametric bootstrap tests were used to statistically evaluate the variance contrasts such that \[H_{0}^{(p)}: D_p \le 0 \quad \text{vs.} \quad H_{A}^{(p)}: D_p > 0, \qquad H_{0}^{(e)}: D_e \le 0 \quad \text{vs.} \quad H_{A}^{(e)}: D_e > 0,\] where \(D_p>0\) and \(D_e>0\) indicate that pROM introduces greater variability across participants and exercises, respectively. This analysis therefore addresses whether lengthened partials increase the absolute magnitude of heterogeneity relative to performing an exercise with full ROM.
In the second analysis, we assessed whether the proportion of total variability attributed to each source (participants versus exercise) differed between ROM conditions. To facilitate this analysis, we the total variance per ROM condition \(g \in \{\textrm{pROM},\textrm{fROM}\}\) can be defined as \[T_g = V_{p,g} + V_{e,g} + \bar{\sigma}_g^2,\] where \(\bar{\sigma}_g^2\) denotes the average within-observation variance under ROM condition \(g\). Condition-specific intraclass correlations (ICCs) were then computed as \[\mathrm{ICC}_{p,g} = \frac{V_{p,g}}{T_g}, \qquad \mathrm{ICC}_{e,g} = \frac{V_{e,g}}{T_g},\] which gives the proportion of total variability explained by participants and exercises, respectively. The following contrasts then compute the relative change in the explained variance: \[\Delta\mathrm{ICC}_p = \mathrm{ICC}_{p,\mathrm{pROM}} - \mathrm{ICC}_{p,\mathrm{fROM}}, \qquad \Delta\mathrm{ICC}_e = \mathrm{ICC}_{e,\mathrm{pROM}} - \mathrm{ICC}_{e,\mathrm{fROM}},\] which we used to assess whether these proportions differed under pROM. One-sided bootstrap tests [32] evaluated whether \(\Delta\mathrm{ICC}_p>0\) or \(\Delta\mathrm{ICC}_e>0\), corresponding to a shift toward greater participant- or exercise-level variation under pROM. A significant result would imply that pROM alters the distribution of heterogeneity, i.e., lengthened partials change the relative contributions of participant- and exercise-level variation to overall performance differences.
| Contrast | Statistic | Estimate | \(p\)-value (one-sided) |
|---|---|---|---|
| Repetition Duration (s) | |||
| Participant-Level Variance | \(D_p\) | 0.539 | \(<0.001^{***}\) |
| Exercise-level Variance | \(D_e\) | 0.444 | \(<0.001^{***}\) |
| Participant ICC Difference | \(\Delta \mathrm{ICC}_p\) | 0.119 | \(0.013^{*}\) |
| Exercise ICC Difference | \(\Delta \mathrm{ICC}_e\) | 0.002 | \(0.243\) |
| Range of Motion (°) | |||
| Participant-Level Variance | \(D_p\) | 657.202 | \(< 0.001^{***}\) |
| Exercise-level Variance | \(D_e\) | 255.850 | \(0.002^{**}\) |
| Participant ICC Difference | \(\Delta \mathrm{ICC}_p\) | 0.0136 | \(0.262\) |
| Exercise ICC Difference | \(\Delta \mathrm{ICC}_e\) | \(-0.0117\) | \(0.727\) |
The results are presented in Table 5 for repetition duration (s) and range of motion (°), as again, these outcomes showed significant fixed effects in Table 3. For repetition duration, both participant- and exercise-level variance contrasts (\(D_p\), \(D_e\)) were significant (\(p<0.001\)), indicating that pROM increases the absolute magnitude of between-participant and between-exercise variability relative to fROM. The participant-level ICC difference (\(\Delta \mathrm{ICC}_p\)) was positive and significant (\(p=0.011\)), suggesting that a larger share of total variability is attributable to individual differences under pROM, whereas the exercise-level ICC difference (\(\Delta \mathrm{ICC}_e\)) was not significant, implying that the exercises’ relative contribution is unchanged. For range-of-motion, both variance contrasts (\(D_p\), \(D_e\)) were again significant (\(p<0.01\)); however, neither ICC contrast reached significance, indicating that the proportional structure of variability remains stable across ROM conditions. Notably, \(D_e>0\) is compatible with the earlier finding that \(\tau_u^2=\tau_v^2\) cannot be rejected, because a positive intercept–slope correlation (\(\rho>0\)) increases \(V_{e,\mathrm{pROM}}\). Taken together, these results indicate that performing lengthened partials amplify variability across both participants and exercises.
The findings from the preceding analyses indicate that pROM does not merely alter mean performance. Instead, it also changes how variability manifests across individuals and exercises. The increased between-participant and between-exercise variance suggests that lengthened partials induce a more heterogeneous pattern of exercise performance as compared with fROM. Nevertheless, the ICC test results reveal proportional stability. That is, although overall variability increases under pROM, the relative contributions of participant- and exercise-level factors remain largely unchanged (note: for repetition duration, the participant share does modestly increase). In sum, these results suggest that lengthened partials amplify existing sources of variation without fundamentally altering their structure or relative influence on performance outcomes.
We next introduce %ROM, defined as the percentage of fROM achieved during a lengthened partial repetition. This metric is compared across exercises to assess whether a relative definition of lengthened partials exists or if it varies according to each exercise’s unique strength curve. This is particularly relevant as it indicates whether there exists a standard definition of the amount of range-of-motion to complete when performing a lengthened partial set as compared to fROM.
For participant \(i\), denote \(\bar{R}_{ieg}\) as the mean range-of-motion across all repetitions for exercise \(e\) and ROM condition \(g\). To obtain %ROM, analyses were conducted on the log-transformed outcomes \[Y_{ieg} = \log(\bar{R}_{ieg}),\] as we aim to determine whether a consistent definition of lengthened partial exercises can be established. The coefficient \(\beta_1\) from 1 now represents the log ratio of pROM versus fROM. Note that, after applying the delta method, the sampling variance is now \(\sigma_{ieg}^2=s^2_{ieg}/(k_{ieg}\bar{R}_{ieg}^2)\).
To estimate exercise-level differences in the effect of pROM relative to fROM, we defined \[\label{exercise95level} \delta_e = \beta_1 + v_e,\tag{2}\] where \(v_e\) denotes the exercise-specific random slope associated with the pROM indicator. Accordingly, \(\delta_e\) represents the estimated pROM effect for exercise \(e\), expressed as a deviation from the overall mean effect \(\beta_1\). A positive \(v_e\) indicates a higher %ROM (i.e., a smaller reduction than the overall mean); a negative \(v_e\) indicates a lower %ROM (a larger reduction). Significance was assessed by testing whether \(\delta_e\) differed from the overall mean effect \(\beta_1\). To control the false discovery rate across multiple pairwise comparisons, the resulting \(p\)-values were adjusted using the Benjamini–Hochberg procedure [32] as we run eight exercise-level hypothesis tests. While the average percentage of full ROM achieved under pROM is given by \(\%ROM = 100 \times \exp(\beta_1)\), an exercise-specific definition \(\%ROM_e\) can be obtained by exponentiating \(\delta_e\) in 2 .
| Exercise | \(\%\)ROM | \(95\%\) CI | \(p\) (bootstrap) |
|---|---|---|---|
| Bayesian Curl | 74.2 | [41.3, 134.1] | 0.313 |
| Cable Pushdown | 58.9 | [32.5, 108.0] | 0.815 |
| Dumbbell Curl | 63.3 | [34.5, 114.2] | 0.686 |
| Dumbbell Overhead Extension | 73.0 | [42.1, 132.3] | 0.305 |
| Dumbbell Row | 54.9 | [29.8, 101.5] | 0.959 |
| Flatpress | 64.0 | [35.7, 118.3] | 0.599 |
| Incline Press | 26.9 | [15.3, 49.3] | 0.008 |
| Lat Pulldown | 49.8 | [26.7, 92.7] | 0.648 |
Results on the exercise-specific effects of pROM, presented in Table 6, indicate that only the Incline Press shows a statistically significant deviation from the overall mean: its pROM effect is more negative than average (%ROM \(= 26.9\), 95% CI \([15.3, 49.3]\), \(p<0.01\)). For all other exercises, the estimated pROM effects are consistent with the average (bootstrap \(p\)-values ranging from \(0.31\) to \(0.96\)). However, the corresponding confidence intervals are wide, indicating that while no systematic exercise-level deviations were detected, the precision of these estimates is limited. Thus, the cross-exercise consistency of %ROM should be interpreted as provisional rather than definitive.
Practically, our findings imply that pROM may be programmed similarly across most exercises, with lengthened partial repetitions corresponding to approximately 56% of the full range of motion. A plausible explanation for the deviation observed in the Incline Press is that bench angle meaningfully alters shoulder mechanics, thereby decreasing effective ROM and shifting the sticking region [33]–[35]. Moreover, bench-press kinematics are highly sensitive to joint angles and bar path near the sticking point [36], and individual differences in morphology (e.g., arm length, shoulder width) are known to affect movement mechanics. These factors may account for the greater ROM loss in incline variations as compared with flat or cable presses [37], [38]. Nevertheless, we caution that this result could partly reflect artifacts of the AI pose estimation pipeline, such as increased landmark occlusion due to equipment obstructions.
This study develops an AI-based pose estimation pipeline to analyze a video dataset from [13], comprising eight upper-body resistance exercises performed by 26 participants using both lengthened partial (pROM) and full range of motion (fROM) conditions. Leveraging the crossed design of the original study, we compared repetition tempo and phase durations between pROM and fROM. We found that pROM not only involved a reduced range of motion (by definition), but also exhibited other performance differences, including slower tempo and shorter eccentric phase durations. We then examined ROM consistency across participants and exercises, finding significantly greater variability during lengthened partials as compared with full ROM. Our analysis suggests that these changes were driven primarily by participant-level differences rather than exercise-specific factors. Finally, we introduced %ROM (the proportion of fROM achieved during pROM) to evaluate whether a consistent empirical definition of lengthened partials could be established. Evidence indicates that pROM can be defined as approximately 56% of full ROM and does not (with the exception of the Incline Press) systematically vary with an exercise’s strength curve.
The use of AI-based pose estimation has gained prominence in domains such as fall detection for older adults [39], exercise recognition for physiotherapy [40], and workout tracking in AI-driven fitness applications [41]. In this study, it is applied to address questions pertinent to science-based strength training communities regarding the definition and characterization of a particular training modality [42]–[45]. Leveraging the data derived from the AI pipeline, we analyzed key aspects of exercise execution, the propagation of variability across ROM conditions, and the extent to which a common empirical definition of lengthened partials could be established across exercises. These findings have important implications for defining, communicating, and prescribing lengthened partials for resistance training. Future research should test the hypothesis that the concentric stimulus produced by pROM and fROM is comparable and that performing approximately 56% of the full movement may be sufficient to elicit similar adaptations across diverse exercise types. Moreover, the greater variability observed in the execution of lengthened partials highlights the need for more precise mechanical descriptions and coaching cues to ensure consistent prescription. More specifically, lengthened partials may be inherently less self-standardizing than full-ROM movements, which emphasizes the need for careful technique monitoring and individualized load adjustments.
The pose estimation pipeline, while objective and reproducible, has several limitations. First, it relies on monocular 2D pose estimation, which infers body landmarks using deep learning models trained on large, heterogeneous datasets. Although effective in controlled conditions, these models can be brittle when applied to occluded, cropped, or visually altered videos. This was evident in our pipeline, as many videos in which repetitions were clearly identifiable to human observers had to be excluded from further analysis. Second, the dataset is drawn from a single prior study with a limited participant pool, restricting the generalizability of the findings. While the large number of exercise repetitions enhances within-subject reliability and statistical power, it does not compensate for the limited demographic diversity of participants or the relatively narrow range of strength training exercises included in the study. Indeed, the original study did not include any lower-body exercises. Given that prior research has suggested different training prescriptions to stimulate upper- and lower-body hypertrophy [46], [47], future work should examine whether variability patterns differ between these regions and whether the data-driven definition of lengthened partials is consistent. Third, using 2D rather than 3D kinematic data introduces projection errors that may reduce biomechanical accuracy, although this trade-off allows for significantly faster and more efficient video processing. Finally, repetition segmentation is sensitive to parameter tuning, As a consequence, despite careful calibration, some video processing errors are inevitable. These limitations call for cautious interpretation of our results, but also underscores the need for future research that is designed with AI-based analysis in mind, such as ensuring that participants remain fully in-frame and incorporating multiple camera-views.