Validity of FuncUseRatio for ecological monitoring of functional arm use after stroke: toward precision neurorehabilitation Authors: Gaël Le Perf, Benoit Pratola, Denis Mottet, Isabelle Laffont, Karima Bakhti Published: 2026 ABSTRACT BackgroundThe use of the paretic arm in activities of daily living after stroke is critical to prevent further decline in function over time. Accelerometers offer the opportunity to monitor arm activity during real-life daily activities in home environments. However, commonly used accelerometric metrics may be influenced by passive movements unrelated to functional arm use. In this study, we evaluate the validity of FuncUseRatio for monitoring activities of daily living after stroke.MethodsTwo experiments were conducted. Experiment 1 enrolled 23 healthy participants who performed 22 arm functional movements representative of activities of daily living. Movements of both arms were simultaneously recorded using wrist-worn accelerometers and a Vicon motion capture system allowing direct comparison between accelerometer-derived forearm orientation and Vicon measurements (gold standard). Experiment 2 included 25 individuals after stroke who performed ecological activities (meal preparation, eating, and dressing). Arm movements were recorded using accelerometers and synchronized video analysis (ground truth), enabling comparison between accelerometer-derived FuncUseRatio and structured video-based arm-use computation.ResultsIn Experiment 1, forearm orientation estimated from accelerometer data showed excellent consistency with Vicon measurements under dynamic conditions required for FuncUse computation. The intraclass correlation coefficients (ICC) were ICC(C,1) = 0.93 for the right forearm (95% CI: 0.85–0.97, p < 0.01) and 0.95 for the left forearm (95% CI: 0.89–0.98, p < 0.01), confirming strong agreement between the two systems. In Experiment 2, FuncUseRatio derived from accelerometer data demonstrated excellent agreement with the structured video-based analysis ratio, considered the ground truth, with (ICC(A,1) = 0.96, 95% CI 0.90–0.98, p < 0.01).ConclusionThe FuncUseRatio is a valid measure for assessing functional arm use during activities of daily living in ecologically relevant settings. Its implementation may enable early detection of reduced paretic arm use, including when the patient is at home, facilitate timely reassessment to understand its underlying causes, and thereby allow the implementation of specific arm rehabilitation strategies before any decline in function or participation occurs.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12984-026-01978-9. Introduction After a stroke, most individuals reduce the use of their paretic arm and compensate by relying more on their non-paretic arm [12]. This shift in behavior, combined with changes in motor performance, leads to a significant decline in participation in leisure activities and daily living, as well as reduced overall satisfaction [9]. Even in patients with good sensory-motor recovery, as assessed by clinical tests, reduced use of the paretic arm may persist, a phenomenon commonly referred to as “non-use” [11]. Non-use creates a vicious cycle whereby reduced arm use leads to diminished task performance and success, rendering functional activities more effortful and further reinforcing avoidance [1]. Given this challenge, objectively capturing the use of the paretic arm during daily activities is critical for preventing further decline in function and participation. In recent years, wearable sensors such as accelerometers have been increasingly investigated for measuring arm activity during activities of daily living [22]. However, commonly used accelerometric metrics, such as Euclidean norm, activity count, or use hours, were originally designed to assess general physical activity and may be influenced by passive movements or motions that do not reflect the functional use of the arm [28]. As a result, discrepancies between the activities of the paretic and non-paretic arms may be underestimated due to the contribution of non-functional movements, as arm swing during walking. Moreover, longitudinal interpretation may be complicated, as it becomes unclear whether observed changes over time reflect true improvement of the arm use or simply variations in whole-body movements [15, 24, 25]. To address these limitations, Leuenberger et al. proposed an algorithm based on inertial sensors designed to exclude arm swing movements during walking thereby isolating functional use of the arm (FuncUse). This approach enables more specific detection of truly functional movements, such as reaching, transporting, or repositioning [27]. This algorithm excludes arm movements with amplitudes smaller than 30° or outside a spatial range of −30° to + 30° around the horizontal. However, the short recording duration of inertial sensors (a few hours) limits their usability in real-world conditions, such as during daily home activities [32]. Recently, Leuenberger et al.’s approach was adapted for use with accelerometers, enabling continuous monitoring of arm use over a seven-day period [5]. This adaptation demonstrated the feasibility of the method in both healthy individuals and stroke participants. Moreover, the FuncUseRatio, which reflects the relative functional use of the two arms, successfully distinguished between the two populations. However, accelerometers measure both gravitational and motion-related accelerations. Since orientation is inferred from the gravity vector, additional motion-induced accelerations can corrupt this measurement, making orientation estimates susceptible to errors under dynamic conditions [26]. How large are these errors under dynamic conditions? Furthermore, the FuncUse and FuncUseRatio metrics have not yet been validated against ground-truth data. Do these metrics truly reflect how individuals use their arms in daily life? To address these questions, we conducted two experiments with distinct objectives: Experiment 1 (Healthy Participants): To establish the concurrent validity of forearm orientation measurements against the Vicon motion capture system, and to assess the reliability of FuncUse.Experiment 2 (Participants with stroke): To examine the validity of FuncUseRatio and FuncUse derived from accelerometers by comparing them with structured video analysis in ecological scenarios. Experiment 1 (Healthy Participants): To establish the concurrent validity of forearm orientation measurements against the Vicon motion capture system, and to assess the reliability of FuncUse. Experiment 1 (Healthy Participants): To establish the concurrent validity of forearm orientation measurements against the Vicon motion capture system, and to assess the reliability of FuncUse. Experiment 2 (Participants with stroke): To examine the validity of FuncUseRatio and FuncUse derived from accelerometers by comparing them with structured video analysis in ecological scenarios. Experiment 2 (Participants with stroke): To examine the validity of FuncUseRatio and FuncUse derived from accelerometers by comparing them with structured video analysis in ecological scenarios. Experiment 1 Method experiment 1 Participants We enrolled 23 healthy participants recruited from Montpellier University students, personnel from the EuroMov Digital Health in Motion (DHM) laboratory, and staff members of the Montpellier University Hospital and the Lamalou-les-Bains Hospital Center. The inclusion criteria were adults aged 18 or older who were fully informed and provided written consent to participate in the study. The exclusion criteria included any arm limitation or pain, refusal to allow anonymous data processing, and the presence of cognitive or psychiatric disorders that could interfere with task performance. Experimental setup The experiment was conducted at the EuroMov Digital Health in Motion (DHM) laboratory and was approved by the Ethics Committee of Montpellier University (approval number: UM 2023-021bis). Participants performed twenty-two functional arm movements representative of activities of daily living (ADLs), inspired by items from the Motor Activity Log [29], as well as one task not specifically involving the arms, based on the Timed Up and Go test [16] and the Dynamic Gait Index [17]. A complete list of all activities is provided in supplementary material (SM Table S1). Between each functional movement, participants were instructed to hold their arms still at their sides for 5 s to facilitate movement segmentation during post-processing. After a 15-minute rest period, the entire procedure was repeated to assess the repeatability of the measurements. During this interval, the sensors were removed to allow the investigator to download and clear the data from the first trial and reconfigure the devices before placing them on the participant again for the second trial. Materials Movement of both arms was simultaneously recorded using wrist-worn accelerometers and a Vicon motion capture system (Fig. 1). The wrist-worn devices (AX3, Axivity, Newcastle Helix, United Kingdom) contain a triaxial accelerometer that records acceleration in all three spatial planes, with a sampling frequency of 50 Hz and a measurement range of ± 8 g. The Vicon system (Oxford Metrics, Oxford, UK) is a real-time motion analysis system. Passive reflective markers were placed on the arms, specifically on the wrists (labels: RWRB: Right Wrist Marker B & LWRB: Left Wrist Marker B) and lateral epicondyles (labels: RELB: Right Elbow & LELB: Left Elbow). These markers reflect infrared light toward a set of CMOS cameras, which capture and digitize their positions. The 3D coordinates of each marker were reconstructed using Vicon software. Data from the Vicon system were recorded at a sampling frequency of 100 Hz. Fig. 1Positioning of the wrist-worn accelerometer and reflective Vicon markers on the participant Positioning of the wrist-worn accelerometer and reflective Vicon markers on the participant Signal processing Signal processing was conducted using Scilab (version 2023.1.0). All data were resampled at 50 Hz, and a dual-pass Butterworth low-pass filter with a 1 Hz cut-off frequency was applied to eliminate high-frequency noise. Forearm orientation relative to the vertical axis was estimated through a transformation of the accelerometers’ data. This transformation relies on gravity, which is a constant acceleration of 1 g (9.81 m/s²) directed downward. In static conditions, this gravitational acceleration is distributed across the three axes of the accelerometer, corresponding to the three spatial planes, depending on the orientation of the forearm. Since the y-axis of the accelerometer is aligned with the forearm axis, trigonometric principles allow the computation of the angle between the gravity vector and the forearm using equation. 1 [5]:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:\left(t\right)\:=\:arccos\left(\frac{{a}_{y}\left(t\right)}{svm\left(t\right)}\right)$$\end{document} where ay(t) is the acceleration along the y-axis at time t, and svm(t) is the scalar vector magnitude of the acceleration, computed at each time step using equation. 2:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:svm\left(t\right)\:=\:\sqrt{{a}_{x}{\left(t\right)}^{2}+\:{a}_{y}{\left(t\right)}^{2}\:+\:{a}_{z}{\left(t\right)}^{2}}$$\end{document} The orientation of the forearm relative to the vertical axis was also measured using data from the Vicon motion capture system. This system provides the Cartesian coordinates of passive markers placed on anatomical landmarks. Specifically, markers were positioned on the wrist and lateral epicondyle (elbow). The forearm vector was defined as the segment connecting these two markers. The polar angle between this vector and the vertical axis was then computed using the following equation:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:\left(t\right)\:=\:arctan\left(\frac{\sqrt{(\varDelta\:{X)}^{2}\:+\:(\varDelta\:{Y)}^{2}}}{\varDelta\:Z}\right)$$\end{document} Where ΔX, ΔY, ΔZ are the coordinate differences between the elbow and wrist markers along the X, Y, and Z axes, respectively. The signals from the Vicon system and the accelerometer were synchronized manually by visually aligning characteristic features (e.g., angular peaks, plateaus) across the time series. To verify the quality of this alignment, a post hoc calculation of the root-mean-square error (RMSE) between the angular signals was performed separately for each side. The RMSE between the two systems was 6.58° for the right forearm and 5.92° for the left forearm, reflecting a high degree of agreement and confirming that the synchronization was adequate for subsequent analyses. Outcome variables The main variable was the forearm orientation relative to the vertical axis, estimated from both accelerometer and Vicon data (see Signal Processing). The post hoc RMSE values of 6.58° (right arm) and 5.92° (left arm) confirmed the adequacy of the synchronization between systems for subsequent analyses. FuncUse was defined in accordance with Leuenberger et al., [15] as the number of 2-second epochs per arm containing a functional movement. An epoch was considered functional if the forearm orientation exhibited an amplitude threshold of at least 30° within a ± 30° orientation range relative to the horizontal (Fig. 2). Fig. 2Schematic representation of the FuncUse parameters based on the initial approach proposed by Leuenberger et al., [15]. A movement was considered functional if, within a 2-second epoch, the forearm orientation exhibited an amplitude threshold ≥ 30° (blue arrow) within a ± 30° orientation range (green area) relative to the horizontal reference axis (blue dashed line) Schematic representation of the FuncUse parameters based on the initial approach proposed by Leuenberger et al., [15]. A movement was considered functional if, within a 2-second epoch, the forearm orientation exhibited an amplitude threshold ≥ 30° (blue arrow) within a ± 30° orientation range (green area) relative to the horizontal reference axis (blue dashed line) In addition, participants’ age, gender, and handedness were recorded to provide a demographic description of the sample. Statistical analysis The statistical analysis was performed on Rstudio (2023.06.0 + 421), with a significance level set at 5%. Sample size According to the method proposed by Zou [33], assuming an expected intraclass correlation coefficient (ICC, ρ) of 0.95 and a lower bound of the 95% confidence interval of 0.8 (the minimum acceptable value for ρ), with two raters per subject (Vicon and accelerometer), a Type I error rate of 5%, and a statistical power of 90%, the required sample size was 21 participants. To account for potential data loss or unforeseen exclusions, we increased the calculated sample size by 10%, bringing the total number of participants to 23. Descriptive analysis Continuous variables are reported as median with their interquartile range (IQR), while categorical data are presented as counts and proportions. Concurrent validity analysis To assess the concurrent validity of forearm orientation relative to the vertical derived from accelerometers, measurements were compared with those obtained from the Vicon motion capture system, which served as the gold standard. Intraclass correlation coefficients (ICCs) were calculated using a two-way mixed-effects model (single measurement, consistency), which is appropriate when the same fixed raters are used across subjects. The consistency type was chosen because systematic differences between the accelerometer and Vicon measurements were expected. ICC values were interpreted according to conventional thresholds: < 0.5 = poor, 0.5–0.75 = moderate, 0.75–0.9 = good, and > 0.9 = excellent [13]. Because participants were instructed to hold their arms still at their sides for 5 s to facilitate movement segmentation during post-processing, and because the objective was to assess validity under dynamic conditions, static periods were excluded from the analysis. A 40° threshold was therefore applied, based on inspection of the static phases, which showed that forearm orientation remained well below this value during instructed immobility. This ensured that only dynamic conditions were retained for analysis. Subject-specific median values were then computed from the filtered dataset. Given the large number of observations per participant, using median values prevented artificial inflation of the effective sample size and provided a more conservative estimate of agreement. Linear regression and Bland–Altman analyses were also performed using these median values. Reliability analysis To evaluate the stability of the FuncUse measurements across two sessions, single-score, two-way mixed-effects ICCs for agreement were calculated independently for the Vicon motion capture system and the accelerometers. This ICC type was chosen because the same measurement system was used for all participants across sessions, and the objective was to assess the absolute agreement of repeated measurements, despite the expected between-session variability inherent to unconstrained functional activities. Results experiment 1 Twenty-three participants were enrolled in the study. The mean age was 31 (IQR : 24.5 ; 38) years, with 15 (65.2%) females and 8 (34.8%) males. Twenty-two (95.7%) participants were right-handed. All participants completed all activities. Three hours and twenty-five minutes of data were usable. Three minutes were excluded because the markers were outside the capture volume, and eight minutes were excluded due to accelerometer data loss. Validity analysis forearm orientation The ICC(C,1) computed from the median values per subject was 0.93 for the right forearm (95% CI: 0.85–0.97; F(22,22) = 28.7; p < 0.01) and 0.95 for the left forearm (95% CI: 0.89–0.98; F(22,22) = 40; p < 0.01), indicating excellent consistency between the accelerometer and Vicon measurements. Bland–Altman plots (Fig. 3) showed a very small systematic bias of -1.95 ± 1.94° for the right forearm and − 3.46 ± 1.84° for the left, and the limits of agreement were narrow. No proportional bias was observed. Linear regression models (Fig. 4) explained 87% and 90% of the variance for the right and left arms, respectively, indicating a strong linear relationship between accelerometer- and Vicon-derived measurements, with positive slopes and significant coefficients of 0.86 and 0.93. The Bland–Altman plots and linear regression models performed on all forearm orientation data are provided in the supplementary material. Fig. 3Bland–Altman plot for forearm orientation. The x-axis represents the median Vicon-derived forearm angle for each subject, and the y-axis represents the difference between the median angles estimated by the Vicon system and the accelerometers. The black dashed line indicates the mean bias, and the red dashed lines represent the 95% limits of agreement (in degrees) Bland–Altman plot for forearm orientation. The x-axis represents the median Vicon-derived forearm angle for each subject, and the y-axis represents the difference between the median angles estimated by the Vicon system and the accelerometers. The black dashed line indicates the mean bias, and the red dashed lines represent the 95% limits of agreement (in degrees) Fig. 4Scatter plot and linear regression of the median forearm orientation for each subject. The x-axis represents the forearm-to-vertical angle estimated by the accelerometers (degrees), and the y-axis represents the forearm-to-vertical angle estimated by the Vicon system (degrees). The solid blue line indicates the linear regression, and the dashed line represents the identity line Scatter plot and linear regression of the median forearm orientation for each subject. The x-axis represents the forearm-to-vertical angle estimated by the accelerometers (degrees), and the y-axis represents the forearm-to-vertical angle estimated by the Vicon system (degrees). The solid blue line indicates the linear regression, and the dashed line represents the identity line Test–retest reliability of FuncUse Despite the inherent variability of freely performed activities, test–retest reliability of FuncUse was robust for both measurement systems. For the Vicon system, ICC(A,1) was 0.84 (95% CI: 0.66–0.93; F(22,22.9) = 11.2; p < 0.01), indicating good agreement between sessions. For the accelerometers, ICC(A,1) was 0.76 (95% CI: 0.51–0.89; F(22,22) = 7.08; p < 0.01), indicating moderate-to-good agreement. Overall, reliability was slightly higher for the Vicon system, while accelerometers showed acceptable reliability within the expected variability of functional movements. Experiment 2 Method experiment 2 Participants This study protocol was approved by the Ouest II Ethics Committee of Angers University Hospital Center (RCB ID: 2024-A00196-41), France. The protocol was registered on ClinicalTrials.gov (NCT06509542). Participants were recruited at Montpellier University Hospital and Lamalou-les-Bains Hospital and were included if they met the following criteria: (1) Adults with hemiparesis secondary to stroke; (2) engagement in ecological activities during occupational therapy sessions; and (3) provision of written informed consent to participate in the study. Exclusion criteria were: (1) the presence of other neurological or musculoskeletal conditions impairing arm function; (2) pregnancy or breastfeeding; (3) lack of informed consent; (4) concurrent participation in another research study with an ongoing exclusion period; and (5) individuals under legal guardianship or without social security coverage. Experimental setup Arm movements were simultaneously recorded during occupational therapy sessions using wrist-worn accelerometers and video cameras. The wrist-worn device integrated three triaxial accelerometers per wrist, capturing acceleration data across all three spatial planes. Participants performed three ecological tasks: meal preparation, eating, and dressing (or another task if they were unable to complete one of the three specified tasks). Sessions were supervised by an occupational therapist for safety, with occasional use of assistive devices, but participants were free to use their paretic arm and perform the activities as they preferred. To ensure full visual coverage, two cameras were positioned frontally and at a three-quarter angle, maintaining visibility of the participants’ forearms throughout the assessment. At the beginning and end of each task, participants performed five complete elevations of the non-paretic arm. These movements served as temporal markers for aligning video recordings with accelerometric data. Materials For accelerometry measurements, the same wrist-worn (AX3, Axivity, Newcastle Helix, United Kingdom) devices utilized in Experiment 1 were employed in Experiment 2. Video recordings of the entire session were captured using two high-definition DJI Pocket 2 cameras (DJI), set to record at a frequency of 50 Hz. Video analysis, including the quantification of functional movements, was conducted using Scilab (version 2023.1.0). Signal processing In Experiment 2, forearm orientation relative to the vertical axis was estimated using the identical signal processing pipeline as in Experiment 1. Leuenberger defined functional movements (FuncUse) as those exceeding 30° amplitude within a -30° to + 30° range relative to the horizontal over 2 s. We extended this definition by considering orientation ranges of ± 30° to ± 50° and amplitude thresholds of 10° to 30°. Based on previous analyses [18], we observed that shorter epochs tended to underestimate movement for the paretic arm, while overlapping windows could lead to overestimation. Therefore, a 2-second epoch was selected, consistent with the approach of Leuenberger et al. [15]. These parameter combinations were systematically explored to quantify their effect on the computed FuncUseRatio. We computed FuncUseRatio in its normalized version (Eq. 3), as it facilitates interpretation and comparability of results: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{FuncUseRatio}}=\frac{{{\mathrm{FuncUs}}{{\mathrm{e}}_{{\mathrm{parectic}}}}\, - {\mathrm{FuncUs}}{{\mathrm{e}}_{{\mathrm{nonparectic}}}}}}{{{\mathrm{FuncUs}}{{\mathrm{e}}_{{\mathrm{parectic}}}}\,{\mathrm{+}}\,{\mathrm{FuncUs}}{{\mathrm{e}}_{{\mathrm{nonparectic}}}}}}$$\end{document} The video signal was segmented into one-minute intervals. For each participant, 30 one-minute intervals were randomly selected for detailed analysis. The FuncUse calculation was performed on the same selected minute to ensure direct comparison with the video analysis. Across all intervals selected, a visual assessment was performed in 2 s-epochs, following the Functional Ability Assessment Behavioral Observation System (FAABOS) method [31]. The rating was based on the Taxonomy of Arm Functional Movements [27], which categorizes arm use into five types: reaching, transporting, repositioning, stabilizing, and inactivity. In this experiment, only the first three movement types were considered functional: Reaching (movement of the arm to touch or grasp an object), transporting (movement or manipulation of an object) and repositioning (movement around an object before or after interaction). The stabilizing and inactivity categories, which do not involve active movement, were excluded from the analysis. Consequently, for each minute of video, a binary score was assigned every two seconds for each arm: 0 for the absence of functional movement and 1 for its presence. Statistical analysis The statistical analysis was performed on Rstudio (2023.06.0 + 421), with a significance level set at 5%. Sample size According to the method proposed by Zou [33], assuming an expected intraclass correlation coefficient (ICC, ρ) of 0.95 and a lower bound of the 95% confidence interval of 0.8 (the minimum acceptable value for ρ), with two raters per subject (Video and accelerometer), a Type I error rate of 5%, and a statistical power of 90%, the required sample size was 21 participants. To account for potential data loss or unforeseen exclusions, we increased the calculated sample size by 20%, bringing the total number of participants to 25. Descriptive analysis Continuous variables are expressed as median and interquartile range (IQR), and categorical data are summarized as counts (proportions). Sensitivity analysis As a complementary step, a sensitivity analysis was conducted to examine the impact of parameter choices (orientation range and amplitude threshold) on FuncUseRatio performance and to identify the optimal parameter set for validity testing. Single-score intraclass correlation coefficients [ICC(A,1), two-way mixed-effects model, agreement definition] were calculated to assess the agreement between accelerometer-derived metrics and the structured video analysis (VideoUseRatio) across orientation ranges of ± 30° to ± 50° and amplitude thresholds of 10° to 30°. Each combination of parameters was evaluated in this way to determine its validity against the video-based reference. As part of the sensitivity analysis, a correlation matrix was constructed to assess the interrelationships between the evaluated parameters. Validity analysis For each parameter combination (orientation range and amplitude threshold), Bland–Altman plots and regression analyses were performed on the FuncUseRatio, using the VideoUseRatio as the ground truth; absolute systematic error was calculated. The same analyses were conducted for FuncUse of the paretic and non-paretic arms, with relative systematic error expressed as a percentage. Results experiment 2 Twenty-five patients were included in the study. The mean age was 63 (IQR = 52 ; 72) years, with 13 females and 12 males. Twenty-two patients were right-handed. Fourteen had an ischemic stroke, and 60% ​​of participants were in the subacute stage after stroke (< 6 months). The Box and Block Test (BBT) scores ranged from 30 (IQR = 8 ; 38) on the paretic side to 49 (IQR = 44 ; 57) on the non-paretic side, indicating a wide spectrum of functional impairment. Table 1 summarizes the characteristics of the study population. Table 1Participant clinical characteristicsVariablesAge (years)median (IQR)63(52 ; 72)Gender, n (%)Male12(48)Female13(52)Handedness, n (%)Right-handed22(88)Left-handed3(12)Stroke type, n (%)Hemorrhagic11(44)Ischemic14(56)Stroke side, n (%)Right11(44)Left14(56)Box and Block score, median (IQR)Paretic side30(8 ; 38)Non-Paretic Side49(44 ; 57)Ratio -0.23(-0.76 ; -0.12) Table 1Participant clinical characteristicsVariablesAge (years)median (IQR)63(52 ; 72)Gender, n (%)Male12(48)Female13(52)Handedness, n (%)Right-handed22(88)Left-handed3(12)Stroke type, n (%)Hemorrhagic11(44)Ischemic14(56)Stroke side, n (%)Right11(44)Left14(56)Box and Block score, median (IQR)Paretic side30(8 ; 38)Non-Paretic Side49(44 ; 57)Ratio -0.23(-0.76 ; -0.12) Participant clinical characteristics Age (years) median (IQR) Gender, n (%) Male Handedness, n (%) Right-handed Stroke type, n (%) Hemorrhagic Stroke side, n (%) Right Box and Block score, median (IQR) Paretic side Thirty minutes of video analysis were performed for all patients, except for one for whom only 28 min were usable. Among the 25 patients, 22 completed three distinct activities, while 3 completed only two due to either a video recording issue (n = 2) or an accelerometer recording problem (n = 1). Twenty-three patients performed the cooking activity, with an average analysis duration of 15.43 ± 5.68 min. All patients performed the eating activity, with an average analysis duration of 10.84 ± 6.83 min. Twenty patients performed the dressing activity, with an average analysis duration of 4.70 ± 2.39 min, four patients performed another activity (e.g., gardening or DIY), because they declined to perform the dressing activity, their analysis duration was 7.00 ± 3.56 min. FuncUseRatio sensitivity analysis We performed a sensitivity analysis to evaluate the algorithm’s performance in computing the FuncUseRatio, using forearm orientation ranges of ± 30°, ± 40° and ± 50° and amplitude thresholds from 10° to 30°. The results were compared with those obtained from the structured video analysis (Fig. 5). While several combinations display overlapping confidence intervals, indicating broadly similar performance, amplitude thresholds of 15° combined with orientation ranges of ± 40° and ± 50° consistently yielded the highest ICC values with relatively narrow confidence intervals (ICC(A,1) = 0.96 [95% CI: 0.90–0.98; F(24,25) = 44.4, p < 0.01] and ICC(A,1) = 0.95 [95% CI: 0.90–0.98; F(24,24.1) = 40, p < 0.01], respectively). These parameter sets were therefore retained as the most suitable candidates for computing the FuncUseRatio, although several configurations achieved comparably high agreement with the video-based ground truth. Fig. 5Sensitivity analysis of the FuncUseRatio according to parameter selection. The left panel illustrates the combinations of amplitude thresholds and orientation ranges relative to the horizontal used to compute the accelerometer-derived FuncUseRatio. The right panel presents the corresponding intraclass correlation coefficients (ICC) between the FuncUseRatio and the VideoUseRatio derived from structured video analysis (ground truth). Values are reported with their 95% confidence intervals Sensitivity analysis of the FuncUseRatio according to parameter selection. The left panel illustrates the combinations of amplitude thresholds and orientation ranges relative to the horizontal used to compute the accelerometer-derived FuncUseRatio. The right panel presents the corresponding intraclass correlation coefficients (ICC) between the FuncUseRatio and the VideoUseRatio derived from structured video analysis (ground truth). Values are reported with their 95% confidence intervals FuncUseRatio validity analysis As mentioned above, the FuncUseRatio showed excellent agreement with the ratio derived from structured video analysis. For an amplitude threshold of 15° combined with an orientation range of ± 50°, the Bland–Altman plot revealed an absolute systematic error close to zero (ASE = − 0.02 ± 0.11), a relative systematic error of 5.31%, and acceptable limits of agreement (Fig. 6). The linear regression model explained 92% of the variance (B = 1.09, t = 17.21, p < 0.01). Results for all other parameter combinations are provided in the Supplementary Material, including the corresponding Bland–Altman plots and regression analyses. Fig. 6Bland–Altman Plot for FuncUseRatio. The x-axis represents the average FuncUseRatio, and the y-axis represents the difference between the VideoUseRatio and the FuncUseRatio. The black dashed line indicates the systematic bias, and the red dashed lines represent the limits of agreement Bland–Altman Plot for FuncUseRatio. The x-axis represents the average FuncUseRatio, and the y-axis represents the difference between the VideoUseRatio and the FuncUseRatio. The black dashed line indicates the systematic bias, and the red dashed lines represent the limits of agreement FuncUse validity analysis The FuncUse measure, using an amplitude threshold of 15° within a ± 50° spatial range around the horizontal axis, showed an ICC(A,1) of 0.81 [95% CI: 0.61–0.91; F(24,24) = 9.86, p < 0.01] for the paretic arm and 0.86 [95% CI: 0.70–0.93; F(24,24.1) = 12.5, p < 0.01] for the non-paretic arm. However, FuncUse systematically underestimated functional use by approximately 25% compared with the 2-second functional use epochs identified through video analysis (Detailed performance results for all FuncUse parameters are provided in the supplementary material). Discussion These experiments demonstrated, to our knowledge for the first time, that forearm orientation measurements obtained from accelerometers during dynamic, functional movement conditions are sufficiently accurate to compute FuncUse. They also showed that the resulting FuncUseRatio reliably reflects how individuals post-stroke use their arms during daily activities. Our findings are consistent with previous evidence indicating that accelerometers can introduce orientation-estimation errors during dynamic movements [19]. However, the magnitude of these errors remained acceptable for computing both FuncUse and FuncUseRatio, as supported by the excellent agreement with video-based analyses in Experiment 2. Consistent with the sensitivity analysis, although orientation ranges of ± 40° and ± 50° showed similarly high levels of agreement, the wider range (± 50°) may be preferable in practice, as it allows the capture of a broader spectrum of functional movements, including movements performed above or below the horizontal plane. This may be particularly relevant in ecological settings where arm use occurs across a wide range of spatial configurations. These results highlight the validity of FuncUse and FuncUseRatio. Their specific algorithm isolates functional arm movements during ADL, in contrast to classical accelerometer metrics, such as activity counts, Euclidean norm, Use Hours, or jerk, which are more suited to measuring general physical activity [2, 3, 4, 21, 23, 31]. These metrics are sensitive to non-functional movements, including arm swing during walking [10, 28], and have generally not been validated against real-world ground truth measures of arm activity, thereby compromising the validity and responsiveness of the measure [25]. To address this issue, [6] proposed combining wrist-worn accelerometers with an additional sensor on the thigh to adjust the intensity thresholds used to classify an epoch as arm use according to body position (lying, sitting, standing) and locomotion (walking). The limitation of this approach is that it requires an extra sensor on the lower limb, unlike our method. More recently, [8] used a machine learning approach to identify functional and non-functional arm movements during ADL in stroke patients. Their results were excellent; however, as stated in the limitations of the paper, the classifiers were specific to each patient and therefore cannot currently be applied in clinical practice. Tran et al., [30] subsequently proposed a deep learning approach with encouraging performance; however, despite a dataset of 35 subjects and a greater number of tasks, generalizability remains a major challenge. In this context, and pending the availability of robust, generalized machine learning models, which require large, extensively labeled datasets, our approach represents a pragmatic and clinically relevant compromise. FuncUse and FuncUseRatio are interesting indicators for assessing arm use during ADL and provide specific information that complements the standard post-stroke examination [22]. These new metrics are particularly important for monitoring arm use, as they could allow early detection of non-use, including when the patient is at home, and facilitate timely reassessment to understand its causes, thereby allowing the implementation of specific arm rehabilitation strategies before any decline in function or participation [1]. This engineering-based approach supports precision neurorehabilitation by providing clinicians and researchers with validated tools to characterize real-life behavior beyond the clinical setting. In addition, these metrics could be integrated into digital health solutions. For example, FuncUse and FuncUseRatio could be implemented in smartphone applications, as tested by Mayrhuber et al., [20], allowing patients to receive regular feedback and encouraging them to use their arm during ADL to achieve their goals (see [7]). Such, this approach appears to be well accepted and usable [14]. The interpretation of these results should consider the following limitations. First, Experiment 1 includes only healthy participants, who typically display larger ranges of motion and higher movement speeds and frequencies compared to individuals post-stroke, potentially increasing dynamic orientation errors. Despite this, we showed that the orientation errors remained reasonable. Additionally, the inherent variability of freely performed activities could be reduced through stricter standardization of the gesture protocol, which would improve test-retest reliability. Second, although the scenarios used in Experiment 2, were ecologically valid and inspired by ADLs, they were conducted in a simulated room and do not fully reproduce real home environments. A next step is to acquire video and accelerometry data from individuals with stroke living in the community, performing in the full spectrum of ADLs, in their natural environments. Third, certain functional movements are not detected due to algorithmic parameters, chosen to exclude arm swing during gait, particularly when the arm is close to a vertical position (e.g., picking an object from the floor, tying shoelaces). Furthermore, wrist-worn accelerometers also show inherent limitations, as fine finger movements or stabilization functions, common in many functional uses of the arm, that do not involve forearm motion are not captured. Fourth, we acknowledge that the FuncUse calculation is sensitive to algorithm parameters in order to ensure the best possible match with the ground truth, and that the parameters we have deemed optimal may be specific to our test protocol. Their generalizability to other datasets, populations, and activity contexts therefore remains to be confirmed in future studies. Finally, future research on FuncUse and FuncUseRatio could further investigate potential false positives, especially in patients who exhibit flexion synkinesis during walking. Performance should be compared with other existing metrics to understand what FuncUse and FuncUseRatio provide. As a final component of the discussion, we provide practical recommendations for researchers and clinicians interested in implementing the FuncUse and FuncUseRatio metrics. Raw accelerometer data should be resampled to 50 Hz and filtered using a dual-pass Butterworth low-pass filter with a 1 Hz cut-off frequency. Forearm orientation should then be computed using Eqs. 1 and 2 (see Methods, Experiment 1). FuncUse should be derived using a 2-second epoch duration, identifying forearm movements greater than 15° within an orientation range of ± 50° relative to the horizontal. The FuncUseRatio should be computed using the normalized formulation described in Eq. 3 (see Methods, Experiment 2). The code used to compute the FuncUse and FuncUseRatio metrics is provided on https://github.com/GaelLeperf/FuncUseMetric. In conclusion, FuncUse and FuncUseRatio are valid measures of arm use during ADLs and for monitoring arm use in the patient’s home. Future clinical research should further evaluate how these new indicators may help prevent or identify early cases of non-use at an early stage and improve the management and targeting of the rehabilitation intervention before any functional deterioration or loss of use occurs. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1. Supplementary Material 2. Supplementary Material 2. Supplementary Material 3. Supplementary Material 3. Supplementary Material 4. Supplementary Material 4. Supplementary Material 5. Supplementary Material 5. Supplementary Material 6. Supplementary Material 6. Supplementary Material 7. Supplementary Material 7. Supplementary Material 8. Supplementary Material 8. Supplementary Material 9. Supplementary Material 9.