Machine Learning-Based Real-Time Detection of Compensatory Trunk Movements Using Trunk-Wrist Inertial Measurement Units

Jannis Gabler; Clément Lhoste; Max Quast; Laura Mayrhuber; Andrea Ronco; Olivier Lambercy; Paulius Viskaitis; Dane Donegan

cs.ROApr 14, 2026

Machine Learning-Based Real-Time Detection of Compensatory Trunk Movements Using Trunk-Wrist Inertial Measurement Units

Authors: Jannis Gabler, Clément Lhoste, Max Quast, Laura Mayrhuber, Andrea Ronco, Olivier Lambercy, Paulius Viskaitis, Dane Donegan

Organizations: Rehabilitation Engineering Laboratory, ETH Zurich, Switzerland · TU Munich, Germany · Neuroscience Center Zurich (ZNZ), Switzerland · Singapore-ETH Centre (CREATE), Singapore

Abstract

Compensatory trunk movements (CTMs) are commonly observed after stroke and can lead to maladaptive movement patterns, limiting targeted training of affected structures. Objective, continuous detection of CTMs during therapy and activities of daily living remains challenging due to the typically complex measurements setups required, as well as limited applicability for real-time use. This study investigates whether a two-inertial measurement unit configuration enables reliable, real-time CTM detection using machine learning. Data were collected from ten able-bodied participants performing activities of daily living under simulated impairment conditions (elbow brace restricting flexion-extension, resistance band inducing flexor-synergy-like patterns), with synchronized optical motion capture (OMC) and manually annotated video recordings serving as reference. A systematic location-reduction analysis using OMC identified wrist and trunk kinematics as a minimal yet sufficient set of anatomical sensing locations. Using an extreme gradient boosting classifier (XGBoost) evaluated with leave-one-subject-out cross-validation, our two-IMU model achieved strong discriminative performance (macro-F1 = 0.80 +/- 0.07, MCC = 0.73 +/- 0.08; ROC-AUC > 0.93), with performance comparable to an OMC-based model and prediction timing suitable for real-time applications. Explainability analysis revealed dominant contributions from trunk dynamics and wrist-trunk interaction features. In preliminary evaluation using recordings from four participants with neurological conditions, the model retained good discriminative capability (ROC-AUC ~ 0.78), but showed reduced and variable threshold-dependent performance, highlighting challenges in clinical generalization. These results support sparse wearable sensing as a viable pathway toward scalable, real-time monitoring of CTMs during therapy and daily living.

Figures & tables

Fig. 1: Experimental setup for data acquisition and tasks conditions. (A) Optical motion capture (OMC) markers (black) were attached to the wrist, upper arm, and trunk, while inertial measurement units (IMU, blue) were co-mounted at the wrist and trunk. RGB cameras (red) provided synchronized video recordings from two viewpoints. Participants performed standardized activity of daily living at pre-defined target locations (“Start”, “T1-T4”) with task instructions displayed on a screen positioned in front of the participant. The axes of the anatomically aligned coordinate frame are indicated (x, y, z). (B) Schematic of the brace condition used to restrict elbow flexion and extension. (C) Schematic of the band condition designed to induce flexor-synergy like behavior.

Panel of Figure 2 — Fig. 2: Example scenarios illustrating arm–trunk coordination strategies during a forward-reaching task and the rationale for labeling compensatory trunk movements. (a) The degree of elbow extension ( $\beta$ ) and shoulder flexion ( $\gamma$ ) is sufficient to reach the target. Trunk flexion is therefore not required to extend the arm’s reach. (b) The degree of elbow extension ( $\beta$ ) and shoulder flexion ( $\gamma$ ) is insufficient to reach the target. In response, the trunk is flexed ( $\alpha\downarrow$ ) to extend the arm’s forward reach. (c) The degree of elbow extension ( $\beta$ ) and shoulder flexion ( $\gamma$ ) is maximized. The trunk is flexed to extend the arm’s forward reach toward the target. Since this trunk flexion is not compensating for limited elbow extension or shoulder flexion, it is not considered compensatory. (d) The degree of elbow extension ( $\beta$ ) and shoulder abduction ( $\gamma$ ) stays significantly below the required range to reach the target. Therefore, the trunk is excessively flexed ( $\alpha\downarrow$ ) as a compensatory strategy to extend the arm’s forward reach. Red annotations indicate trunk flexion angle ( $\alpha$ ), elbow extension ( $\beta$ ), and shoulder flexion/abduction ( $\gamma$ ).

Fig. 3: Overview of the signal processing and modeling pipeline. Inertial measurement units (IMU) and optical motion capture (OMC) signals are transformed into multiple kinematic representations (*variables in italic are only present in OMC data), segmented into overlapping windows, and characterized using statistical, location similarity (pairwise comparison of kinematic streams to capture coordination, e.g., wrist–trunk), and smoothness metrics. Model training and evaluation are conducted within a nested leave-one-subject-out cross-validation (LOSO-CV) scheme with balanced sample weighting to mitigate class imbalance. Final evaluation on patient data, together with post-hoc SHAP analysis [ 34 ] , supports preliminary clinical applicability and provides insight into model interpretability.

Fig. 4: Classification performance across optical motion capture marker configurations in able-bodied participants (n=10). Boxplots show fold-wise F1 scores for the three classes (Calibration, Movement: No Trunk Compensation, and Movement: Trunk Compensation) obtained using leave-one-subject-out cross-validation. Overlaid points indicate individual participant performance. Statistical comparisons on Movement: Trunk Compensation across marker configurations are indicated by brackets (* $p_{corr}<$ 0.05, n.s, not significant). Results for all classes and pairwise model comparisons are reported in Tab VI .

Fig. 5: Two-IMU-based model performance. (A) Boxplots show fold-wise F1 scores for optical motion capture (OMC) and inertial measurement unit (IMU) based setups using wrist and trunk signals in able-bodied participants ( $n=10$ ). Overlaid points indicate individual participant performance. (B) Mean one-vs-rest receiver operating characteristic (ROC) curves across able-bodied participants ( $n=10$ ) for the three classes demonstrate strong discriminative performance of the IMU model.

Fig. 6: Detailed analysis of IMU-based model performance. (A) Row-normalized confusion matrix summarizing class-wise prediction outcomes (recall) for Calibration (Calib), Movement: No Trunk Compensation (Mov: No TC), and Movement: Trunk Compensation (Mov: TC), aggregated across all test folds. (B) Boxplots depict fold-wise F1 scores across able-bodied participants ( $n=10$ ), stratified by movement task category. Overlaid points indicate individual participant performance.

Fig. 7: Explainability analysis using Shapley Additive Explanations (SHAP) to identify the top 10 discriminative features between Movement: No Trunk Compensation (Mov: No TC) and Movement: Trunk Compensation (Mov: TC). (A) Distributions of signed $\Delta$ SHAP values indicating directional class support, where positive values favor Mov: TC and negative values favor Mov: No TC. Feature values are color-coded from low (blue) to high (red) based on their relative magnitude. (B) Separation score expressed as Mean $|\Delta$ SHAP $|$ per feature summarizing overall discriminative importance. Bars are color-coded by feature origin (trunk, wrist, interaction).

Fig. 8: Detection performance with clinical dataset. (A) Mean one-vs-rest receiver operating characteristic (ROC) curves across across patients performing tasks with the most affected arm, with wrist sensor located on the most affected limb ( $n=4$ ) (B) Representative 60 s time-series excerpt showing ground-truth and predicted class labels for a patient performing conventional upper-limb therapy. Ground-truth annotations are shown as a black line, and model predictions as colored dots for the three classes (Movement: Trunk Compensation, Movement: No Trunk Compensation, and Calibration).

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. 9: Experimental workspace and task setup. (a) Schematic of the workspace showing the fixed start position (10 cm from the table edge) and four target locations. Target 1 (T1) corresponds to the participant’s maximum reaching distance, while Targets 2–4 (T2–T4) are placed at predefined radial distances (25–30 cm) with 60 ∘ angular separation. (b) Overview of task categories: (1) planar reaching and transport; (2) drinking and pouring; (3) elevated reaching and transport; (4) planar sliding; (5) fine object manipulation; and (6) continuous movements.

Task Category	Objects & Setup	Movement Tasks
Planar Reaching & Transport	Single cup at Start	– Reaching to T1–T4 (Task No. 1–4) – Cup transport between Start and T1–T4 (Task No. 5–8) – Sequential transport: Start $\rightarrow$ T2 $\rightarrow$ T3 $\rightarrow$ Start and Start $\rightarrow$ T3 $\rightarrow$ T2 $\rightarrow$ Start (Task No. 9–10) – Continuous reaching between T1–T4 in pseudo-random order (Task No. 11–12)
Drinking & Pouring	Two cups positioned to the left and right of T1	– Water pouring (right $\leftrightarrow$ left) (Task No. 13–14) – Drinking from right and left cup (Task No. 15–16)
Elevated Reaching & Transport	One large block at T1 and two smaller blocks on either side	– Elevation of small block onto large block with $180^{\circ}$ flip and return (Task No. 17–18) – Transport over large block to opposite side and return (Task No. 19–20)
Planar Sliding	Single card at Start	– Two-handed sliding: Start $\rightarrow$ T2 $\rightarrow$ T3 $\rightarrow$ Start and Start $\rightarrow$ T3 $\rightarrow$ T2 $\rightarrow$ Start (Task No. 21–22) – Two-handed sliding: Start $\rightarrow$ T1 $\rightarrow$ Start and Start $\rightarrow$ T4 $\rightarrow$ Start (Task No. 23–24) – Single-handed sliding to T1–T4 with $180^{\circ}$ card flip (Task No. 25–28) – Single-handed sequences: Start $\rightarrow$ T2 $\rightarrow$ T3 $\rightarrow$ Start and Start $\rightarrow$ T3 $\rightarrow$ T2 $\rightarrow$ Start (Task No. 29–30) – Continuous multi-target sliding sequences (Task No. 31–32)
Fine Object Manipulation	Five-drawer box at T1 and a coin placed to its right	– Coin manipulation across drawers (opening, closing, insertion, transfer) (Task No. 33–34) – Sequential opening and closing of all drawers (Task No. 35–36)
Continuous Movements	—	– Hand-biking motion (20–30 s) (Task No. 37) – Swimming motion (20–30 s) (Task No. 38)

Appendix

TABLE I: Overview of task categories, objects, and movement tasks defined relative to target locations T1–T4 in the data acquisition protocol for able-bodied participants.

Label	Description
Movement: Trunk Compensation	Movement Pattern: Visible trunk displacement in the sagittal plane relative to the pelvis. Elbow extension and shoulder flexion remain below the required range, and trunk flexion compensates for limited reach. Onset: Initiation of forward trunk displacement with limited shoulder and elbow motion. Offset: Return to upright posture and stabilization.
Movement: No Trunk Compensation	Movement Pattern: Arm movement occurs while trunk displacement remains within the functional range required for task execution. No compensatory trunk motion is observed. Onset: Visible arm movement while the trunk remains stable. Offset: Arm returns to calibration position and posture stabilizes.
Calibration	Movement Pattern: Arm rests in a bent position close to the body with no visible trunk or wrist movement. Neutral alignment of shoulders, pelvis, and spine. Onset: Completion of previous task and stabilization in calibration posture. Offset: Initiation of movement via visible displacement of wrist, shoulder, or trunk.

Appendix

TABLE II: Labeling guidelines for the three movement classes, including movement characteristics and onset/offset definitions.

Fig. 10: Coordinate frames used for IMU-based anatomical calibration: Local sensor frames are attached to each IMU and vary with segment motion, while the VQF estimates orientations in a global world frame with arbitrary yaw. A standardized calibration pose is used to define a sensor-specific anatomically aligned frame that matches the OMC recording frame, allowing consistent interpretation of sensor orientations.

Modality	Kinematic data streams	Frame
IMU	Wrist/trunk acceleration (3-axis & magnitude)	Local & Calibrated
	Wrist/trunk angular velocity (3-axis & magnitude)	Local & Calibrated
	Absolute wrist/trunk orientation (3-axis)	Calibrated
	Relative wrist–trunk orientation (3-axis & 3D rotation angle)	Calibrated
OMC	Wrist/trunk/upper-arm acceleration (3-axis & magnitude)	Recording
	Wrist/trunk/upper-arm angular velocity (3-axis & magnitude)	Recording

Appendix

TABLE III: Summary of sensing modalities and available kinematic data streams after pre-processing, together with their respective coordinate frames. Three-axis quantities are represented along the Cartesian axes (x/y/z) or rotational axes (roll/pitch/yaw). Magnitude denotes the Euclidean norm of the corresponding vector signal. The 3D rotation angle represents the total angular displacement between two orientations.

Metric Category	Measures
Univariate channel-wise statistics	Mean, standard deviation (std), minimum (min), maximum (max), range, root mean square (RMS), mean absolute deviation (MAD), trend, skewness, kurtosis
Bivariate channel-wise similarity metrics (applied on acceleration, absolute orientation, linear/angular velocity)	Maximum normalized cross-correlation, normalized dynamic time warping distance (DTWD)
Smoothness metrics (applied on acceleration and linear velocity)	Logarithmic dimensionless jerk (LDJ) for wrist & trunk + LDJ ratio between wrist & trunk

Appendix

TABLE IV: Summary of feature metric categories applied to the kinematic data streams and derived descriptors listed in Tab. III . Univariate metrics are computed independently for each input signal, whereas bivariate metrics quantify relationships between pairs of signals. Channel-wise metrics are computed separately for each signal component (e.g., x/y/z for vector quantities or roll/pitch/yaw for Euler angles), with each component treated as an individual channel (i.e., a single dimension of a multi-dimensional signal) within a kinematic data stream. For 3D vector signals, the Euclidean norm is additionally treated as an extra channel. All metrics are computed within each analysis window of the sliding-window segmentation and serve as input features for the classification model.

Configuration	$\mathbf{F1_{\text{macro}}}$	MCC	Calib			Mov: No TC			Mov: TC
			Prec	Rec	F1	Prec	Rec	F1	Prec	Rec	F1
IMU (W+T)	0.80 $\pm$ 0.07	0.73 $\pm$ 0.08	0.92 $\pm$ 0.06	0.86 $\pm$ 0.13	0.88 $\pm$ 0.07	0.85 $\pm$ 0.08	0.81 $\pm$ 0.12	0.82 $\pm$ 0.07	0.64 $\pm$ 0.21	0.82 $\pm$ 0.14	0.68 $\pm$ 0.11
OMC (W+T+UA)	0.86 $\pm$ 0.05	0.81 $\pm$ 0.08	0.95 $\pm$ 0.03	0.92 $\pm$ 0.08	0.93 $\pm$ 0.05	0.88 $\pm$ 0.07	0.88 $\pm$ 0.07	0.88 $\pm$ 0.06	0.75 $\pm$ 0.12	0.83 $\pm$ 0.13	0.78 $\pm$ 0.09
OMC (W+T)	0.85 $\pm$ 0.06	0.79 $\pm$ 0.09	0.95 $\pm$ 0.04	0.91 $\pm$ 0.10	0.92 $\pm$ 0.05	0.89 $\pm$ 0.07	0.84 $\pm$ 0.12	0.86 $\pm$ 0.08	0.72 $\pm$ 0.17	0.86 $\pm$ 0.10	0.77 $\pm$ 0.10
OMC (T)	0.66 $\pm$ 0.12	0.50 $\pm$ 0.13	0.72 $\pm$ 0.08	0.68 $\pm$ 0.26	0.65 $\pm$ 0.16	0.73 $\pm$ 0.14	0.62 $\pm$ 0.11	0.66 $\pm$ 0.10	0.60 $\pm$ 0.24	0.83 $\pm$ 0.16	0.65 $\pm$ 0.17
OMC (W)	0.78 $\pm$ 0.06	0.71 $\pm$ 0.08	0.95 $\pm$ 0.03	0.91 $\pm$ 0.10	0.92 $\pm$ 0.06	0.86 $\pm$ 0.07	0.77 $\pm$ 0.09	0.81 $\pm$ 0.06	0.54 $\pm$ 0.13	0.77 $\pm$ 0.09	0.62 $\pm$ 0.09

Appendix

TABLE V: Able-bodied classification performance across sensing configurations using wrist (W), trunk (T), and upper-arm (UA) information. Metrics are reported as mean

\pm

standard deviation.

Model A	Model B	Class	$p$	$p_{\mathrm{corr}}$
W+T+UA	W+T	Calib	0.232	1.000
W+T+UA	W+T	Mov: No TC	0.193	1.000
W+T+UA	W+T	Mov: TC	0.625	1.000
W+T+UA	T	Calib	0.002	0.035
W+T+UA	T	Mov: No TC	0.002	0.035
W+T+UA	T	Mov: TC	0.004	0.039

Appendix

TABLE VI: Pairwise models comparison (using wrist (W), trunk (T), and upper-arm (UA) information) using Wilcoxon signed-rank tests (Holm-corrected p-values). Sig.* corresponds to

p_{\mathrm{corr}}<0.05

Patient	$\mathbf{F1_{\text{macro}}}$	MCC	$\mathbf{AUC_{\text{macro}}}$	Calib			Mov: No TC			Mov: TC
				Prec	Rec	F1	Prec	Rec	F1	Prec	Rec	F1
P03 (Right)	0.51	0.35	0.85	0.15	0.34	0.21	0.90	0.96	0.93	0.67	0.27	0.38
P18 (Left)	0.40	0.28	0.73	0.96	0.39	0.55	0.34	0.91	0.50	0.39	0.09	0.15
P63 (Left)	0.34	0.26	0.82	0.96	0.15	0.27	0.40	0.99	0.57	0.96	0.11	0.20
P31 (Right)	0.45	0.25	0.72	0.86	0.31	0.44	0.48	0.85	0.61	0.28	0.34	0.30
Avg $\pm$ Std	0.43 $\pm$ 0.06	0.28 $\pm$ 0.04	0.78 $\pm$ 0.06	0.73 $\pm$ 0.35	0.30 $\pm$ 0.09	0.37 $\pm$ 0.14	0.53 $\pm$ 0.22	0.93 $\pm$ 0.06	0.65 $\pm$ 0.16	0.57 $\pm$ 0.28	0.20 $\pm$ 0.11	0.26 $\pm$ 0.09

Appendix

TABLE VII: Clinical generalization results using sensor placed on the most affected arm, during tasks involving most affected limb.

Wearable Sensing

Wearable sensing uses body-worn devices to measure motion, physiology, and interaction for health, activity, and assistive applications.

See all 130 papers

Sep 11, 2026

Hybrid Physics-AI Framework of Body Center of Mass Dynamics from Wrist-Worn Sensors

Wrist-worn IMU has been widely used for daily-life health monitoring. Yet, it does not fully represent whole-body dynamics, for which the body center of mass (COM) is considered the physiological reference standard. Therefore, this work proposes a simplified kinematic model (KM), which is designed to map the wrist IMU to the COM acceleration. It is built upon several reductive assumptions that enable the solvability of the dynamic equations based on wrist IMU measurements alone. This work further proposes three types of hybrid AI modeling methods, namely human kinematic model-based neural network (HKM-NN) models, to leverage the power of both grey-box and black-box modeling. The HKM-NN methods include serial learning (ser-) and two approaches of simultaneous learning (sim1- and sim2-). The proposed models are trained and tested using our dataset, which includes wrist IMU measurements and ground-truth COM measurements from 10 healthy volunteers during six gait activities and sit-to-stand (SS) transitional movement. The results demonstrate the feasibility of estimating COM acceleration from wrist IMU measurements. Our KM model yields satisfactory results, with an error ranging from 6.7% to 12.5% for gait activities and 5.6% for the SS. In comparison with the KM model, our HKM-NN models significantly enhance the performance, achieving 5.3% to 9.3% errors for gait activities, and the best error of 3.9% for the SS. In addition, the HKM-NN models demonstrate distinct robustness characteristics under noisy test conditions, with sim1-/sim2- generally maintaining greater robustness under Gaussian perturbations, while the KM model exhibits comparatively strong robustness under salt-and-pepper noise. These findings highlight the importance of combining biomechanical structure with data-driven learning for wearable sensing applications operating under imperfect and noisy measurement conditions.

Shuhao Que, Valentina Breschi, Ying Wang

Electrical Engineering, University of Twente, Drienerlolaan 5, Enschede, 7522 NB, Netherlands. · Electrical Engineering, Eindhoven University of Technology, Groene Loper 19, Eindhoven, 5612 AP, Netherlands.

Gait Analysis Wearable Sensing Scientific ML

Ask PDF

Feb 24, 2026

Event-Driven On-Sensor Locomotion Mode Recognition Using a Shank-Mounted IMU with Embedded Machine Learning for Exoskeleton Control

This work presents a wearable human activity recognition (HAR) system that performs real-time inference directly inside a shank-mounted inertial measurement unit (IMU) to support low-latency control of a lower-limb exoskeleton. Unlike conventional approaches that continuously stream raw inertial data to a microcontroller for classification, the proposed system executes activity recognition at the sensor level using the embedded Machine Learning Core (MLC) of the STMicroelectronics LSM6DSV16X IMU, allowing the host microcontroller to remain in a low-power state and read only the recognized activity label from IMU registers. While the system generalizes to multiple human activities, this paper focuses on three representative locomotion modes - stance, level walking, and stair ascent - using data collected from adult participants. A lightweight decision-tree model was configured and deployed for on-sensor execution using ST MEMS Studio, enabling continuous operation without custom machine learning code on the microcontroller. During operation, the IMU asserts an interrupt when motion or a new classification is detected; the microcontroller wakes, reads the MLC output registers, and forwards the inferred mode to the exoskeleton controller. This interrupt-driven, on-sensor inference architecture reduces computation and communication overhead while preserving battery energy and improving robustness in distinguishing level walking from stair ascent for torque-assist control.

Mohammadsaleh Razmi, Iman Shojaei

Align Technology, Inc., San Jose, CA 95134, USA

Wearable Sensing Human Activity Recognition On-Device Inference

Ask PDF

May 30, 2026

An explainable hierarchical self attention-based approach for tremor detection in the time domain

Tremor is a common movement disorder associated with conditions like Parkinson's disease and Essential tremor, traditionally diagnosed through expert clinician assessment. Current automated detection methods rely on frequency-domain features informed by clinical expertise. In this work, we present an explainable, two-stage hierarchical framework for tremor detection in the time domain that learns tremor patterns directly from 3D kinematic marker time-series data across entire tremor-provoking trials. Our framework combined a deep convolutional and long short-term memory network to learn tremor representations from short, discrete, non-overlapping time segments of kinematic time series data from trials, which are then processed by a vision transformer that models their long-term temporal dynamics of time segment features for trial (session) level classification. Evaluated across nine body parts, the framework achieved F1-scores of 0.594 - 0.947 depending on body parts (average: 0.765), falling short of the frequency-domain state-of-the-art performance (0.909) while requiring minimal preprocessing. Attention weights and gradient-based class activation maps (Grad-CAM) identified time-domain features of tremor across body parts. This proof of concept demonstrated the feasibility of data-driven time-domain modeling for tremor detection across anatomically diverse body parts, while reducing reliance on expert-engineered spectral features and providing posthoc interpretability of temporal and anatomical patterns of tremor.

Timothy Odonga, Jeanne M. Powell, Mark Saad +5

Department of Biomedical Informatics, School of Medicine, Emory University, Atlanta, GA, USA · Jean and Paul Amos Parkinson’s Disease and Movement Disorders Program, Department of Neurology, School of Medicine, Emory University, Atlanta, GA, USA · Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology, Atlanta, GA, USA

Time Series Classification Time Series Representation Learning

Ask PDF

Abstract

Wearable Sensing

Explore similar work

Hybrid Physics-AI Framework of Body Center of Mass Dynamics from Wrist-Worn Sensors

Event-Driven On-Sensor Locomotion Mode Recognition Using a Shank-Mounted IMU with Embedded Machine Learning for Exoskeleton Control

An explainable hierarchical self attention-based approach for tremor detection in the time domain