Thor F. Besier
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Thor F. Besier has a total of 365 co-authors. Here are the top co-authors by number of shared publications:
- David G. Lloyd
Australian Centre for Precision Health and Technology (PRECISE), Griffith University
39 shared publications Scott L. Delp
Affiliation Departments of Mechanical Engineering and Bioengineering, Stanford University, Stanford, California, United States of America
28 shared publications- Gary S. Beaupre
VA Palo Alto Health Care System
19 shared publications Garry E. Gold
Department of Bioengineering, Stanford University, Stanford, CA, USA
15 shared publicationsMichael Fredericson
Department of Orthopaedic Surgery, Stanford University, Stanford, CA, USA
15 shared publications- Justin Fernandez
Auckland Bioengineering Institute, University of Auckland, Auckland, New Zealand
14 shared publications
Latest Publications
- Open Access
- 28 September 2026
- Published by Elsevier in SSRN Electronic Journal
Accuracy of Personalised Paediatric Femur and Tibia Finite Element Modelling Under Varying Imaging Conditions Using Statistical Shape and Density Models
Paediatric bone mechanics exhibit substantial inter-subject variability driven by growth-related changes in geometry and bone mineral density, motivating subject-specific finite element (FE) modelling. However, development of these models is challenged by limited imaging availability in children. This study aims to evaluate how imaging modality and axial coverage influence the mechanical accuracy of paediatric femoral and tibial FE models when geometry and material properties are predicted from statistical shape and density models (SSDMs). Using a large cohort of paediatric CT scans (n = 330, ages 4-18 years), SSDMs were constructed to characterise three-dimensional bone geometry and bone mineral density variation in the tibia and femur. Three different predictive frameworks were implemented to reconstruct individual bone shape and density under varying input conditions: 1) non-image-based (partial least square regression, PLSR using demographic and anthropometric information), 2) image-based using shape only (SO), and 3) imaging-based using shape and density (SD). The SO and SD approaches were predicted with full or 15% axial coverage. Predicted FE outcomes were compared with the CT-based reference models. Across all strategies, average FE errors remained within 5-8% for von Mises stress and 1-5% for first and third principal strains. SD fittings consistently outperformed SO fittings and PLSR-based approaches in reproducing CT-based FE outcomes, while maintaining robust performance with partial coverage. PLSR provided a radiation-free alternative with performance similar to SO fitting with partial coverage. These findings highlight that SSDM-based modelling can generate personalised FE models from limited imaging inputs, enabling lower-radiation biomechanical assessment with potential applications in implant design and surgical planning.
- 28 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAccuracy of Personalised Paediatric Femur and Tibia Finite Element Modelling Under Varying Imaging Conditions Using Statistical Shape and Density Models
Paediatric bone mechanics exhibit substantial inter-subject variability driven by growth-related changes in geometry and bone mineral density, motivating subject-specific finite element (FE) modelling. However, development of these models is challenged by limited imaging availability in children. This study aims to evaluate how imaging modality and axial coverage influence the mechanical accuracy of paediatric femoral and tibial FE models when geometry and material properties are predicted from statistical shape and density models (SSDMs). Using a large cohort of paediatric CT scans (n = 330, ages 4-18 years), SSDMs were constructed to characterise three-dimensional bone geometry and bone mineral density variation in the tibia and femur. Three different predictive frameworks were implemented to reconstruct individual bone shape and density under varying input conditions: 1) non-image-based (partial least square regression, PLSR using demographic and anthropometric information), 2) image-based using shape only (SO), and 3) imaging-based using shape and density (SD). The SO and SD approaches were predicted with full or 15% axial coverage. Predicted FE outcomes were compared with the CT-based reference models. Across all strategies, average FE errors remained within 5-8% for von Mises stress and 1-5% for first and third principal strains. SD fittings consistently outperformed SO fittings and PLSR-based approaches in reproducing CT-based FE outcomes, while maintaining robust performance with partial coverage. PLSR provided a radiation-free alternative with performance similar to SO fitting with partial coverage. These findings highlight that SSDM-based modelling can generate personalised FE models from limited imaging inputs, enabling lower-radiation biomechanical assessment with potential applications in implant design and surgical planning.
- Open Access
- 17 June 2026
- Published by Wiley in Developmental Medicine & Child Neurology
Machine learning approaches in automated infant General Movements Assessment: A scoping review
Aim: To evaluate and synthesize the literature on machine learning and deep learning methodologies in automated infant General Movements Assessment (GMA) for early prediction of cerebral palsy (CP). Method: A scoping review was conducted using the framework of Arksey and O'Malley and reported in accordance with the PRISMA/PRISMA‐ScR guidelines. Literature published from 2015 to February 2025 was searched across major databases. Included studies were analysed for key factors commonly identified as influencing GMA classification performance. Results: Forty‐three studies met the inclusion criteria. The most common analysis pipeline involved deep learning‐based markerless pose reconstruction, followed by classification, typically using support vector machines (machine learning) and convolutional neural networks (deep learning). Smaller datasets (12–38 infants) yielded greater accuracy (up to 100%) in maximum general movement classification than larger datasets (557–776 infants), which achieved up to 93.8%. All CP classification studies reported accuracies above 90%. Interpretation: Although deep learning methods generally outperform traditional machine learning approaches, reported performance varies substantially across studies and is shaped by factors such as data quality and diversity, recording conditions, pose estimation methods, feature representation, and classifier design. Unlike earlier reviews, this work provides a comprehensive synthesis of recent automated GMA literature, explicitly linking methodological choices across data acquisition, pose estimation, feature extraction, and classification to differences in reported performance, generalizability, and clinical interpretability. By consolidating these interdependent sources of variability into a single clinically oriented framework, this review offers a stand‐alone reference to the field and provides guidance for developing more robust and clinically meaningful automated GMA systems for early detection of CP.
- 17 June 2026
- Published by Wiley in Developmental Medicine & Child Neurology
Open AccessMachine learning approaches in automated infant General Movements Assessment: A scoping review
Aim: To evaluate and synthesize the literature on machine learning and deep learning methodologies in automated infant General Movements Assessment (GMA) for early prediction of cerebral palsy (CP). Method: A scoping review was conducted using the framework of Arksey and O'Malley and reported in accordance with the PRISMA/PRISMA‐ScR guidelines. Literature published from 2015 to February 2025 was searched across major databases. Included studies were analysed for key factors commonly identified as influencing GMA classification performance. Results: Forty‐three studies met the inclusion criteria. The most common analysis pipeline involved deep learning‐based markerless pose reconstruction, followed by classification, typically using support vector machines (machine learning) and convolutional neural networks (deep learning). Smaller datasets (12–38 infants) yielded greater accuracy (up to 100%) in maximum general movement classification than larger datasets (557–776 infants), which achieved up to 93.8%. All CP classification studies reported accuracies above 90%. Interpretation: Although deep learning methods generally outperform traditional machine learning approaches, reported performance varies substantially across studies and is shaped by factors such as data quality and diversity, recording conditions, pose estimation methods, feature representation, and classifier design. Unlike earlier reviews, this work provides a comprehensive synthesis of recent automated GMA literature, explicitly linking methodological choices across data acquisition, pose estimation, feature extraction, and classification to differences in reported performance, generalizability, and clinical interpretability. By consolidating these interdependent sources of variability into a single clinically oriented framework, this review offers a stand‐alone reference to the field and provides guidance for developing more robust and clinically meaningful automated GMA systems for early detection of CP.
- 1 April 2026
- Published by British Editorial Society of Bone & Joint Surgery in The Bone & Joint Journal
- Vol. 108-B (4) , 478-489
Have we come to the end of the patient-reported outcome measure (PROM)?
Aims: The purpose of this study was to evaluate early outcomes of patients who underwent robotic-assisted total knee arthroplasty (RA-TKA) and conventional TKA using wearable sensors and traditional patient-reported outcome measures (PROMs). Methods: This was a prospective, matched, parallel cohort study of 100 patients with symptomatic end-stage knee osteoarthritis undergoing primary TKA (44 RA-TKA, 56 conventional TKA). Functional outcomes were assessed using wearable inertial measurement units (IMUs). IMU-based outcomes included impact load, impact asymmetry, maximum knee flexion angle, and bone stimulus. PROMs, including Oxford Knee Score, EuroQol five-dimension five-level questionnaire, EuroQol visual analogue scale, and Forgotten Joint Score, were evaluated at preoperative baseline and weeks 2 to 6 postoperatively. Results: By postoperative week 6, RA-TKA patients showed improved function compared with the conventional TKA group, with increased maximum knee flexion angle (118° (SD 6.6°) vs 113° (SD 5.4°); p = 0.038), improved symmetry with regard to limb loading (82.3% (SD 16.1) vs 22.4% (SD 9.5); p < 0.001), increased cumulative impact load (146.6% (SD 16.7) vs 37% (SD 8.1); p < 0.001), and increased bone stimulus (25.1% vs 13.6%; p = 0.003). RA-TKA patients demonstrated an earlier return to symmetrical limb loading compared with conventional TKA by postoperative week 3. There were no significant differences in PROMs between the two groups at any timepoint. Conclusion: RA-TKAs using personalized alignment were associated with faster return of symmetrical gait and improved range of motion when assessed using IMUs compared with conventional TKA, which were not detected by traditional PROMs during the early postoperative period. Cite this article: Bone Joint J 2026;108-B(4):478–489.
- 1 April 2026
- Published by British Editorial Society of Bone & Joint Surgery in The Bone & Joint Journal
- Vol. 108-B (4) , 478-489
Have we come to the end of the patient-reported outcome measure (PROM)?
Aims: The purpose of this study was to evaluate early outcomes of patients who underwent robotic-assisted total knee arthroplasty (RA-TKA) and conventional TKA using wearable sensors and traditional patient-reported outcome measures (PROMs). Methods: This was a prospective, matched, parallel cohort study of 100 patients with symptomatic end-stage knee osteoarthritis undergoing primary TKA (44 RA-TKA, 56 conventional TKA). Functional outcomes were assessed using wearable inertial measurement units (IMUs). IMU-based outcomes included impact load, impact asymmetry, maximum knee flexion angle, and bone stimulus. PROMs, including Oxford Knee Score, EuroQol five-dimension five-level questionnaire, EuroQol visual analogue scale, and Forgotten Joint Score, were evaluated at preoperative baseline and weeks 2 to 6 postoperatively. Results: By postoperative week 6, RA-TKA patients showed improved function compared with the conventional TKA group, with increased maximum knee flexion angle (118° (SD 6.6°) vs 113° (SD 5.4°); p = 0.038), improved symmetry with regard to limb loading (82.3% (SD 16.1) vs 22.4% (SD 9.5); p < 0.001), increased cumulative impact load (146.6% (SD 16.7) vs 37% (SD 8.1); p < 0.001), and increased bone stimulus (25.1% vs 13.6%; p = 0.003). RA-TKA patients demonstrated an earlier return to symmetrical limb loading compared with conventional TKA by postoperative week 3. There were no significant differences in PROMs between the two groups at any timepoint. Conclusion: RA-TKAs using personalized alignment were associated with faster return of symmetrical gait and improved range of motion when assessed using IMUs compared with conventional TKA, which were not detected by traditional PROMs during the early postoperative period. Cite this article: Bone Joint J 2026;108-B(4):478–489.
- Open Access
- 5 February 2026
- Published by ASME International in Journal of Biomechanical Engineering
- Vol. 148 (5) , 1-29
Deciphering the “Art” in Modeling and Simulation of the Knee Joint: Model Benchmarking
Given the strong ties to data sharing and the responsible use of resources, reproducibility of modeling and simulation practice is of paramount importance in science. Computational models in orthopedics provide insight into healthy and injured joint mechanics and can inform clinical decision-making. The KneeHub project investigated the influence of modelers' decisions and thus their “art” in simulation and modeling; five teams developed and calibrated knee models using the same experimental data. Model benchmarking evaluated the predictive ability of the models under loading scenarios that were not considered in the development and calibration process. The objective of this study was to evaluate the accuracy of predictions of knee-specific joint biomechanics for benchmark scenarios of simulating a resected anterior cruciate ligament (ACL) using models of one knee and a combined pivot shift loading using models of another knee. The models predicted the major trends in kinematics and kinetics; however, differences were observed in comparison to experimental data and between teams. Model-to-experiment root-mean-square (RMS) errors were up to 6.6±2.4 mm in anterior–posterior (AP) translation, 13.5±12.9 deg in internal–external (IE) rotation, and 5.3±3.4 deg in varus–valgus (VV) rotation; errors were largest in internal–external rotation, and standard deviations reflected differences between teams. While calibrated models were tuned to a similar set of conditions (albeit with different decisions), the optimized stiffness and reference length/strain of ligament structures may not fully reproduce the contributions of these structures to joint kinematics that were measured experimentally in the benchmark scenarios. As researchers often extend models beyond the conditions used to calibrate them, quantifying model accuracy and limitations with benchmarking represents a crucial step toward reproducibility and can help establish best practices for credible modeling in our community.
- 5 February 2026
- Published by ASME International in Journal of Biomechanical Engineering
- Vol. 148 (5) , 1-29
Open AccessDeciphering the “Art” in Modeling and Simulation of the Knee Joint: Model Benchmarking
Given the strong ties to data sharing and the responsible use of resources, reproducibility of modeling and simulation practice is of paramount importance in science. Computational models in orthopedics provide insight into healthy and injured joint mechanics and can inform clinical decision-making. The KneeHub project investigated the influence of modelers' decisions and thus their “art” in simulation and modeling; five teams developed and calibrated knee models using the same experimental data. Model benchmarking evaluated the predictive ability of the models under loading scenarios that were not considered in the development and calibration process. The objective of this study was to evaluate the accuracy of predictions of knee-specific joint biomechanics for benchmark scenarios of simulating a resected anterior cruciate ligament (ACL) using models of one knee and a combined pivot shift loading using models of another knee. The models predicted the major trends in kinematics and kinetics; however, differences were observed in comparison to experimental data and between teams. Model-to-experiment root-mean-square (RMS) errors were up to 6.6±2.4 mm in anterior–posterior (AP) translation, 13.5±12.9 deg in internal–external (IE) rotation, and 5.3±3.4 deg in varus–valgus (VV) rotation; errors were largest in internal–external rotation, and standard deviations reflected differences between teams. While calibrated models were tuned to a similar set of conditions (albeit with different decisions), the optimized stiffness and reference length/strain of ligament structures may not fully reproduce the contributions of these structures to joint kinematics that were measured experimentally in the benchmark scenarios. As researchers often extend models beyond the conditions used to calibrate them, quantifying model accuracy and limitations with benchmarking represents a crucial step toward reproducibility and can help establish best practices for credible modeling in our community.
- 1 January 2026
- Published by Human Kinetics in Journal of Applied Biomechanics
- p. 1-8
Evaluating the Accuracy of Low-Cost Scanning Technology to Reconstruct Full-Body Morphology
Accurate reconstruction of full-body morphology is critical in many fields. This study evaluated the accuracy of 3 low-cost scanning technologies for capturing 3D body surface geometry: LiDAR sensors, ZOZOSUIT2, and ZOZO_v3. Participants were scanned using laser scan, LiDAR, and ZOZO’s methods. The scans were fitted to STAR models for comparison. The low-cost technologies captured 3D body surface geometry in an adult population, with RMSE values of 11.33 (4.39) mm (LiDAR), 7.13 (1.65) mm (ZOZOSUIT2), and 6.61 (1.93) mm (ZOZO_v3). All 3 methods demonstrated no gender bias in reconstruction accuracy.
- 1 January 2026
- Published by Human Kinetics in Journal of Applied Biomechanics
- p. 1-8
Evaluating the Accuracy of Low-Cost Scanning Technology to Reconstruct Full-Body Morphology
Accurate reconstruction of full-body morphology is critical in many fields. This study evaluated the accuracy of 3 low-cost scanning technologies for capturing 3D body surface geometry: LiDAR sensors, ZOZOSUIT2, and ZOZO_v3. Participants were scanned using laser scan, LiDAR, and ZOZO’s methods. The scans were fitted to STAR models for comparison. The low-cost technologies captured 3D body surface geometry in an adult population, with RMSE values of 11.33 (4.39) mm (LiDAR), 7.13 (1.65) mm (ZOZOSUIT2), and 6.61 (1.93) mm (ZOZO_v3). All 3 methods demonstrated no gender bias in reconstruction accuracy.
- Open Access
- 4 December 2025
- Published by Elsevier in Journal of Biomechanics
- Vol. 195, 113107
An investigation into the redefinition of the alpha angle for 3D imaging
- 4 December 2025
- Published by Elsevier in Journal of Biomechanics
- Vol. 195, 113107
Open AccessAn investigation into the redefinition of the alpha angle for 3D imaging
- 28 October 2025
- Published by Cassyni
Modelling skeletal muscle for orthopaedic application
Characterising the mechanical environment of skeletal tissues is important for understanding injury and disease and critical to designing medical devices or personalising treatment interventions. Neuromusculoskeletal models provide a framework for estimating muscle and joint contact forces that are consistent with observed motion and measured external forces. Skeletal muscles (and tendons) are typically represented as line-segment force actuators when solving multi-body dynamics, using phenomenological, lumped parameters to describe force-length and force-velocity relationships (Hill-type models). However, this approach does not capture the spatial arrangement, fibre orientation and mechanical interactions of muscles with bones and connective tissues. This talk discusses opportunities to improve models of skeletal muscle with application to orthopaedics.
- 28 October 2025
- Published by Cassyni
Modelling skeletal muscle for orthopaedic application
Characterising the mechanical environment of skeletal tissues is important for understanding injury and disease and critical to designing medical devices or personalising treatment interventions. Neuromusculoskeletal models provide a framework for estimating muscle and joint contact forces that are consistent with observed motion and measured external forces. Skeletal muscles (and tendons) are typically represented as line-segment force actuators when solving multi-body dynamics, using phenomenological, lumped parameters to describe force-length and force-velocity relationships (Hill-type models). However, this approach does not capture the spatial arrangement, fibre orientation and mechanical interactions of muscles with bones and connective tissues. This talk discusses opportunities to improve models of skeletal muscle with application to orthopaedics.
- Open Access
- 27 October 2025
- Published by Elsevier in Journal of Biomechanics
- Vol. 194, 113034
Robust workflow for diaphyseal cortical bone thickness calculation
- 27 October 2025
- Published by Elsevier in Journal of Biomechanics
- Vol. 194, 113034
Open AccessRobust workflow for diaphyseal cortical bone thickness calculation
- 5 February 2025
- Published by Wolters Kluwer Health in Medicine & Science in Sports & Exercise
- Vol. 57 (7) , 1376-1382
Knee Ligament Moment Arms Scale with Epicondylar Width and Might Contribute to Ligament Injuries in Females
Introduction: Females are prone to knee ligament injuries compared with males. Moment arms describe the mechanical advantage of ligaments to stabilize the knee from injurious moments. Compared with males of the same stature, females have a smaller femoral epicondylar width, which we hypothesized would reduce their knee joint ligament moment arms in the frontal plane. Methods: To calculate varus–valgus ligament moment arms, we created 26 patient-specific finite element models (17F; 9M) and simulated an axial load of half body weight. Tibial cartilage contact pressures were obtained and used to define the medial and lateral peak contact pressure, which were assumed to be the point of rotation in the frontal plane. The varus and valgus moment arms of the anterior cruciate ligament (ACL), medial collateral ligament (MCL), and lateral collateral ligament (LCL) were calculated as the perpendicular distance between the point of peak contact pressure and the centroid of each ligament. Multiple linear regression models with sex, height, and epicondylar width as independent variables were produced. Results: Females had smaller varus–valgus ligament moment arms compared with males, with a mean difference of 7 to 9 mm observed by the MCL and LCL about the lateral and medial contact points, respectively (~12% smaller, P < 0.001). The female cohort exhibited ACL moment arms about the medial and lateral contact points that were ~4 mm less than the male cohort ( P < 0.05). Correlations between varus–valgus ligament moment arm and epicondylar width were found for all three ligaments using either the medial or lateral points of peak contact pressure. Conclusions: These findings support our hypothesis that females have smaller varus–valgus ligament moment arms compared with males, which could contribute to the higher rate of female ligament injuries compared with males.
- 5 February 2025
- Published by Wolters Kluwer Health in Medicine & Science in Sports & Exercise
- Vol. 57 (7) , 1376-1382
Knee Ligament Moment Arms Scale with Epicondylar Width and Might Contribute to Ligament Injuries in Females
Introduction: Females are prone to knee ligament injuries compared with males. Moment arms describe the mechanical advantage of ligaments to stabilize the knee from injurious moments. Compared with males of the same stature, females have a smaller femoral epicondylar width, which we hypothesized would reduce their knee joint ligament moment arms in the frontal plane. Methods: To calculate varus–valgus ligament moment arms, we created 26 patient-specific finite element models (17F; 9M) and simulated an axial load of half body weight. Tibial cartilage contact pressures were obtained and used to define the medial and lateral peak contact pressure, which were assumed to be the point of rotation in the frontal plane. The varus and valgus moment arms of the anterior cruciate ligament (ACL), medial collateral ligament (MCL), and lateral collateral ligament (LCL) were calculated as the perpendicular distance between the point of peak contact pressure and the centroid of each ligament. Multiple linear regression models with sex, height, and epicondylar width as independent variables were produced. Results: Females had smaller varus–valgus ligament moment arms compared with males, with a mean difference of 7 to 9 mm observed by the MCL and LCL about the lateral and medial contact points, respectively (~12% smaller, P < 0.001). The female cohort exhibited ACL moment arms about the medial and lateral contact points that were ~4 mm less than the male cohort ( P < 0.05). Correlations between varus–valgus ligament moment arm and epicondylar width were found for all three ligaments using either the medial or lateral points of peak contact pressure. Conclusions: These findings support our hypothesis that females have smaller varus–valgus ligament moment arms compared with males, which could contribute to the higher rate of female ligament injuries compared with males.
- Open Access
- 8 January 2025
- Published by Springer Nature in Experimental Brain Research
- Vol. 243 (1) , 1-16
The role of muscle synergies and task constraints on upper limb motor impairment after stroke
This study explores the role of task constraints over muscle synergies expression in the context of upper limb motor impairment after stroke. We recruited nine chronic stroke survivors with upper limb impairments and fifteen healthy controls, who performed a series of tasks designed to evoke muscle synergies through various spatial explorations. These tasks included an isometric force task, a dynamic reaching task, the clinical Fugl-Meyer (FM) assessment, and a pinch task. Electromyographic data from 16 upper limb muscles were collected during each task, alongside intermuscular coherence (IMC) measurements during the pinch task to assess neuromuscular connectivity. The findings confirm that motor impairment is inversely related to the diversity of muscle synergies, with fewer synergies and more stereotypical synergy structures observed post-stroke. The study further reveals that the nature of motor tasks significantly affects the number of identifiable muscle synergies, with less constrained tasks revealing a broader array of synergies. These findings highlight the importance of carefully selecting motor tasks in the context of clinical research and assessments to understand a patient’s motor impairment, thus aiding in developing tailored rehabilitation strategies.
- 8 January 2025
- Published by Springer Nature in Experimental Brain Research
- Vol. 243 (1) , 1-16
Open AccessThe role of muscle synergies and task constraints on upper limb motor impairment after stroke
This study explores the role of task constraints over muscle synergies expression in the context of upper limb motor impairment after stroke. We recruited nine chronic stroke survivors with upper limb impairments and fifteen healthy controls, who performed a series of tasks designed to evoke muscle synergies through various spatial explorations. These tasks included an isometric force task, a dynamic reaching task, the clinical Fugl-Meyer (FM) assessment, and a pinch task. Electromyographic data from 16 upper limb muscles were collected during each task, alongside intermuscular coherence (IMC) measurements during the pinch task to assess neuromuscular connectivity. The findings confirm that motor impairment is inversely related to the diversity of muscle synergies, with fewer synergies and more stereotypical synergy structures observed post-stroke. The study further reveals that the nature of motor tasks significantly affects the number of identifiable muscle synergies, with less constrained tasks revealing a broader array of synergies. These findings highlight the importance of carefully selecting motor tasks in the context of clinical research and assessments to understand a patient’s motor impairment, thus aiding in developing tailored rehabilitation strategies.