Fabian Kahl
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Fabian Kahl has a total of 44 co-authors. Here are the top co-authors by number of shared publications:
- Lara Marie Reimer
Institute for Digital Medicine, University Hospital Bonn, Bonn, Germany
3 shared publications - Stephan Jonas
University Hospital Bonn
3 shared publications - Leon Nissen
University Hospital Bonn
2 shared publications Alexander Hanke
University Hospital Bonn, University of Bonn
1 shared publicationAlexander Lange
German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany
1 shared publication- Anja Ophey
Medical Psychology | Neuropsychology and Gender Studies, Center for Neuropsychological Diagnostics and Intervention (CeNDI), Faculty of Medicine and University Hospital Cologne, University of Cologne Cologne Germany
1 shared publication
Latest Publications
- Top Cited PapersOpen Access
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (7) , 8470-8484
A Comparative Assessment of Accuracy in Video-Based Monocular Human Pose Estimation Frameworks
In human pose estimation, a comprehensive evaluation of state-of-the-art frameworks is necessary to advance both research and practical applications. This paper presents a thorough review of state-of-the-art 2D and 3D human pose estimation frameworks, analyzing 118 papers and four GitHub repositories, with a focus on frameworks made since 2019. The following frameworks are chosen based on predefined inclusion criteria: AlphaPose, Detectron2, MediaPipe, MeTRAbs, MHFormer, MMPose, MoveNet, OpenPifPaf, OpenPifPaf-vita, OpenPose, PoseFormerV2, rtmlib, StridedTransformer-Pose3D, ultralytics (YOLOv8), ViTPose, and YOLOv7. This paper evaluates these 16 frameworks on an existing, unpublished dataset consisting of exercise videos recorded with a monocular RGB camera and synchronized gold-standard motion capture data. The dataset includes videos of nine individuals performing eight exercises, recorded from two camera views with different planar angles. The analysis evaluates joint angle performance of the frameworks using weighted mean absolute error and weighted intraclass correlation coefficient as quantitative metrics. MeTRAbs emerged as the best overall framework, while AlphaPose, rtmlib, and YOLOv7 were the top 2D performers. The used code is available open source.1.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (7) , 8470-8484
Top Cited PapersOpen AccessA Comparative Assessment of Accuracy in Video-Based Monocular Human Pose Estimation Frameworks
In human pose estimation, a comprehensive evaluation of state-of-the-art frameworks is necessary to advance both research and practical applications. This paper presents a thorough review of state-of-the-art 2D and 3D human pose estimation frameworks, analyzing 118 papers and four GitHub repositories, with a focus on frameworks made since 2019. The following frameworks are chosen based on predefined inclusion criteria: AlphaPose, Detectron2, MediaPipe, MeTRAbs, MHFormer, MMPose, MoveNet, OpenPifPaf, OpenPifPaf-vita, OpenPose, PoseFormerV2, rtmlib, StridedTransformer-Pose3D, ultralytics (YOLOv8), ViTPose, and YOLOv7. This paper evaluates these 16 frameworks on an existing, unpublished dataset consisting of exercise videos recorded with a monocular RGB camera and synchronized gold-standard motion capture data. The dataset includes videos of nine individuals performing eight exercises, recorded from two camera views with different planar angles. The analysis evaluates joint angle performance of the frameworks using weighted mean absolute error and weighted intraclass correlation coefficient as quantitative metrics. MeTRAbs emerged as the best overall framework, while AlphaPose, rtmlib, and YOLOv7 were the top 2D performers. The used code is available open source.1.
- Open Access
- 15 May 2026
- Published by Wiley in Movement Disorders
- Vol. 41 (8) , 2189-2200
Wearable Movement‐Tracking for Prodromal Parkinson's Disease Detection: A Cross‐Country Validation Study
Background: Models trained on accelerometer data have been proposed for detecting prodromal Parkinson's disease (PD). However, uncertainties in diagnosis timing in the UK Biobank (UKBB) may affect generalizability to other cohorts. Objectives: The aim of the study was to evaluate the performance of previously published models for prodromal PD detection in other international cohorts. Methods: We applied the models to data from German and British cohorts of individuals with isolated or idiopathic rapid eye movement sleep behavior disorder and healthy controls. We compared hourly acceleration patterns and classification performance across cohorts. Results: The British cohort exhibited visually similar activity patterns to UK Biobank but weaker statistical differences and reduced model performance. The German cohort showed no significant group differences and lower performance. No pair of cohorts demonstrated statistical equivalence. Conclusions: Models trained on UK Biobank data may capture early clinical disease rather than universal prodromal markers. Prospective validation in well‐characterized cohorts is essential before clinical translation. © 2026 The Author(s). Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.
- 15 May 2026
- Published by Wiley in Movement Disorders
- Vol. 41 (8) , 2189-2200
Open AccessWearable Movement‐Tracking for Prodromal Parkinson's Disease Detection: A Cross‐Country Validation Study
Background: Models trained on accelerometer data have been proposed for detecting prodromal Parkinson's disease (PD). However, uncertainties in diagnosis timing in the UK Biobank (UKBB) may affect generalizability to other cohorts. Objectives: The aim of the study was to evaluate the performance of previously published models for prodromal PD detection in other international cohorts. Methods: We applied the models to data from German and British cohorts of individuals with isolated or idiopathic rapid eye movement sleep behavior disorder and healthy controls. We compared hourly acceleration patterns and classification performance across cohorts. Results: The British cohort exhibited visually similar activity patterns to UK Biobank but weaker statistical differences and reduced model performance. The German cohort showed no significant group differences and lower performance. No pair of cohorts demonstrated statistical equivalence. Conclusions: Models trained on UK Biobank data may capture early clinical disease rather than universal prodromal markers. Prospective validation in well‐characterized cohorts is essential before clinical translation. © 2026 The Author(s). Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.
- Open Access
- 1 February 2026
- Published by SAGE Publications in Digital Health
- Vol. 12
Tablet-based SDMT: Digital metric validation and within-task learning in MCI and dementia
Background: Scalable, process-sensitive tools are needed to address the gap between neuropathological onset and clinical recognition of cognitive decline. Tablet-based assessments can capture behavioral and temporal data beyond conventional test metrics, but require validation in target populations. Objective: To evaluate whether a tablet-based Symbol Digit Modalities Test (SDMT) and a response-time-based Performance Index capture clinically relevant cognitive performance in a memory-clinic population with mild cognitive impairment (MCI) or dementia. Methods: In a cross-sectional validation study, forty-one adults (mean age 73.7 years, 51% scoring below the CERAD cutoff for likely major cognitive impairment) completed a tablet-based SDMT alongside standard neuropsychological assessments (CERAD battery, MoCA). Digital response counts were compared with oral SDMT scores, the Performance Index was evaluated against clinical reference measures, and exploratory analyses examined within-task response-time patterns across repeated symbol exposures. Results: Digital and oral SDMT scores showed strong rank-order agreement ( ρ = 0 . 77 – 0 . 80 , p < . 001 ), with systematically lower absolute digital scores, indicating preserved inter-individual ranking but non-interchangeable modalities. The Performance Index and correct response count showed comparable associations with MoCA ( ρ = 0 . 59 and 0 . 62 ) and CERAD ( ρ = 0 . 49 for both; all p ≤ . 001 ); their convergence was attributable to near-ceiling response accuracy (mean 95.6%, excluding skipped items). Exploratory analyses revealed within-task response-time facilitation across repeated symbol exposures. Conclusion: The tablet-based SDMT preserves inter-individual ranking relative to the oral format, and the Performance Index generalizes to a memory-clinic population. Within-task response-time facilitation was observed even among participants with impaired declarative memory. Whether it reflects implicit or procedural learning rather than explicit mediation remains a hypothesis for controlled study.
- 1 February 2026
- Published by SAGE Publications in Digital Health
- Vol. 12
Open AccessTablet-based SDMT: Digital metric validation and within-task learning in MCI and dementia
Background: Scalable, process-sensitive tools are needed to address the gap between neuropathological onset and clinical recognition of cognitive decline. Tablet-based assessments can capture behavioral and temporal data beyond conventional test metrics, but require validation in target populations. Objective: To evaluate whether a tablet-based Symbol Digit Modalities Test (SDMT) and a response-time-based Performance Index capture clinically relevant cognitive performance in a memory-clinic population with mild cognitive impairment (MCI) or dementia. Methods: In a cross-sectional validation study, forty-one adults (mean age 73.7 years, 51% scoring below the CERAD cutoff for likely major cognitive impairment) completed a tablet-based SDMT alongside standard neuropsychological assessments (CERAD battery, MoCA). Digital response counts were compared with oral SDMT scores, the Performance Index was evaluated against clinical reference measures, and exploratory analyses examined within-task response-time patterns across repeated symbol exposures. Results: Digital and oral SDMT scores showed strong rank-order agreement ( ρ = 0 . 77 – 0 . 80 , p < . 001 ), with systematically lower absolute digital scores, indicating preserved inter-individual ranking but non-interchangeable modalities. The Performance Index and correct response count showed comparable associations with MoCA ( ρ = 0 . 59 and 0 . 62 ) and CERAD ( ρ = 0 . 49 for both; all p ≤ . 001 ); their convergence was attributable to near-ceiling response accuracy (mean 95.6%, excluding skipped items). Exploratory analyses revealed within-task response-time facilitation across repeated symbol exposures. Conclusion: The tablet-based SDMT preserves inter-individual ranking relative to the oral format, and the Performance Index generalizes to a memory-clinic population. Within-task response-time facilitation was observed even among participants with impaired declarative memory. Whether it reflects implicit or procedural learning rather than explicit mediation remains a hypothesis for controlled study.
- Open Access
- 27 January 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Leveraging machine learning for digital gait analysis in ataxia using sensor-free motion capture
Background: Gait disturbances are the clinical hallmark of ataxia. Their severity is assessed within a well-established clinical scale, which only allows coarse scoring and does not reflect the complexity of individual gait deterioration. We investigated whether sensor-free motion capture enables to replicate clinical scoring and improve the assessment of gait disturbances. Methods: The normal walking task during clinical assessment was videotaped in 91 ataxia patients and 28 healthy controls. A full-body pose estimation model (AlphaPose) was used to extract positions, distances, and angles over time while walking. The resulting time series were analyzed with four machine learning (ML) models, which were combinations of feature extraction (tsfresh, ROCKET) and prediction methods (XGBoost, Ridge). First, in a regression and classification approach, we trained the ML models on reconstructing the clinical score. Second, we used explainable AI (SHAP) to identify the most important time series. Third, we investigated time series features to study longitudinal changes. Results: Gait disturbances are assessed with high accuracy by ML models, slightly improving human rating (i) in the categorial prediction of the clinical score (F1-score best model: 63.99%, human: 60.57% F1-score), (ii) in the detection of subtle changes (pre-symptomatic patients, clinically rated unimpaired are differentiated from HC with a F1-score of 75.96%) and (iii) in the detection of longitudinal changes over time (Pearson’s correlation coefficient model: −0.626, p < 0.01; human: −0.060, not significant). Conclusions: ML-based analysis shows improved sensitivity in assessing gait disturbances in ataxia. Subtle and longitudinal changes can be captured within this study. These findings suggest that such approaches may hold promise as potential outcome parameters for early interventions, therapy monitoring, and home-based assessments.
- 27 January 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Open AccessLeveraging machine learning for digital gait analysis in ataxia using sensor-free motion capture
Background: Gait disturbances are the clinical hallmark of ataxia. Their severity is assessed within a well-established clinical scale, which only allows coarse scoring and does not reflect the complexity of individual gait deterioration. We investigated whether sensor-free motion capture enables to replicate clinical scoring and improve the assessment of gait disturbances. Methods: The normal walking task during clinical assessment was videotaped in 91 ataxia patients and 28 healthy controls. A full-body pose estimation model (AlphaPose) was used to extract positions, distances, and angles over time while walking. The resulting time series were analyzed with four machine learning (ML) models, which were combinations of feature extraction (tsfresh, ROCKET) and prediction methods (XGBoost, Ridge). First, in a regression and classification approach, we trained the ML models on reconstructing the clinical score. Second, we used explainable AI (SHAP) to identify the most important time series. Third, we investigated time series features to study longitudinal changes. Results: Gait disturbances are assessed with high accuracy by ML models, slightly improving human rating (i) in the categorial prediction of the clinical score (F1-score best model: 63.99%, human: 60.57% F1-score), (ii) in the detection of subtle changes (pre-symptomatic patients, clinically rated unimpaired are differentiated from HC with a F1-score of 75.96%) and (iii) in the detection of longitudinal changes over time (Pearson’s correlation coefficient model: −0.626, p < 0.01; human: −0.060, not significant). Conclusions: ML-based analysis shows improved sensitivity in assessing gait disturbances in ataxia. Subtle and longitudinal changes can be captured within this study. These findings suggest that such approaches may hold promise as potential outcome parameters for early interventions, therapy monitoring, and home-based assessments.
- Open Access
- 15 May 2025
- Published by IOS Press
- Vol. 327, 462-466
XGBOrdinal: An XGBoost Extension for Ordinal Data
We propose XGBOrdinal, an extension of XGBoost designed for ordinal classification problems commonly found in fields like medicine, where outcomes are often represented as scores, scales, stages, or grades. The proposed approach builds on the theoretical method introduced by Frank and Hall (2001) to transform an ordinal classification problem into a series of binary classification problems. Evaluated on multiple datasets, XGBOrdinal outperformed XGBClassifier and XGBRegressor, as well as existing ordinal methods. The implementation is fully compatible with GridSearchCV and RandomizedSearchCV, making it a scalable and efficient solution for handling ordinal data in machine learning pipelines. The used code is available open source (https://github.com/digital-medicine/XGBOrdinal).
- 15 May 2025
- Published by IOS Press
- Vol. 327, 462-466
Open AccessXGBOrdinal: An XGBoost Extension for Ordinal Data
We propose XGBOrdinal, an extension of XGBoost designed for ordinal classification problems commonly found in fields like medicine, where outcomes are often represented as scores, scales, stages, or grades. The proposed approach builds on the theoretical method introduced by Frank and Hall (2001) to transform an ordinal classification problem into a series of binary classification problems. Evaluated on multiple datasets, XGBOrdinal outperformed XGBClassifier and XGBRegressor, as well as existing ordinal methods. The implementation is fully compatible with GridSearchCV and RandomizedSearchCV, making it a scalable and efficient solution for handling ordinal data in machine learning pipelines. The used code is available open source (https://github.com/digital-medicine/XGBOrdinal).