Head movement dynamics in dystonia: a multi-centre retrospective study using visual perceptive deep learning

Robert L. Peach, Maximilian U. Friedrich, Lara Fronemann, Muthuraman Muthuraman, Sebastian R. Schreglmann, Daniel Zeller, Christoph Schrader, Joachim K. Krauss, Alfons Schnitzler, Matthias Wittstock, Ann-Kristin Helmers, Steffen Paschen, Andrea A. Kühn, Inger Marie Skogseid, Wilhelm Eisner, Joerg Mueller, Cordula Matthies, Martin M. Reich, Jens Volkmann, Chi Wang IpNpj Digit. Medicine 2024 · DOI

TL;DR

This paper introduces a novel deep learning framework that analyzes head movement dynamics from standard clinical videos to objectively quantify dystonia severity and treatment effects. By moving beyond static head angle measurements, this approach captures intricate kinematic "pathosignatures" that are more sensitive to disease progression and therapeutic interventions like deep brain stimulation.

What Problem Does This Paper Address?

Dystonia, a neurological movement disorder, is currently assessed using clinical rating scales that, while practical, suffer from significant limitations. These scales often rely on simplified, static measurements (like head angle deviations) and exhibit substantial inter-rater variability. This lack of objective, granular, and reproducible measures hinders our understanding of dystonia's underlying neurobiology, limits the monitoring of treatment efficacy, and complicates the development of personalized therapies. The intricate spatiotemporal nature of dystonic movements, involving complex oscillations and overflow to other body regions, is largely missed by these traditional methods.

Methods

The core of this work is a "visual perceptive deep learning framework." This framework combines two key components:

  • Head Pose Estimation: It utilizes a pre-trained model (MediaPipe) to track facial landmarks and derive static head angles from attempted neutral head positions. This part is standard for many computer vision applications.
  • Movement State Classification: A custom-trained convolutional neural network (CNN) is developed to classify head movement states (e.g., "face forward," "tilt left") from video frames. This model was trained on a dataset of healthy controls and cervical dystonia patients.

By integrating the outputs of these two models, the researchers engineer a rich set of kinematic features. These go beyond simple angles to capture:

  • Correlations and Symmetries: How different movement axes interact and how symmetrical movements are.
  • Oscillatory Features: Amplitude and dominant frequency of head oscillations (tremor).
  • Entropy Measures: Complexity and irregularity of movement patterns.
  • Harmonic Strengths: Related to tremor frequencies.

These features are extracted from both static "null position" segments and the full video, allowing for the analysis of both static deviations and dynamic movement patterns.

Datasets

The study utilized retrospective, longitudinal video data from three multi-center cohort studies.

  • Cervical Dystonia Cohort: 86 patients with cervical dystonia.
  • Generalised Dystonia Cohort: 30 patients with generalised dystonia.
  • Healthy Controls: 22 individuals for comparison and model training.

The data comprised clinical videos of patients undergoing dystonia assessments, including those with and without Deep Brain Stimulation (DBS). The videos were re-scored by blinded experts using the Toronto Western Spasmodic Torticollis Rating Scale (TWSTRS) severity item and a TWSTRS-2 tremor subscore.

Key Insights and Ideas

  • Beyond Static Angles: The most significant insight is that dynamic kinematic features, not just static head angles, are crucial for accurately assessing dystonia severity and treatment response. These dynamic features capture aspects like tremor and movement overflow that are missed by traditional scales.
  • Kinematic Pathosignatures: The framework can identify distinct "pathosignatures" – unique combinations of kinematic features – that are highly sensitive to disease progression and the effects of neuromodulation (like DBS).
  • DBS Effects on Dynamics: Deep brain stimulation appears to restore movement regularity primarily on sub-second timescales, suggesting it impacts the phasic aspects of dystonic movements more than tonic ones.
  • Generalised vs. Cervical Dystonia Differentiation: Specific kinematic features, particularly those related to oscillations and movement correlations, can effectively differentiate between cervical and generalised dystonia subtypes.
  • Clinical Relevance: The identified kinematic features are not just statistically significant but also correlate with clinical assessments and are more strongly modulated in patients who show a better treatment response.

Key Results

  • Correlation with Clinical Scores: Computer vision-derived head angle excursions showed a strong correlation (r ≥ 0.66) with clinically assigned TWSTRS scores across all principal axes of motion.
  • DBS Impact: DBS treatment led to significant reductions in both static head angle deviations and dynamic kinematic features, with larger effect sizes for some kinematic features compared to static angles.
  • Predictive Power: A linear model combining static head angles and kinematic features provided more accurate predictions of overall dystonic severity (mean absolute error 4.79) than using head-angle deviations alone (mean absolute error 5.63).
  • Tremor Correlation: Oscillatory kinematic features correlated with clinically assigned tremor scores, suggesting they capture a key aspect of dystonia.
  • Subtype Differentiation: Seven kinematic features significantly differentiated between generalised and cervical dystonia patients, with generalised dystonia patients exhibiting larger values for these features.

Limitations and Future Work

  • Video Focus: The study was limited to videos of the upper body, neglecting dystonic phenomena in other regions. However, cervical dystonia is common, making head kinematics a valuable starting point.
  • Tremor Measurement: While the framework captured tremor amplitudes from full videos for ecological validity, this introduced natural variability compared to studies focusing solely on null positions.
  • DBS Parametrization: The study did not incorporate information on DBS parameters (lead location, stimulation fields), which could influence therapy response and potentially affect the observed differences between therapy states.
  • Heterogeneity: The retrospective integration of data from multiple centers might introduce heterogeneity not fully controlled for.
  • Monogenetic Dystonias: The possibility of undetected monogenetic dystonias in the sample could introduce unexplained variance.

Future work could involve incorporating DBS parametrization, expanding to whole-body movement analysis, and further validating the framework in prospective studies and diverse clinical settings. The authors suggest that this approach can bridge gaps between clinical and experimental neuroscience, strengthen translational research, and advance precision neurology.

Summary updated

BibTeX
@article{peach2024head,
  author = {Robert L. Peach and Maximilian U. Friedrich and Lara Fronemann and Muthuraman Muthuraman and Sebastian R. Schreglmann and Daniel Zeller and Christoph Schrader and Joachim K. Krauss and Alfons Schnitzler and Matthias Wittstock and Ann-Kristin Helmers and Steffen Paschen and Andrea A. Kühn and Inger Marie Skogseid and Wilhelm Eisner and Joerg Mueller and Cordula Matthies and Martin M. Reich and Jens Volkmann and Chi Wang Ip},
  title = {{Head movement dynamics in dystonia: a multi-centre retrospective study using visual perceptive deep learning}},
  journal = {Npj Digit. Medicine},
  year = {2024},
  doi = {10.1038/s41746-024-01140-6},
}
Figure