Evidence for Phenotype-Driven Disparities in Freezing of Gait Detection and Approaches to Bias Mitigation

Timothy Odonga; Jeanne M. Powell; Claire Bae; Yumeng Tan; Christine D. Esper; Stewart A. Factor; J. Lucas McKay; Hyeokhyen Kwon

eess.SPJan 29, 2025

Evidence for Phenotype-Driven Disparities in Freezing of Gait Detection and Approaches to Bias Mitigation

Authors: Timothy Odonga, Jeanne M. Powell, Claire Bae, Yumeng Tan, Christine D. Esper, Stewart A. Factor, J. Lucas McKay, Hyeokhyen Kwon

Organizations: Emory University · Georgia Institute of Technology

Abstract

Freezing of gait (FOG) is a debilitating feature of Parkinson's disease (PD) and a common cause of injurious falls. Recent advances in wearable-based human activity recognition (HAR) have enabled FOG detection, but bias and fairness in these models remain understudied. Bias is defined as a systematic error leading to unequal outcomes, while fairness refers to consistent performance across subject groups. Biased models could systematically underserve patients with specific FOG phenotypes or demographics, widening care disparities. We evaluated the bias and fairness of state-of-the-art HAR models for FOG detection across FOG phenotypes and patient demographics using 4 public wearable datasets (145 subjects and 59.8 hours of recording) with multiple on-body sensor locations and temporal annotations on FOG episodes across various behavioral test protocols. Four mitigation approaches were assessed: conventional methods (threshold optimization and adversarial debiasing) and two transfer learning-based approaches (multi-site transfer and fine-tuning pre-trained models). Fairness was primarily quantified using the demographic parity ratio (DPR) and the equalized odds ratio (EOR). HAR models exhibited substantial bias (DPR & EOR < 0.8) across age, sex, disease duration, and critically, FOG phenotype. This phenotype-specific bias is particularly concerning as tremulous and akinetic FOG may require different clinical management. Conventional methods failed: threshold optimization (DPR=-0.126, EOR=+0.063) and adversarial debiasing (DPR=-0.008, EOR=-0.001). In contrast, multi-site transfer learning improved fairness (DPR=+0.037, p<0.01; EOR=+0.045, p<0.01) and performance (F1-score=+0.020, p<0.05), though neither metric reached 0.8. Transfer learning across diverse datasets may partially mitigate these disparities, helping wearable-based FOG monitoring benefit all individuals with PD.

Figures & tables

Figure 1: Model bias evaluation pipeline for FOG detection. In this work, we analyze the group fairness of the model based on protected attributes in FOG and PD, including FOG phenotype, sex, age, and disease duration.

Training Dataset	Test Dataset	Sensor locations
DeFOG + Daphnet	DCS FOG	Lower back
tDCS FOG + Daphnet	DeFOG	Lower back
Daphnet	De Souza	Ankle, Thigh, Shank
De Souza	Daphnet	Ankle, Thigh, Shank

Table 2: Training and testing dataset pairs used for leave-one-dataset-out cross-validation for transfer learning

Attribute	Daphnet	De Souza	tDCS FOG	DeFOG
Age (years)	66	69	69	69
Disease Duration (years)	12.5	7	9	13

Table 3: Median values for continuous subject attributes

Figure 2: F1-score, Demographic Parity Ratio (DPR), and Equalized Odds Ratio (EOR) for each model across the multi-site datasets prior to applying bias mitigation approaches. Dashed horizontal lines represent the fairness threshold (0.8 for DPR and EOR). Each dot represents one iteration (a single seeded run of k-fold cross-validation) averaged across its folds.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1: Confusion matrix between IMU-derived and clinician-annotated FOG phenotypes, pooled across the DeFOG and tDCS FOG (502 recordings, 2,877,915 samples).

Dataset	Medication state	Disease severity	Cognitive status	FOG severity
Daphnet	$\checkmark$ (On or Off medication state provided)	$\checkmark$ (Hoehn and Yahr score)	—	—
De Souza	$\checkmark$ (Levodopa dosage in mg)	$\checkmark$ (Hoehn and Yahr score, total UPDRS-II and UPDRS-III scores)	$\checkmark$ (MoCA score)	$\checkmark$ (N-FOG-Q)
DeFOG	$\checkmark$ (On and Off medication state provided)	$\checkmark$ (total UPDRS-III in on and off states)	—	$\checkmark$ (N-FOG-Q)
tDCS FOG	$\checkmark$ (On and Off medication state provided)	$\checkmark$ (total UPDRS-III in on and off states)	—	$\checkmark$ (N-FOG-Q)

Appendix

Table S7: Clinical metadata availability differs across datasets, which constrains the attributes available for subgroup fairness evaluation.

\checkmark

indicates that the attribute is reported at the subject level in the dataset’s accompanying clinical documentation; — indicates that the attribute is not reported, and no inference should be drawn about whether it was assessed during data collection. Instruments and scales are given in parentheses.

				Demographic Parity Ratio (DPR)				Equalized Odds Ratio (EOR)			Equality of Opportunity Difference (EOD)
Dataset	Model	Pretrain dataset	F1-score	Sex	Age	Disease duration	FOG phenotype	Sex	Age	Disease duration	FOG phenotype
DeFOG	A2	tDCS FOG + Daphnet	0.582 $\pm$ 0.032	0.445 $\pm$ 0.152	0.445 $\pm$ 0.236	0.464 $\pm$ 0.182	0.088 $\pm$ 0.049	0.251 $\pm$ 0.142	0.226 $\pm$ 0.195	0.285 $\pm$ 0.104	0.200 $\pm$ 0.107
DeFOG	A3	tDCS FOG + Daphnet	0.574 $\pm$ 0.024	0.330 $\pm$ 0.051	0.430 $\pm$ 0.116	0.390 $\pm$ 0.069	0.073 $\pm$ 0.025	0.148 $\pm$ 0.092	0.220 $\pm$ 0.172	0.203 $\pm$ 0.047	0.172 $\pm$ 0.067
tDCS FOG	A2	DeFOG + Daphnet	0.411 $\pm$ 0.005	0.053 $\pm$ 0.135	0.044 $\pm$ 0.085	0.042 $\pm$ 0.116	0.061 $\pm$ 0.168	0.000 $\pm$ 0.000	0.000 $\pm$ 0.000	0.000 $\pm$ 0.000	0.002 $\pm$ 0.005
tDCS FOG	A3	DeFOG + Daphnet	0.450 $\pm$ 0.059	0.249 $\pm$ 0.337	0.249 $\pm$ 0.350	0.218 $\pm$ 0.313	0.202 $\pm$ 0.259	0.126 $\pm$ 0.157	0.182 $\pm$ 0.288	0.151 $\pm$ 0.225	0.051 $\pm$ 0.090
Daphnet	A2	De Souza	0.402 $\pm$ 0.153	0.530 $\pm$ 0.525	0.460 $\pm$ 0.520	0.418 $\pm$ 0.565	0.340 $\pm$ 0.525	0.389 $\pm$ 0.480	0.290 $\pm$ 0.465	0.304 $\pm$ 0.487	0.321 $\pm$ 0.530

Appendix

Table S8: Model performance and fairness on performing zero-shot transfer on A2 (DeepConvLSTM) and A3 (Masked Transformer). Mean

\pm

95% confidence interval over 5 runs of leave-one-dataset-out cross-validation (5 random seeds). Each run’s value is the mean over its 5 folds. In folds where the model detected no FOG episodes, the ratios (DPR, EOR) are undefined and are counted as 0.

Subject	Recordings (n)	Akinetic (s)	Tremulous (s)	Total FOG (s)	Total recording (s)	Tremulous (proportion of total FOG)	Proportion of Tremulous over total recording	Proportion of Akinetic over total recording
Daphnet (n = 10)
S01	2	1.0	103.6	104.6	1,900.0	0.990	0.055	0.001
S02	2	0.0	181.4	181.4	1,415.0	1.000	0.128	0.000
S03	3	1.5	285.0	286.6	2,010.1	0.995	0.142	0.001
S04	1	0.0	0.0	0.0	2,070.0	—	0.000	0.000
S05	2	62.9	411.6	474.5	2,090.1	0.867	0.197	0.030

Appendix

Table S9: FOG type distribution per subject. Durations are derived from the annotated samples at each dataset’s native sampling rate (Daphnet 64 Hz, DeFOG 100 Hz, De Souza and tDCS FOG 128 Hz). Tremulous (proportion of total FOG) is undefined (—) for subjects with no annotated freezing.

Algorithmic Fairness

Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.

See all 419 papers

Sep 8, 2026

Supervised Cross-Modal Feature Alignment for Zero-Wearable Freezing of Gait Detection in Parkinsonism

Objective assessment of Freezing of Gait (FoG) in Parkinson's disease (PD) relies predominantly on wearable Inertial Measurement Units (IMUs). While IMUs provide optimal kinematic precision, mandatory sensor attachment restricts continuous clinical deployment. Conversely, unobtrusive vision-based alternatives suffer substantial classification errors during turning-in-place tasks, where geometric self-occlusion degrades deterministic skeletal coordinates and obscures the high-frequency precursors required for FoG detection. To resolve these physical observation limits, we propose a supervised cross-modal subspace distillation framework. During optimisation, pre-trained kinematic data from IMU sensors and contextual clinical metadata act as oracles to guide a deployable visual architecture. By incorporating joint velocity and acceleration derivatives, utilising a confidence-based gating mechanism, the visual model mitigates some of the tracking errors during occlusion events. Empirical evaluations confirm this latent alignment transfers the predictive fidelity of hardware sensors directly into the visual representation, yielding

85.5\%

accuracy, and

82.4\%

balanced accuracy. All the while maintaining a vision only model at inference.

Aryan Singh, Chandan Biswas

NeuroAI Fusion Labs Kolkata, India

Cross-Modal Knowledge Distillation Cross-Modal Alignment Gait Analysis

Ask PDF

Sep 9, 2026

Freezing of Gait Prediction Under Spatial Occlusion: An IMU-Supervised Cross-Modal Distillation Approach

Parkinson's disease is a progressive neurodegenerative disorder characterised by gradual deterioration of movement control. Automated freezing-of-gait (FOG) detection supports the objective assessment of gait-related motor impairment. Two common approaches are used for FOG prediction: (i) analysing video recordings of the patient's movements and (ii) analysing data collected using inertial measurement unit (IMU) wearable sensors attached to the patient's lower limbs. Video-based approaches may suffer detection errors during continuous turning-in-place tasks because the lower limbs undergo substantial geometric self-occlusion, degrading pose-estimation accuracy. IMU-based approaches are generally less affected by visual occlusion; however, they are difficult to deploy outside clinical or laboratory settings, as the sensors must be attached securely and remain in place throughout the assessment. Motivated by this, we propose a cross-modal subspace distillation framework to mitigate the limitations of unimodal FOG detection by combining IMU accuracy with video-based practicality. We extract invariant latent topologies from a pre-trained kinematic oracle to structurally supervise a non-encoded visual architecture during training. To resolve periods of severe spatial occlusion, a dual-stream visual model probabilistically fuses skeletal graph nodes and continuous spatial pixels, dynamically shifting reliance to uninterrupted pixel boundaries as joint tracking confidence drops. Evaluated against a public, multi-modal sequence dataset of Parkinson's individuals executing continuous

360^\circ

turns, empirical results demonstrate that applying sensory boundary topologies strictly mitigates tracking evaluation entropy. Our constrained optimisation confirms that highly precise FOG prediction bounds can be achieved over zero-wearable inference environments.

Chandan Biswas, Aryan Singh, Anabik Pal

NeuroAI Fusion Labs Kolkata, India · Indian Institute of Science Education and Research Berhampur, Odisha

Cross-Modal Knowledge Distillation Gait Analysis Parkinson's Disease Multimodal Sensor Fusion

Ask PDF

Aug 13, 2026

Towards Context-Aware Clinical Motion Understanding in Daily Living at Home: Freezing of Gait Detection with Egocentric Vision

Understanding motion in daily living requires context beyond kinematics, because similar inertial patterns during activities of daily living (ADLs) can reflect intentional stopping, object interaction, or pathological movement impairment. Egocentric vision provides task-related context that may help disambiguate these cases. We investigate this challenge through freezing of gait (FOG) detection in Parkinson's disease (PD), a symptom strongly influenced by contextual factors during ADLs. Using synchronized egocentric video, wearable IMUs, and expert-annotated FOG labels collected from 13 PD participants in their homes, we evaluate frozen representations from pretrained ego-video and time-series foundation models, alongside an IMU-based TCN trained from scratch, under leave-one-subject-out evaluation. The IMU-based TCN achieved the strongest event-detection performance, reaching 42.3 F1 and 83.0 AUROC, compared with 32.6 F1 and 77.2 AUROC for V-JEPA2 ego-video features. Although ego-video alone did not outperform IMU-based sensing, it showed above-chance discrimination, and qualitative analyses suggest that egocentric vision may capture FOG-relevant information independent of IMUs. Together, these results support the use of pretrained ego-video representations to add contextual information to wearable-sensor-based clinical motion understanding in daily living.

Vayalet Stefanova, Diwas Lamsal, Margot Genbrugge +5

Department of Electrical Engineering (ESAT), KU Leuven, Leuven, Belgium · Neurorehabilitation Research Group (eNRGy), Department of Rehabilitation Sciences, KU Leuven, Leuven, Belgium · Institute of Interdisciplinary Exercise Science and Sports Medicine, Medical School Hamburg, Hamburg, Germany +2

Gait Analysis Parkinson's Disease Egocentric Vision Multimodal Sensor Fusion

Ask PDF

Abstract

Algorithmic Fairness

Explore similar work

Supervised Cross-Modal Feature Alignment for Zero-Wearable Freezing of Gait Detection in Parkinsonism

Freezing of Gait Prediction Under Spatial Occlusion: An IMU-Supervised Cross-Modal Distillation Approach

Towards Context-Aware Clinical Motion Understanding in Daily Living at Home: Freezing of Gait Detection with Egocentric Vision