Sensor-Derived Parameters from Standardized Walking Tasks Can Support the Identification of Patients with Parkinson’s Disease at Risk of Gait Deterioration

Francesca Boschi, Stefano Sapienza, Alzhraa A. Ibrahim, Magdalena Sonner, J Winkler, Bjoern M. Eskofier, Heiko Gaßner, Jochen KluckenBioengineering 2026 · DOI

TL;DR

This study demonstrates that combining wearable sensor data with clinical assessments can accurately predict which Parkinson's Disease (PD) patients are at risk of worsening gait impairments. By analyzing gait parameters like velocity and stride length, researchers developed a machine learning model that outperforms purely clinical assessments, offering a more objective and proactive approach to patient management.

What Problem Does This Paper Address?

Parkinson's Disease (PD) is characterized by progressive motor deficits, with gait impairments being a major cause of disability and reduced quality of life. Current clinical assessments, like the UPDRSIII gait score, are subjective, rater-dependent, and lack the granularity to precisely track subtle changes or predict future deterioration. This makes it challenging for clinicians to tailor interventions effectively and identify patients who need more intensive support to prevent falls and maintain mobility. The paper aims to bridge this gap by exploring how objective, sensor-derived gait data, when combined with clinical information, can provide a more accurate and prognostic assessment of gait progression in PD patients.

Methods

The researchers employed a two-pronged approach: statistical analysis and machine learning (ML).

Statistical Analysis: They compared clinical and sensor-derived gait parameters at baseline across three groups of patients: "Improvers" (those whose UPDRSIII gait score improved), "Stables" (no significant change), and "Deteriorators" (those whose score worsened). This was done both for the entire cohort and for subgroups stratified by their baseline UPDRSIII gait score.

Machine Learning: Three Random Forest (RF) models were trained to predict these gait progression trajectories:

  1. Clinical features only: Using standard clinical variables.
  2. Sensor-derived gait features only: Using parameters extracted from wearable sensors.
  3. Combined features: Integrating both clinical and sensor data.

The models were evaluated using Area Under the Curve (AUC) and accuracy, with feature selection performed using the ReliefF algorithm.

Datasets

The study utilized a dataset of 111 longitudinal visit pairs from 71 individuals diagnosed with Parkinson's Disease. These data were collected between May 2010 and June 2020 at the University Hospital Erlangen in Germany. The dataset included:

  • Clinical data: Age, disease duration, Hoehn and Yahr (H&Y) stage, gender, levodopa equivalent daily dose (LEDD), and UPDRSIII scores (total, gait, bradykinesia, postural stability, arise-from-chair).
  • Sensor-derived gait data: Collected during a 4x10m walking test using IMU sensors on shoes. Parameters included gait velocity, swing time, stance time, stride length, initial contact (IC) angle, terminal contact (TC) angle, maximum lateral excursion, maximum sensor lift, and their coefficients of variation (CV).

No new benchmark datasets were introduced; the study used existing patient data.

Key Insights and Ideas

  • Stratification is Key: Simply looking at all PD patients together masks important differences. Stratifying by baseline UPDRSIII gait score revealed clearer distinctions in gait parameters between progression groups.
  • Clinical Context Matters: While sensor data is valuable, it's most powerful when interpreted alongside clinical information. The best predictive model combined both, suggesting a synergistic effect.
  • Subtle Gait Changes Predict Deterioration: Even in patients with mild baseline gait impairments (UPDRSIII score 1), subtle differences in gait velocity, stride length, and stance time variability could distinguish between those who would improve, stay stable, or worsen.
  • Stables are Hard to Predict: The "Stable" group proved to be the most challenging to classify, indicating a need for more nuanced features or models to capture this intermediate trajectory.

Key Results

  • Combined Model Dominance: The machine learning model combining clinical and sensor-derived gait features achieved the highest performance, with an AUC of 0.82 ± 0.01 and an accuracy of 67.21 ± 1.92%.
  • Clinical Model Performance: The model using only clinical features performed well (AUC = 0.78 ± 0.01, accuracy = 60.72 ± 3.07%).
  • Sensor-Only Model Lagged: The model relying solely on sensor data showed the lowest performance (AUC = 0.62 ± 0.01, accuracy = 44.59 ± 1.86%).
  • Significant Baseline Differences: Improvers, compared to Stables and Deteriorators, showed significantly higher gait velocity and stride length, and lower stance time variability, particularly in the UPDRSIII gait score = 1 subgroup.
  • Important Features: For the best combined model, UPDRSIII gait, TC angle, stride length, and gait velocity were among the most important features for prediction.

Limitations and Future Work

  • Observational Study Limitations: The study design was observational, and the time interval between visits was not standardized, which could influence results. Detailed treatment regimens were also not available.
  • Sample Size for Severe Group: The subgroup with UPDRSIII gait score of 2 at baseline was small (18 individuals), potentially limiting the generalizability of findings for this group.
  • Confounding Factors: Factors like age, sex, comorbidities (e.g., diabetes, cognitive decline), and social support were not fully controlled for or explored, but are known to affect gait.
  • Circularity in ML: The UPDRSIII gait score was used both as a predictor and to generate labels, which could inflate model performance. However, the authors argue this was necessary to evaluate prognostic value and that comparison with the clinical model mitigates bias.
  • Future Work: The authors suggest further validation in larger cohorts and the development of decision-support systems to aid clinicians. They also highlight the need to incorporate other factors like sensory function, real-world gait assessment, and cognitive status in future research.

Summary updated

Abstract

<b>Background:</b> People with Parkinson's disease suffer from gait impairments. Clinical scales provide a limited and rater-dependent assessment of gait. Wearable sensors allow an objective characterization by capturing rhythm, pace, and signature patterns. This study investigated if sensor-derived gait parameters have prognostic value for short-term progression of gait impairments. <b>Methods:</b> A total of 111 longitudinal visit pairs were analyzed, where participants underwent clinical evaluation and a 4 × 10 m walking test instrumented with wearable sensors. Changes in the UPDRSIII gait score between baseline and follow-up were used to classify participants as Improvers, Stables, or Deteriorators. Baseline group differences were assessed statistically. Machine-learning classifiers were trained to predict group membership using clinical variables alone, sensor-derived gait features alone, or a combination of both. <b>Results:</b> Significant between-group differences emerged. In participants with UPDRSIII gait score = 1, Improvers showed higher median gait velocity (0.81 m/s) and stride length (0.80 m) than Stables (0.68 m/s; 0.70 m) and Deteriorators (0.59 m/s; 0.68 m), along with lower stance time variability (3.10% vs. 4.49% and 3.75%; all p<0.05). The combined sensor-based and clinical model showed the best performance (AUC 0.82). <b>Conclusions:</b> Integrating sensor-derived gait parameters with clinical score can support the identification of patients at risk of gait deterioration in the near future.
BibTeX
@inproceedings{boschi2026sensor,
  author = {Francesca Boschi and Stefano Sapienza and Alzhraa A. Ibrahim and Magdalena Sonner and J Winkler and Bjoern M. Eskofier and Heiko Gaßner and Jochen Klucken},
  title = {{Sensor-Derived Parameters from Standardized Walking Tasks Can Support the Identification of Patients with Parkinson’s Disease at Risk of Gait Deterioration}},
  booktitle = {Bioengineering},
  year = {2026},
  doi = {10.3390/bioengineering13020130},
}
Figure