Modeling Personalized Difficulty of Rehabilitation Exercises Using Causal Trees

Nathaniel Dennler, Zhonghao Shi, Uksang Yoo, Stefanos Nikolaidis, Maja J. Mataric

Abstract

Rehabilitation robots are often used in game-like interactions for rehabilitation to increase a person's motivation to complete rehabilitation exercises. By adjusting exercise difficulty for a specific user throughout the exercise interaction, robots can maximize both the user's rehabilitation outcomes and the their motivation throughout the exercise. Previous approaches have assumed exercises have generic difficulty values that apply to all users equally, however, we identified that stroke survivors have varied and unique perceptions of exercise difficulty. For example, some stroke survivors found reaching vertically more difficult than reaching farther but lower while others found reaching farther more challenging than reaching vertically. In this paper, we formulate a causal tree-based method to calculate exercise difficulty based on the user's performance. We find that this approach accurately models exercise difficulty and provides a readily interpretable model of why that exercise is difficult for both users and caretakers.

Introduction

Stroke rehabilitation requires long-term, repetitive exercise to regain mobility, but maintaining patient motivation is a significant hurdle. While rehabilitation robots can assist by gamifying exercises, most current systems employ "one-size-fits-all" difficulty settings, such as simply increasing reach distance. However, stroke survivors exhibit highly individualized impairments; one patient may struggle with vertical reaches while another finds lateral distance more challenging. This research addresses the need for Personalized Functional Difficulty, enabling robots to mathematically model and adapt to the unique "difficulty map" of each patient rather than relying on generic presets.

Methodology

The researchers formulated difficulty estimation as a causal inference problem, leveraging the Challenge Point Framework (CPF). This framework distinguishes between nominal difficulty (inherent task difficulty) and functional difficulty (difficulty relative to a specific individual's skill level).

The core innovation is the application of Causal Trees and Causal Forests. Typically used in econometrics to estimate heterogeneous treatment effects, these models are used here to determine the "effect" of a stroke on a person's reaching time across a 3D workspace. The model partitions the workspace into distinct regions (leaves of a tree), where each leaf represents a specific area where difficulty is statistically distinct. This decision-tree structure ensures the model is explainable, allowing clinicians to identify specific spatial thresholds that drive difficulty for a patient.

Experimental Setup

The study utilized two primary datasets:

  1. Neurotypical Baseline: 10 healthy participants performed 100 randomized reaching tasks to establish a "nominal" baseline for reach duration.
  2. Post-Stroke Dataset: 15 stroke survivors performed the same tasks, completing up to 300 reaches across multiple sessions.
The experimental setup included a Kinova JACO2 robot arm to move targets and a Lux AI QTRobot to provide social interaction and instructions.

The hardware setup involved a Kinova JACO2 assistive robot arm for target placement and a Lux AI QTRobot for social assistance, providing instructions and encouragement to the participants.

Results and Discussion

The Causal Forest model significantly outperformed standard machine learning baselines, including XGBoost, Random Forests, and Neural Networks. It achieved a Mean Squared Error of 0.300 $s^2$ and an explained variance of $r^2 = 0.826$, indicating it could explain over 82% of the variance in patient performance.

Causal Trees generate personalized difficulty maps, as seen for PID 22 and PID 26, identifying unique regions of high (yellow) and low (purple) additional reach time.

Beyond accuracy, the Causal Tree approach provides superior stability and interpretability. Unlike "black-box" models, it produces smooth, logical difficulty maps that identify specific "pockets" of difficulty in 3D space. This spatial granularity allows for more nuanced exercise prescriptions tailored to the individual's motor deficits.

Conclusion

This work demonstrates that jointly modeling patient and baseline data through causal inference is more robust to noise than traditional subtraction methods. While currently limited to "time to reach" as a single metric, future work aims to incorporate movement quality via computer vision and scale the approach to more complex activities of daily living. This interpretable, personalized modeling represents a significant step toward more effective and motivating robotic rehabilitation.

Summary updated

Abstract

Rehabilitation robots are often used in game-like interactions for rehabilitation to increase a person's motivation to complete rehabilitation exercises. By adjusting exercise difficulty for a specific user throughout the exercise interaction, robots can maximize both the user's rehabilitation outcomes and the their motivation throughout the exercise. Previous approaches have assumed exercises have generic difficulty values that apply to all users equally, however, we identified that stroke survivors have varied and unique perceptions of exercise difficulty. For example, some stroke survivors found reaching vertically more difficult than reaching farther but lower while others found reaching farther more challenging than reaching vertically. In this paper, we formulate a causal tree-based method to calculate exercise difficulty based on the user's performance. We find that this approach accurately models exercise difficulty and provides a readily interpretable model of why that exercise is difficult for both users and caretakers.
BibTeX
@inproceedings{dennler2025modeling,
  author = {Nathaniel Dennler and Zhonghao Shi and Uksang Yoo and Stefanos Nikolaidis and Maja J. Mataric},
  title = {{Modeling Personalized Difficulty of Rehabilitation Exercises Using Causal Trees}},
  booktitle = {ICORR},
  year = {2025},
  url = {https://openreview.net/forum?id=hajENMEFT5},
}
Figure