🤖 machine learning

Adaptive Group-Based Counterfactual Explanations for Time-Series Rehabilitation Data

This paper proposes a two-stage framework utilizing Learnable Gate methods to generate adaptive, group-based counterfactual explanations for multivariate time-series rehabilitation data, effectively aligning interpretability with clinical reasoning by producing sparse, biomechanically coherent guidance at the muscle-group level rather than the individual sensor channel level.

Original authors: Emmanuel C. Chukwu, Rianne M. Schouten, Monique Tabak, Mykola Pechenizkiy

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Emmanuel C. Chukwu, Rianne M. Schouten, Monique Tabak, Mykola Pechenizkiy

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to fix a broken machine, but instead of a mechanic, you have a super-smart computer that tells you why it broke. The problem is, the computer's explanation is like a list of 48 tiny screws, saying, "Tighten screw #3, loosen screw #12, and wiggle screw #47." While technically correct, this is confusing and overwhelming for a human trying to fix the machine. You don't think in terms of individual screws; you think in terms of "the engine block" or "the transmission."

This paper tackles that exact problem, but instead of a machine, the "machine" is a human body moving during rehabilitation exercises (like squats or walking), and the "screws" are data from sensors strapped to the body.

Here is a simple breakdown of what the researchers did:

The Problem: Too Many Tiny Details

In physical therapy, patients wear special sensors (called IMUs) on their legs. These sensors record movement from 8 different spots, and each spot sends 6 different signals (like speed and rotation). That's 48 separate data streams all at once.

When a computer tries to explain why a patient's squat looks "wrong," standard methods look at all 48 streams individually.

The Result: The computer says, "Change the data on the left ankle sensor at second 3, and the right knee sensor at second 5."
The Issue: This is like telling a chef to "add a pinch of salt to the left side of the pot and a drop of pepper to the right side." It's scattered, confusing, and doesn't make sense to a doctor or patient who thinks in terms of whole body parts (e.g., "Your right hamstring is too tight").

The Solution: Grouping the Sensors

The researchers proposed a new way to think about the data. Instead of looking at 48 individual streams, they grouped them into 16 logical "buckets" based on anatomy and movement type.

Bucket 1: All the "speed" sensors on the right thigh.
Bucket 2: All the "rotation" sensors on the left shin.
And so on.

They call this Group-Based Counterfactual Explanations. Instead of asking, "Which specific screw is wrong?" they ask, "Which part of the machine needs adjustment?"

How It Works: The "Smart Gate"

The paper introduces a clever two-step system to find the best fix:

The Detective (Shapley Ranking): First, the computer acts like a detective. It looks at the bad movement and uses a mathematical tool (Shapley values) to guess which "buckets" (groups of sensors) are the most likely culprits. It narrows the list down from 16 groups to the top suspects.
The Gatekeeper (Learnable Gates): This is the paper's main innovation. Imagine a set of gates controlling water flow to different parts of a garden. The computer learns to open the gates for the groups that actually need changing and close the gates for the ones that are fine.
- It doesn't just guess; it learns which gates to close while it tries to fix the movement.
- This ensures the final explanation only mentions the specific body parts that need help, ignoring the rest.

The Results: A Clearer Prescription

The researchers tested this on a dataset called KneE-PAD, which contains data from patients doing squats, knee extensions, and walking. They compared their new "Group" method against the old "Individual Channel" method.

Old Method (M-CELS): Like a messy list of 23 different screws to tweak. It worked, but the explanation was cluttered.
New Method (Learnable Gate): Like a clear instruction: "Focus on your right hamstring and left ankle."
- Fewer Changes: The new method changed significantly fewer "buckets" (about 27% fewer groups) while still successfully fixing the computer's prediction.
- Better Logic: The changes made sense biomechanically. For example, when fixing a squat, the system correctly identified that the patient needed to adjust their weight transfer, rather than randomly tweaking sensors on unrelated body parts.
- Speed: It was just as fast (or faster) at generating these explanations.

The Bottom Line

The paper claims that by organizing sensor data into logical groups (like muscle groups) and using a "smart gate" to select only the necessary ones, they can generate clearer, more concise, and more useful explanations for rehabilitation.

Instead of a doctor getting a confusing list of 20 random sensor adjustments, they get a focused, anatomical explanation (e.g., "Reduce activity in the right hamstring during the squat") that aligns with how humans actually understand the body. This makes the AI's advice much easier to trust and act upon.

Technical Summary: Adaptive Group-Based Counterfactual Explanations for Time-Series Rehabilitation Data

Problem Statement
In clinical rehabilitation, particularly when analyzing movement via multi-sensor Inertial Measurement Units (IMUs), machine learning models often lack interpretability for domain experts. Clinicians reason in terms of semantic feature groups (e.g., muscle groups, joint segments, and synergies) rather than individual sensor channels. Standard Counterfactual Explanation (CE) methods for time-series data typically operate at the channel level. This approach leads to four critical issues in high-dimensional IMU settings (e.g., 48 channels from 8 sensors):

Combinatorial Explosion: The search space scales with channels times time steps, often resulting in dense, scattered perturbations.
Biomechanical Incoherence: Channel-wise edits can violate physical constraints, producing implausible sequences (e.g., isolated axis changes ignoring joint kinematics).
Temporal Fragmentation: Dispersed edits create "temporal confetti" rather than coherent intervention windows.
Lack of Hierarchy: Explanations fail to map system-level concepts (muscle groups) down to specific sensors.

Existing methods rarely encode sensor grouping or segment-level abstractions, limiting their clinical utility.

Methodology
The authors propose Adaptive-MO, a two-stage framework for generating group-based counterfactuals in high-dimensional time-series data. The method integrates domain-informed grouping with adaptive multi-objective optimization.

Domain-Informed Grouping:
The framework defines sensor groups based on anatomical locations and signal modalities. Specifically, it adopts a modality-based grouping strategy where accelerometer and gyroscope signals are treated as separate groups for each muscle (yielding 16 groups from 48 channels). This distinguishes linear acceleration from rotational kinematics while preserving anatomical relationships.
Shapley-Adaptive (SA) Group Ranking:
To reduce search complexity, the method first estimates channel-level Shapley values (via GradientSHAP) and aggregates them to group-level importance using the maximum absolute value. The top- $k$ most influential groups are selected to restrict the optimization search space. However, experiments showed that static SA ranking preserves validity but fails to enforce sufficient group-level sparsity.
Learnable Gate (LG) Mechanism:
To address the sparsity limitation, the authors introduce Learnable Gates. These are trainable per-group relevance parameters ( $\gamma_g$ ) jointly optimized with the perturbation masks.
- Parameterization: Counterfactuals are generated by blending the input with a nearest unlike neighbor (guide) using a continuous mask modulated by the gates.
- Regularization: The gates are regularized with a combined $\ell_1$ penalty (to drive unused gates to zero) and a binarization penalty (to push gates toward 0 or 1). This allows the model to dynamically prune irrelevant groups during optimization, resulting in a sparse, binary subset of groups.
Multi-Objective Optimization:
The framework optimizes a total loss function comprising four terms:
- Validity: Ensuring the modified input predicts the target class.
- Sparsity: Minimizing the magnitude of the mask (channel-level) and the number of active groups.
- Temporal Smoothness: Enforcing smoothness via a higher-order total variation norm.
- Gate Regularization: Encouraging the selection of a minimal set of groups.
  An adaptive weighting scheme shifts emphasis from validity to sparsity once a confident target prediction is achieved.

Key Contributions

Adaptive-MO Framework: A novel framework combining Shapley-based group ranking with dynamic sensor group selection for structured CEs in high-dimensional time series.
Learnable Gate (LG) Mechanism: A method using trainable gates to automatically select a sparse, interpretable, and clinically meaningful set of sensor groups, optimized jointly with perturbation masks.
Empirical Validation: Systematic evaluation on the KneE-PAD dataset (31 patients, 8 IMUs, 3 exercises: squat, knee extension, gait) demonstrating that group-structured counterfactuals yield concise, muscle-level corrective guidance.

Results
Experiments compared the proposed LG variants against the channel-level M-CELS baseline (a state-of-the-art method for multivariate time series).

Validity and Success Rate: The best-performing variant, LG-SHAP pruned, achieved a counterfactual success rate of 94.7% (vs. 90.0% for M-CELS) with high target confidence (80.8%).
Sparsity: LG methods significantly reduced explanation complexity. LG-SHAP pruned modified an average of 8.2 modality groups (a 27% reduction compared to M-CELS's 11.2) and 18.0 channels (a 22% reduction). The LG-fixed prune variant achieved even higher sparsity (5.2 groups) with a modest trade-off in validity.
Plausibility and Smoothness: Group-structured perturbations maintained temporal smoothness comparable to M-CELS, indicating no introduction of artificial discontinuities.
Exercise-Specific Patterns: LG produced exercise-specific, selective activation patterns aligned with biomechanics (e.g., targeting quadriceps/hamstring co-contraction for knee extension), whereas M-CELS exhibited diffuse, broad activation across groups.
Efficiency: Generation times for LG variants (8.0–9.7s) were comparable to or faster than M-CELS (10.0s).

Significance and Claims
The paper claims that the proposed framework enhances interpretability without sacrificing counterfactual quality. By aligning the explanation structure with clinical reasoning (muscle groups and joint segments), the method produces actionable explanations for rehabilitation movement analysis.

Clinical Alignment: The group-structured counterfactuals yield concise, muscle-level corrective guidance (e.g., "reduce right hamstring activity") that matches the hierarchical reasoning of physical therapists.
Structural Improvement: The work demonstrates that dynamic feature group-level optimization yields more interpretable and reliable explanations than channel-level methods in IMU-based rehabilitation.
Design Choice: The paper highlights sensor grouping as a theoretically grounded and practically important design choice for time-series counterfactuals.

Limitations and Future Work
The authors note that experiments were limited to the KneE-PAD dataset with a specific number of patient recordings. The sensor grouping is hand-crafted for lower-limb IMUs. Future work is suggested to include automatic group discovery, multimodal fusion (e.g., with sEMG), and real-time deployment in telerehabilitation, as well as user studies with clinicians to assess trust and actionability.