acceptodds
Under review as a conference paper at ICLR 2027

FoGConcept: Identifying Interpretable Movement Concepts for Freezing-of-Gait Detection from Multi-Scale IMU Signals

Abstract

Freezing of gait (FoG) is a disabling motor symptom of Parkinson's disease, characterized by an episodic inability to initiate or sustain effective stepping. Current inertial measurement unit (IMU)-based models can detect FoG accurately, but they fail to explain their predictions in terms of human-understandable movement concepts. A major contribution of this work is to identify nine human-understandable movement concepts encoded by a multi-scale FoG model, including gait initiation, rhythmic bilateral gait, trembling in place, shuffling steps, and pivot turns. These concepts recur across subjects, are functionally linked to the model's FoG decisions, and exhibit emergent specialization across short and long temporal contexts. Accordingly, we propose FoGConcept, a mechanism-based interpretability framework for learning movement concepts from multi-scale IMU signals. Domain experts independently recognize 88.9% of the identified concepts, while the proposed model achieves an F1 score of 94.2% for FoG detection. Our findings show that accurate FoG detection can be understood through human-recognizable movement patterns, providing new insight into what a detection model learns to distinguish FoG from non-FoG movement.

Primary area:
interpretability and explainable AI
Keywords:
Freezing-of-Gait, AI for Health, Sparse Autoencoders, Interpretability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_eJ3JOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript aims for the interpretability for freezing of gait (FoG) detection. It involves a mult-scale Transformer for FoG detection with wearable signals. An SAE is used to analyze the model's internal states, and nine human-understandable movement concepts are identified with observations such as gait initiation, rhythmic bilateral gait and tremblinginplace. Clinical experts were able to recognize 88.9% of the identified concepts, and an F1 score of 94.2% for FoG detection was achieved.

Critical strengths
  • This is a reasonble framework for analyzing FoG events. Some findings demonstrate some common FoG gaiting patterns during the clinical observation, which indicate the modeling itself learns critical patterns for prediction.
  • Clinicians are involved into this study to very the findings. And the main results seem to be consistent statistically.
Critical weaknesses
  • My main concern is regarding the findings, which can be trivial. The discovered concepts do not really reveal something new. They are established FoG gait patterns that have been recognized by the domain already.
  • The clinical significance is not enough. The work mainly discover the concepts relevant to the anchor points. Yet, this is post-hoc. The explaniability in this study may not be able to uitlize for a deeper step: pre-alarming of FoG event.
Questions for authors

Besides the weaknesses part, I have the following questions/comments.

  • As the authors mentioned, 'representative, clearly observable features' are dopted for this study. I thought these are more relevant to some common knowledge/concepts to the FoG gait understanding. Yet, the remaining hints may be more critical and insightful. I was wondering any observations from this part that provides some new knowledge. I understand the authors mention this as an limitation of the current work. But this can be important to reveal real new insights to the community.
  • The current participants involved in this study seems to be not many. I was wondering about the representativeness of the population.
  • Some fundamental causes of the FoG pattern might not be explaniable if with motion sensors only. My understanding is that the FoG happens with some particular scenarios, like the motivation to design the TUG test protocol (with the visual hints). It can be relevant to the environment while the proposed framework is not able to discover.
  • There are various architectures for human motion understanding including the FoG recognition. Many different modalities are invovled. This proposed framework may not be able to used in a broader range of methods.
5 comments

Response to Weakness 1 & Q1: Our goal is not to invent new clinical gait categories. In fact, requiring the discovered concepts to be unfamiliar or clinically novel would work against the purpose of interpretability as unfamiliar or newly invented concepts would make the model harder for clinicians/patients to interpret. The objective is to determine whether a black-box FoG detector internally represents movement structure that humans can recognize and clinically interpret. The nine reported concepts are therefore meaningful precisely because they emerge from the model without being predefined, labeled, or supervised as concepts, yet align with recognizable gait phenomena such as gait initiation, shuffling, trembling in place, and turning. Therefore, the correspondence with established FoG phenomena is itself evidence supporting the explanation. It shows that these clinically recognizable movement patterns are encoded as functionally relevant internal features that influence the detector’s FoG prediction. Our contribution is to propose a reproducible pipeline that uncovers and validates these concepts inside FoG detection models, not to propose a new clinical taxonomy of FoG.

Response to Weakness 2: The current paper is an interpretability study, and we do not view this as limiting its clinical relevance. Our goal is to identify which movement patterns a FoG detector relies on and make a previously black-box model understandable in clinically meaningful terms. The explanations themselves can provide clinically useful information. FoG is heterogeneous: patients may exhibit different manifestations, such as trembling in place or shuffling, and may freeze during different movement contexts such as gait initiation, ongoing walking, or turning. Identifying which concepts characterize an individual patient’s episodes could therefore support a patient-specific FoG profile, helping clinicians and caregivers understand how that patient tends to freeze and potentially informing more personalized management or intervention strategies. Our work provides the interpretable information needed for such patient-specific profiling and future personalized intervention. The same framework can also be extended to pre-FoG models for early-warning applications, and the underlying FoG detection model itself can be used for pre-FoG detection.

Response to Q2: We have now additionally repeated the complete FoGConcept analysis on a new dataset, FoG-STAR. All nine primary movement-concepts recur on the new dataset as well (full analysis in response to reviewer 5XZp, Q1). Collecting and annotating FoG data is challenging because FoG is intermittent, must be observed in people with Parkinson's disease, and requires extended expert reviews to obtain reliable labels. The study now evaluates 45 participants across two independent PD cohorts, which we believe provides a substantial basis for this type of clinically annotated FoG interpretability analysis and addresses the concern about reliance on a single cohort.

Response to Q3: FoG can also be influenced by environmental factors that are not directly observable from IMU signals. Our contribution is a reproducible interpretability pipeline for wearable FoG detection models, aimed at explaining how their learned sensor representations encode human-understandable movement patterns. In this setting, FoGConcept identifies which gait patterns, such as gait initiation, turning, shuffling, and trembling in place, are represented by the model and functionally contribute to its FoG decisions. This movement-level interpretation is important because IMU-based FoG detectors make their predictions from movement signals, yet their learned representations are difficult to understand. Our results show that these representations can be decomposed into recognizable movement concepts that recur across subjects and datasets and are functionally linked to model predictions. Environmental or cognitive mechanisms require models that additionally observe those sources of information, such as visual context or other multimodal signals. Extending FoGConcept to such multimodal models could reveal how movement concepts interact with environmental or contextual factors and is an important direction for future work. We will clarify in the revision that our present contribution explains the learned movement mechanisms of wearable FoG detection, and is not to provide a complete causal account of why FoG occurs.

Response to Q4: We added several experiments to test whether FoGConcept depends on the specific model configuration used in the primary study. In summary, our FoGConcept pipeline can be applied to different datasets, input windows and model backbones, and result in a similar concept set as the primary study (please see full analysis in response to reviewer 5XPz Q1 and Q2). These results show that FoGConcept is not tied to one model. At the framework level, FoGConcept operates on the learned representation of a FoG detector, regardless of its architectures. It decomposes that representation into sparse features, identifies features that functionally affect the detector through steering, and characterizes the movement patterns associated with those features. The procedure therefore does not rely on a particular feature index, model, or a fixed concept dictionary, and can be applied to any FoG detection methods. For different sensing modalities or placements, we do not expect identical concepts because the observable information itself changes. For example, bilateral ankle IMUs directly expose left- and right-foot motion, whereas wrist sensors cannot represent foot motions in the same way. Our FoG-STAR lower-back analysis confirms this boundary, that the pipeline remains applicable, but the resulting concept vocabulary reflects the information available at that placement. We therefore frame FoGConcept as a reproducible interpretability framework for learned wearable-sensor representations, with broader evaluation across architectures and multimodal sensing as an important future direction.

Dear Reviewer eJ3J,

We would like to thank you for taking the time to review our paper and for your insightful comments. We have addressed all the comments and suggestions you made. In particular, we clarified that our contribution is not to propose new clinical gait categories, but to provide a reproducible interpretability pipeline that reveals clinically recognizable movement concepts learned by a black-box FoG detector without concept-level supervision. We also substantially expanded the evaluation in response to your concerns about representative population and generalizability. In addition, we clarified the clinical role of the framework, the distinction between movement-level interpretation and causal explanation of FoG, and the potential extension to multimodal and pre-FoG settings. As we are approaching the end of the discussion period, please kindly let us know if you have any additional concerns. We truly appreciate this opportunity to improve our work and would be most grateful for any feedback you could give to us. If you do not have further questions, we are curious if you could consider raising our score. Thank you very much!

Reviewer_LYsPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FoGConcept suggests an effective method of interpretable movement features with SOTA results. It identifies 9 movement concepts using sparse autoencoders and achieves the best results I know (94.2% F1 score). The paper is clear and easy to follow.

Critical strengths

The interpretability and the expert-method mechanism. The method could generalize to other domains. State of the art results. Well written, well structured, and easy to follow.

Critical weaknesses

I like this paper, but its main weakness is that it is evaluated on a single small dataset. External validation, or at least many more subjects, could make it much more significant.

Questions for authors

Is there any reason you evaluate it only on this small dataset?

2 comments

Response to Weakness 1 & Q1: Please see our response to Reviewer 5XZp, Weakness 1. We have now repeated the complete FoGConcept analysis on a new dataset, FoG-STAR. It contains 22 participants with Parkinson’s disease and collects data using different IMU hardware, sampling rate, sites, and protocol. All our nine concepts recur in this new dataset. These results extend the analysis beyond the original cohort and show that the discovered movement concepts generalize across datasets. The original VCU dataset was selected because it provides the combination required for this study with 23 participants with Parkinson’s disease, 528 expert-annotated FoG episodes, bilateral ankle IMUs, and synchronized video for movement interpretation. Obtaining this combination is difficult in FoG research because episodes are intermittent and require clinical annotation. Thus, while the original cohort is limited in size, it provides substantial event-level supervision for concept discovery. With the added FoG-STAR experiment, the study now evaluates FoGConcept on two independent PD cohorts comprising 45 participants. We added the cross-dataset results in Sec. 4.5 and provide the full FoG-STAR concept analysis in Appendix F.

Thank you for your response. I'm maintaining my score as 3

Reviewer_5XZpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FoGConcept, an interpretability framework for freezing-of-gait detection from multi-scale IMU signals. The core idea is to decompose a strong multi-scale detector into sparse latent features and identify human-understandable movement concepts that support FoG decisions. Experiments report 94.2% F1 for FoG detection, and 88.9% of the discovered movement concepts can be recognized by domain experts.

Critical strengths
  1. The paper aims higher than common FoG explanations that only look at sensor channels or local heatmaps. Looking for movement-level concepts such as initiation, shuffling, or trembling is closer to how clinicians think about FoG. 2.Expert recognition of the discovered concepts gives the analysis more weight than purely automatic labeling of latent units. Together with the finding that concepts repeat across subjects, makes the interpretability claim more solid.
Critical weaknesses
  1. Almost all evidence comes from one IMU FoG dataset. It is unclear whether the same concepts would appear with different sensors, placements, patient groups, or backbones. Without that check, the concepts may be tied to specific IMU or dataset.

  2. Concept finding still depends a lot on analysis choices and human judgment. It is unclear whether another analysis setup would find a different but equally plausible set of concepts, and whether the concept list is complete or 'orthogonal'.

Questions for authors
  1. Do the same concepts still appear if the temporal scales change, or if a different IMU setup(placement or brand/model) is used?
  2. Steering finds 38 model-relevant SAE features, but only 9 get clear names. What are the other 29? The backbone itself is shallow. Could the unreadable features be mixed patterns because the backbone representation is too limited? Did a deeper backbone, higher sampling rate, or wider time windows make more relevant features identifiable? Please add ablations if possible.
  3. How were the current window lengths determined? They follow a clear power pattern (1.28 / 2.56 / 5.12 s). Is this design well justified, or mainly inherited for convenience? Does the pipeline filter raw-signals? IMU data often have noise and drift, but the paper does not describe filtering. If a filter is used (or should be used), how is it coupled with the window-length design?
4 comments

Response to Weakness 1 and Q1: We have conducted additional experiments to directly test this concern. First, we repeated the complete FoGConcept pipeline on a new dataset, FoG-STAR (Borzì et al., 2026). It contains 22 people with Parkinson’s disease collected at different sites using different IMU hardware, a different source sampling rate (60 Hz versus 100 Hz), different tasks, and independent clinical annotation. Using the bilateral ankle sensors, FoGConcept recovers all nine movement-concepts identified on the original dataset. We further tested whether the recovered concepts depend on the exact temporal contexts. We tested using a 10.24s input length. The wider-context models again recover all 9/9 concept families on both VCU and FoG-STAR. For sensor placement, we additionally evaluate FoG-STAR using its lower-back IMU. We do not expect exact concept identity across placements because different sensor locations observe different aspects of gait biomechanics. For example, a single trunk sensor cannot directly represent left- versus right-foot strike in the same way as bilateral ankle sensors. Accordingly, the lower-back representation contains fewer directly corresponding movement concepts, particularly for foot-specific patterns. Our target detector uses bilateral ankle IMUs because the interpreted concepts describe gait information directly observable from foot motion. A substantially different placement, such as wrist sensors, does not observe the same bilateral foot information and therefore need not yield an identical concept set. We clarify this distinction between pipeline reproducibility and sensor-invariant concept identity in the revision. In all, these experiments show that FoGConcept provides a reproducible interpretability pipeline and recovers a comparatively stable set of human-understandable movement concepts across datasets, hardware, and substantially different temporal contexts, while allowing expected variation when the sensing representation itself changes. We added these results in Sec. 4.5, Reproducibility Across Datasets and Model Settings, with the cross-setting concept summary in Table 2, the full FoG-STAR concept atlas and sensor-placement analysis in Appendix F, and the wider-context results in Appendix F.2.

Response to W2: Yes, another analysis setup can potentially produce a different set of interpretable concepts. But our contribution is not to claim that FoGConcept recovers a unique concept vocabulary, but that it provides a reproducible pipeline for discovering functionally relevant, human-understandable movement concepts. The additional experiments in Response 1 quantify this variability: the independent FoG-STAR analysis recovers 9/9 concept families from the primary analysis, both wider-context models recover 9/9, and the 12-layer models recover 7/9 while each exposing one additional interpretable pattern. Thus, the exact concept set can vary by one or two concepts across substantially different representations, while the major movement-concepts remain comparatively stable. The analysis procedure is designed to reduce dependence on subjective judgment. Features are first selected solely by their functional effect on the FoG prediction, before their movement semantics are examined. We then inspect all selected features using the same characterization procedure, group redundant or closely related features into the same movement-concept family, and leave features without a sufficiently consistent pattern unlabeled. The resulting concepts are subsequently evaluated on disjoint participants and by independent domain experts. The manuscript already evaluates recurrence, prototype consistency, and pairwise specificity on held-out subjects. Regarding completeness and orthogonality, we do not claim that the nine concepts exhaust all movement information relevant to FoG detection. Some model-relevant features remain difficult to associate with a consistent human-understandable movement pattern, and additional concepts can emerge under different representations, as observed in our 12-layer experiments. Exhaustive concept discovery therefore remains future work. The reported concepts are strongly distinct from one another as redundant SAE features are grouped into the same concept family; among the nine representative concept directions, the median pairwise absolute cosine similarity is only 0.067, indicating low geometric overlap; and their associated raw-signal patterns remain distinguishable on held-out subjects, with median pairwise-specificity AUC 0.875. We describe the concepts as approximately orthogonal and behaviorally distinct, but not strict mathematical orthogonal. We clarify these points in the revised manuscript, add the cross-setting reproducibility analysis in Sec. 4.5 and Appendix F.

Response to Q2: The 9 reported concepts are not hand-picked. We inspected all 38 using the same characterization procedure. Of the remaining 29 features, 11 capture redundant or closely related realizations of one of the nine reported concepts, while 18 do not exhibit a sufficiently consistent movement pattern across high-activation windows to support a reliable human-understandable label. We group redundant features, and leave ambiguous features unlabeled because we don’t want to force an interpretation. The original detector already uses a 6-layer Transformer encoder, so it is a substantial representation for FoG detection. Nevertheless, we directly tested whether greater backbone depth makes the previously ambiguous features more interpretable by doubling each encoder to 12 Transformer layers while keeping the temporal inputs, training procedure, SAE configuration, steering procedure, and evaluation protocol fixed. On both VCU and FoG-STAR, the 12-layer models recover 7/9 original concepts. The deeper VCU model additionally identifies Unilateral Freezing, where one foot repeatedly attempts to move while the contralateral foot remains planted and unable to move; the deeper FoG-STAR model identifies Standing Still, where both feet remain stationary and the feature activates during the stationary interval rather than at movement onset. We also tested substantially wider temporal contexts (10.24s). On both datasets, wider windows result in more steering-relevant features from, yet the model still recovers the same 9/9 concepts. Thus, more steering-relevant SAE features do not necessarily correspond to more distinct human-understandable movement concepts; many features can represent related or redundant realizations of the same movement pattern. For sampling rate, FoG-STAR provides an independent test with different hardware and a 60-Hz source sampling rate and still recovers all nine concepts. These experiments show that the exact concept set can vary slightly across learned representations, but the major movement-concepts remain comparatively stable across different datasets and model settings. Our contribution is therefore a reproducible interpretability pipeline that consistently recovers human-understandable movement concepts, and we provide a relatively stable set of movement concepts that explain FoG detection models. We do not claim that every model configuration must produce the identical latent vocabulary. We have added the 38-feature breakdown and representative redundant/ambiguous examples to the appendix; added the wider-window and 12-layer analyses to Appendix F, and show representative examples of the newly found Unilateral Freezing and Standing Still in Fig. 5 in the modified PDF.

Response to Q3: The three window lengths were chosen to span progressively broader gait context while remaining suitable for FoG detection. The 1.28s base window was selected because it captures approximately a complete gait cycle even for slow walking while preserving rapid transitions into FoG (Koltermann et al, 2024). The longer scales are also consistent with window lengths used in prior FoG work. 2.56s windows have been used in deep-learning FoG detectors (Camps et al., 2018). Prior systematic analysis found peak FoG-detection performance at approximately 3 – 5s (Kwon et al., 2014), reflecting the benefit of observing multiple gait cycles. We therefore extend the base context by 2× and 4× to 2.56s and 5.12s, providing local, intermediate, and sustained gait context in a nested temporal hierarchy. The powers-of-two progression is a convenient implementation choice; we do not claim these exact durations are uniquely optimal. We intentionally do not use substantially longer windows in the primary model. Kwon et al. evaluated windows up to 10s and found that windows longer than approximately 4s can dilute short-duration FoG patterns and increase online detection delay; they recommend shorter windows when responsiveness is important. Our additional 10.24s experiment nevertheless recovers all nine concepts, showing that FoGConcept itself is not dependent on the original window lengths; the shorter time windows are retained because it better matches the real-time detection setting. For preprocessing, the aligned IMU channels are normalized before model input to reduce differences in scale across sensor axes, participants, and bilateral devices. During synchronized video annotation, clinicians additionally mark unusable recording portions as cut intervals; windows containing more than 70% cut samples are excluded (Koltermann et al., 2024). The remaining raw motion signal is preserved for the model, so the preprocessing does not introduce a filter cutoff that must be coupled to the selected window lengths. We clarify these preprocessing and window-selection details in the revision.

References: Borzì, L., Demrozi, F., Bacchin, R.A. et al. A multi-level annotated sensor dataset of gait freezing manifestations and severity in Parkinson’s disease. Sci Data 13, 305 (2026). https://doi.org/10.1038/s41597-026-06645-1 K. Koltermann et al., "Gait-Guard: Turn-aware Freezing of Gait Detection for Non-intrusive Intervention Systems," 2024 IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE), Wilmington, DE, USA, 2024, pp. 61-72, doi: 10.1109/CHASE60773.2024.00016. Camps, Juli, et al. “Deep Learning for Freezing of Gait Detection in Parkinson’s Disease Patients in Their Homes Using a Waist-Worn Inertial Measurement Unit.” Knowledge-Based Systems, vol. 139, 2018, pp. 119–131. Elsevier, https://doi.org/10.1016/j.knosys.2017.10.017. Kwon, Y., Park, S. H., Kim, J. W., Ho, Y., Jeon, H. M., Bang, M. J., Jung, G. I., Lee, S. M., Eom, G. M., Koh, S. B., Lee, J. W., & Jeon, H. S. (2014). A practical method for the detection of freezing of gait in patients with Parkinson's disease. Clinical interventions in aging, 9, 1709–1719. https://doi.org/10.2147/CIA.S69773

Thank you for the detailed response. My concerns have been addressed to a large extent, and the score will be raised accordingly.

Comments on the paper

1 comment
Area_Chair_DbsNOpenReview ↗

Author-Reviewer Discussion

Dear Reviewers:

Thank you for your thoughtful reviews. The authors have responded to your reviews. Please check their response and reach out to them if you have any further questions. Please act as soon as possible, as the discussion period with the authors closes on Nov 18.

Best regards,

  • AC.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.