Deep Learning 1D-CNN-Based Ground Contact Detection in Sprint Acceleration Using Inertial Measurement Units

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents a clear application of an InceptionTime-style 1D-CNN to detect sprint ground-contact (IC/TC) events and estimate contact times from bilateral shank IMUs, using high-speed video as ground truth, and reports very strong agreement, including 100% precision/recall after postprocessing with mean Hausdorff errors below 6 ms.

• The literature review is generally current and the motivation for using deep learning in sprint acceleration is well argued; however, the paper would benefit from a clearer statement of what is genuinely novel beyond applying an established backbone to this specific label set and sensor configuration.
• The athlete-wise data split is appropriate for assessing subject generalization, but the robustness claims would be stronger if the authors justified why two held-out athletes are sufficient and if uncertainty were reported across subjects, not only across repeated runs.
• Although hyperparameter ranges and stopping criteria are described, some architecture and implementation details remain underspecified (e.g., equation notation, padding/stride choices, normalization placement, random seeds), which makes exact replication difficult despite the provided block diagram.
• The reported perfect detection performance depends on a post hoc removal of very short predicted contacts (<48 ms); key results should therefore be shown both with and without this heuristic, or the rule should be integrated into model selection to avoid test-set–specific tuning.
• The conclusion would read as more balanced if it explicitly constrained the claims to the tested sprint protocol, sensor placement, and limited cohort, and clearly outlined what additional validation is required for broader or real-world deployment.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. The current training/validation set consists of 10 athletes (2 runs × 2 legs), while the test set includes only 2 athletes. This is acceptable for demonstrating methodological feasibility, but it makes the manuscript’s claim of “robust generalization and transferability to real-life application” somewhat overstated.

2. Ground contact (GC) is annotated using three 250 fps high-speed cameras plus manual identification of initial contact and toe-off frames, then refining the touchdown instant based on maximal Achilles/heel displacement, followed by cross-correlation with the IMU aRES signal. In the Discussion, the authors themselves acknowledge that this synchronization approach at 250 Hz may still introduce non-systematic errors. It is recommended to add one sentence at the end of Section 2.2 “Data collection and preparation” or in the Discussion to describe the potential impact of synchronization error on training and evaluation, for example: “We acknowledge that residual synchronization errors at 250 Hz may introduce label noise on the order of one frame, which could slightly underestimate the true model performance.” If any simple repeated-annotation/inter-rater agreement information is available (e.g., a subset of steps independently annotated by two raters), it would be helpful to add one sentence to strengthen the reliability of the labels.

3. The authors perform a grid search over window size ∈ [25, 200] and stride ∈ [5, 30], and finally select window = 100 and stride = 15; the window length roughly covers one full gait cycle (GC + flight). This design of using a complete cycle as temporal context is reasonable, but the rationale is not clearly articulated in the manuscript.

4. Regarding the current post-processing: the 12-frame threshold is essentially a heuristic rule derived from the prior knowledge that the shortest GC duration is > 90 ms. While it removes short spikes, it may also affect truly short contacts (even though GC is not extremely short in the acceleration phase used here). This implies that the reported “100% detection” is achieved by the combination of the deep model and this rule-based post-processing, rather than by the end-to-end capability of the model alone.

5. The presentation of the statistical results could be made slightly more structured.

6. For Tables 2–4, which contain quite a lot of information, it is recommended to explicitly state in the text: “The mean absolute error of GC time was around 5–6 ms (<5% MAPE) in both VAL and TEST”; and “LOA were approximately ±15 ms, comparable to or better than state-of-the-art IMU-based methods under similar conditions.”

7. It is recommended to add the following sentence in the Discussion: “While the 1.5 kg towing resistance might have slightly altered sprint mechanics, we expect its effect on the relative timing of GC events to be minimal. Future work should nevertheless confirm the findings under fully free sprinting conditions.”

8. “learns to down-weights perturbed channels” → “learns to down-weight perturbed channels”

9. “the last classification” → “the final classification layer” or “the final classification decision”

10. Introduction: “its’ subsequent toe-off or terminal contact (TC)” → “its subsequent toe-off or terminal contact (TC)”

11. “the trunk gets into his most upright position” → use a gender-neutral form, e.g., “its most upright position” or “the trunk reaches its most upright position”

12. Discussion: expressions such as “its’ performance metrics” should consistently be written as “its”.

13. “1D CNN” and “1D-CNN” both appear in the manuscript; it is recommended to unify the terminology to “1D-CNN” (including in the title, abstract, main text, and figure captions).

14. “ground contact” vs “Ground Contact (GC)”: provide the abbreviation GC at the first occurrence, and then use GC throughout the rest of the manuscript, to avoid ambiguous uses such as “GCs” where it is unclear whether it refers to events or time intervals.

15. “1D-CNN accurately detected 100% of ground contacts” → it is recommended to add “after post-processing”.

16. Author Contributions: “supervision, Felix Fried, T.M., J.E-N..;” appears to contain a typo in the first author’s name (earlier it is “Felix Friedl”) and also has an extra period.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised manuscript presents an InceptionTime-style 1D-CNN to detect IC/TC events and ground-contact times from bilateral shank IMUs during the first 15 m of sprint acceleration, using high-speed video as reference and achieving near-perfect agreement.

• Acknowledging the limitation of having only two TEST athletes is appropriate; however, the resulting performance claims remain sensitive, and reporting uncertainty at the athlete level (not only pooled steps), ideally complemented by a leave-one-athlete-out or repeated subject-wise split, would better support robustness.
• The paper suggests potential real-time field deployment, but no evidence is provided on inference latency or computational load; including per-window inference time, hardware specifications, and sensitivity to 250 Hz resampling would bring the conclusions in line with the experimental evidence.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

NO other concerns.

Author Response

Thank you!