Stochastic Sparse Sampling: A Framework for Variable-Length Medical Time Series Classification

1. This paper uses a single-channel approach for SOZ detection, justifying it by citing the challenges posed by varying numbers of iEEG channels across iEEG recordings. However, this reason does not demonstrate that single-channel analysis is more effective. In clinical practice, SOZ and early propagation zones involve multiple channels with timing differences in abnormal discharges [1]. The authors do not provide sufficient theoretical or experimental support to explain why a single-channel approach would effectively capture these critical distinctions.
2. This paper defines SOZ detection as a variable-length time series classification (VTSC) task. Still, the authors do not clarify the specific benefits of VTSC over general anomaly detection for this application. Their justification, that “effective treatment requires analysis of variable-length signals,” lacks concrete references or explanations relevant to variable-length signals, weakening the rationale for using VTSC in this context.
3. The proposed method is quite simple. How does it differ from other downsampling or sparse sampling approaches? Model output still provides a global classification result, which seems no different from a standard classification model. The paper should compare its approach with more methods that share a similar motivation and more details of related works to clarify its position within the research field.
4. The authors state in the abstract that their method outperforms "state-of-the-art (SOTA) baselines across most medical centers." However, none of these baselines were designed for the SOZ detection task or medical time series data.Instead, all baseline models were built for general time series or other sequential data forecasting tasks, not even for time series classification. The authors should have used baselines developed for iEEG or EEG signal analysis, such as methods in references [2-4], or, at the very least, models designed for time series classification or anomaly detection [4,5].

[1] Li et al., Neural fragility as an EEG marker of the seizure onset zone, Nature Neuroscience, 2021.

[2] Tang et al., Self-Supervised Graph Neural Networks for Improved Electroencephalographic Seizure Analysis, ICLR, 2022.

[3] Luo et al., Exploring Adaptive Graph Topologies and Temporal Graph Networks for EEG-Based Depression Detection, IEEE Transactions on Neural System and Rehabilitation Engineering, 2023.

[4] Rikuto et al., SplitSEE: A Splittable Self-supervised Framework for Single-Channel EEG Representation Learning, ICDM, 2024.

[5] tang et al., Omni-Scale CNNs: a simple and effective kernel size configuration for time series classification, ICLR, 2022.

[6] Lu et al., Out-of-Distribution Representation Learning for Time Series Classification, ICLR, 2023.

The motivation to save computational resources is well-intentioned, though I am concerned about its practicality in actual training. For a given set of window sizes with corresponding independent models, even if only subsets of window sizes are selected during training, the space complexity of the models remains unchanged. This approach primarily improves training speed without reducing memory requirements. Additionally, using an independent model for each patch length may not be optimal for memory efficiency. A shared backbone across different patch lengths could be a more effective choice for memory savings. Overall, the method resembles an enhanced version of MTST [1], employing a random subset of models with varying patch lengths during training.

Moreover, while the paper’s title refers to medical time series, only a single seizure dataset is used for evaluation. Expanding the evaluation to include additional datasets would strengthen the claim of generalizability. The ablation study could also benefit from a deeper investigation into multi-scale learning with various patch lengths. For instance, exploring which combinations of patch lengths yield the best performance would be informative. Additionally, the impact of stochastic sparse sampling should be assessed in detail. For a given list of patch lengths, how do memory usage and running time between training with and without stochastic sparse sampling? Lastly, a recent work, Medformer[2], should be compared in baseline methods, as it is also designed for medical time series classification using multi-scale patching. A discussion on the differences between this method and Medformer would also be valuable for highlighting the unique aspects of the proposed approach.

[1] Multi-resolution Time-Series Transformer for Long-term Forecasting [2] Medformer: A Multi-Granularity Patching Transformer for Medical Time-Series Classification

Main concern is the single domain/task used to test the method. While the single domain is very interesting, there is something different in the fact that the seizure periods are themselves randomly occurring throughout the time series. In other tasks, long-term dynamics of the time series may require extracting patterns through time rather than this which is more akin to multiple instance learning where the search is for any evidence of positive class.

The lack of other tasks weakens the generality of the method, but I don't have the perfect case of variable length it is hard to say where there would

• The proposed approach is simple, which is perfectly fine, but lacks a more thorough analysis, at least empirical. For a reader with a use case in mind, it is difficult to assess if this method is appropriate.
• Another drawback is that this approach has only been tested on one data set. To assess its generalization capability, the authors could test their method on additional publicly available time series datasets.
• Section 4.5 on qualitative interpretation needs to (re)written to be more convincing as it lacks much information or comments. Interpretability is one of the contributions listed in the introduction.