Efficient Personalized Adaptation for Physiological Signal Foundation Model

Thank you for your extensive feedback. We especially appreciate your effort to meet the conference's rigorous review criteria. We kindly address your questions as follows.

W1: Scope of adaptation time.

For W1: Thanks for your thoughtful comments. We separated the two phases of pre-training and adaptation, considering a practical cloud-edge scenario. If utilizing a traditional pre-trained TSFM, the TSFM is first pre-trained with massive data in a venue with sufficient resources (cloud). Then, the TSFM is used to be trained on local physiological signals, and its training time is the adaptation time we defined. We did not count the original pre-training of TSFM into its adaptation time. Similarly, for our method, pre-training can be achieved using public physiological signals and a venue with sufficient resources. Our adaptation time includes generator inference and TSFM inference. Therefore, we respectfully disagree that LoRA dataset preparation and DiT training are part of adaptation time.

W2: Discussion of important ablation study and baselines.

For W2: We sincerely thank you for your detailed suggestions. We would like to address your concerns by showing additional ablation results for our method. We adopt the TSFM with local LoRA fine-tuning, TSFM with a prototype-based classifier, ECG-FM on arrhythmia diagnosis, along with two given row results in Table 5 as references. From the results, we can find that if the training is allowed, pre-trained TSFM could show great power in most cases after fine-tuning. Considering the variant without LoRA, we could acknowledge that the generated LoRA weights could provide a comparable improvement with direct training. A prototypical network is also a valuable approach as it could surpass general baselines but still be lower than the adapted weight-based strategy (Ours, fine-tuning TSFM). It may require more computing cost in similarity(metric) calculation, and could be easily affected by imbalanced data. More alignment or neural-collapse tricks may help to improve this approach. Metric learning is a promising future work to explore. We will also add the complete results of ECG-FM to Table 3 in the revised manuscripts.

W3: Clarifications on generator selection.

For W3: In parameter generation tasks, on the one hand, these small models may not be able to fully demonstrate their generalization ability when applied to more complex tasks and parameter spaces. On the other hand, previous methods do not support conditional high-performance parameter generation, and the novel DiT-based conditional neural network diffusion has better generation results. The architecture of DiT has great expressive power in diffusion tasks, especially in conditional cross-modal applications such as text-to-image and text-to-video. Unlike hypernetwork, which takes the model parameters as input and generates parameters, we directly map the data feature space to the parameter space. In addition, the hypernetwork needs to perform backpropagation based on the loss of the backbone network, which will be costly in large model scenarios. We conducted a validation experiment, taking the condition and noised model parameters as input, and measuring the Euclidean distance between the output model parameters and the original input parameters. The results show that the generation effect of the method based on conditional GAN or MLP is far behind DiT. As the SOTA cross-modal generator, we adopted DiT. It is also worthwhile to find the trade-off between cost and quality in future work.

W4: Typos.

For W4: Thank you for this astute observation. We apologize for any confusion caused and have carefully revised it.

W5: Clarifying the data partition and definition of personalization.

For W5: The experiments are conducted on corresponding signals in a subject-independent setting. We assign subjects to train/val/test partitions. One subject’s data won’t appear across train/val/test sets. The inference of the generator is based on testing sets, where the train and val sets are not used for our method but for baselines, for fairness. Compared with the generalized TSFM, our personalization refers to obtaining a model parameter that adapts to local data for new local tasks and data features. Because the generalized model may not work well on specific tasks, and fine-tuning a large foundation model will also be costly, realizing model personalization in a lightweight way is meaningful.

We sincerely thank you for your valuable comments, and we are grateful for the time and effort you have invested in reviewing our work. Below, we provide a point-by-point response to address each of your concerns:

W1: Clarifying the data partition.

For W1: Thanks for your valuable comments. It is at the patient level. Physiological signal data are collected from individual subjects. The experiments are conducted on corresponding signals in a subject-independent setting. We divide the subjects into training, validation, and test sets in a 3:1:1 ratio. One subject’s data won’t appear across train/val/test sets. It may arouse concerns about how it can be considered a personalized adaptation (for baselines) without training on the data of a specific local patient. In practice, local test data must be unlabeled, and the model training is through existing labeled data, and then the model is applied to the patients who need to be diagnosed. So, this setting is reasonable and in line with the existing related works' paradigm.

W2: Hyperparameter search.

For W2: We test the best learning rate, diffusion steps, chunk size and rank of the adapter for our methods. For baselines, we generally follow the original hyperparameters and keep necessary settings fair, such as data partition.

W3: Discussion on the setting and selection of the generator.

For W3: The architecture of DiT has great expressive power in diffusion tasks, especially in conditional cross-modal applications such as text-to-image and text-to-video, where the original DiT is applied. Small models like MLP have difficulty in cross-modal generation and conditional generation. In difficult tasks such as parameter generation, the latent space generated by MLP may be poorly representative. Therefore, how to achieve a trade-off between the lightness and performance of the generator is a future work worth exploring. In this work, we mainly applied the SOTA cross-modal generator DiT to achieve high-quality parameter generation. At the same time, the transformer complies with the scaling law, a feature that will help the model to generalize only by amplifying parameters. We conduct an analysis of the generation quality on weights, considering the Euclidean distance between input weights(noised) and generated weights. As demonstrated in the following results, at the end of the training of DiT, the results of distance reached 2.32 to 4.10 for different tasks, proving superior generating performance than conditional GAN and MLP. In addition, MLP and CGAN require huge numbers of rounds to converge, more than 1000 epochs, while DiT does not.

W4: Detailed ablation study on model fine-tuning.

For W4: We sincerely thank you for your insightful feedback. We evaluate the TSFM with local LoRA fine-tuning, and the results are in the following table. We also add the existing results from Table 5, which includes ours without LoRA generation, and our method. It can be seen that training TSFM on local data could achieve superior performance in most cases. While our generated LoRA weights could also provide a comparable improvement to direct training. Without training or generated weights, the pre-trained TSFM is hard to fit the specific domain knowledge well.

W5: Typo and presentations.

For W5: We appreciate the reviewer's keen eye and thorough review. We have carefully revised the typos and will provide a more rigorous expression of Figure 3 in the revised version of our work.

W6: Suggestions on related work.

For W6: Thanks for your constructive suggestions. Due to the length limit, we have to adopt the current organization. It is truly significant to place the related work section in the main body for better understanding by the readers. We promise to reorganize it. We will also be adding more extensive related work, such as privacy protection-related work.

We sincerely appreciate for your comprehensive and valuable review, particularly given the meticulous standards during the review of this venue. We appreciate the opportunity to address your concerns.

W1: Clarification on the privacy issue.

For W1: We apologize for missing a detailed discussion on privacy preservation. Privacy constraint is one of our contributions. A generic pre-trained time series foundation model is hard to fit the diverse local tasks and data. Uploading patients’ data to the cloud for robust pretraining may raise privacy concerns. Our proposed method achieves this goal by data isolation, that is, sensitive patients’ data are preserved locally. Due to the removal of data exposure, privacy issues could be solved. The most related work includes some time series foundation models and federated learning. Brant-X [1] adopts the EEG foundation model Brant-2 as a basis, which is pre-trained on 4TB of private brain signal data. This approach requires private data and large-scale pre-training on it. If the medical entity has sufficient computing resources, it could be reasonable, otherwise, it may need to transfer private data to a robust cloud. On the other hand, federated learning preserves privacy by exchanging models, not data, which is similar to our paradigm. [2] considers using FL to train a foundation model for medical time series. Multiple rounds of exchange on the large models may lead to remarkable communication costs, and models may be attacked by model poisoning attacks from history gradient updates. While our work focuses on the personalized TSFM with a privacy guarantee. Due to the length limitation of the rebuttal, we promise to add a more comprehensive privacy-preserving literature on related work and our method.

[1] Zhang, Daoze, et al. "Brant-X: A Unified Physiological Signal Alignment Framework." KDD 2024.

[2] Ali, Mahad, et al. "Fine-Tuning Foundation Models with Federated Learning for Privacy Preserving Medical Time Series Forecasting." IEEE EMBC 2025.

W2: Definition 2.1 and Proposition 3.1.

For W2: Thanks for your valuable comments. Definition 2.1 follows the standard definition of neural collapse. It is not a contribution, but an illustration.

The proof for Proposition 3.1 is given as:

$*Proof.*$ Suppose $C$ class prototype vectors $p_1, p_2, \dots, p_C \in \mathbb{R}^d$ satisfy: Unitization: $|p_i| = 1$, symmetric inner product constraint: $p_i^T p_j = -\frac{1}{C-1}, \forall i \neq j$. Define the Gram matrix $G \in \mathbb{R}^{C \times C}$, where:

$$G_{ij} = p_i^T p_j = \begin{cases} 1 & \text{if } i = j, \newline -\frac{1}{C-1} & \text{if } i \neq j. \end{cases}$$

The matrix has the following properties: the diagonal elements are 1 and the off-diagonal elements are $-\frac{1}{C-1}$. The matrix rank is $d$, and it is a symmetric semi-positive matrix (because the vector is in d-dimensional space). The above Gram matrix corresponds to the Simplex Equiangular Tight Frame. Its core features are: all off-diagonal elements are equal, that is, the angles between vectors are consistent, the vectors are evenly distributed in the feature space, and the minimum interval between classes is maximized. This structure is the only configuration that satisfies symmetry and minimizes the similarity between classes. The angle between any two vectors $\theta$ satisfies:

$$\cos\theta = p_i^T p_j = -\frac{1}{C-1}.$$

In classification problems, the decision boundary is determined by the geometric relationship of the prototype vectors. For a linear classifier, the decision boundary for two categories $i$ and $j$ is:

$$x \in \mathbb{R}^d \mid (p_i - p_j)^T x + \frac{\|p_j\|^2 - \|p_i\|^2}{2} = 0.$$

Since $\|p_i\| = \|p_j\| = 1$, the boundary is simplified to:

$$(p_i - p_j)^T x = 0.$$

The margin $\gamma$ between the two boundaries is:

$$\gamma= \frac{2}{\|p_i - p_j\|}.$$

Calculate $\|p_i - p_j\|^2 = 2(1 - p_i^T p_j) = 2\left(1 + \frac{1}{C-1}\right)$, so:

$$\gamma = \frac{2}{\sqrt{2\left(1 + \frac{1}{C-1}\right)}} = \sqrt{\frac{2(C-1)}{C}}.$$

When all class prototypes meet the symmetry condition, the intervals between all classes are consistent and reach the maximum possible value, so the decision boundary is optimal.

We will add this proof to the appendix in the revised manuscripts.

Thanks for recognizing the value of our work. We are grateful for your thorough feedback, especially considering the massive requirements of the review. We hope the following comments could address your questions:

W1: Relation to hypernetwork.

For W1: In a macro sense, our method and the hypernetwork both provide adaptive parameters for another backbone network. Specifically speaking, according to Figure 1(b) of the reference, the hypernetwork also needs to be trained together with the main model, calculate the performance loss of the main model, and perform gradient backward propagation on the hypernetwork. Unlike the hypernetwork, we do not update the hypernetwork based on the performance of the backbone model, which is costly in our context. Our paper directly maps the data feature space to the parameter space, aiming to establish a mapping between LoRA parameters and data features. Our method pursues better generation quality and aims to learn an effective and representative latent space. This is logically different from the traditional hypernetwork and is designed to meet the challenges of our scenario to achieve local training-free adaptation. In general, our method has similar macroscopic goals to hypernetworks, but their specific implementations are quite different. We will include a discussion of this aspect in the main text and provide more detailed related work.

W2: Discussion of applying to multi-task scenarios.

For W2: Thank you for raising this important concern. We focus on general physiological signals classification in this work. When the target task changes, e.g. transferring to a time series forecasting task, our approach may need to adjust the pre-training on the generator to build a new ability for new tasks. Another possible way is to expand the type of input condition, which could be a significant future work to extend our methods to multi-task scenarios.

W3: The impact of training sample size for the generator.

For W3: We kindly note that we have included the impact of training sample size on the generator in Section 4.3, Impact of training samples. To fully enhance the ability of the generator, we adopt multiple ways to expand the training sets in the data preparation phase, including partial dataset, partial class, and random selection of subjects. Given the different input of data, the generalized ability of the generator could be more diversified.

W4: Definition of the adaptation time.

For W4: We apologize for any confusion caused. For our proposed method, Figure 5 includes the time to generate LoRA weights and TSFM inference. For baselines, the adaptation time refers to the local training time. Because our method does not need to fine-tune the foundation model, while the baselines need to train and update the model. Backpropagation on a large model is time-consuming. We only have a single-step inference to generate LoRA, which takes little time compared to multiple-round training.