Efficient Personalized Adaptation for Physiological Signal Foundation Model

Authors: Chenrui Wu, Haishuai Wang, Xiang Zhang, Chengqi Zhang, Jiajun Bu

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that integrating generated models with TSFM enhances performance, and transferability, and reduces the need for additional sensitive data training. In Table 1-4, we evaluate the performance of our proposed method against diverse baseline methods... 5. Experiments 5.1. Experimental Setup 5.2. Main Results 5.3. Ablation Study
Researcher Affiliation Academia 1Zhejiang Key Laboratory of Accessible Perception and Intelligent Systems, College of Computer Science and Technology, Zhejiang University. 2School of Computing Science, Simon Fraser University. 3Department of Computer Science, The University of North Carolina at Charlotte. 4Department of Data Science and Artificial Intelligence, Hong Kong Polytechnic University. Correspondence to: Haishuai Wang <EMAIL>.
Pseudocode Yes Algorithm 1 Lo RA generator training
Open Source Code No This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none of which we feel must be specifically highlighted here.
Open Datasets Yes Sleep-EDF dataset (Kemp et al., 2000) is a public dataset... DREAMER (Katsigiannis & Ramzan, 2017)... The MIT-BIH arrhythmia dataset (Moody & Mark, 2001)... The FOG dataset (Li, 2021)... In Table 6, we provide statistical information on collected public physiological signals, mainly from (Zhang et al., 2024; Qiu et al., 2023) and Pyhsio Net (Goldberger et al.).
Dataset Splits Yes We randomly sample 60% of the data for training, 20% for validation, and 20% for testing, following the existing work (Zhang et al., 2024), for all evaluation tasks.
Hardware Specification No Our Physio PFM conducts generator training on the server, which requires about 20GB of GPU memory, which is feasible for servers with sufficient computing power. In local adaptation, Di T inference only occupies 3 GB of GPU memory, reaching a remarkable balance between accuracy, speed, and Consumption.
Software Dependencies No For generative model architecture, we adopt GPT-2 (Radford et al., 2019) as the diffusion transformer with 12 layers. During training, we use Adam W with a batch size of 64, a learning rate of 4e-4, 1000 diffusion steps, and a linear noise scheduler ranging from 0.0001 to 0.012.
Experiment Setup Yes For generative model architecture, we adopt GPT-2 (Radford et al., 2019) as the diffusion transformer with 12 layers. During training, we use Adam W with a batch size of 64, a learning rate of 4e-4, 1000 diffusion steps, and a linear noise scheduler ranging from 0.0001 to 0.012. we divide the Lo RA weights into chunks by layer, and the size of each chunk is 576. We set the rank of the adapter as 4. For the pre-trained time series foundation model, we adopt the 6-layer GPT2-based backbone (Radford et al., 2019; Liu et al., 2024b), pre-trained by UTSD datasets (Liu et al., 2024b).