Research map / Time series / Foundation models for time series
Efficient Personalized Adaptation for Physiological Signal Foundation Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PhysioPFM, which pre-trains a LoRA-adapted time series foundation model on public physiological signals and trains a diffusion transformer to generate personalized LoRA weights from local shapelet conditions, evaluated on four physiological classification tasks.
Problem
Adapting time series foundation models to physiological signal classification in clinical settings, where local computing is limited, patient data is private, and local data is imbalanced, so generic TSFMs underperform task-specific methods.
Approach
Train a diffusion transformer to map shapelet-based representations of local data to LoRA adapter weights for a frozen TSFM, so personalization needs only generator inference locally, with no fine-tuning on private data; neural-collapse-inspired anchored classifier loss handles class imbalance during public-data LoRA training.
Key results
- On Sleep-EDF sleep-state detection, PhysioPFM reaches 86.39% accuracy, 4.49 points above the best baseline MOMENT (81.90%).
- On MIT-BIH arrhythmia diagnosis, PhysioPFM achieves 89.94% accuracy, 3.39 points above the second-best method DeepArr (86.55%).
- On FoG detection, PhysioPFM achieves 81.32% accuracy versus 74.71% F1 for the Extra Tree baseline's best listed metric; reported as SOTA over all baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PhysioPFM for personalized adaptation | GPT-2-based TSFM · Sleep-EDF | accuracy: 81.90±1.42 -> 86.39±2.54 | Helped |
| Use PhysioPFM for personalized adaptation | GPT-2-based TSFM · DREAMER valence | accuracy: 71.29±3.13 -> 73.97±2.36 | Helped |
| Use PhysioPFM for personalized adaptation | GPT-2-based TSFM · DREAMER arousal | accuracy: 80.59±2.72 -> 82.19±2.73 | Helped |
| Use PhysioPFM for personalized adaptation | GPT-2-based TSFM · DREAMER dominance | accuracy: 81.26±2.54 -> 84.45±4.35 | Helped |
| Use PhysioPFM for personalized adaptation | GPT-2-based TSFM · MIT-BIH arrhythmia | accuracy: 86.55±2.62 -> 89.94±1.53 | Helped |
| Use PhysioPFM for personalized adaptation | GPT-2-based TSFM · Freezing of Gaits | accuracy: 73.95±2.44 -> 81.32±2.65 | Helped |
| Add anchored-classifier loss (ablation) | PhysioPFM · Sleep-EDF | accuracy: 82.47±2.01 -> 86.39±2.54 | Helped |
| Add anchored-classifier loss (ablation) | PhysioPFM · MIT-BIH arrhythmia | accuracy: 84.30±1.52 -> 89.94±1.35 | Helped |
| Add anchored-classifier loss (ablation) | PhysioPFM · Freezing of Gaits | accuracy: 78.56±3.78 -> 81.32±2.65 | Helped |
| Set anchored-classifier temperature to τ=0.1 instead of the default (ablation) | PhysioPFM · Sleep-EDF, MIT-BIH arrhythmia, and Freezing of Gaits | accuracy: 84.64±1.92; 86.10±2.16; 79.45±3.96 vs. 86.39±2.54; 89.94±1.35; 81.32±2.65 | Hurt |
| Set anchored-classifier temperature to τ=10 instead of the default (ablation) | PhysioPFM · Sleep-EDF, MIT-BIH arrhythmia, and Freezing of Gaits | accuracy: 81.76±2.31; 85.52±2.84; 78.37±1.60 vs. 86.39±2.54; 89.94±1.35; 81.32±2.65 | Hurt |
| Use local shapelets rather than public-task shapelets (ablation) | PhysioPFM · Sleep-EDF, MIT-BIH arrhythmia, and Freezing of Gaits | accuracy: 85.94±2.05 -> 86.39±2.54; 87.43±1.65 -> 89.94±1.35; 79.96±3.21 -> 81.32±2.65 | Helped |
Surprising findings
- Larger LoRA rank does not help: increasing rank from 8 to 16 gave no further gain, and smaller ranks achieved acceptable performance.
- Textual prompts performed worse than shapelet prompts as DiT conditions, since no robust pre-trained text-to-time-series encoder exists.
Limitations
- Generator training requires about 20GB GPU memory on a server and depends on public physiological datasets that may not cover all signal types.
- Evaluated on only four classification tasks with small-scale datasets and no public code or weights stated.
- Ablation numbers are inconsistent across tables (e.g., MIT-BIH full-model value differs between Table 3 and Table 5), limiting clarity of reported gains.
Experimental setup
Datasets: Sleep-EDF DREAMER MIT-BIH arrhythmia FOG dataset UTSD
Benchmarks: Sleep-EDF DREAMER MIT-BIH arrhythmia FOG dataset
Metrics: accuracy macro F1 kappa precision recall AUC
Baselines: Informer FEDformer SimMTM TimesNet PatchTST iTransformer OneFitsAll Time-LLM MOMENT SleepFM SleepDG DeepArr
Builds on: LoRA Time-LLM One Fits All (GPT4TS) Diffusion Transformer (DiT) Train-Once-for-All Personalization
Closest papers by meaning
- Scaling Wearable Foundation Models (ICLR 2025 · similarity 0.84)
- Towards Neural Scaling Laws for Time Series Foundation Models (ICLR 2025 · similarity 0.83)
- Specialized Foundation Models Struggle to Beat Supervised Baselines (ICLR 2025 · similarity 0.82)
- OATS: Online Data Augmentation for Time Series Foundation Models (NeurIPS 2026 · similarity 0.82)
- Less is More: Unlocking Specialization of Time Series Foundation Models via Structured Pruning (NeurIPS 2025 · similarity 0.82)
- Multi-Scale Finetuning for Encoder-based Time Series Foundation Models (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.