🚀 NormWear: A Foundation Model for Multivariate Wearable Sensing of Physiological Signals.

🚀 [News! Mar 2026] We are pleased to announce that our paper has been accepted by ACM Transactions on Computing for Healthcare (HEALTH) and is now In Press!

✨ Introduction

This is the official implementation of the paper Toward Foundation Model for Multivariate Wearable Sensing of Physiological Signals.

Time-series foundation models have the ability to run inference, mainly forecasting, on any type of time series data, thanks to the informative representations comprising waveform features. Wearable sensing data, on the other hand, contain more variability in both patterns and frequency bands of interest and generally emphasize more on the ability to infer healthcare-related outcomes. The main challenge of crafting a foundation model for wearable sensing physiological signals is to learn generalizable representations that support efficient adaptation across heterogeneous sensing configurations and applications. In this work, we propose NormWear, a step toward such a foundation model, aiming to extract generalized and informative wearable sensing representations. NormWear has been pretrained on a large set of physiological signals, including PPG, ECG, EEG, GSR, and IMU, from various public resources. For a holistic assessment, we perform downstream evaluation on 11 public wearable sensing datasets, spanning 18 applications in the areas of mental health, body state inference, vital sign estimations, and disease risk evaluations. We demonstrate that NormWear achieves a better performance improvement over competitive SoTA baselines with different modeling strategies. In addition, leveraging a novel representation-alignment-match-based method, we align physiological signals embeddings with text embeddings. This alignment enables our proposed foundation model to perform zero-shot inference, allowing it to generalize to previously unseen wearable signal-based health applications.

📈 Usage

Download:

# Clone the repository
git clone git@github.com:Mobile-Sensing-and-UbiComp-Laboratory/NormWear.git

The python libraries dependencies are specified in NormWear/dependencies.txt, with an example bash script NormWear/config_env.sh.

The pretrained model checkpoint can be found in Release.

Extracting Encoder Embeddings

An example showing how to get signal embedding using NormWear:

import torch
torch.uint64 = torch.int64
torch.uint32 = torch.int32
torch.uint16 = torch.int16

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "mosaic-laboratory/normwear",
    trust_remote_code=True
)

# demo input/output
batch_num = 2 # bn
num_channel = 3 # nvar
sequence_length = 256 # L
x = torch.rand(batch_num, num_channel, sequence_length)
outpack = model(x)

# log
for k in outpack:
  print(k, outpack[k].shape)
# above should output:
# spec torch.Size([2, 3, 3, 255, 65]), # bn, nvar, cwt_channels, L, num_scales
# enc_out torch.Size([2, 3, 365, 768]) # bn, nvar, num_patches, embed_size
# NOTE: the embedding of the 1st patch in enc_out is the [CLS] embedding

Zero shot inference

The following code is for the model downloaded from the github repo. We are working on integrating the zero-shot module into Huggingface model. The instruction will be updated soon.

import torch
from NormWear.zero_shot.msitf_fusion import NormWearZeroShot

# config
device = torch.device('cpu')

### init model ##################################################################
# weight_path = "weight path here"
# msitf_ckpt = "fusion weight path here"

model = NormWearZeroShot(
    weight_path=weight_path, 
    msitf_ckpt=msitf_ckpt
).to(device)

### generate data ###############################################################
sampling_rate = 64
x = torch.rand(2, 3, sampling_rate*2).to(device) # test example: 2 samples, 3 sensor, sequence length of 2 seconds

# setup question and options
task_of_interest = [
    "Is the person doing good?"
]

options = [
    "The person is doing so so",
    "The person is under stress.",
    "The person is having fun.",
]

### Inference ###################################################################
# text encoding
txt_embed = model.txt_encode(task_of_interest+options) # [4, 2048]

query_embed, option_embed = txt_embed[:1, :], txt_embed[1:, :]

# signal encoding
signal_embed = model.signal_encode(x, query_embed, sampling_rate=sampling_rate) # bn, nvar, P, E

# calculate distance
pred_prob = model.inference(signal_embed, option_embed) # bn, num_options
pred = torch.argmax(pred_prob, dim=1).numpy()

### check output ################################################################
for i in range(len(x)):
    print("Sample {} inference: {}".format(i, options[pred[i]]))

# log
print("Input shape:", x.shape) # [2, 3, 128]
print("Text embed shape:", txt_embed.shape) # [4, 2048]
print("Signal embed shape:", signal_embed.shape) # [2, 3, P, 768]
print("Inference probability shape:", pred_prob.shape) # # [2, num_options]

For the best performance, conduct prompt engineering on the query could be a good practice.

🔥 Pre-training

The instruction below is for working with pipeline from Github repo https://github.com/Mobile-Sensing-and-UbiComp-Laboratory/NormWear

To run the evaluation on the downstream datasets, run the following command:

python3 -m NormWear.pretrain_main

Required Parameter	Type	Default	Description
--batch_size	<int>	4	Number of input samples for each round of forward and backpropagation.
--epochs	<int>	400	Number of epochs.
--blr	<float>	1e-3	base learning rate: absolute_lr = base_lr * total_batch_size / 256
--warmup_epochs	<int>	10	Number of epochs to increase learning rate from a small number to blr.
--target_len	<int>	388	Sequence length of the reconstructed time series
--weight_decay	<float>	5e-2	Regularization coefficient of L2 penalty on model parameters.
--mask_scheme	<str>	random	Masking scheme for pretext task (choose from 'random' and 'struc')
--mask__prob	<float>	0.8	Masking ratio for random masking
--mask_t_prob	<float>	0.6	Masking ratio along time axis of the scalogram.
--mask_f_prob	<float>	0.5	Masking ratio along scale axis of the scalogram.
--remark	<string>	""	Name to mark the current experimental trail.

An example bash command would be:

torchrun --nproc_per_node=4 --master_port=12345 -m NormWear.pretrain_main --batch_size 256 --epochs 100 --warmup_epochs 4 --remark test_run

❄️ Downstream Evaluation

To run the evaluation on the downstream datasets, run the following command:

python3 -m NormWear.downstream_main

Required Parameter	Type	Default	Description
--model_name	<string>	normwear	Supported models are [stats, chronos, clap, tfc, normwear]
--model_weight_dir	<string>	""	Path to the model checkpoint, only normwear need this parameter.
--group	<int>	0	Run a group of downstream tasks. The group can be customized in NormWear/downstream_main.py
--data_path	<string>	../data	Root path for where the downstream data is placed.
--num_runs	<int>	1	Number of repetition for running the evaluation of each task.
--prepare_embed	<int>	1	Run the inference and save the embeddings before sending them to evaluation. Set to 1 to overwrite previous saved embedding if there are any.
--remark	<string>	""	Name to mark the current experimental trail. By default, all the embeddings are stored under a subfolder named as --model_name under the folder of each downstream data.

An example bash command would be:

CUDA_VISIBLE_DEVICES=0 python3 -m NormWear.downstream_main --model_name normwear --model_weight_dir data/results/normwear_last_checkpoint-15470-correct.pth --group 0 --data_path ../data --num_runs 1 --prepare_embed 1 --remark test_run

:floppy_disk: The processed clean downstream datasets can be downloaded from here.

✏️ For adding a downstream dataset, please following the format:

name_of_new_dataset/
├── sample_for_downstream/
│   ├── name_of_sample_0.pkl
│   ├── name_of_sample_1.pkl
│   ├── name_of_sample_2.pkl
│   └── ...
└── train_test_split.json

where the content in each file requires content in the following format:

# data content in name_of_sample_i.pkl
{
    "uid": "", # subject ID identifier
    "data": np.array(sensor_data_here).astype(np.float16), # float numpy array with shape of [num_channels, sequence_length]
    "sampling_rate": 64, # put the correct sampling rate of the sensor signal here. 
    "label": [ # label for each task on this dataset. 'class' for classification and 'reg' for regression.
        {"class": label}, 
        {"reg": label},
        ...
    ]
}

# data content in train_test_split.json
{
  'train': ["name_of_sample_i.pkl", "name_of_sample_j.pkl", "name_of_sample_k.pkl", ...],
  'test': ["name_of_sample_l.pkl", "name_of_sample_m.pkl", "name_of_sample_n.pkl", ...]
}

🔥 Finetuning model for downstream datasets

To run the evaluation on the downstream datasets, run the following command:

python -m main_finetune --ds_name <dataset_name> --checkpoint <weight_path>

For multiple gpus, run:

torchrun --nproc_per_node=<number of device> main_finetune.py --world_size <number of device> --ds_name <dataset_name> --checkpoint <weight_path>

📈 Summary of Results

📝 Citation

The updated citation will be available shortly with journal information from ACM HEALTH.

If you find NormWear model useful for your research, please consider citing the associated paper:

@misc{luo2024foundationmodelmultivariatewearable,
      title={Toward Foundation Model for Multivariate Wearable Sensing of Physiological Signals}, 
      author={Yunfei Luo and Yuliang Chen and Asif Salekin and Tauhidur Rahman},
      year={2024},
      eprint={2412.09758},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2412.09758}, 
}

📃 License

This project is licensed under the Apache License 2.0.