Turning Silver into Gold: Domain Adaptation with Noisy Labels for Wearable Cardio-Respiratory Fitness Prediction

Deep learning (DL) has been widely applied to many healthcare applications, such as sleep stage classification, stress detection, and fitness prediction (Yildirim et al., 2019; Jaques et al., 2017; Sakr et al., 2018). Collecting high-quality labels (i.e., gold-standard) for healthcare applications may require extensive efforts and can be particularly time-consuming: as a consequence, most of the existing datasets are small-scale. Developing DL models with such small-scale datasets often leads to poor performance and lack of model generalization (Raschka, 2018). At the same time, large-scale datasets with imprecise labels (i.e., silver-standard datasets) are generally available from inexpensive wearable sensors. For instance, polysomnography (PSG) sensors are used as the gold-standard measurement for sleep stages monitoring. PSG is mainly utilized in a controlled environment since they it requires dedicated equipment and the involvement of clinicians. As a result, such high-quality (gold-standard) datasets are usually small-scale (O’Reilly et al., 2014). Recently, Electrocardiogram (ECG) sensors embedded in wearable devices have gained momentum due to their affordability and high portability, leading to the generation of less-accurate but large-scale (sliver-standard) datasets (Lee et al., 2022).

High quality labels are crucial for the development and validation of robust clinical models. The labels of silver-standard datasets often contain estimation noise due to less accurate collection scheme and, as they are collected from larger populations, they are characterized by distribution shifts, which makes validation against gold-standard data difficult. Therefore, a natural question arises: Can we leverage noisy large-scale sliver-standard datasets to improve DL model validation on gold-standard datasets?

Transfer learning (TF) is the natural candidate for this problem  (Xu et al., 2020). However, it is difficult for it to handle the problem where the distribution between two tasks is not similar. Recently, domain adaptation (DA) has achieved some initial success on solving the problem of the distribution mismatch between the source and target domains (Patricia and Caputo, 2014). Specifically, adversarial-based DAs demonstrated state-of-the-art performance on reducing the difference between data distributions (Tzeng et al., 2017; Hassan Pour Zonoozi and Seydi, 2022), for example for medical imaging tasks (Venkataramani et al., 2018). However, adversarial-based DAs mainly focus on discriminating the source and target domains as a binary classification task but ignore the fine-grained information that resides within the distribution shifts. To this end, in this paper we introduce a multi-discriminator model to simultaneously learn the coarse-grained and fine-grained domain information under the DA scheme.

In particular, we focus on the challenging health application of cardiorespiratory fitness (CRF) prediction employing maximal oxygen consumption (VO₂max) as the ground truth. CRF is a vital indicator of Cardiovascular disease (CVD) (Laukkanen et al., 2001), a leading cause of death globally (Kaptoge et al., 2019). Collecting gold-standard VO₂max labels through maximal exercise tests is difficult because it requires participants to reach exhaustion while on a treadmill, using a face mask with a computerized gas analysis system to monitor ventilation and expired gas fractions, as shown in Figure 1. This is not practical for particular groups, such as the elderly, for instance. Therefore, an alternative submaximal VO₂max test (Gonzales et al., 2020) promises to capture fitness levels with lower accuracy. Although the gold-standard and silver-standard data capture similar information, they suffer from distribution shift.

In our work, we propose a novel unsupervised domain adaptation framework to improve the gold-standard fitness prediction by leveraging large-scale noisy VO₂max data.

This work introduces a novel unsupervised domain adaptation framework that utilizes multi-discriminator during adversarial training. Herein, the section discussed details of our framework.

The overall model architecture and multi-discriminator training scheme are shown in Figure 1. In general, our framework consists of two training stages, pre-training on the source domain ${D_{s}}$ and an adversarial-based training stage on the ${D_{t}}$ (target domain). In the first stage, we build a pre-trained model with noisy health-related labels in a fully supervised setting. Specifically, the neural network is built through multiple stacked bidirectional Gated Recurrent Unit network (GRU) layers to extract time-series features’ embedding. After that, Multilayer Perceptron (MLP) layers are followed to extract associated metadata representation.

After pre-training, we fine-tune the large-scale feature embedding on ${D_{t}}$ after incorporating samples from ${D_{s}}$. Besides, a novel multiple discriminator-based adversarial domain adaptation learning scheme is trained to distinguish domain labels ${y_{c}}$ and ${y_{d}}$ and estimate ${y}$.

Multi-discriminator Adversarial Training. Our framework uses a combination of coarse-grained and fine-grained discriminators to aid the adversarial training for regression tasks. The coarse-grained discriminator (${D_{c}}$) is similar to other regression DA (de Mathelin et al., 2020; Zhao et al., 2018) and follows the DANN style. In particular, ${D_{c}}$ mainly focus on discriminating rough domains, which uses a binary domain classifier to discriminate the source of data points (where 0 represents the data comes from the ${D_{s}}$ and 1 from ${D_{t}}$). However, the simple binary classes cannot fully represent the data label distribution of each domain. Therefore, we add a more fine-grained discriminator (${D_{f}}$) to discriminate general distribution differences to aid the adversarial training.

For adversarial training, the shared encoder first leverages the pre-trained model and extract general feature. Then discriminators are trained to differentiate the domain labels. For ${D_{c}}$, we optimize ${l_{CSE}}$, a standard cross-entropy loss, to classify the coarse domain source labels ${y_{c}}$. We can force the extractor to learn a general feature by maximizing such divergence. After that, a ${D_{f}}$ further maximizes the nuance changes among the sample and updates the discriminator by discriminating the domain distribution label using Gaussian Negative Log-Likelihood (GLL) loss ${l_{GLL}}$ (Paszke et al., 2019). After that, this multiple two-game adversarial training between the feature encoder and discriminator finally converges into balance. As a result, the feature extractor learns cross-domain knowledge from the multiple discriminators and thus benefits from the main regression prediction task. The whole framework is optimized by the total loss ${D_{L}}$, which is defined as:

We use ${L}$ to minimize the loss of the predictor while maximizing the loss of domain discriminators. In detail, $\alpha$ is used to scale down the predictor loss to the same level as predictors, $\lambda_{1}$ and $\lambda_{2}$ control the relative weight of the discriminator loss, and $\lambda_{1}$ + $\lambda_{2}$ = 1.

Datasets. The source domain comes from the silver-standard measurement study (Fenland), including 12,425 participants. It contains a combined heart rate and movement signal from chest ECG sensor Actiheart and noisy VO₂max labels collected from submaximal exercise tests (Lindsay et al., 2019). The target domain ${D_{t}}$ represents the gold-standard measurement dataset BBVS, which is a subset of 191 participants from the Fenland study with directly measured ground-truth VO₂max (Gonzales et al., 2021) during the maximal exercise tests. In the BBVS study, participants need to wear a face mask to measure respiratory gas measurements (Rietjens et al., 2001) to experience exhaustion tests. Further, movement and heart rate are collected using the Actiheart (CamNtech, Papworth, UK) sensor. The details of features and the dataset are in Appendix A

Training Strategy. We evaluate UDAMA on these two datasets by first pre-training a model on the ${D_{s}}$ Fenland. Second, we develop the adversarial training framework with multi-discriminators on the BVS ${D_{t}}$. For each domain, feature choices are same as previous study (Spathis et al., 2022). After the adversarial adaptation, we predict VO₂max on the held-out test set of BVS using 3-fold cross-validation using UDAMA. Within each fold, the dataset is split into 70% training and 30% testing consisting only of target domain samples.

In this work, we proposed the unlabeled domain adaptation via multi-discriminator adversarial training framework (UDAMA) to address the problem of training datasets with various levels of label noise and distribution shifts. By leveraging a large-scale noisy silver-standard dataset and adversarial-based domain adaptation, our proposed method alleviates the domain shit problem and improves the performance on the challenging VO₂max prediction task. Future work includes generalizing UDAMA to other healthcare applications, especially regression tasks involving high-dimensional time-series or wearable-sourced data.

This work is supported by ERC through Project 833296 (EAR) and by Nokia Bell Labs through a donation.