My Health Sensor, My Classifier – Adapting a Trained Classifier to Unlabeled End-User Data

Sleep apnea is a common yet severely under-diagnosed sleep related breathing disorder. The overall goal of our work is to enable “anybody” to use low-cost pre-screening solutions at home in which consumer electronics, like smart phones and smart watches, are used for data acquisition and Machine Learning for data analysis. We use as foundation for our research data from a large clinical study, called A3 study, at the Oslo University Hospital and St. Olavs University Hospital. In this study, sleep monitoring data from several hundred patients is collected and analyzed. A portable sleep monitor certified for clinical diagnosis (i.e., Nox T3 [1]) has been used for data acquisition. Currently, we achieve with a convolutional neural network (CNN) trained on this data a classification accuracy of approximately 80% [2]. However, we cannot guarantee that this model can generalize well to new data from an end-user, because of potential domain shifts. Such domain shifts can for example be caused by characteristics of individual end-user sleep data that are not represented in the A3 dataset and quality issues. The latter is based on the fact that a sleep monitor that is certified for clinical diagnosis produces data with substantially higher quality than that of data collected by end-users at home with consumer electronics. Furthermore, personalization can improve classification in many cases, because individuals are different in terms of physiology, prevalence, sensor placement and the same signals might be measured with different sensors, for example different smartwatch brands.

Domain Adaptation (DA) has recently been successfully applied to address the problem of domain shift. DA aims to improve learning for a predictive task at a target, assuming a source and a target domain for which the distributions of source and target differ [3, 4]. The predictive tasks are the same across the two domains. A sub-field of DA is unsupervised DA, for which labels are provided only for the source data [5], but not for the target data. The vast majority of unsupervised DA literature focuses on non-medical, easy to label, visual open data, without specific usage constraints.

Existing DA solutions cannot be applied to create a classifier that is adapted to the end-user data, because (1) end-users might not want to give us their data for privacy reasons and we do not have the resources to support many end-users, (2) privacy regulations do not permit to share the A3 dataset (e.g., with end-users or third parties), and (3) source and target data are processed together in the majority of existing DA solutions.

This scenario can be generalized. Assume a host which has trained a model with labeled data for classification, but the model cannot be directly used on data collected by an end-user due to the presence of domain shift. Both entities, i.e., host and end-user, do not want (or cannot) share their data with the other entity, as it is very often the case in the medical domain. Thus, the only way to create a personalized classifier for the end-user is DA without direct access to the labeled source data (See Figure 1 for an example of this case, which we investigate in this work, in comparison to a baseline case). Furthermore, it must be considered that the end-user potentially has less computing resources than the host (e.g., lack of dedicated hardware components, lack of sufficient GPU memory to load the data, etc.). As a result even if the host is willing to share her data, performing joint training on the end-user could be problematic.

To address these issues, we introduce a method called Step-wise Increasing target dataset Coverage based on high confidence(SICO) to efficiently perform DA at the end-user. SICO performs unsupervised DA with the use of only a classifier trained on the source domain and unlabeled data from the target domain. One of the fundamental ideas in this work is to release only the classifier trained on the source domain, since extracting personal information from a classifier, especially for time series data, is harder than having direct access to the true data of the individual. At the same time, we take into account the possible lack of hardware resources, since training will be done at the end-user. By using only a single trained classifier and only the data from the end-user, the proposed approach is less resource intensive than normal unsupervised DA.

In this work, we include the following contributions: (1) we introduce SICO, a DA technique which leverages the neuronal excitation of the output neurons as a means to iteratively select high confidence regions to train with. The goal of this process is to incrementally address the existing domain shift, and generalize well to the data of the end-user; (2) we apply the proposed approach on the real-world problem that we are interested in, namely sleep apnea detection and we show its effectiveness for different physiological sensors and datasets; (3) we use the proposed approach to perform DA for a different type of data and for a different task (i.e., digit classification), showcasing its generalizability and providing a comparison basis with related literature.

The rest of the paper is organized as follows: Section 2 presents the proposed approach. Section 3 compares SICO with related literature from unsupervised DA and time-series DA. We place our related work in this location since we discuss methodological differences between SICO and other methods, and we want the reader to know how our approach works before reading the related work. Section 4 describes the experiments we performed. Section 5 presents the choices of important hyper-parameters of SICO, and the results from our experiments. In Section 6 we discuss the key empirical insights. Finally, Section 7 concludes this paper.

In this section, we describe the proposed approach, and provide insights about how and why it works. Furthermore, we investigate important properties, and analyze choices for the principal components of SICO.

There is a large body of work in DA, mainly focused on the visual domain, which is to a large extent covered by existing surveys [9, 10]. In one of these surveys [10] Wang et al. separate DA based on the type of learning (supervised, semi-supervised, unsupervised), whether or not the feature space is the same in source and target domains, and whether the work performs one-step or multi-step DA. We follow a similar separation but due to space limitations we focus on the works that are mostly related to ours. Thus, we separate on the basis of supervised or unsupervised DA, with more emphasis on unsupervised DA, and for the unsupervised DA case whether the proposed work has direct or indirect access to the source data (see Figure 3). Furthermore, we separate for visual and time-series applications or methods. For the works which resemble our work, we compare and discuss about specific similarities and differences.

For the case of unsupervised visual DA, a large body of literature use variations of adversarial DA for visual applications [11, 13, 14, 15, 16, 17]. A good example of an advanced technique that also includes adversarial elements is CyCADA [18]. It is a technique for unsupervised DA which uses a team of models trained on a unified loss which consists of task losses, cycle-consistent loss [19], GAN losses, and semantic consistency loss. The goal is to adapt representations at pixel and feature level while enforcing cycle and semantic consistency loss. Other works for visual applications that do not directly include adversarial training are [20, 21]. Additionally, works exist which, like our proposed method, make use of pseudo-labelling on the target domain [22, 23, 24, 25, 26, 27]. However, all of these approaches take advantage of both source and target domain data during their training phase. Furthermore, they use different methods for classifier adaptation on the conditional distribution of the target domain during their training phase, like entropy minimization, minimization of the conditional distribution’s MMD between source and target, pseudo-labelling with multiple classifiers, and so on.

As mentioned, most of the above techniques perform unsupervised DA with access to labeled source data and unlabeled target data. However, we focus on DA at the end-user with access only to unlabeled target data and a trained classifier at the source data and labels, making it an inherently harder problem. Despite these additional restrictions, SICO yields an on average comparable or superior performance relative to several other well-established unsupervised DA works for the digit datasets. As such, one key advantage of SICO relative to other well-established unsupervised DA approaches is that SICO is inherently designed to handle the source-free DA case. This means that for SICO to work joint training with the source and target data is not necessary.

Additional works that provide a solution to unsupervised DA without direct access to the labeled source data include [28, 29, 30]. Wulfmeier et al. [28] address the problem of degraded model performance due to continuously shifting environment conditions. They develop incremental adversarial DA with which they redefine DA as a stream of incrementally changing domains, to enable a classifier to adapt for example to changing weather conditions or day night circle. Additionally, they propose an extension to use only target data assuming an additional GAN generator which has learned the encoded feature marginal distribution of the source data. Similarly, our method uses the source data indirectly via \(h_{src}\) , but we investigate the normal DA scenario, and not a case of transitioning incremental domain shifts. Li et al. [29] introduce Adaptive Batch Normalization (AdaBN), and use only a classifier and the target data. They standardize each neuron’s output with the average and standard deviation of its output from the images in the target domain. This standardization ensures that each neuron receives data from a similar distribution, no matter which domain the data come from. Using this idea, Zhang et al. [31], train a convolutional network with layers of various kernel sizes for fault diagnosis on raw vibration signals. They then perform AdaBN for the CNN model trained in the source domain. This approach does not train with the target domain data, and is only a form of cross domain normalization for the trained classifier. As a result we hypothesize that it does not take full advantage of the knowledge potential of the target sample. Several recent methods have been developed to address the problem of Source-free DA, like [32], where a multi-term adversarial objective with additional contrastive and rotation loss terms is used to achieve Source-free DA, or [33], which aims to achieve universal Source-free DA (where labels from source and target domains can both include domain-private labels). SICO differs from these works as we do not use adversarial objective, and also, for the scenario we study in this work, our source-target label-set relationship is closed-set (same labels in both domains). Additionally, Yang et al. [30] introduce Generalized Source Free DA (G-SFDA). With G-SFDA, a network is pre-trained on the data and labels of the source domain. Then the pre-trained model is trained in an unsupervised manner with the data of the target domain. Specifically, the model is trained on the target domain via identifying the K-closest neighbors of each datapoint based on feature similarity. A class homogeneity loss on each datapoint is imposed based on the classes of their neighbors. Additionally, to avoid forgetting source-specific or target-specific patterns sparse domain attention is used, where source and target masks hinder forward and backward passes for data from each domain. Hence, specific source and target information flows are imposed to the learning process of the model. Kim et al. [34] introduce a similar technique for Source-free DA. By using a trained model from the source domain (feature extractor and classifier), they train a new feature extractor and two new classifiers which adapt to the target data. One of the two classifiers maintains a source loss with the use of the source classifier, mainly for regularization. The other classifier performs the actual DA via periodically labelling an adaptive low-entropy prototype set per-class. These sets serve as class-representatives through which a per-class similarity metric is established for all target data. An additional confidence-based filtering mechanism is applied on the target data with which a reliable sample is established only when the prototypes of the most similar class are closer than prototypes of the second most similar class. The final loss takes into account both the source loss and the established self loss defined via the aforementioned process. One advantage of SICO in comparison to this approach is that inherently SICO is a simpler approach that needs less memory on a best case scenario. That is since SICO does not need a feature bank and a score bank which are needed in G-SFDA to identify and compare the closest neighbors, or sufficient memory for utilization of multiple models, as is the case for [34]. This is an important advantage in our case, since the scenario we are interested in is inherently resource-constrained (i.e., local device of the end-user patient).

Finally, Chidlovski et al. [35] explore the idea of source-free DA, given access to either a few representative source samples (specifically class means) or a set of classifiers. In the first case, a stacked Marginalized Denoise Autoencoder (sMDA) [36] is used. The class means are fed jointly with the target domain data in the sMDA. The denoised class means are then used to classify the target domain examples, with the use of a weighted softmax distance. In the second case, the output of the cassifiers is jointly fed with the target samples to the sMDA and then the reconstructed classifier’s output is used to predict the class of the respective target domain sample. Both of these approaches differ substantially from our approach. Specifically, SICO does not utilize a generative denoise framework, and more importantly, the target data modify the classifiers and not directly the labels.

For the case of unsupervised DA for time-series data, Purushotham et al. [37], use Variational Recurrent Neural Nets [38], and employ adversarial DA to train in order to achieve DA from the latent representations. In [39], Aswolinskiy et al. propose Unsupervised Transfer Learning via Self-Predictive Modelling. In the proposed approach, a linear transformation Q is learned that minimizes the error identified by a self-predictive model between the source and the transformed by Q target domain. Then a classifier trained in the source domain is applied in the transformed target domain. We assigned this work under the umbrella of unsupervised DA since, the authors present a general DA framework and because it fits the data access criteria we specify. Other works include [31, 40, 41]. Chai et al. [40] perform EEG DA for sentiment recognition by using a linear transformation function that matches the marginal distributions of the source and target feature spaces. They additionally employ a pseudo-labelling approach to transfer confident target data to the transformed source domain which relates to our own approach for choosing confident samples. Natarajaan et al. [41] use DA in order to mitigate the negative effects from the lab-to-field transition for cocaine detection using wearable ECG.

In the context of supervised DA, Persello et al. [42, 43], use Active Learning for DA for classification of remote sensing images. This approach loosely relates to ours, in the sense that it iteratively trains the classifier with a differentiating training set. However, it assumes access to the source data (for data points to be removed from the training set), and a user in the role of supervisor, which assigns labels for chosen samples.

Finally, other works that are related in the context of our work include [44, 45, 46]. Fawaz et al. [44] use transfer learning for time series classification. They evaluate on the UCR archive for all possible combinations, and use a method that relies on Dynamic Time Warping to measure the similarity between the different datasets. Regarding CL, both works included apply DA in the context of semantic segmentation. Zou et al. [45] utilize pseudolabels and self-training to learn the new domain. Contrary to our work, they use a loss which contains terms for the loss of the source and the target domain plus a \(L_1\) -regularization term on the pseudolabels. They optimize in a two-step fashion, optimizing first the pseudolabels and then the classifier weights. Zhang et al. [46] derive their curriculum by separating between the hard task of semantic segmentation from the easy task of learning high-level properties of the unknown labels. They minimize a joint loss which includes terms from both domains. For their easy task they infer the target labels and landmark pixel labels by utilizing training and labelling in the source domain (e.g., retrieving the nearest source neighbors for a target image and transfer the labelling).

The main goal of this work is to perform successful domain adaptation for bio-sensor signals for the purpose of sleep apnea detection. Additionally, we complement the physiological datasets with datasets for digit classification, i.e., USPS, MNIST, and SVHN for two reasons: (1) the majority of the related literature uses these datasets and so we aim to compare SICO with related works on their premises; (2) we want to investigate the generalizability of our approach for different types of data.

In this section we present our results. First we discuss how important hyper parameters affect SICO. Then we present our main results in two sets of experiments. The first set of experiments focuses on sleep apnea detection with physiological sensors. The application scenario that we are interested in focuses on the transition from high quality sensor data to low quality sensor data. Additionally, we investigate combinations for which the low quality data are used as \(D_{src}\) , in order to get a more complete picture of the capabilities of the proposed algorithm. We focus on the NAF signal for the combinations \(A3\rightarrow AE ,A3\rightarrow MB\) , \(MB\rightarrow AE\) , \(AE\rightarrow MB\) , \(AE\rightarrow A3\) , and evaluate the abdominal respiration (Resp A), the chest respiration (Resp C) and the oxygen saturation (Sp02) sensor signals for the \(A3\rightarrow AE\) and \(AE\rightarrow A3\) combinations. The second set of experiments evaluates SICO for digit classification and showcases comparison with several related works for three combinations that are commonly used in literature, i.e., \(M\rightarrow U\) , \(U\rightarrow M\) , \(S\rightarrow M\) .

Devices: Please note that for the sleep apnea set of experiments, the devices used are of the same type across the different datasets ( \(AE,MB\) and A3) i.e., nasal thermistors for the NAF signal, Respiratory Inductive Plethysmography (RIP) (Chest-Abdomen belts) for the Resp A and C signals, and pulse oximeter for the measurement of oxygen saturation Sp02.

Metrics: For the Digit Classification experiments, we use accuracy as performance metric since it is commonly used in related literature for the particular task, assuming a well-balanced dataset. For the apnea detection experiments, we use the kappa coefficient [56] as performance metric since it better captures performance characteristics in a single metric than accuracy, as it takes into account the possibility that two annotators (real labels and predictions in our case) agree by chance. For completion, we present the accuracy, specificity, and sensitivity results in Appendix B. All experiments are repeated five times, and we present the average results and the standard error.

From the above evaluation, we observe that the success of SICO on \(D_{tg}\) depends on how well the initial \(h_{src}\) can generalize on \(D_{tg}\) . This observation occurs directly from Equation (2), since the error of \(h_{src}\) , recursively propagates for all \(\Delta _i\) which constitute the total \(\Delta _n\) . Thus, if \(h_{src}\) does not perform well on \(D_{tg}\) this has a direct repeated negative impact on the whole SICO algorithm. Interestingly, even for the cases for which \(h_{src}\) performs really bad on \(D_{tg}\) (mainly for the cases which include MB and A3 as \(D_{tg}\) ) \(h_{tg}\) is still able to outperform \(h_{src}\) . This potentially happens due to the fact that we expect better performance for high-confidence regions than the average performance, assuming that we trust \(h_{src}\) adequately. This means that a smaller error propagates through the algorithm, than for the case in which we have a lower confidence region. This insight is one of the main inspirations for this work.

However, if we use too few data points as our confidence region, then the classifier trained on this region will not be able to generalize well enough to new data, and will have higher generalization error for the new regions. Thus, there exists a trade-off between confidence and trust on the one hand, and generalization capability on the other hand. This needs to be taken into account to determine how strict the criterion \(C\lbrace h\rbrace\) should be, as this decides how many datapoints a new \(D_{CLi}\) will have.

A natural question about SICO is why should we use a new model in each step (reinitialize), and not use the same model. We made this design choice since we want the subsequent models constructed by SICO to focus on the features of the new domain. As such, the use of a new model at each step, serves as a step towards the minimization of domain-specific biases carried by \(h_{src}\) (or its labelling) as SICO progresses.

We identify two main limitations with the current design of SICO. First, the fact that we need to identify appropriate criterion C in order to decide which datapoints to choose to include in each new \(D_{CLi}\) . This choice can be problematic as it adds additional complexity to the learning task and increases the difficulty of the a priori algorithmic design. A potential extension to SICO, which could resolve the aforementioned issue, is to apply a probabilistic criterion, e.g., based on the max class probability of each datapoint, instead of using a “hard” criterion C as a threshold of acceptance for a given datapoint. This is a more natural approach that better captures the probabilistic nature of density estimation that the classifiers perform, and importantly removes the need of tuning C. As such, we hypothesize that it can yield even better results than the original hard acceptance/rejection method. However, in preliminary experiments (see Section 5.1.1) we were not able to showcase improvement in performance in relation to the original method.

Second, though the use of SICO guarantees that no data will be exchanged between the lab/hospital and the end-user patient, \(h_{src}\) is trained on the lab dataset. As such, the use of information leakage attacks against \(h_{src}\) can potentially leak information about the lab dataset. To obtain protection from such attacks, we can, e.g., strategically clip and add noise to the gradients of \(h_{src}\) during training. With such an application, certain dataset traits could become obfuscated and theoretical guarantees like differential privacy can be obtained. Alternatively, we can adversarially train \(h_{src}\) so that certain adversarial defences can be learned to obtain protection against certain information leakage attacks.

The primary motivation for our work is to enable end-users to create personal classifiers for health applications without labeled data. In particular, we foresee a collaboration in which a host releases a classifier \(h_{src}\) , and the end-user (i.e., a patient), uses her data and the classifier to create a new personalized classifier which is adapted to the domain of the end-user. For example, for the sleep apnea case, if we have a person whose datapoints (i.e., one-minute windows) are different due to a variety of potential factors than the datapoints \(h_{src}\) has been trained with, we can expect that by applying a suitable method we can get a better performing classifier for the particular individual.

In this work we achieve this by performing SICO. SICO iteratively adapts classifiers to the new domain based on high-confidence data from the previous classifiers, and without the need of the source data and labels. Based on our scenario, we are more interested in the case of performing DA at the end-user, but obviously SICO can also be applied at the host. We apply SICO for the case of sleep apnea detection, and use a large real-world clinical dataset for its evaluation. Additionally, we experiment with two open sleep monitoring datasets and the MNIST, SVHN, and USPS datasets for digit classification. By this, we achieve (1) reproducibility, (2) demonstrate the generalizability of SICO for another task, and (3) achieve comparability with related work. The results from these experiments show consistently better performance of \(h_{tg}\) in comparison to \(h_{src}\) , as expected. Depending on the dataset, we get an increase in kappa of up to 0.24. For the task of digit classification DA, SICO also achieves a consistently good performance. Despite the additional limitations of the scenario that we investigate, the performance of several well-established related works is lower than the presented results of this paper.

As a next step, we are interested in investigating how well can the technique be applied if the source classifier is trained under differential privacy guarantees. Another interesting application to investigate is the combination of the proposed approach with a style transfer method with the goal of increasing the DA capabilities of the approach for tasks containing different types of sensors. Furthermore, since the digit classification experiments indicate good generalizability for SICO, it would be interesting to apply SICO on other image-based health applications (like, e.g., MRI classification on various tasks, or telemedicine applications) in cases where it is appropriate, i.e., existence of a domain shift and inability of data sharing. Finally, assuming that we choose proper criterion C, we hypothesize that we could apply SICO to properly adapt in a dataset which is balanced with different class frequencies in relation to the source dataset. This, in the context of condition detection, means that we adapt from patients who do not have many pathological datapoints to patients for which the condition is more expressed.

In Appendix A, we discuss additional design decisions made during the algorithm.

Appendix B shows the Accuracy, Sensitivity, and Specificity for the Apnea detection DA experiments. A characteristic which was not identifiable from the kappa values is the balancing effect SICO has on specificity and sensitivity. When using \(h_{src}\) , in many cases we obtain high specificity at the expense of high sensitivity (e.g., NAF: \(A3\rightarrow MB\) , NAF: \(AE\rightarrow MB\) ). We observe that for \(h_{tg}\) this effect is minimized. This means that \(h_{tg}\) obtains increased sensitivity compared to \(h_{src}\) . This is a positive characteristic since sensitivity (i.e., the percentage of the correctly detected apneic minutes) plays a crucial role in the context of sleep apnea detection as low sensitivity can lead to a false negative diagnosis, making the system inherently untrustworthy.