The ubiquitous nature of wearables generates considerable potential for data collection and analysis but comes with the issue of protecting user privacy. Some important privacy considerations are protecting patient confidentiality [1], protecting against security breaches [2], and protecting against researchers using user data in a way that the user did not intend [3,4]. Individuals may be more willing to participate in studies and disclose information if these concerns are alleviated [5,6]. Privacy breaches can occur when data servers are compromised but can also occur when data are shared for legitimate purposes by well-intentioned members of the medical community [7]. In this paper, we use privacy to denote the aspects related to protecting the identity, personal information, and use of the data of users.

Mobile health (mHealth) data are often related to the cognitive, behavioral, and affective states of users, making such data highly sensitive. Therefore, we want to ensure that individuals’ confidential health information is not leaked to others. Wearables passively record a range of medically relevant data, such as temperature, heart rate, and electrodermal activity (EDA). As people may carry these items with them throughout the day, this allows for high-frequency collection of data from more people, who may have a greater variety of health conditions, than ever before. Such rich data collection opens up the possibility of using increasingly powerful but data hungry machine learning methods in the analysis of these data [8].

In this paper, we apply predictive machine learning models on the publicly available Wearable Stress and Affect Detection (WESAD) data set published by Schmidt et al [9]. In particular, we focus on models that can be trained by fitting a function, potentially nonlinear, to the data using gradient descent and variants thereof. Many of these models make few assumptions about the structure of the underlying data-generating process. To improve privacy, we propose leaving each user’s data on their personal device and training our models using federated learning. In federated learning, there is no single server that contains all users’ information. Instead, model training occurs on each individual’s device, and only model parameter updates leave the user’s device. This allows for more user privacy by maintaining data only on individual user devices. In addition, as explained later, federated learning is still able to take advantage of some shared information across individuals. Thus, it can alleviate some of the concerns in analyzing mHealth data, such as user heterogeneity and privacy preservation.

Before further describing federated learning, we establish some common terminology. The goal is to produce a model, often a neural network, trained using data from many individuals. Each individual will have a mobile device, which we call a user device—in the networking literature, these are called clients—and we will not make a distinction between an individual and their user device. Each user device has its own private data, and the data will never leave the user device. Storing private data on each user device instead of uploading them centrally is the source of improved privacy provided by federated learning. We will also have a single device controlled by the researchers that can communicate with each user device, and we will call the former device the server. The server will coordinate the training procedure and store a copy of the model but will not have any private data uploaded onto it. The training process is iterative, as is common for many machine learning models. Each time the model parameters on the server change, we will call that one server training round (or server training iteration).

In Figure 1, we compare how a model is trained on a central server (top) with how a model is trained using federated learning (bottom). In federated learning, one server training round comprises the following three distinct parts:

The server then repeats this process by selecting a new cohort of user devices for each server training round. Thus, the data from every user device contribute to the training of the model. Algorithm 1 (Figure 2) presents the pseudocode, and a full algorithmic description is provided in Multimedia Appendix 1.

The model trains using updates from the data on all user devices; however, individual, private, data remain only with each user device. This protects the privacy of users in several ways. Under federated learning, private data are significantly less vulnerable to the central server being compromised. The server never contains any raw private data and generally contains only the current model parameters. This protects users from being subject to adverse consequences of failings in the server’s security. Another benefit for users under federated learning is private data cannot be reused for other purposes at a later date. It cannot be shared or released either accidentally or in good faith but with inadequate anonymization measures. Finally, federated learning gives users the freedom to withdraw access to their data—both past and future—at any time without relying on guarantees from the researchers that any previously collected data will be deleted.

Federated learning has been successfully applied in complex real-world applications, most notably in training the next-word prediction model for Google’s Keyboard [10]. We refer interested readers to excellent surveys of federated learning [11-13] and federated learning software and data sets [13].

There has been substantial research on privacy in mHealth and mood prediction using machine learning. Protecting the data of individual users stored in mHealth apps is important because the apps contain personal information about the user, information used to make treatment decisions, and information that could be misused for financial gain, among other concerns [4]. Users’ willingness to disclose personal information in mobile data varies according to their demographics and personal characteristics [14-16]. Concerns about privacy vary among different age groups; for example, some older people are less willing to use mHealth services [6,14] and are more likely to cite privacy concerns [6]. As such, it is important to consider the target population when considering the amount of privacy guarantee needed. Increased privacy concerns about health information technologies reduce patients’ willingness to share information and reduce their positive attitudes toward the technologies [14,17]. In the same vein, reducing privacy concerns makes patients more comfortable in sharing their health information [5,17]. Sometimes, individuals are willing to give up some privacy if they consider mHealth technology as beneficial; however, this depends on the sensitivity of the information they are providing [18]. Due to these privacy concerns, there are many innovations for increasing patient trust and privacy protection [19].

Many researchers have used machine learning methods, including neural networks, random forests, and support vector machines, to predict outcomes such as depression, mood, and stress using mobile data [20-26]. Various physiological and behavioral features help predict these outcomes, such as skin temperature, EDA, 3-axis accelerometer data, mobility, sleep, and self-reported histories [20-23,27,28]. Detecting aberrations in some of these measures can help identify early signs of mental illness [16].

Personalization techniques in machine learning allow parts of a model to be specific to each user while other parts are shared between users, allowing both the sharing of data to estimate certain parameters and fitting to unmeasured covariates of each individual. To an extent, both unique traits among different individuals and commonalities across users are helpful in prediction. Some studies have taken advantage of this and grouped users based on common traits, finding that this improved mood prediction accuracy [8,24,29]. In the same vein, accounting for individual user differences is valuable in prediction using machine learning [8]. There is growing recognition within the mHealth literature [30] that there are trade-offs between individual and centrally trained models and that often there are reasons to make them work in tandem.

In the Methods section, we describe federated learning in the context of mHealth by addressing some of the issues raised in the mHealth literature. Specifically, we discuss the application of federated learning using neural networks to make predictions from mHealth data. We address practical considerations for researchers who consider using federated learning for their study. We then evaluate the performance of federated learning on an mHealth data set to predict subjects’ affective states.

Federated learning is a nascent field within machine learning [31], which arose, at least partially, in response to the opportunities and difficulties presented by personal devices that can record information and perform computation [11]. As such, many practical challenges in mHealth, such as privacy [32,33], user heterogeneity [34], user hardware constraints [35,36], and even incentivizing voluntary participation [37], are already being studied by the machine learning community. The medical science community has also proposed ways to share information while preserving privacy, such as developing models in a distributed manner [38,39], using a federated patient hashing framework for similar patient matching [40], and using collaborative privacy-preserving training to detect protected health information in text [41]. A recent paper describes how federated learning can be useful to various stakeholders in the mHealth field [42].

Federated learning provides a middle ground between extremes in privacy and data utilization. One extreme is collecting data centrally on the server and performing all analyses and training there. The other extreme is individual training, in which each user device trains a completely separate model on its own data.

In centralized learning, we take all individual users’ raw data and store them in one location. This unlimited access gives researchers the most flexibility when analyzing the data—clearly any analysis that can be performed with limited data access can be performed with unlimited data access—and the best chance of extracting useful insights from the data. However, it is likely that some individuals are reluctant to have all of their raw data stored in a central location, as the central location could be hacked or because individuals feel that their data are too sensitive to be released. A model trained using federated learning is useful because it is trained without putting all individuals’ raw data together in one place. Not only does this help alleviate concerns about data privacy, but it can also help potential subjects be more willing to engage in studies, as their raw data will not be sent to a foreign central location. As we show in our case study, a model trained in a federated way performs almost as well as a centrally trained model. We refer to the former as the federated model and the latter as the server model.

In individual training, each user device trains a separate model using only their own data; therefore, no information, besides potentially their final model, leaves the user device. We refer to this as the individual model. This provides a very high level of privacy. However, it prevents information sharing across users during model training. Federated learning takes advantage of the information shared across users, which can be especially useful when analyzing health care data. Each person’s characteristics and health care need personalization; however, there is also a lot of useful information that can be shared across people, as every person’s distribution is not entirely different. In training, the algorithm can learn from similar users and apply this knowledge when predicting another user’s outcome. It is significantly easier to use a federated model to make predictions for a new patient because there is a single model that is trained using data from many people, and we do not need previous data from the new patient. In this way, the federated model can do better than the individual model. Thus, federated learning can help improve the prediction accuracy while preserving privacy.

The decision to use federated learning must be made when designing the study so that one leaves the data on the user devices and trains the model using federated learning. If data have already been collected from each user device and stored together on a server, privacy has been violated, and we cannot use federated learning to retroactively make the experiment privacy preserving. A small burden might need to be placed on user devices to use federated learning. To minimize disruptive usage of user device bandwidth and compute resources, the local update on the user device should take place while the user device is connected to a nonmetered internet such as Wi-Fi, and the device should be idle and plugged into a charger. To avoid inducing bias in the training process by the overuse or underuse of certain user devices because of their passive adherence to these restrictions, we may need to request that user devices be available for local training during certain times of the day. The structuring of these times should be considered during the experimental design phase. In addition, the server should follow certain protocols to minimize the privacy loss. The model updates may contain some information about the data on the user device; however, the privacy loss from these updates can be mitigated by not storing or viewing any locally updated parameters during each server training round. Similarly, information about when each user device participated in the training should also not be stored.

Although concerns about user privacy deter users from sharing health information, the perceived effectiveness of information security reduces these concerns [5,17]. However, federated learning is not the only tool for privacy, and it does not alone provide perfect privacy. Depending on the privacy requirements of the study, federated learning may or may not be the best method to use. We provide a few examples to illustrate cases where federated learning may not be the correct method.

Federated learning is not a silver bullet against privacy loss, and it is not a black-box tool that can be simply tacked onto an existing study. However, when properly integrated into the study design, federated learning can significantly reduce the loss in users’ privacy while incurring only a small reduction in model accuracy.

To assess the practicality of using federated learning on mHealth data, we compared its predictive performance with that of other predictive machine learning models. We used the WESAD data set for this purpose [9]. We had a 3-class classification problem using physiological and motion data to predict whether each subject was, at the time, performing a task designed to elicit a neutral, stressed, or amused affective state.

We used the WESAD data set because it is publicly available and thus easily accessible by other interested researchers, and it is in the University of California, Irvine, Machine Learning Repository as a data set that measures stress in users using wearables. The data were collected from 15 subjects and contained physiological and motion data measured simultaneously by a wrist device and a chest device during specific tasks designed to capture 3 different affective states: neutral, stress, and amusement. There was approximately 36 minutes of data for each subject. Using 30-second windows for all 15 subjects, 1087 windows were generated in total. Of these windows, 53.45% (581/1087) were collected during the baseline task, 29.99% (326/1087) during the task designed to elicit stress, and 16.56% (180/1087) during the task designed to elicit amusement.

The authors of the data set found that using physiological and motion data from the chest-worn device was more informative than using physiological and motion data from the wrist-worn device in the 3-class classification problem [9]. The chest-worn and wrist-worn devices differed slightly in the modalities they measured. Thus, we restricted our analysis to the data collected using the chest-worn device. These include electrocardiogram, EDA, electromyogram, respiration, body temperature, and 3-axis acceleration (x-axis, y-axis, and z-axis) measurements collected at 700 Hz. For additional information, we refer interested readers to the original WESAD paper [9].

The authors of the WESAD paper used feature extraction to identify useful features for their predictive models, which often included the mean, SD, minimum, and maximum of measurements. For simplicity, we used these four summary statistics, calculated over 30-second windows, of each of the 8 measurements for each subject as the features in our models. The code to extract our features was derived from Matthew Johnson’s GitHub repository [46]. The limitations of our analysis include whether the subject was in the intended affective state and the set of features used.

As we had access to each subject’s data, we were able to use a server model. However, in many situations, researchers may not want to access each subject’s raw data for privacy reasons. As true federated learning would not store raw data on a central server, we could not use a true federated model and instead use simulated federated learning. Nonetheless, we saw this as a feature and not a limitation. As we had the server data, we could compare the accuracy between our simulated federated model and the truth in the centralized data. If we had implemented true federated learning, we would not be able to compare performance with the centralized, or server, model, or compare performance with the individual model. Much of published federated learning research uses federated learning simulations for experimental results [31,32,34,36].

This paper discusses the advantages and challenges of using federated learning as a predictive model in mHealth data collection and demonstrates an empirical example in which federated learning could have been effective. Furthermore, it shows empirical evidence that a federated model can make accurate predictions, is compatible with personalization, and has the added privacy advantage of not storing raw data from individual users. The personalized server model performed the best on the chosen data set, followed closely by the personalized federated model. As a federated model offers more privacy for users than a server model, there is evidence to suggest that federated learning may be a valuable option for collecting and analyzing sensitive mHealth data.

The decision of whether to use federated learning is based on a combination of factors, including how much privacy is required, how much data are available, and what resources are available to implement the federated model. Each of the models we tested has advantages and disadvantages. The server model uses all raw data stored in one place for easy access and future use. However, this poses the risk of breaching user privacy if the server is compromised and the risk of the data being accessed and used in a way that the users did not intend. An individual model can maintain user privacy by keeping the data on the user’s device and requires less engineering to implement than a federated model. However, the individual model did not perform as well as the personalized server model and personalized federated model in our tests. Nonetheless, if a researcher has a lot of data from each user, using an individual model can be useful for prediction, and a federated model may be unnecessary. The federated model provides more privacy than a server model, as well as reasonably accurate predictions, but it requires some design work before collecting data and software engineering to implement. Moreover, although federated learning provides an added degree of privacy, it does not guarantee privacy, as explained in the Methods section. To achieve stronger privacy guarantees in federated learning, a substantial amount of software engineering is needed.

Federated learning is used for predictive modeling; therefore, implementing it limits the research questions that can be answered from the data. Researchers may often wish to pursue research questions aside from prediction; in the future, it would be interesting to extend federated learning to estimate treatment effects. Other future work includes developing ways to handle missing data for time series in the context of federated learning. In addition, Python TensorFlow has recently released capabilities to apply federated learning in their TensorFlow Federated package, and it would be interesting to compare it with our implementation.

Previous research in the mHealth field has explored privacy-preserving methods. Some of these studies do not involve neural networks and federated learning [38-40]. One study applied a methodology similar to federated learning [41], although it selectively updated parameters from individual users, whereas we used multiple local updates between transmission rounds. As federated learning started in the machine learning community, there have been papers in the computer science and engineering fields regarding collaborative privacy-preserving learning tested on mHealth data [33,51]. Our work bridges the mHealth and machine learning communities by discussing practical considerations for mHealth researchers who want to consider implementing federated learning. Furthermore, we evaluate the effectiveness of federated learning on a publicly available mHealth data set. Our code is available on GitHub to encourage reproducibility of our results.

As mentioned earlier, many mHealth researchers have applied machine learning methods to predict mood and stress [20-23,25,26]. Research has shown that accounting for both unique traits among different individuals and commonalities across users is valuable in prediction [8,24,29]. We demonstrate the use of federated learning as a privacy-preserving method that also has these advantages.

This paper discusses federated learning and its importance in the field of mHealth. For example, federated learning provides added protection to preserve patient confidentiality and inhibits the use of data in a way that the participants in a study did not intend. As federated learning does not store raw data from individual users on a central server, there is no possibility of a central server being hacked and raw data leaked. This provides more privacy to users when recording sensitive data.

Protecting user privacy is critical in mHealth. Having more privacy protection measures in place may encourage people to participate in a study and to be more willing to disclose information that is useful for treatment and research. Federated learning offers additional data privacy and can overcome some common challenges in mHealth data by addressing user heterogeneity and taking advantage of commonalities across users. As such, federated learning has considerable potential to help advance mHealth research.