Sleep is a significant contributor to leading a healthy lifestyle. Each day, most people go to sleep without any idea about how their night's rest will be or how they can leverage their data to improve it. For an activity that humans spend near a third of their life doing, there is a surprising amount of mystery around it. Despite current research, creating personalized sleep models in real-world settings has been challenging. Existing literature provides several connections between daily activities and sleep quality. Unfortunately, these insights do not generalize well in many individuals. For these reasons, it is essential to create a data-driven personalized sleep model. This research proposes a sleep model that captures causal relationships between daily activities and sleep quality and presents the user with specific feedback recommendations to improve sleep quality. Using N-of-1 experiments on longitudinal user data and event mining, the model generates a probabilistic understanding between lifestyle choices (exercise, eating, circadian rhythm, environmental selection) and their respective impact on sleep quality. Our experimental results identified and quantified relationships while extracting confounding variables through a causal framework. We then utilize the generated model to provide lifestyle recommendations to optimize sleep outcomes in a context-aware health recommendation system.
HealthRecSys ’20, September 26, 2020, Online, Worldwide © 2020 Copyright for the individual papers remains with the authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). This volume is published and copyrighted by its editors. 1
Lifestyle factors have a high impact on our health outcomes. Our eating, sleeping, and movement patterns determine large parts of our short and long-term health [Hills et al. 2015; Nag 2020; Nag et al. 2018; Shochat 2012]. Keeping track of how we behave in diferent contextual situations, and the impact these behavioral patterns have on our health is dificult for medical professionals and individuals. At the same time, with the increasing prevalence of chronic diseases such as diabetes and hypertension, understanding the efect of lifestyle on diferent aspects of our health becomes a critical research challenge [Lee et al. 2017; Sarris et al. 2014]. The rising popularity of wearable and IoT devices provide us with an opportunity to address this problem computationally. There is a considerable body of research dedicated to finding and logging life events from multimodal data streams [Gurrin et al. 2014; Oh and Jain 2017; Sellen and Whittaker 2010]. A multitude of consumer devices such as smartwatches and smart home systems measure aspects of our daily life as events and data streams and control our local environment [Casino et al. 2018]. Using the data streams and events captured by these devices, we can find recurring behavioral patterns and associated health outcomes to create an explainable rule-based model of the person [Pandey et al. 2018]. Explainability is a desired quality in health prediction and recommendation systems as it can verify the quality of predictions and builds user engagement and trust in the system.
We can use such a model to provide the right guidance at the right time for health management. Health recommendation systems provide us a way to apply cybernetic principles to manage a person’s health [Nag et al. 2017]. Using lifestyle interventions, we can build a navigation system that guides us through our day much in the same way that modern navigation systems inform drivers about the most optimum path towards their destination [Nag and Jain 2019]. We need to design a context-aware personal recommendation system that changes the person’s context with every event that happens during the day. The dynamic context allows us to provide an optimal recommendation at every point of the day, and calibrate the recommendations as diferent events occur.
In this work, we present a sleep recommendation system that considers various lifestyle factors as contextual variables and possible recommendations for optimizing a sleep parameter. We create a rule-based model to understand the efects of diferent lifestyle factors (such as exercise during the day, and mealtimes) on sleep parameters. These rules are used in a recommendation system framework for providing the most efective interventions at any time throughout the day. These interventions could be lifestyle recommendations presented to the user or a command to one of the IoT or smart devices that control the user’s environment (e.g., HVAC systems, Light bulbs, Music or ambient sounds). 2
Our work spans across context-aware and health recommendation systems, sleep specific monitoring and prediction, and causal event mining. We review the current approaches and limitations below. 2.1
The field of sleep and health recommendation systems is relatively new. Studies have explored the pitfalls of using the conventional recommendation systems for health and devised alternatives using entity properties and relationships [López-Nores et al. 2012]. Context-awareness is an essential quality for health recommendation systems [Schäfer et al. 2017]. Context-aware recommendation systems (CARS) have been explored in diferent domains [Villegas et al. 2018]. CARS have traditionally incorporated context information in collaborative filtering models in one of three ways, 1) Contextual Pre-filtering, 2) Contextual Post-filtering, and 3) Contextual Modelling [Adomavicius and Tuzhilin 2015]. There can be diferent types of contextual information relevant to a recommendation system. These usually fall into one of the following categories: temporal, location, individual (user characteristics), activity (about the activity), and relational (when multiple entities are involved)[Villegas et al. 2018]. We have adopted a contextual modeling approach and incorporated the contextual information in the rule-based model itself. Multiple studies have explored personalized recommendation systems for diferent aspects of user-health, such as diet [Khan et al. 2019]. These utilize diferent learning techniques to develop personalized models for individuals but usually lack explainability, which is an important characteristic of health recommendation systems. 2.2
Polysomnography is considered to be the gold standard for understanding sleep quality. The test records various metrics such as brain waves, oxygen levels in the blood, heart rate, breathing, and eye and leg movements [Kushida et al. 2005]. It requires a sleep expert and multiple medical sensors. The study’s accuracy does come at the cost of needing too many resources and equipment to be performed every night reliably. Actigraphy is another popular technique used to measure sleep quality. It measures sleep quality using a wearable device (e.g., a watch) by recording movement during a sleep event. Its simplicity comes at the cost of accuracy, as it can only infer sleep quality via movement measurements. For our study, we used a combination of actigraphy and sound to record movements during sleep events. Several sleep applications use audio from the phone mic to record and report sleep quality (e.g., SleepCycle, Sleepscore).
Previous works have attempted to predict sleep quality using smartphone mic data in conjunction with machine learning [Min et al. 2014]. Other studies have attempted to use actigraphy graphs and utilize Deep Learning to predict sleep quality [Sathyanarayana et al. 2016]. There are even studies that attempt to forgo the idea of tracking sleep and use factor graph models based on daily activity to predict sleep quality with 78% accuracy [Bai et al. 2012]. While these models are useful, they do not address the individual variability in sleep and do not provide a way to incorporate diferent lifestyle aspects. 2.3
Current literature has made many attempts to identify daily activities that afect a person’s sleep quality. Studies have shown that sleep and exercise are related, and higher physical activity levels can lead to better sleep latency [Yang et al. 2012] [Kline 2014][Kelley and Kelley 2017]. A systematic review has also shown that dietary patterns and the types of food eaten throughout the day lead to better sleep quality and duration [St-Onge et al. 2016]. The environmental factors (temperature and humidity) are also vital to our sleep duration and quality [Troynikov et al. 2018]. From these studies and many more, it is clear that choices made throughout the day afect the quality of the next night’s sleep. 3
Creating a model of the person’s behavior and health is central to building personalized health recommendation systems. In this work, we present an approach to build a rule-based explainable model for predicting sleep outcomes in diferent contextual situations. We apply event mining [Jalali 2016] principles to perform N-of-1 experiments on a user’s data [McDonald et al. 2017] that allows us to find causal relationships between diferent lifestyle events and biological outcomes. The process is described in figure 1.
Event mining allows us to discover and specify patterns between diferent events in a person’s life. We utilize these patterns to create hypotheses that might describe a person’s behavior. A hypothesis needs to specify the intervention event and the associated confounding factors that afect the relationship between the intervention and the outcome. The confounding factors are specified using the temporal delay operator, Δ[, ], that relates the events that occur within the specified time interval [, ]. The confounding factors are specified as a set of patterns between the lifestyle events and the outcome. Thus, a hypothesis would be specified as −→ , where we want to measure the causal efect of intervention on the outcome event while controlling for the events specified by the set of patterns . These patterns and hypotheses can be derived from existing knowledge and human intuition, allowing us also to leverage the results of population studies performed in clinical and controlled settings.
Combining the event mining operators with causal inference principles allows us to perform N-of-1 experiments on the user’s longitudinal data. We utilize the candidate hypothesis specification to create diferent subsets of data based on the values of co-occurring confounding events. These subsets minimize the variance in the outcome due to confounding factors and mimic a version of the do-operator [Pearl 2009]. We can use diferent statistical techniques, such as linear regression or t-tests to find the efect of the intervention event on the outcome within each subset. This allows us to ifnd the efect of the intervention on the outcome in an unbiased manner, and if we can capture all the confounding variables in the set of patterns, we would obtain the causal efect of the intervention on the outcome. The result of the test is stored as a conditional rule that uses confounding variables and the intervention event as the predicate. The distribution of the outcome events in the subset is used to make a prediction.
A set of these contextual rules can be used to predict health outcomes. We would need to find the most relevant rule by matching the user’s current context with the set of rules and utilize it to make the prediction. We describe it in more detail in the next section. A rule-based model, while lacking in complexity, ofers the advantage of explainability and online training. Every prediction and recommendation generated from this model can be explained using the associated contextual factors, thus eliminating recommendations based on spurious relationships. This is an essential characteristic of health models and recommendation systems. As the user behavior changes over time, the model needs to adapt to the changing user parameters and trained using the latest observed data. We can easily update the rules by updating the outcome variable’s distribution whenever the rule matches the user’s current context. 4
The rule-based health model allows us to find the health outcomes in diferent contextual situations. The user’s activities during the day (such as exercise, meals, work-related stress) and their local environmental parameters (such as temperature, humidity, and ambient light and sounds) determine these contextual variables. Thus, we can utilize this model in a recommendation system setting to determine the set of parameters (both user behavior and environmental variables) for optimizing a health outcome (e.g., sleep quality).
Every action taken by the user and every environmental exposure changes the user’s health state [Nag 2020], which changes the context for future actions and recommendations. We need to retrieve the relevant events from the user’s events and data streams that impact their health state [Pandey et al. 2020]. Diferent contextual parameters are defined as aggregations of these events. For example, Total Screen Time during the day is an important confounding factor for understanding an individual’s sleeping habits. It can be determined by aggregating the duration of all the screen activity events (such as working, watching TV, and social media activity) during the day. These aggregations can be performed using events based triggers encoded as condition-action rules. As new events appear in the person’s events log, the retrieved events can be aggregated to change the user’s live context parameters. We can use the latest context values to provide a set of recommendations that would optimize the user’s health outcomes.
We match the live contextual parameters for the person with the contextual parameters of the various rules present in the model. If the current context matches multiple rules, then we utilize the rule with the highest likelihood of the desired outcome. Once we have identified the rules that match the current context, we can utilize the unmatched contextual parameters and the intervention event to find the set of parameters that can lead to the optimal outcome. We can either present the recommendation to the user (if the recommendation is an action to be taken by the user) or communicate with a smart device that controls the user’s environmental context (e.g., smart home devices, HVAC systems, smart bulbs). The recommendation system produces a set of actions that would maximize the likelihood of the optimal outcome; thus, the proposed recommendation system is diferent from typical recommendation systems as the recommendation consists of multiple items. Since any event during the day can change the user’s context, the recommendations are recomputed anytime an event changes the user’s context. This process is depicted in figure 2. Thus, at any point during the day, the recommendation system would provide a list of timestamped actions to be performed by diferent agents (the user, or an automated device) to optimize the sleep outcomes. 5
We ran experiments to create a personal rule-based model for optimizing a person’s sleep quality metrics by providing lifestyle and local environmental recommendations. We utilized data collected by one individual for more than two years using readily available consumer applications and wearable and IoT devices. We performed two sets of experiments on the collected dataset to create and evaluate the model. The first set of experiments find the average causal efect of input variables on sleep quality metrics. We used Welch’s t-tests and a p-value of 0.05 to determine statistical significance. The second set of experiments tested the prediction accuracy for a static pre-trained model vs. an online training model. The data set includes exercise and lifestyle parameters for a 31 year old male collected continuously over 2 years via the user’s Garmin Fenix 5 smart watch, their smartphone, and an IoT sensor that collected local temperature and humidity values. Sleep Cycle was primarily used to keep track of sleep events. Apple Health Kit was used to help compile sleep quality measures recorded by the Garmin smart watch, daily step counts, and daily floors climbed. The acclerometer measures of the smartwatch and the audio recordings of Sleep Cycle were used to create sleep quality measures. Strava was used to keep track of exercise events. An image based food log recorded feeding times with phone camera metadata, and a SensorPush IoT sensor was used to collect temperature and humidity during sleep events. All of these data sources were then temporally matched to record lifestyle events that took place throughout the day accurately.
We used the thresholds mentioned in Table 1 and Table 2 to convert the data streams to relevant events for the event mining analysis. We used nine lifestyle/environmental events: Previous Night’s Sleep Quality Measures, Exercise Minutes in the Day, Interval Between Eating and Sleeping, Minutes Awake Between Sleep Events, Temperature, and Humidity when going to bed. The possible output events are sleep quality measures (Table 1). We used 70% of the data to build the model, and 30% of the data to test the model. The train-test split was created based on temporal order. 5.2
We perform diferent N-of-1 tests on the user’s data to find the average efects of diferent lifestyle and environmental events on sleep quality parameters while controlling for other lifestyle parameters as confounding factors. We treat one of the input event’s possible values as the baseline and compare the distribution of the outcome variable for other values of the event with the baseline distribution. If changing the input event value causes a significant change in the outcome distribution while controlling for confounding variables, the rule is deemed significant. This gives us the causal efect of different values of an input event on the observed outcome. We repeat this experiment while controlling for diferent sets of variables and aggregate the causal efects to find the input event’s average causal efect. If the diference is not found to be significant, then we merge Variables
the two distributions and use the combined distribution at the time of contextual matching.
The results of these experiments are in Figure 3. One interesting observation is that an average temperature(60-67 ) seems to improve every sleep quality measure except for sleep latency. This is an important observation as it shows that not all quality measures are correlated with each other and that an improvement in one does not necessarily equate to an improvement in all other sleep quality measures. Another interesting insight is that exercise improves sleep latency the most. On average, we can tell that exercising a lot will reduce sleep latency by 10.5 minutes with just a small workout will help reduce sleep latency by an average of 8 minutes. 5.3
We also want to demonstrate the contextual matching of rules and test the accuracy of the model’s predictions, as that will determine the eficacy of any recommendations we generate. We train a linear regression model corresponding to every rule in the model, and use the data subset that matches the rule to train the model. This model is then used to predict sleep outcomes for situations matching with the rule.
We used two training strategies for the prediction model; 1) Pretrained static models, and 2) the warm start online training. We expect that over time the user’s sleep behavior would change, and thus an online learning strategy would eventually start outperforming the pre-trained model.
The model’s input features are Exercise minutes during the day, Feeding Time, Time Awake, Humidity, and Temperature while going to bed. We match the user’s context with the context of the rules, and the most significant rule that matches the context is used to provide the recommendation.
We create a set of contextual variables at the end of the day for each day in the dataset. These values are then used to find a matching rule. If multiple matches are found, then we used the rule with a higher statistical significance. The linear model associated with the matched rule would then be used to predict the sleep outcome parameter. The key diference between the pre-trained and online models is that the online model would be updated continuously using the data in the test set. This way, the online model has the opportunity to adapt to the user over time. The results of the model predictions are in 4. The results illustrate an improvement in the performance of the online model over the pre-trained model. Eventually, we expect the online model would achieve a much smaller MSE as it adapts to the changing sleep behavior exhibited by the user. 6
We have shown the need for and created a sleep model that utilizes event mining and causal inference principles to provide useful feedback about the relationships between lifestyle events and sleep quality. With enough data, this model can be potent. Coupled with the context-aware health recommendation system, it can give people control over their sleeping habits that have not been previously possible. The insights from the model are easily understandable and should promote user engagement as the recommendations are not coming from a black-box model but are simple relationships between daily habits. The context-aware recommendation approach allows us to provide recommendations at diferent points during the day. Even if the user fails to follow any recommendations, we can provide them with a new set of recommendations and modify their local environment to best suit their sleep requirements. This helps us move from recommendation-based guidance to navigational guidance.
Although this framework incorporates many useful data sources and provides useful insights into users’ sleep behavior, there are many ways to improve. Many other lifestyle factors afect our sleep outcomes but are not included in our study, such as stress and nutritional intake. Our events based framework would allow us to include these events and data streams with minimal additional efort.
We have proposed a recommendation system to optimize one health outcome. However, in a real-world application, the users may want to optimize multiple outcomes simultaneously. This can be an exciting opportunity for the recommendation systems research community, and we hope to stimulate future work expanding to a larger user base and with diferent applications.