SleepNet: Attention-Enhanced Robust Sleep Prediction using Dynamic Social Networks

Contagion is ”rapid communication of an influence (such as a doctrine or emotional state)” (con, 2023). Contagion in behavior is the production of similar behavior in response to interaction with another person (Hatfield et al., 1992; Nickerson, 2021; Centola, 2018). One example is the tendency to mimic and synchronize expressions, vocalizations, postures, and movements with those of another person. Multiple works have explored how health outcomes such as obesity (Christakis and Fowler, 2007), tuberculosis (Klovdahl et al., 2001), pneumonia (Meyers et al., 2003), and severe acute respiratory syndrome (Meyers et al., 2005), can spread in a network suggesting the existence of contagion in health behaviors (Smith and Christakis, 2008). The work in (Mednick et al., 2010) suggests the contagion of sleep behavior in a social network. An empirical study was conducted using data collected from students in grades 7–12 in the National Longitudinal Study of Adolescent Health (Udry, 2003) over 8 years. A social network of 8349 adolescents was extracted to study the relationship between sleep behavior and drug usage. The study found clusters of poor sleep behavior; statistical analysis revealed that the probability of a person sleeping less than seven hours goes up by $11\%$ if their friend also sleeps less than seven hours. Further findings indicated that high network centrality negatively impacts future sleep behavior. We extend this work to predict next-day sleep duration labels at a day-to-day resolution.

The literature closest to the proposed work was presented in (Liu et al., 2019) that leveraged social networks for health label prediction (Liu et al., 2019). Characteristics about the position of a user in the weekly social networks extracted from smartphone data (SMS exchanged), such as centrality, were used as input features to a binary classifier to predict whether an individual was depressed/anxious or not. We expand this work by integrating rich physiological, smartphone, and weather data in the prediction model. Our architecture also significantly differs from this work: instead of using network metrics as model input, we actively use the network to perform information aggregation from multiple users and exploit contagion at a deeper level.

In this section, we discuss the sensing technologies used to extract sleep behavior information.

We pose a problem of predicting sleep labels for the next day based on wearable and mobile data for the current and past few days. Formally, the multi-modal data with $m$ features for a given day $k$ is represented by the vector $x[k]\in\mathbb{R}^{m}$ and the corresponding sleep duration in minutes is represented by $s[k]$. The feature matrix $X[k]$ is composed of past $L$ days of feature data, where $L$ is an indicator of temporal memory, and the sleep label $y[k]$ is generated from sleep duration $s[k]$ (mins). The recommended sleep duration for young adults is between $8-10$ hours (Paruthi et al., 2016; Hirshkowitz et al., 2015b, a), so we choose a threshold of 8 hours for differentiating between good and poor sleep behavior. Also, the sleep duration distribution after preprocessing is relatively balanced around 8 hours. The sleep duration for our collected dataset had a mean of 7.2 hours (std= 2.1 hours).

The objective of this work is to develop a data-driven model $f(.)$ parameterized by $\theta$ that can predict sleep labels from the feature matrix $X$ and participants’ social network represented by $\mathcal{G}$,

We extract participants’ social networks from mobile phone data to analyze their impact on sleep behavior. While an alternative could be using social media friend networks, these networks are not always indicative of real interaction. Participants may be connected to people they don’t interact with, or to those who do not influence their behavior. Additionally, social media networks are larger and noisier, leading the model to aggregate irrelevant information and negatively impact prediction accuracy.

We validated social interactions among study participants using data from pre- and post-study social network surveys. These data allowed us to quantify the closeness between participants and create graph networks, which are detailed in Supplementary Material section E.2. Our findings affirm the presence of social interactions and phone communication among the study participants.

We represent a participant’s social network information as a graph network (Arney, 2013). Graph $\mathcal{G}$ is composed of two main components: nodes $\mathcal{V}$ and edges $\mathcal{E}$, $\mathcal{G}=(\mathcal{V},\mathcal{E})$. The nodes, also known as vertices, are participants who were part of the study at a given time. Since the study was conducted in seven cohorts, we cluster all participants in each cohort as part of the same graph in the initial processing. The edges are connections between nodes and are indicative of how close two participants are. We model the edges as weighted links such that the weights are proportional to their interaction. For convenience with mathematical operations, we utilize the matrix representation of a graph known as an adjacency matrix $\boldsymbol{A}$ where the value at $i^{th}$ row and $j^{th}$ column is represented by $A_{ij}$,

And $w_{ij}$ represents the weight of an edge between node $i$ and node $j$, where $i,j\in\{1,2,..,|\mathcal{V}|\}$.

We construct two graphs: the call graph $\mathcal{G}_{c}$, represented by adjacency matrix $\boldsymbol{A^{c}}$, and the SMS graph $\mathcal{G}_{s}$, represented by $\boldsymbol{A^{s}}$. We design them based on call and SMS metadata, respectively, collected over a time interval $[0,T]$. Representing an incoming call of duration $d$ seconds from participant $\mathcal{V}_{i}$ to $\mathcal{V}_{j}$ at time $t$ by $\mathcal{C}_{ij}[t]$ ,

For text messages, we consider two types of incoming messages: normal SMS with a text message body (Class 1 message), and Flash SMS with no message body (Class 0 message). Denoting an SMS from participant $\mathcal{V}_{i}$ to $\mathcal{V}_{j}$ at time $t$ by $S_{ij}[t]$,

Prioritizing Class 1 messages because of their stronger interaction $w_{1}>w_{2}$, we construct the SMS graph,

In the experimental evaluation, $w_{1}$ and $w_{2}$ were chosen to be 1 and $0.5$, respectively. Class 1 messages are given greater importance than Class 0 messages due to the presence of real conversations in Class 1 exchanges indicating a stronger interaction. For each node in the adjacency matrix, there is associated feature data $X$ and label data $y$. Only phone data is used to construct these networks; pre and post-study surveys do not contribute to these graphs.

We develop a system that integrates information from multiple participants in a social network such that the contagion in sleep behavior can be utilized to better predict the next day’s sleep duration.

Multiple experiments are conducted to evaluate different aspects of the proposed model. Firstly, we quantify the improvement in prediction, if any, provided by aggregating information from multiple participants of a social network. We further investigate the impact of different network sizes and temporal memory on prediction performance. The preliminary evaluation with both SMS and Call graphs did not find much difference in the performance. However, we prefer to report the result for SMS graphs because they are more dense, thus containing a wider distribution of network topologies (and therefore centralities). This is helpful in obtaining insights into how the user’s position in the network impacted the model’s robustness.

The performance over a randomly chosen test set is evaluated in terms of accuracy score and the process is repeated 25 times for each graph size and temporal memory. A similar procedure is repeated for Leave-One-Cohort-Out (LOCO). The mean accuracy score and the standard deviation in accuracy across the trials are reported in Table 2. ANOVA reveals that differences between models’ means for both types of the split are statistically significant (Random split p-value$=5.2x10^{-8}$, LOCO p-value$=6.3x10^{-15}$). Additionally, the Kruskal-Wallis non-parametric test also confirms that there is a significant difference among different models (Random split p-value$=3.2x10^{-6}$, LOCO p-value$=4.5x10^{-9}$).

LSTM alone performs the worst among all models. Adding CNN slightly improves the performance because it uses the kernel to observe and slide across chunks of data from multiple participants. However, since there is no systematic mechanism for aggregating information from multiple participants, the improvement is very small. CONV-LSTM is also the most unstable among all given the high variation in the accuracy score. The proposed method GCN-LSTM incorporates additional information about inter-participant relationships and therefore the performance is improved significantly. Our final model, GAN-LSTM further incorporates attention while aggregating information from multiple participants resulting in the best performance. The variance in the performance of the proposed models is also small. ANOVA confirms that the difference in the performance of all four models is statistically significant. Further pair-wise comparison between proposed models and benchmarks through Tukey HSD results in all adjusted p-values$<\alpha=0.05$ as shown in Table 2. This establishes that the proposed models perform statistically significantly better than the benchmarks.

The average performance reported in the previous subsection masks the impact of varying network sizes. A larger graph with dense edges is expected to have larger information aggregation than a smaller graph and that can significantly impact the accuracy of predictions. Thus, to characterize this impact, we fix temporal memory and vary graph size, evaluating the model in a random test-train split loop.

Figure 4(a) shows the impact of network size with a short temporal memory of 3 days. Initially, increasing the network size improves performance for all 4 models. However, for GAN and GCN, the performance plateaus at N = 15 before dropping due to irrelevant aggregation from unimportant nodes in larger graphs. The difference between GCN and GAN suggests that attention helps mitigate this effect. LSTM does not show significant improvement with larger graphs due to the lack of aggregation. CONV-LSTM improves as graph size increases, plateauing at N = 15. For a longer sequence length of 7 days (Supplementary material Fig. 10(a)), smaller network sizes (N = 5) benefit from longer memory, and attention-based models outperform GCN by a wider margin.

To explore the impact of temporal memory L, we fix the graph size and vary L from 3 to 9 days. In Figure 4(b), for L = 3, graph aggregation significantly improves proposed models compared to benchmark models. As sequence length increases, benchmark models perform better while GCN’s improvement is minimal. GAN shows the highest gain in performance. The results for a larger network are presented in Supplementary material Fig. 10(b). When the network size is large, the graph-based models perform really well at $L$ =3, but longer memory has a detrimental effect. In summary, very large networks and long memory lead to unstable behavior due to excessive aggregation from irrelevant nodes and past memories.

For better interpretability and to gain insight into how much individual modalities contribute to the proposed model GAN-LSTM’s predictions, we compute feature importance. Since each unit in deep networks applies a non-linear operation to multiple inputs, visualization of weights provides little insight into which feature contributes the most. Additionally, our input has two dimensions: features and the number of nodes. Therefore, we utilize saliency maps (Molnar, 2020) to compute which parts of the input contribute the most to the model’s predictions. To this end, first, the model is trained, and then for each test input, a saliency map is obtained by computing a normalized gradient for the input. The average map is obtained by taking the average across all input samples. Since we are interested in feature importance only, we collapse the map across the number-of-nodes dimension by taking the mean. For generalization, we repeat this process using the LOCO split described in section 4.4.2 and report the mean and $95\%$ confidence interval across all trials.

The feature importance analysis reveals distinct standout features within each modality. However, given the vast number of features, it becomes impractical to visually represent the importance of every individual feature. Therefore, we opt to showcase the top ten features from each modality. This approach allows us to compare the significance of different modalities and gain insights into individual features that exert the most influence.

The top features from each modality are presented in Fig. 5. Comparing different modalities, physiological features hold the most importance. In particular, skin conductance (SC) and skin temperature during or just before sleep are the most significant. Among phone features, phone screen usage patterns emerge as the most influential. Additionally, weather attributes like moon phase, cloud cover, temperature, and humidity exhibit similar levels of importance.

We further investigate the impact of the temporal module by removing it and observing the drop in performance. Our findings highlight that the temporal module aids in improving performance in proposed architectures and its impact is more pronounced in GCN-LSTM than GAN-LSTM. More details in Supplementary Material section D.

The data are perturbed at two levels, feature level and participant level. For feature level, we assume that the feature $f_{i}$, where $i\in\{1,2,..m\}$, follows an univariate Gaussian distribution, $\mathcal{N}(\mu_{f_{i}},\sigma_{f_{i}})$. After estimating the distribution parameters from data, perturbations are created by replacing feature data with randomly sampled values from the learned distribution with extended standard deviation. Note that if we create aggressive perturbations by replacing data with values that the feature does not often take, then those would be detected as outliers and filtered out in the preprocessing phase. Instead, we create noise and/or missingness by imputing values that would not be detected in the preprocessing phase or would be actually used for imputation if data are missing.

We learn the parameters $(\mu_{f_{i}},\sigma_{f_{i}})$ of this distribution (mean and standard deviation), using the maximum likelihood estimation (MLE) (Robert, 2014). The derivation for the estimated parameters is provided in supplementary material B.2, We expand the distribution slightly by increasing the variance by a factor of three for generalization. Once the distributions of features are learned, we investigate three cases. For ease of understanding, let us overview the original state of the system as shown in Fig. 6(a). The yellow circles represent the participants $u_{j}$ connected in a network and each participant has an associated feature vector $x$ consisting of $f_{i}$ where $i\in\{1,2,..,m\}$.

In the first scenario shown in Fig. 6(b), we consider feature-level perturbation across all participants. First, a random subset $\tilde{I}_{f}$ of size $\tilde{m}<m$ is selected from feature index set $I_{f}=\{1,2,..,m\}$. Then, for each feature $f_{p}$, where $p\in\tilde{I}_{f}$, the values are imputed with randomly selected data from the corresponding distribution $\mathcal{N}(\mu_{f_{p}},\sigma_{f_{p}})$. This procedure is repeated for all participants in the network. In this example, the first 2 features are chosen at random, $f_{2}$ and $f_{4}$. Then the features are perturbed in all participants as shown by red participant nodes.

In the second case, we investigate a more complex scenario. Like previous scenario, first a random subset $\tilde{I}_{f}$ of size $\tilde{m}<m$ is selected from feature index set $I_{f}=\{1,2,..,m\}$. Then a random subset of $\tilde{k}<N$ participants $\tilde{I}_{u}$ is chosen from the set of participants in the graph network $\mathcal{G}$ of size $N$. In the end, perturbation is only applied to participants $u_{q}$ where $q\in\tilde{I}_{u}$. In both scenarios, perturbations are only applied to the current-day data, and past memory is not perturbed. In the third case, we vary perturbations along temporal memory.

To characterize the impact of perturbations on model performance, we utilize the accuracy score as the metric. It is important to randomize and repeat the process multiple times to counter sensitivity to the choice of features and participants. First, the model is trained on unperturbed training data. Then, the perturbation is applied to test data and the accuracy score is computed for all test samples. The data are randomly perturbed again $H_{p}$ times and in each iteration, the performance metrics are computed. The original trained model is sensitive to parameter initialization and training-test split of data. To alleviate this sensitivity, we repeat the process of random splitting, training a new model and testing it on perturbed data $H_{t}$ times. Finally, for a given set of design parameters, the perturbation process is repeated $H_{t}\times H_{p}$ times. The average scores and $95\%$ confidence interval are reported.

To further investigate how network topology and a participant’s location in the network impact its performance, we evaluate the importance of a node and its accuracy score. We characterize node importance by its centrality in the graph network (Arney, 2013). To this end, we use two centrality metrics, degree, and eigenvalue centrality. Degree centrality assigns higher importance to nodes that have a large number of neighbors. However, it does not account for the cascade effect resulting from the fact that a node can also be important if it is connected to influential nodes. Eigenvalue centrality quantifies the influence of a node in a network by measuring the node’s closeness to influential parts of a network. More details on these metrics are provided in Supplementary Material B.1.

The centrality metrics are categorized into three categories: alone, small, and large. Alone nodes are not connected to anyone or have very small eigenvalue centrality ($C_{e}<10^{-4}$) and are mainly affected by perturbations in their own features. Small centrality nodes have small degree ($0<C_{d}<4$) or eigenvalue centrality ($10^{-4}<C_{e}<0.3$). Centrality scores greater than these thresholds are considered large. For each category, evaluation metrics are computed for a given trial and then average results over multiple ($\approx 500$) trials are reported.

In the first experiment, we perturb randomly selected features in inputs for all participants in a graph of size 15 and calculate performance. By increasing the percentage of perturbed features, we average the results over 500 trials, as shown in Fig. 7(a). Features were perturbed in increments of 25 (e.g., 100, 125, 150, …), and then converted to percentages for better understanding. At x=0 (no perturbation), all models perform well. As perturbations increase, performance drops, with LSTM showing overall poor performance. CONV-LSTM exhibits a sharp performance decline and is the least robust with numerous perturbed features. GCN also shows decreasing performance with more disturbance. However, attention-based GAN remains highly robust against perturbations.

In the second experiment depicted in Fig. 6 (c), instead of all participants experiencing the same perturbation, a random subset of participants in a network is chosen. While aggregating information from multiple participants has shown promising improvement in prediction accuracy, it also makes the predictions of the entire network more vulnerable to perturbations if a few participants in the network have noisy/missing data. Thus, we present how the proportion of perturbed participants impacts the performance in Fig. 7(b) by perturbing $50\%$ of the features in a varying number of participants. As observed in Fig 7(b) LSTM experiences very high errors consistently. CONV-LSTM shows a steep increase in error as more and more participants are perturbed.

The results presented in Figures 7 represent an average impact on the entire network. However, network topology also plays a role in how the model for some participants is more robust to noisy or missing feature data than others. In order to segregate these different parts of the network, we report the average accuracy score of different degree category nodes when 5 out of 15 participants experience perturbations in $50\%$ features, both chosen randomly, in Fig.8(a). The average performance without any noise for each model is represented by dashed lines. The category ’alone’ indicates nodes that are not connected to anyone. The drop in performance in the ’alone’ category results from the vulnerability of a participant to noise in its own features. Both GCN and CONV-LSTM suffer from a significant drop in performance for alone nodes. For nodes with a small degree, CONV-LSTM experiences a drop while other models are robust to it. Finally, for participants with a large number of direct neighbors, all models observe a drop in accuracy score because they have a higher chance of being closely connected to a node that is perturbed. We further increase the perturbations in Fig.8 to 13 participants. The drop in performance is higher than what is observed in Fig.8(a), especially for large degree nodes.

We use another centrality metric called eigenvalue centrality to segregate different parts of the network. First, a small number of participants i.e., 5 are perturbed in Fig.9(a). For participants who are alone or have extremely small eigenvalue centrality ($C_{e}<10^{-4}$), there is a sharp drop in performance. For small eigenvalues, only CONV-LSTM suffers from a drop in the accuracy score. For large eigenvalue, the behavior is unexpected and instead of a decrease, there is an increase in average performance. To investigate it further, we raised the number of perturbed participants to 13 (Fig.9(b)). For the ’alone’ category, the drop in performance is slightly larger than before. For small eigenvalue centrality, we observe unexpected behavior, where performance is slightly improving for graph-based models. Finally, for large centrality nodes, there is a significant decrease in performance. This category is the only one in which even attention-based models are not robust. These results indicate that more connections to influential nodes can lead to unstable predictions.

Finally, we investigate how perturbations in different parts of the sequence impact the performance. Performance declines linearly for CONV-LSTM and GCN with increasing sequence perturbations, while attention-based models demonstrate robustness to these changes.Detailed results are provided in Supplementary section C.2.1.

While existing literature has provided evidence for contagion through statistical testing (Udry, 2003), there exists a notable gap in the field: a lack of quantitative modeling that hinders its seamless integration into learning systems. Our proposed models have the capability to discern and incorporate underlying contagious behaviors in the prediction task. This is especially valuable because of the inherent scarcity of health and wellbeing labels, such as user emotions, in contrast to the high-resolution features readily obtainable through wearables and phones. This sparsity often necessitates some kind of downsampling of features, resulting in information loss. Because of the same limitation in our dataset, we were constrained to extracting statistical descriptions from multiple features rather than utilizing raw data. Leveraging contagion through graph networks offers a novel approach to harnessing the potential of wearable and smartphone data. This approach provides complementary information to the model, effectively reducing the loss of valuable information.

While comparing different models’ performance in Table 2, we observe an improvement when the model aggregates data from multiple participants in a certain pattern, e.g. when employing CNN. However, the extent of this gain is limited due to the absence of information about the connections between these participants. Thus, the inclusion of graph networks significantly enhances the model’s predictive capabilities, providing indirect evidence of contagion. Nonetheless, it is crucial to acknowledge that not all friends or neighbors contribute equally, nor do they exert an equal magnitude of contagion. Therefore, when the model focuses its attention solely on important nodes using GAN, we not only witness performance improvements but also a substantial increase in model robustness as shown in Fig. 7 and 9.

Our framework paves the way for testing different networks in digital health systems. Graphs provide a simple yet powerful representation of multiple participants and links between them that can be exploited in numerous ways. In addition to the social network, graphs can be constructed from similar lifestyle patterns(Zhang et al., 2021) (e.g., mobility, eating, and exercise habits) to complement the multi-modal prediction. The architecture proposed in this work can handle temporal graphs as well as those with dynamic sizes. This means that as the user’s social interactions change, the model is able to adapt to the changing contagious behavior.

Robustness analysis highlighted that graph integration helps make the model more robust to data perturbations only when attention is paid to important nodes. Relying solely on the graph neural network for combining information from other nodes can make some nodes more vulnerable to noisy or missing data. Interestingly, the impact of perturbations on the performance varies depending on a node’s connectivity. As shown in Fig. 8 and 9, nodes directly connected to a substantial number of peers experience a less pronounced decline in performance compared to those linked to a select few highly influential nodes. Firstly, nodes with higher degrees exhibit lower accuracy relative to their counterparts with fewer connections. The attention mechanism helps mitigate this effect to some extent and also makes the model more robust to perturbations. However, the users with high eigenvalue centrality experience a large drop in performance even with GAN. Thus, not all neighbors are equal, and different friends impact a node’s performance differently based on their own influence in the network.

Network size plays a pivotal role for two reasons: firstly, we want to capture all nodes that might have a contagious effect on user’s behavior. Therefore, we vary the graph size from 5 to 20 and observe the performance in Fig. 4(a). Our findings revealed a performance plateau at N=15. Larger graphs tended to introduce irrelevant data aggregation, although this could be partially alleviated by the implementation of an attention mechanism. This tradeoff can be optimized by carefully designing network size which is customized based on the user’s sociality, demographics, and lifestyle. For instance, in our study, all participants were university students who knew other people in the study; for these participants, we expect interactions with a larger number of fellow students. However, for specialized groups like caregivers, who may have extensive interactions within smaller circles (mainly care receivers), a smaller network size may be required for reliable performance. Graph-based models tend to be more sensitive to perturbations in neighboring nodes’ features due to their aggregation mechanism.

This challenge can be addressed using GEDD which enables us to optimize network sizes, thereby mitigating the adverse effects of large networks and noisy data while generating graphs tailored to the available dataset.

While constructing social networks is relatively straightforward, their effective use in health label prediction presents multiple challenges. This is exemplified in Figs. 8 and 12, where GCN performs poorly for nodes with a large number of direct neighbors that are also more susceptible to perturbations. However, the introduction of attention mechanisms significantly mitigates irrelevant data aggregation, resulting in improved performance. This underscores the critical role of attention in our approach. The attention mechanism in this work is based on feature correlation. It can be further improved by utilizing complementary information from study surveys or customized based on network topologies. We noticed in our analysis that nodes with large eigenvalue centrality are more prone to noise. This insight can be used to do a weighted normalization of attention weight or thresholding the maximum number of important nodes when the node degree is very large.

The collected dataset was a human-centric dataset posing challenges such as a limited number of participants. This limitation makes it impossible to capture the entire social network of each participant (e.g., friends, family, coworkers) in one study. Therefore, the maximum graph size in our evaluation is only 20, which does not represent very complex graph typologies. However, despite these limitations, the proposed architecture shows promising results and highlights the potential of social network integration in digital health systems. While this work focuses only on sleep duration, it can be easily extended to other health and well-being labels such as mood, emotion, and eating habits. Another limitation of this dataset is that all the participants were young adults at one university which results in sampling bias. In future works, the system needs to be tested with other diverse groups of the population. Moreover, the acquisition of ground truth involved user questionnaires, a method susceptible to subjective biases including recall limitations. These inherent biases can be mitigated in future works through the integration of sleep-tracking technologies (Rahman et al., 2015), such as smartwatches and under-mattress sleep trackers. The study was limited to metadata about calls, SMS, and screen on/off only. In future works, data from messenger or social media apps can be utilized to further expand the social networks used in the model.

Another limitation of this system is that it is completely centralized. Information from all users is directed to one centralized model for making predictions. In order to deploy this system in large-scale real-life networks, this issue should be resolved through an elaborate distributive computing solution(Witt et al., 2022). Furthermore, the robustness analysis makes assumptions about the distribution of features in order to perturb them. We also used similar assumptions for standardization and removing outliers in data preprocessing. In future works, this assumption can be relaxed and domain knowledge about feature distribution can be utilized to further improve this analysis.

• •

  Selective Sampling: In multi-user architecture, complete feature information about all users is required by the system. As observed in the robustness analysis, the proposed model can produce reasonable performance even when some users are missing over $50\%$ data. The drop in performance strongly depends on where in the network a node is located. This insight can be extended in future work to identify influential nodes in the network. These nodes can either be tapped for the collection of existing modalities or additional information that can help improve the performance of the entire network.

• •

  Active Mobile Health: In systems where health labels are monitored, tracked, or predicted from non-obtrusive data sources, the role of the system is passive. On the other hand, in an active mHealth system, in addition to monitoring and predicting, the system can also provide interventions to regulate the behavior of interest. Scaling this to a network level, where the users are interconnected and exert influence on each other’s behaviors, important nodes can be identified to provide possible points of intervention for regulating the behavior of the entire network.

• •

  Lifestyle recommendations: The proposed system can be extended to provide lifestyle recommendations to rectify unhealthy behaviors. A unique feature of this extension would be that in addition to changes in daily habits, it can also give recommendations about social interactions conducive to enhanced health and overall wellbeing. Such recommendations would need to be tested before implementation.