A scalable semantic framework for IoT healthcare applications

Abstract

IoT-based systems for early epidemic detection have not been investigated yet in the research. The state-of-the art in sensor technology and activity recognition makes it possible to automatically detect activities of daily living (ADL). Semantic reasoning over ADLs can discover anomalies and symptoms for disorders, hence diseases and epidemics. However, semantic reasoning is computationally rather expensive and therefore unusable for real-time monitoring in large scale applications, like early epidemic detection. To overcome this limitation, this paper proposes a new scalable semantic framework based on several semantic reasoning techniques that are distributed over a semantic middleware. To reduce the number of events to process during the semantic reasoning, a complex event processing (CEP) engine is used to detect abnormal events in ADL and to generate the associated symptom indicators. To demonstrate real-time detection and scalability, the proposed framework integrates a new extension of ADLSim, a discrete event simulator that simulates long-term sequences of ADL.

Appendix 1: Evaluation of epidemic simulation in ADLSim

Fig. 6

Results from experiments set 1

To obtain reliable results from experiment sets 2 and 3, it is necessary to first ensure that our simulation models and SI queries work as intended. This section presents a set of experiments designed for this purpose, and discusses the results.

1.1 Experiment goals and design

The main goals of these experiments are to show that (1) the epidemic model behaves as intended, (2) the output from the SI queries correspond to the impact of the epidemic model, and (3) that the results vary realistically with the values of key parameters (listed in Table 1).

We conduct two experiments, Experiment 1.1 and 1.2. Each experiment includes one sub-experiment per combinations of the three ED and three DIP parameters shown in Table 1. This enables us to study epidemics ranging from very mild to very severe. We conduct each sub-experiment three times with different seeds and which results in $3\times 3\times 3=27$ sub-experiments per experiment. In Experiment 1.1, we use highly conservative thresholds for SI event emissions. SI values are emitted only when they change to “very low” or “very high”. In Experiment 1.2, we increase the sensitivity of the queries to emit events whenever the value changes from “normal” to any other value. This way we can study the impact on the number of events emitted for further semantic reasoning. In addition, we generate every simulated day 24-h SI statistics regarding the entire apartment complex.

1.2 Results and discussions

To get a comprehensive understanding of our results, we present statistics across all runs and parameter values as well as plots that provide a deeper look at individual runs for a sub-set of the results.

Figure 6 presents the results from experiment set 1. Graphs (a), (b), (c), and (d) present results from the first run in Experiment 1.1 with ED = 30 and DIP = 1.5% (left) and 3% (right). The epidemic begins at simulated Day 20. Graphs (a) and (b) show the complex-wide 24-h SI statistics (y-axes) for all simulated days (x-axes), and Graphs (c) and (d) show the number of apartments (y-axes) with SI values “very low” and “very high” for all simulated days (x-axes). Graphs (e) and (f) show statistics across all runs for all parameter values from Experiment 1.1 and 1.2. They show the average (lines) and SD (error bars) of the number of emitted SI events (left y-axes) across all three runs per combination of DIP (x-axes) and ED (with different line colors). Note the differences in the left y-axes in Graphs (e) and (f). The box plots show the average number of the infected individual across all three runs (right y-axes) per combination of DIP (x-axes) and ED (with different box colors).

The results in Graphs (a), (b), (c), and (d) show a clear impact on the complex-wide SI of the simulated epidemic, namely that the SI values during the epidemic have significantly larger absolute values during the epidemic as well as before and after the epidemic. This effect is also clearly larger with DIP = 3% than with DIP = 1.5%. We also see that only the queries affected by the disease are impacted, i.e., the bottom five queries. Notice that although the epidemic lasts from Day 20 to Day 50, the main impact on the complex-wide SI values is found somewhat later. Three factors contribute to this. The most important factor is that a certain amount of time after the epidemic onset is required for a sufficient number of individuals to be infected to have a significant impact on the aggregate activity magnitudes in the whole apartment complex. Second, due to the gradual onset of symptoms, 3 days pass after an individual is infected until the full effect of the symptoms is manifested. Finally, at the end of the epidemic period, a certain number of individuals are still symptomatic for some time until recovery, which shows up in the SI queries. There are occasional, abnormal SI values before and after the epidemic explained by natural variations in behavior. These are trivially exposed as false positives by jointly considering a larger set of samples and/or samples from multiple SI queries. These results thus indicate that the complex-wide SI queries respond correctly to the simulated epidemic.

The results in graphs (c) and (d) show that also the apartment-specific SI values are as expected, i.e., the period with the epidemic results in a significantly larger number of apartments with indications of decreased activity levels reduced intake of food, and increased toileting. The effect is again clearly more pronounced with the largest DIP. We can thus conclude that also the apartment-specific SI work as intended.

Figure 6e, f show that the least restrictive SI thresholds (“high” and “low”, Fig. 6e) yield an order of magnitude more SI events than the most restrictive thresholds (“very high” and “very low”, Fig. 6f). This has a significant impact on the number of resources required for subsequent analysis. We also see that the epidemic is clearly detectable by visual inspection in graphs (a) and (b) even with the most restrictive thresholds. It is therefore advisable in this scenario to employ highly restrictive SI thresholds to minimize the emission rate of SI events. The box plots finally show that the number of infected apartments agrees well with Eq. 1, and that these numbers are proportional to the number of emitted SI. Thus, we find that the SI thresholds and the epidemic models work as intended.

Summary and key insights: Results from Experiment Set 1 demonstrate that the epidemic model and all aspects of the SI queries work as intended. We also find that using the most restrictive SI emission thresholds can significantly reduce the number of events emitted and thus save considerable amounts of event processing resources while preserving clearly detectable changes in SI values during the epidemic. We therefore decide to use these SI emission thresholds in experiment set 2.