Quality of Information Metrics for Body Sensor Design

Figure 1: data flow in a BSNBSNs perform long-term, continuous, remote monitoring of physiologic and bio kinematic information. Because of varying computational, storage and communication capabilities at different layers of the BSN, system designers must make design choices that trade-off information quality with resource consumption and battery lifetime. Given these trade-offs there is the possibility that the information provided to the health practitioner may deviate from what was originally sensed, causing the doctor to perform the wrong diagnosis. These trade-offs are based on traditional measures of data quality like RMSE, but those don’t correlate well with the notion of information quality for a particular application. Objective metrics of information distortion and its effect on decision making are necessary to help designers to make more informed trade-off, and to help practitioners understand the kind of information provided by the BSN.

Body Sensor Networks (BSN) continuously monitor patients in a long period of time outside clinical settings. The benefit of such systems comes with many challenges: balancing the need to provide the best possible information (which requires more computation, storage and communication capabilities) with the need to provide long BSN lifetime (which requires minimization of resource usage). In a typical BSN, data is sensed (and possibly processed) at the node which transmits it to an aggregator. Both devices are usually on the patient. The aggregator streams the same data to a base station to be further processed and presented to the practitioner. Nodes are usually resource constrained (in computation, storage, communication, and energy capabilities). The aggregator has more resources, but still with some limits. The base station has the most resources available. This variation in capabilities necessitates various design choices in order to meet application goals of presenting the best quality information. As can be seen in the figure above, distortion introduced to the data in a BSN can be intentional (lossy compression) and unintentional (accumulation error due to integer arithmetic instead of fixed or floating point). 

Traditional ways to measure the quality of a signal are statistical (RMSE, PRD) or information theoretic (SNR). These metrics are information agnostic (assign equal weight to all data) and work only when close-to-perfect signal reconstruction is needed. Depending on the application, only certain features have to be preserved in the physiologic signal. The figure below gives an example:

(a) is the continuous signal that the node is sensing, (b) is the output of the BSN, after processing the signal. The only features of interest are: a1t12, and a2. Those are preserved in (b) therefore we would assign the BSN's design a high quality of information value, while traditional metrics would assign it a bad value.

It is therefore important to assess how design points (and the distortion they cause) of the BSN impact medical decisions. What is needed is application-specific metrics that can relate system operating points to an objective measure of their impact on practitioners’ decisions. In [1] we present a methodology for developing these metrics, which we term quality of information (QoI) metrics.

 


References