Medical diagnostic-based sensor selection

Abstract

Abstract-Wearable sensing systems have facilitated a variety of applications in Wireless Health. Due to the considerable number of sensors and their constant monitoring these systems are often expensive and power hungry. Traditional approaches to sensor selection in large multisensory arrays attempt to alleviate these issues by removing redundant sensors while maintaining overall sensor predictability. However, predicting sensors is unnecessary if ultimately the system needs only to quantify diagnostic measurements specific to the application domain. We propose a new method for optimizing the design of medical sensor systems through diagnostic-based bottom-up sensor selection. We reduce the original sensor array from ninety nine to twelve sensors while maintaining a prediction error rate of less than 5% over all diagnostic metrics in our testing dataset.

Fault-Tolerant and Low-Power Sampling Schedules for Localized BASNs

“…As we shall see, in comparison to the CICA-based algorithm that selects a fixed subset of sensors of cardinality, , to be sampled at every epoch [14], our approach offers the added advantage that fewer than sensors may be sampled at some epochs. This translates to a further savings in energy over this algorithm.…”

“…By the time the algorithm alters course to increase coverage, the competition may be too high. We resolve this by running the PCSS algorithm in the context of an iteratively refinement strategy for sensor coverage, that we call the Coverage Improving Iterative Refinement (CIIR) algorithm and outline in Table III. The CIIR algorithm eliminates those sensors from consideration for coverage at each successive run of the PCSS algorithm, which have failed to meet a coverage threshold, , over previous runs (lines [12][13][14]. This is achieved by removing them from the senPred structure and a corresponding cleanup of the predMap, redSet, and sensSamp structures.…”

“…In [14] and [19] the authors proposed an algorithm to further relax the tradeoff between sensing fidelity and "covering" sensor subset size by defining the fidelity only in terms of diagnostic accuracy. However, extending these solutions to provide fault tolerance by constructing multiple distinct sensor subsets is hindered by multiple factors.…”

“…For reasons that will be made clear in Subsection IV-D, PCSS also takes sensSamp and predMap as inputs that can be assumed to be empty for now. Note that we are only interested in a sensor's measurement at epochs when the metric observation likelihood matrix, , indicates that the likelihood is nontrivial (lines 5-9), and when the measurement has not already been covered (lines [14][15][16][17]. This latter point is noteworthy as we do not want to value sensor at an epoch for its ability to predict sensor 's readings at an important epoch , if has already been covered by a different pair that was selected into sensSamp.…”

“…By sampling measurements at each stride based on the schedule in sensSamp, the diagnostic metric is composed, based on a model, from the metric values at the sensors covered in predMap. In our case, similar to [14], we use multiple linear regression as the model as it provides a good fit. In the case of sensor faults, redSet is consulted to alter predMap, and possibly sensSamp, so that diagnostic metric composition can be continued at the same level of accuracy.…”