Instrumentation & Measurement Magazine 23-4 - 45

original signal. The algorithm has been effectively employed
for ECG anomaly detection, as described in [18].

Event Detection

Fig. 5. Golomb codewords based on union sets. Each codeword is composed
by a fixed-length prefix, whose value identifies the union set, and a variable
length suffix. B is the ADC resolution in bits.

vector compression algorithm is characterized by a very low
computational cost, and it can be effectively implemented on
multi-sensor systems with very limited resources.

Feature Extractions
A real-time monitoring system must detect significant events
that are usually associated with changes in some significant
signal feature. To implement an event-based system, therefore,
the signal received at smartphone level needs to be analyzed
to extract related features. Two methods are briefly introduced
next: one based on compressed measurements and the second
one based on a sub-string search algorithm.

Feature Extraction Based on Compressed
Measurements
It has been shown in the literature that feature extraction from
compressively sampled signals can be carried out directly on
compressed measurements [15]. Working with compressed
measurements has the advantage that no signal reconstruction
by the data collection unit is needed, however the target feature needs to be known in advance and an associated template
needs to be created.
Detection of a given waveform within a compressed vector can be obtained by creating first a "compressed template"
for that waveform shape, that is then correlated with the
compressed vector. Let ψi be the time-domain shape of the
waveform that we want to identify. We construct the N×N
matrix Ψ containing the one-shifted version of the considered template: Ψ = ψ 1T ,ψ 2T ,... ,ψ NT . The cross-correlation


between the acquired N-sample vector x and the Ψ matrix is:
−1
rx ,Ψ = y , ΘΘT ΘΨ where y is the M-sample compressed vector received by the smartphone.
A suitable threshold on the correlation value allows the desired feature to be detected and localized. A full example is
discussed in the following.

(

)

Method Based on Strings: HOT SAX
HOT SAX [18] is a widely employed algorithm that involves
the transformation of a sample sequence into a string, after which efficient sub-string search algorithm are applied.
Rare or uncommon strings are associated to anomalies in the
June 2020

This final step in a real-time monitoring application consists
of the detection of possible events of interest. In general, an
event-driven architecture (EDA) system is composed of producers
and consumers, where the former are devices that can generate events and the latter are elements reacting to events. An
event is defined as a significant change in the state of the monitored system, and it propagates from producers to consumers
through a communication channel. It is very useful, in some
applications, to store raw data acquired by physical system
while an event takes place and send them to a main unit for further, more detailed processing.
Depending on their significance, events generated locally
may be logged, notified to the user or immediately forwarded
to medical staff. Finite state machines (FSM) are one of the simplest implementations of an event detector. In the simplest
case, only two states are defined: a "normal" state that does
not involve transmission; and a "detected anomaly" state, that
is entered when an anomaly is detected, where the generation
of events is enabled.
System state is determined from the analysis of the set of
previously identified signal features. When a particular combination of features points to potential anomalies, or when a
condition needs reporting to the medical staff, an event is generated. Classification algorithms, usually called classifiers, are
often employed for this aim. Their ability to provide reliable
and trusted results depends on the available level of expertise
encoded in the classifier.

Example
This section presents an example of event-driven architecture
for a biomedical IoT system. For this purpose, we consider the
system presented in [19] that is composed of a low-cost ECG
sensor, a local processing device-i.e., a smartphone-and a
remote unit, through which experienced medical personnel
can access data from each monitored patient.
Data are initially compressed by the sensing device, according to the compressive sampling paradigm, using a Bernoulli
sensing matrix. The generating polynomial employed in [19]
has order 20 and results in the occupation of just 20 memory
bytes to support the compression algorithm. Total memory
space required by this solution corresponds to a temporary
vector of N samples, plus the few bytes needed for the sensing matrix. This is more than what is needed by a predictive
compression algorithm but can be justified by the scope of the
applications, particularly ones that require higher accuracy
information.
Data acquired from the sensing unit can be transmitted to
a smartphone where they are analyzed to detect any meaningful state change that would lead to the generation of events.
In this case, state changes refer to ECG waveform anomalies,
which require detection of variations in any of the three main
parts of a trace, namely, the QRS complex, the T-wave and the

IEEE Instrumentation & Measurement Magazine 45

Instrumentation & Measurement Magazine 23-4

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 23-4