ADAPTIVE NIBP LOW-PULSE DETECTION

Detecting Low-pulse Adaptively

Ricardo Henrique Gracini Guiraldelli and Ricardo Luis de Azevedo da Rocha

Polytechnic School, University of S

ao Paulo, Prof. Luciano Gualberto Avenue, tv. 3, 158 , S

ao Paulo, Brazil

Keywords:

Non-invasive blood pressure, Low pulse amplitude, Adaptive decision trees, Adaptive technology.

Abstract:

This paper presents a proposal of an adaptive device for low-pulse detection on NIBP measurements using

an adaptive decision tree algorithm (AdapTree) and a probabilistic methodology, besides featuring learning

related to expert knowledge.

1 INTRODUCTION

Non-invasive blood pressure (NIBP) is a must-have

measure on critical and day-by-day diagnosis, provid-

ing fast and contiguous feedback about general health

aspects from patients through a simple method of ar-

terial pulse observation (electronically, pneumatically

or acoustically way), supporting the medical staff de-

cisions over one treatment. But its “noninvasibility”

has some side effects such as amortized signal, what is

oddly problematic on low pulse amplitudes cases. For

instance, it has been the “neck” of NIBP and the in-

creasingly dependence on automatic electronic mea-

surement just enlarged the problem.

As as example, there is a correlate research from

(Lin et al., 2003) that deals with noise compensation

generated by movements artifacts in NIBP measures.

It uses fuzzy logic to calculate the smoothness of os-

cillometric curve. We are using adaptive technology,

on the other hand, somehow complementing it.

Aiming this speciﬁc problem, the present work

proposes a technique for low pulse amplitude recog-

nition and detection, as well as dynamic learning of

these cases, seeking to improve the facility brought

by NIBP, particularly on intensive care units (ICUs).

For text comprehension, this paper is organized as

follows: from section 2 to section 4 we introduce the

background techniques behind our proposal; at sec-

tion 5 we describe the designed solution and at the

6the section we summarize our conclusion and expec-

tations.

2 NON-INVASIVE BLOOD

PRESSURE

Measurement of the arterial pressure is of greater im-

portance on patient’s diagnostics because of the high

density of information about the body condition —

especially hemodynamic system — it provides. Al-

though the ideal arterial pressure is made invasively,

this methodology is too much costly in health as-

pect to the patient and a very risky procedure. For

this reason some non-invasive blood pressure (NIBP)

methods were developed, like auscultatory, Korotkoff

sounds and oscillometric, the latter one we will focus

on.

The oscillometric method is widely used on au-

tomatic biomedical devices because it is independent

of a human expert for proceeding with the measure-

ment and it is extremely simple, based on descendant

step-pressures (as shown on ﬁgure 1). At the end of

the measure, the signals collected allow the plot of

a pressure (mmHg) × pulse amplitude) chart, where

is possible to extract the mean arterial pressure (P

the pressure where occurs the global maximum of the

curve. Once we have the P

, we can calculate the sys-

tolic pressure (P

) and diastolic pressure (P

) by the

equation that follows:

· P

(1)

Although the oscillometric method seems as an

ideal one, it is less efﬁcient when the measure presents

artifacts or low pulse amplitudes. Even thou it contin-

ues to be the best method for automatic biomedical

devices nowadays.

113

Gracini Guiraldelli R. and de Azevedo da Rocha R. (2010).

ADAPTIVE NIBP LOW-PULSE DETECTION - Detecting Low-pulse Adaptively.

In Proceedings of the Third International Conference on Biomedical Electronics and Devices, pages 113-116

DOI: 10.5220/0002741101130116

 SciTePress

Figure 1: NIBP curve through oscillometric method

(Frankenreiter, 1990).

3 PROBABILISTICS METHODS

The probabilistic techniques have widespread use on

artiﬁcial intelligence because of their capability of in-

formation inference from small (or even lack of) data

about the decision subject and are sustained by math-

ematical formalisms giving, this way, some reliable

information.

Nowadays these techniques are even more popu-

lar because of the “smart software” development, i.e.

software that executes a whole (and complex) task

without rational human intervention (e.g. buying sug-

gestion software at online bookstores). And these ap-

plications are getting so evolved that they are already

at some specialized ﬁelds.

Resulting from that era of software development,

a probabilistic method called certainty factor (Pearl,

1988) was designed focusing on bringing some ex-

pert knowledge to probabilistic reasoning, especially

medical support one. Even more, it was thought to be

simple to implement and provide fast inferences.

Certainty factor works with a set of hypothesis-

evidences rules based on expert knowledge (interde-

pendent or not) that, interconnected, infer some sys-

tem’s information. For reasoning purposes, each hy-

pothesis or evidence has a certainty degree (c

) such

that −1 ≤ c

≤ 1, pointing out total unbelief (−1) or

total belief (1) on the events veracity.

4 ADAPTIVE TECHNOLOGY

For complex systems development, some ﬂexibility

during decision making is an essential requirement.

Today, biological systems are almost the unique ones

that have this ability but some (mathematical) adap-

tive formalisms were developed, and they are capable

of changing themselves at runtime (Neto, 1993; Shutt,

1995).

At the present work, adaptivity means dynamic

modiﬁcation of a rules set that controls a speciﬁc de-

vice, i.e. given a device, your transition function (or

rules set) is dynamically modiﬁed at runtime of such

device. It’s a rough simpliﬁcation of the Adaptive

Technology concept presented on (Neto, 1993), but

keeps the main meaning of it.

Adaptive technology has wide uses, especially

in context change systems, as voice recognition and

user-personalized systems (as biomedical software).

In the attempt to satisfy the wide niche of applica-

tions, a number of adaptive devices were developed,

all of them based on pre-existing and known models;

e.g., adaptive ﬁnite automata(Neto, 1993), statecharts

(Neto et al., 1998), decision tables (Pedrazzi et al.,

2005) and trees(Pistori and Neto, 2002; Pistori and

Neto, 2003). We are going to focus particularly on

the last one.

4.1 AdapTree

When classifying data, decision trees are efﬁcient de-

vices for the task; based on this concept, there are

numbers of algorithms(Quinlan, 1996).

Traditional decision trees needs some training

over a solved body of “problem cases”. This train-

ing is a huge limitation for systems that have a learn-

ing requirement and, for instance, there are some al-

gorithms like ID3 that envisions some kind of learn-

ing based on re-training or become almost ineffective,

though. Targeting to improve it, the AdapTree (Pis-

tori and Neto, 2003) algorithm was developed using

adaptive technology, putting together decision trees

and adaptive ﬁnite state automata.

AdapTree still requires some training, receiving

an input string with an additional data ﬁeld for class

representation (ﬁgure 2), what we will call as static

learning. When changing from training mode to clas-

siﬁcation mode, we just need to suppress that addi-

tional data ﬁeld for classiﬁcation under the classes

previously trained.

Whenever it ﬁnds a “problem case” that has not

matched a classiﬁcation pattern at the tree set, Adap-

Tree calls an statistical mechanism based on relative

observed frequency up to the moment and the string

sequence already read, concluding the most probable

class for that input data.

Although it seems too simple, AdapTree has been

well positioned on benchmarking tests(Pistori and

Neto, 2003) among several well-known algorithms,

showing a great beneﬁt/cost relation for a large num-

ber of projects.

BIODEVICES 2010 - International Conference on Biomedical Electronics and Devices

114

Figure 2: Three stages of AdapTree: the beginning, after a

string read a, after a string read aS, respectively.(Pistori and

Neto, 2003).

5 PROPOSAL

Once given the entire theoretical basis for understand-

ing the problem we aimed, this section formalizes a

proposal for the low-amplitude pulse problem (that

we will refer as low-pulse).

Low-pulse is a condition of extremely difﬁcult de-

tection and though many A/D transducers have the

capability to detect that amplitude, it is not safe to

keep that sensibility activated because many environ-

mental interferences (like patient’s movement) may

be erroneously translated as pulse signal. But some

classes of special patients commonly presents this

condition, when they are under drug or pathologi-

cal effects, or even naturally, possibly introducing

complications for their correct treatment and diagno-

sis. Analyzing some NIBP measures bank

, we were

able to design a simple class-feature classiﬁcation that

keeps cause-consequence relationship with low-pulse

as shown below:

Table 1: Classes related with low-pulse.

Class Feature

Patient Neonate

Patient Elderly

Place ICU

Place Surgical Ward

External Modules Ventilation

External Modules Anesthesia

It is important to note that the above information

are available on ICU monitor once they’re used to

conﬁgure alarms, cuff pressures and other safety re-

This NIBP bank of measures is private and it is exclu-

sive of an ICU monitor manufacturer.

sources. Hence, the classiﬁcation is appropriate to use

on (almost) every ICU monitor without restrictions.

Although table 1 seems too simpliﬁed for some

signiﬁcant improvement, it’s enough for problem def-

inition without negative side effects on quality pulse

detection; doubtful datum won’t be necessary, in this

research scope, for an appropriate solution deﬁnition.

Once we have the classes of probably low-pulse

occurrence cases, the expected step is to realize some

kind of algorithmic classiﬁcation of them making

possible a correct low-pulse classiﬁcation based on

those input datum. Previously, on section 4, we have

seen that decision trees are the natural mechanism

for knowledge classiﬁcation; ID3 or C4.5 algorithms

could be used but, as discussed early, biomedical

equipments need some personalization characteristics

for each new patient, hence, given this requirement,

AdapTree ﬁts as the best solution.

Using AdapTree, we should train the algorithm

with real and classiﬁed data group (training body) for

some rule determination before the classiﬁer (deci-

sion tree) usage. In the same way it would be done if

we were using some other conventional decision tree

algorithm. With a large enough database, we could

use it as a test body and reach a signiﬁcant advan-

tage once the (statistically) majority of cases would

be covered with this kind of training.

If we cover the majority number of cases, why do

we need AdapTree for? We need it for those cases

in which common decision tree algorithms are unable

to classify the complementary group, converging to

the dynamic learning region of AdapTree; this kind

of learning is usually made by statistical inference,

based on counting.

Even thou literature benchmark (Pistori and Neto,

2002) shows its great performance, we must not use

the counting mechanism because it is not based on

medical knowledge for classifying data. We will use

instead a probabilistic technique already used on med-

ical decision auxiliary software, the certainty factor

— which may be referenced as inference machine —

, developing a hermetic rules set based on specialist

knowledge (hypothesis relations) and statistical infor-

mation (hypothesis weights).

Therefore, we propose the set of hypothesis of the

inference machine as the classes presented on 1 and

the development of the relations as showed on ﬁg-

ure 3.

5.1 Connecting the Devices “Low-pulse

Detector” and “Pulse Detector”

Once we have deﬁned above the low-pulse detector

device, now it must be connected with the traditional

ADAPTIVE NIBP LOW-PULSE DETECTION - Detecting Low-pulse Adaptively

115

Figure 3: Hypothesis-inference machine relationship.

pulse detector device already implemented and in use

on the ICU monitor.

Focusing on minimum system interference, we

propose an event-driven (Ferg, 2006) mechanisms

connection: after a speciﬁc no pulse timeout, the main

module (which contains the pulse detector) requests,

through an event, that the low-pulse identiﬁer (or in-

ference machine) evaluates and returns its certainty

factor as other data, for example new pulse amplitude

levels (for detection). Figure 4 shows this connection.

Figure 4: Communication interface between pulse identiﬁer

and inference machine.

A great advantage of this kind of communication

between the modules is that it keeps the information

segregated in each module, isolating possible soft-

ware bugs from one module to affect the whole NIBP

system and exposes this new module (low-pulse de-

tector) as a system improvement, either.

6 CONCLUSIONS AND FURTHER

WORK

This work proposes an evolution on apparently stag-

nated research ﬁeld of NIBP (Rolfe, 1979), dealing

with risk low-pulse cases through the use adaptive

technology and probabilistic methods. The addition

of these two features on NIBP measurements brings

personalized diagnosis to automatic measure biomed-

ical devices.

A numerical validation is planned to verify the

methodology on the “ﬁeld work” of the system and

explore the opportunity to look for improvements.

These data will be provided by a brazilian ICU mon-

itor manufacturer, based on their large real measure-

ments database and their certiﬁed and validate pulse

detector algorithm.

Another step to be taken is to complete the solu-

tion, so we will also propose the design of methods

for recalculating the (maximum and minimum) cuff

pressure limits providing, this way, a complete solu-

tion for low-pulse detection.

REFERENCES

Ferg, S. (2006). Event-driven programming: Introduction,

tutorial, history. Online.

Frankenreiter, M. (1990). Method for measuring blood

pressure and apparatus for automated blood pressure

measurements.

Lin, C.-T., Liu, S.-H., Wang, J.-J., and Wen, Z.-C. (2003).

Reduction of interference in oscillometric arterial

blood pressure measurement using fuzzy logic. Trans-

action on Biomedical Engineering, 50(4):432–441.

Neto, J. J. (1993). Contributions to Compiler Design

Methology. PhD thesis, Escola Polit

ecnica da Uni-

versidade de S

ao Paulo. In Portuguese.

Neto, J. J., de Almeira Jr., J. R., and dos Santos, J. M. N.

(1998). Synchronized statecharts for reactive systems.

Proceedings of the IASTED International Conference

on Applied Modelling and Simulation. Proceedings

of the IASTED International Conference on Applied

Modelling and Simulation.

Pearl, J. (1988). Probabilistic Reasoning in Intelligent

Systems: Networks of Plausible Inference. Morgan

Kauffman.

Pedrazzi, T., Tchemra, A., and de Azevedo da Rocha, R. L.

(2005). Adaptive decision tables - a case study of their

application to decision-taking problems. Proceedings

of International Conference on Adaptive and Natural

Computing Algorithms - ICANNGA 2005. Proceed-

ings of International Conference on Adaptive and Nat-

ural Computing Algorithms - ICANNGA 2005.

Pistori, H. and Neto, J. J. (2002). Adaptree - proposal of an

algorithm for decision tree induction based on adap-

tive technology. Anais Conferłncia Latino Americana

de Informtica - CLEI 2002. In Portuguese.

Pistori, H. and Neto, J. J. (2003). Decision tree induction

using adaptive fsa. CLEI Eletronic Journal, 6.

Quinlan, J. R. (1996). Learning decision tree classiﬁers.

ACM Computing Surveys.

Rolfe, P., editor (1979). Non-Invasive Physiological Mea-

surements. Academic Press.

Shutt, J. N. (1995). Self-modifying ﬁnite automata: Power

and limitations. Technical Report, (WPI-CS-TR-95-

4).

BIODEVICES 2010 - International Conference on Biomedical Electronics and Devices

116