ACTIVE MONITORING USING REAL-TIME METRIC LINEAR
TEMPORAL LOGIC SPECIFICATIONS
Gabor Simko and Janos Sztipanovits
Institute for Software Integrated Systems, Vanderbilt University, 1025, 16th Ave S, Nashville, U.S.A.
Keywords: Real-time monitoring, Temporal logic, Active monitoring, Policy monitoring, Control design.
Abstract: Monitoring temporal relationships among events in event streams has wide scale applicability in health
information systems. From detecting violations of privacy policies in message sequences to diagnosing
conditions in physiological data streams real-time event monitoring of temporal invariants is becoming an
important tool for system design. We developed an Active Real-Time Event Monitoring and Integration
System (ARTEMIS) capable of integrating event streams and monitoring the existence of temporal
invariants among events expressed in a safety fragment of metric first-order temporal logic (MFOTL). The
paper discusses the mathematical foundations of the monitor, and demonstrates the application concepts in a
physiological alarm generator and clinical information workflow system.
1 INTRODUCTION
The concept of policies is a widely used abstraction
in health information system design. Policies may be
defined with different purposes. For example,
privacy policies (Bartha et al., 2006) express
restrictions on information flows among actors of a
care delivery environment. Alert policies define the
rules for signalling alerts in clinical environments.
Treatment policies capture the decision rules for
applying and ordering treatment activities. These
policies share some common characteristics, namely,
all of them can be modelled by formal logic and
most of the time they contain temporal relationships.
Huge difference is found however in time scales
(from seconds to years), and whether they are only
passive monitors or active integrators of event
streams. Policies defined for passive monitoring
express logical and temporal invariants over event
streams. The passive monitors observe the event
sequences and indicate if the invariants are violated.
However, most treatment policies and several alert
policies need active participation, for generating new
events and integrating policy groups via the
generated events (Figure 1). This is necessary for
requesting actions (e.g. approvals in privacy policies
or tests in treatment policies) that may be time and
resource consuming and must be scheduled only by
demand. The request events
{
,…
}
and the
outcome of the requested activities{
,…
} may be
used for integrating other policies in the monitoring
process. The interplay between the different policies
leads to the concept of Active Real-Time Event
Monitoring and Integration System, (ARTEMIS)
which monitors different events, displays or logs
policy violations and starts up activities, which may
also affect other policies.
Figure 1: Schema of Active Real-Time Monitoring
System: large arrows denote the access points to external
resources, small arrows represent data flow. Monitor can
activate Tester Unit on demand, and the feedback
mechanism is provided to the Monitor.
Policies may be modelled with two different
approaches, rule-based and statistical based. Due to
space limitations we do not discuss statistical based
methods here. Temporal logic is often used to
370
Simko G. and Sztipanovits J..
ACTIVE MONITORING USING REAL-TIME METRIC LINEAR TEMPORAL LOGIC SPECIFICATIONS.
DOI: 10.5220/0003768703700373
In Proceedings of the International Conference on Health Informatics (HEALTHINF-2012), pages 370-373
ISBN: 978-989-8425-88-1
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
describe rule-based policies (Basin, Klaedtke, and
Muller, 2010) and to monitor events in a rule-based
fashion. Our choice based on algorithmic
considerations fall on a subset of temporal logic
called Metric First Order Temporal Logic (MFOTL)
(Basin et al., 2008). MFOTL can be used to specify
a broad set of complex temporal constraints, while
real-time operation and reasonable computation
complexity is achievable. We built our monitor
based on the concepts defined in their work, but over
an extended MFOTL
I
language and using a
significantly different monitoring algorithm.
The structure of the paper is the following: in
Section 2 we introduce MFOTL
I
language and
MFOTL
I
monitoring. Section 3 details the concept
of ARTEMIS, and in section 4 we show a clinical
workflow example. Finally, in section 5 we give the
conclusions.
2 BACKGROUND
2.1 Syntax of MFOTL
I
Language
We define the logic MFOTL
I
as a superset of
MFOTL (Basin et al., 2008), such that point and
interval based semantics are both supported.
MFOTL
I
may also be considered as the union of
metric temporal logic (MTL
N
) (Alur and Henzinger,
1991); (Koymans, 1990) and metric interval
temporal logic (MITL
[a,b]
) (Nickovic and Maler,
2007); (Alur et al., 1991) extended with predicates
and quantification. The past temporal modalities are
interpreted on possibly infinite intervals, while the
future temporal modalities are bounded. Description
of the MFOTL
I
language is based on the work of
Basin et al., (2008) and Nickovic and Maler,
(2007).
Formulae of MFOTL
I
are inductively defined in
Backus-Naur Form by the following grammar:
∶=
|
¬ 
|
∧ 
|
∃.
|
| 
,
where I is a possibly singular time interval and t is a
basic term, i.e. a function compared to a constant
value, or a boolean predicate value. The operators
represent the standard negation, conjunction,
existence, until and since operators, respectively.
Based on the basic formulae we can express
other standard logic operators and constants, such as
true (), false (), disjunction (
∨ 
) and
universality (∀.). Also, we can express release
and trigger operators, eventually and always
operators, and their past versions, once and
historically:
≔¬(¬
¬
)
≔¬
(
¬
¬
)
I
≔⊤
I
≔⊤
I
≔F
I
≔F
Here we illustrate the MFOTL
I
language with a
frequently arising example: the policy declares that a
patient p had to give their consent in the last 8 days
to disclose their lab results prior to the disclosure
taking place. The purpose of the monitor is to detect
if illegal disclosure happened. The policy is easily
expressed using MFOTL
I
:
disclosed
t
(x) →
[8days,0]
consent
t
(x)
If this expression is not satisfied at any time t with
respect to patient x, the policy was violated for that
patient.
2.2 MFOTL
I
Monitoring Algorithm
The main idea behind online monitoring is to
incrementally build an inner representation of
previous states without storing all the unnecessary
details. A possible solution is to introduce auxiliary
relations describing these past states. Satisfiability in
the current state is answered by evaluating only
these additional relations, i.e. answering the
satisfiability of a first-order logic expression. Our
monitoring algorithm works by transparently
building and evaluating these auxiliary relations.
3 ARTEMIS ARCHITECTURE
ARTEMIS systems are built from three different
kinds of components (Figure 1), sources, testers, and
monitor. Sources are independent event generators
(e.g. measurement devices or audit systems), which
set up relations and functions used by the MFOTL
I
monitor. The sources send all the necessary
information to the monitor, which automatically
extracts and stores the relevant data. The monitor
continuously checks whether the policies are
satisfied and on satisfaction the appropriate testers
(e.g. treatment procedures, lab tests) are activated.
Finally, the results of testers are fed back to the
monitor, which may lead to other coupled actions.
The controller actions are declared in the form of
Horn clauses (head(x) body(x)), where the head
is the action and body is the conditions leading to
the action. We can define any number of such
expressions as long as the body of the Horn clause is
temporal sub-formula domain independent (Basin et
al., 2008).
The expressions are ordered which defines the
ACTIVE MONITORING USING REAL-TIME METRIC LINEAR TEMPORAL LOGIC SPECIFICATIONS
371
Figure 2: Snapshot from our test application showing the relations of SIRS alert. Left side shows the relations of
ARTEMIS, right side shows the ground-truth data (signal high represents true, signal low represents false).
order of evaluation. Expressions may refer to each
other: any expression may use the past values of any
other expression including themselves or the current
results of any expressions defined earlier in the
order.
ARTEMIS shows several advantages over
systems built with low level languages (Table 1).
The fact that policies are well represented in formal
logic leads to a compact, readable and easily
maintainable policy code in ARTEMIS. The
optimized monitoring algorithm results in high
performance and optimal resource management.
Furthermore, extensibility is easily achieved by
using formal logic conjunction and disjunction
operators without touching any of the previously
written code.
Table 1: Advantages of ARTEMIS.
ARTEMIS with
MFOTL
monitoring
Traditional system
built with a low level
language
Policy code
complexity
Very compact Highly complex
Performance
Automatically
high, optimal
High, if optimized
Extensibility Easily extensible
Extensible, but
complicated
Maintenance
Easy
maintenance
Cumbersome
maintenance
Comparison to rule-based workflow management
systems like Drools does not show this great
difference. The main advantage of ARTEMIS is its
MFOTL engine which can evaluate temporal logic
with complex temporal expression. Although Drools
is very efficient to describe simple temporal policies,
it cannot handle compound temporal operators,
which may seriously limit its applicability in some
scenarios (e.g. signal processing).
4 EVALUATION
We evaluated ARTEMIS using the initial phases of
the sepsis alert and treatment protocol (Figure 2). By
monitoring several physiological data of patients we
could express the Systemic Inflammatory Response
Syndrome (SIRS) alert system (Shapiro et al., 2006).
After SIRS alert was issued, ARTEMIS had to
request the approval of a doctor to begin the sepsis
treatment protocol. In case the approval arrived, a
lab test request was issued to analyse additional
conditions. Only after the receipt of approval and lab
tests could the sepsis treatment start (Dellinger et al.,
2008).
The SIRS alert protocol (Shapiro et al., 2006)
defines validity ranges for physiological data. In
case the measured function is outside the validity
range, the measured data is abnormal. Abnormal
body temperature and white-blood cell count are
major criteria, abnormal respiration rate and heart
rate are minor criteria. The protocol defines two
kinds of alerts: high priority alerts are issued if two
major criteria were met in the last 24 hours; low
priority alerts are issued if at least one major and one
minor criterion were met in the last 24 hours, and no
alert were issued in the last 24 hours.
Our system contained four measured functions
interpreted on patients: temp (temperature), wbc
(white-blood cell count), rr (respiratory rate) and hr
(heart rate). Using these functions we could express
the policy as seen in Table 2. The derived abnormal
functions were satisfied for a patient x, if their
measurement was abnormal (i.e. out of normal
range). majorCriteria was true, if at least one major
criterion held, minorCriteria was true if at least on
minor criterion held. Based on these criteria we
could define highPriority and lowPriority, which
were satisfied when the high priority or low priority
alert requirements (see above) were met
HEALTHINF 2012 - International Conference on Health Informatics
372
Table 2: Sepsis treatment policy expressed in ARTEMIS.
(x) wbc(x) < 4000 | () > 20000
(x) (x) < 96.8 | (x) > 100.4
(x) (x) > 20
(x) ℎ(x) > 90
(x) 
(
x
)
∨(x)
(x) 
(
x
)
∨(x)
ℎℎ(x)
[
,
(x) ∧
[
,
(x)
(x)
[
,
m
(
x
)
[
,
min(x)
ℎ(x) ℎℎ(x)
(x)

(
x
)
∧¬
[,)

(
x
)
∨
(
x
)
(x)
ℎ
(
x
)
∨ 
(
x
)
∧¬
[
,
)
()

_
(x) 
(
x
)

_
(x)
[
,

(
x
)
∧
_
(x)
issueHighAlert and issueLowAlert signalled, when a
high priority or low priority alert had to be sent out
for patient x. For any patient x, when we reached
SIRSalert, we initialized the Verification tester (once
in every 24 hours, because the test takes significant
amount of time; also note that this is not part of the
original protocol, we used it for demonstration
purposes only) for that patient. Relation approval
signalled the results of the verification. If approval
was received, the patient was sent to lab tests by
issuing Sepsis_lab. On the arrival of lab tests,
relation labtests_done was updated. If lab tests were
done within 24 hours after the approval of sepsis
treatment, we could start the sepsis treatment action
signalled by Sepsis_treatment.
Even though the performance of the algorithm
heavily depends on several factors, we simulated
SIRS alert with 100000 distinct events affecting 100
patients leading to 25074 issued SIRS alert to
demonstrate the order of magnitude. The average
performance was 0.21ms / event.
5 CONCLUSIONS
We showed the concept of Active Real-Time Event
Monitoring and Integration System (ARTEMIS)
which is an extension of traditional real-time
monitoring systems with active participation from
the part of the monitor. As a workflow management
system ARTEMIS competes with systems like
Drools, but supports a broader and more expressive
set of temporal expressions which might show
immediate advantages in signal processing
scenarios.
REFERENCES
Barth, A., Datta, A., Mitchell, J. C., Nissenbaum, H.,
2006. Privacy and contextual integrity: Framework
and applications. SP’06, pp.184-198.
Basin, D., Klaedtke, F., Muller, S., 2010. Monitoring
security policies with metric first-order temporal logic.
SACMAT, pp.23-34.
Basin, D., Klaedtke, F., Muller, S., Pfitzmann, B., 2008.
Runtime monitoring of metric first-order temporal
properties. FSTTCS, 2, pp.49-60.
Chomicki, J., 1995. Efficient checking of temporal
integrity constraints using bounded history encoding.
TODS, 20(2), pp.149-186.
Alur, R., Henzinger, T., 1991. Logics and models of real
time: A survey. REX Workshop on Real Time: Theory
in Practice, pp.74-106.
Koymans, R., 1990. Specifying real-time properties with
metric temporal logic. Real-Time Systems, 2, pp.255-
299.
Nickovic, D., Maler, O., 2007. AMT: A property-based
monitoring tool for analog systems. FORMATS, 4763,
pp. 304-319.
Alur, R., Feder, T., Henzinger, T. A., 1991. The benefits
of relaxing punctuality. Tenth Annual Symposium on
Principles of Distributed Computing, pp.139-152.
Shapiro, N. I., Howell, M. D., Talmor, D. et al., 2006.
Implementation and outcomes of the Multiple Urgent
Sepsis Therapies (MUST) protocol. Critical Care
Medicine, 34, pp.1025-1032.
Dellinger, R., Levy, M., Carlet, J. et al., 2008. Surviving
Sepsis Campaign: International guidelines for
management of severe sepsis and septic shock.
Intensive Care Medicine, 34, pp.17-60.
ACTIVE MONITORING USING REAL-TIME METRIC LINEAR TEMPORAL LOGIC SPECIFICATIONS
373