Prospective Trials of Intelligent Alarm Algorithms for Patient Monitoring

aroocarmineAI and Robotics

Oct 29, 2013 (4 years and 8 months ago)


Prospective Trials of Intelligent Alarm Algorithms for Patient Monitoring

Ying Zhang
and Christine L. Tsien, M.D., Ph.D.

Laboratory for Computer Science, MIT, Cambridge, MA
Brigham and Women’s Hospital/Massachusetts
General Hospital Harvard Affiliated Emergency Medicine Residency, Boston, MA

Abstract. “Intelligent” alarm algorithms have the
potential to improve monitoring systems for critical
care by safe reduction of false alarms and better
integration of physiological signals. Evaluation is a
key process in the development of these algorithms.
This study explores a new evaluation approach:
prospective trials in real time. Preliminary results
indicate that this approach may allow better data
correlation, collection of algorithm-specific
information, and adaptive learning in real time.

Background. Patient monitoring systems for critical
care should sound an alarm when the patient needs
immediate attention. Previous studies have shown,
however, that as many as 86% of alarms in the
Intensive Care Unit (ICU) are false or clinically
irrelevant [1]. Most alarm algorithms of state-of-the-
art monitors process individual physiological signals
separately and detect clinical events based on signal
thresholds. These may have been adequate had their
inputs accurately represented patients’ vital signs; but
the signals (e.g., ECG waveforms) are often noisy,
and the numerics (e.g., heart rate) derived from them
may not be accurate or available.
Intelligent alarm algorithms that employ AI
techniques have the potential to reduce false alarm
rates by incorporating multi-parameter signal analysis
and trend detection into the decision process [2]. The
conventional approach for evaluating these
algorithms is to try them on an existing database of
monitor data. Most databases, however, lack the
clinical event annotations necessary for correlating
physiological signals with their annotations due to
problems such as inadequate annotations and poor
data synchronization. Thus, results from most trials
of intelligent alarm algorithms are still largely
speculative. This study explores an alternative
approach to testing intelligent alarm algorithms that
processes the numerics: prospective trials along with
simultaneous collection of physiological data and
clinical annotations in real time.

System. A laptop is connected to a bedside monitor
(HP Viridia CMS) via an RS232 interface. The
communication protocol is based on a client/server
model. The laptop collects numerics and alarm status
from the monitor and stores them in the first dataset
of an SQL database. A customized application runs
intelligent alarm algorithms on the numerics and
records their alarm decisions in the second dataset of
the database. A trained observer at the bedside
questions physicians and/or nurses about the patient’s
state when the monitor or the algorithm displays an
alarm, and records her findings and other clinical
events in the third dataset of the database as clinical
annotations through an interactive user interface.
Each dataset is stamped with the laptop’s system time
to ensure time synchronization and accurate
correlation of the datasets.

Evaluation. This system has been implemented at
the Multidisciplinary ICU (MICU) at Children’s
Hospital in Boston. It has been used to conduct
prospective trials of an intelligent alarm algorithm
based on decision trees. The physiological data and
clinical annotations are well synchronized. These
trials are capable of capturing some false negatives.
They may also provide resources necessary for
intelligent alarm algorithms to perform adaptive
learning in real time, i.e., automatic adjustment of
algorithm parameters according to patterns in an
individual patient’s physiological data.

Conclusions. Prospective trials of intelligent alarm
algorithms in real time may provide a more effective
evaluation of such algorithms than can be done by
retrospective analysis alone. They also allow unique
data integration and analysis that can support
functionality such as adaptive learning; this enables
alarm algorithms to be patient-specific. This study is
in progress; it will bring to surface more benefits and
challenges of intelligent alarm algorithm
development and help researchers better understand
and use physiological signal patterns in critical care.

Acknowledgments. The authors would like to thank
M. Curley, J. Fackler, I. Kohane, D. Martin, A.
Randolph, P. Szolovits, and the MICU nurses at
Children’s Hospital in Boston.

1. Tsien CL, Fackler JC. Poor prognosis for
existing monitors in the intensive care unit. Crit
Care Med 1997; 25: 614-619.
2. Tsien CL. TrendFinder: Automated detection of
alarmable trends. Laboratory for Computer
Science Technical Report 809, Massachusetts
Institute of Technology, July 2000.