ABSTRACT

An Acute Hypotensive Episode (AHE) is the sudden onset of a period of sustained low blood
pressure and is one of the most critical conditions in Intensive Care Units (ICU). Without timely
medical care, it can lead to irreversible organ
and death. By identifying patients at risk for
this complication, adequate medical intervention can save lives and improve patient outcomes. In
this paper we study the problem of identifying patients at risk for AHE. We cast the problem as a
supervised classication task and design a novel dual{boundary classi_ca- tion algorithm. Our
algorithm uses only past blood pressure measurements of the patients thereby being much
simpler than many existing methods that use multiple sources of data. It can also be used in
online or batch mode which is advantageous in an ICU setting. Our extensive experiments on
1700 patients' records demonstrate that the algorithm signicantly outperforms existing
approaches in predictive accuracy, sensitivity and specicity. It can identify patients at risk for
AHE with nearly 95% accuracy up to 120 minutes before the episode begins.
Introduction

Hypotension is also an important mortality predictor for patients with cardiovascular

abnormalities . Hypotension is a precursor to septic shock, the second most common cause of
death in ICU patients in the United States and a study shows that mortality in such cases
depends criti-cally on the duration of hypotension before treatment .If patients at risk for AHE
are identi_ed in advance, then timely intervention can prevent a number of complications
and death.Automated systems for predicting AHE and other compli-cations in ICU patients are
becoming increasingly important given the enormous di_culties of manually monitoring the
abundance of data collected for each patient and the short-age of quali_ed clinical sta_ in ICUs .
Several works in the past have attempted to build automated machine learn-ing based systems
that can identify patients at risk for AHE in advance. Section (2.1) outlines the results of these
stud-ies.
This paper presents a novel supervised classi_cation based system for identifying patients at risk
for AHE. There are three advantages of our algorithm over the state-of-the-art:_ Our algorithm
achieves signi_cantly higher accuracy, sensitivity and speci_city compared to existing algorithms, in our experiments. The performance improve-ment is observed for predictions made up
to 120 min- utes in advance. Today most systems can predict the onset of AHE only up to 30
minutes in advance (at an accuracy that is inferior to our algorithm's) which is often insu_cient
for adequate medical intervention. _ Unlike several existing systems that use multiple sources of
data, we use only blood pressure measurements to predict AHE. Blood pressure is monitored
easily and widely in ICUs and so, a predictor based on only blood pressure measurements is easy
to deploy. Thus, we avoid the problems of using heterogeneous sources of data (such as other
vital signs, laboratory records,

nursing charts etc.) each of them fraught with con-founding issues like di_ering sampling
frequencies, lack of synchronization, missing values, instrument errors and errors in entry._ Our
algorithm can be used both in o_ine and online modes. As more data from new patients as well
as existing patients arrive, the classfier updates itself in an online manner using only the newly
available data (i.e. without re{training itself with the entire dataset).

Existing System
Existing system use a large number of ways in which features can be generated for the data
A widely used approach in signal processing is to trans-form the input time series into feature
vectors (in a different space) and perform classification on the transformed features. We
investigate the following transformations:
_ Reconstructed Phase Space (RPS) transforms
_ Discrete Wavelet transforms
_ Transformations based on Statistical features
RPS based transforms were used for signal classification. In their approach, the original time
series signals are transformed into RPS matrices which rep-resent the data in a topologically
equivalent space
Discrete wavelet transforms are widely used in biomedical signal processing to transform timeseries data into features that capture both frequency and temporal information. We use
Daubechies wavelets to obtain the wavelet features.

Disadvantages

1) Cannot achieve better performance and accuracy

2) Lack in missing value Estimations

Proposed System

The proposed system presents a novel supervised classification based system for identifying
patients at risk for AHE. There are three advantages of our algorithm over the state-of-the-art:
Our algorithm achieves significantly higher accuracy, sensitivity and specificity compared to
existing algo-rithms, in our experiments. The performance improve-ment is observed for
predictions made up to 120 min- utes in advance. Today most systems can predict the onset of
AHE only up to 30 minutes in advance (at an accuracy that is inferior to our algorithm's) which
is often insuffcient for adequate medical intervention._ Unlike several existing systems that use
multiple sources of data, we use only blood pressure measurements to predict AHE. Blood
pressure is monitored easily and widely in ICUs and so, a predictor based on only blood pressure
measurements is easy to deploy. Thus, we avoid the problems of using heterogeneous sources of
data (such as other vital signs, laboratory records, nursing charts etc.) each of them fraught with
con- founding issues like differing sampling frequencies, lack of synchronization, missing
values, instrument errors and errors in entry.This algorithm can be used both in offine and online
modes. As more data from new patients as well as existing patients arrive, the classifier updates
itself in an online manner using only the newly available data (i.e. without re{training itself with
the entire dataset).

Advantages

1) High Accuracy
2) Avoid problems in heterogeneous source of data

Literature Survey
1) A methodology for prediction of acute hypotensive episodes in ICU via a risk scoring
model including analysis of ST-segment variations.

The aim of this study is to detect Acute Hypotensive Episodes (AHE) and Mean Arterial Pressure
Dropping Regimes (MAPDRs) using ECG signal and Arterial Blood Pressure waveforms. To
meet this end, the QRS complexes and end-systolic end-diastolic pulses are first extracted using

two innovative Modified Hilbert Transform-Based algorithms namely as ECGMHT and

BPMHT. A new smoothing algorithm is next developed based on piecewise polynomial fitting to
smooth the fast fluctuations observed in RR-tachogram, systolic blood pressure (SBP) and
diastolic blood pressure (DBP) trends. Afterwards, in order to consider the mutual influence of
parameters on the evaluation of shock probability, a Sugeno Adaptive Network-based Fuzzy
Inference System-ANFIS is trained using Hasdai et al. (J Am Coll Cardiol, 35: 136143, 2000)
parameters as input, with appropriate membership functions for each parameter. Using this
network, it will be possible to incorporate the possible mutual influences between risk
parameters such as heart rate, SBP, DBP, ST-segment episodes, age, gender, weight and some
miscellaneous factors to the calculation of shock occurrence probability. In the next step, the
proposed algorithm is applied to 15 subjects of the MIMIC II Database and AHE and MAPDRs
(MAP 60 mmHg with a period of 30 min or more) are identified. As a result of this study, for a
sequence of MAPDRs as long as 20 min or more, there will exist a consequent high peak with
the duration of 34 min in the corresponding probability of cardiogenic shock diagram.

Disadvantages

1) May results in false prediction

An investigation of pat- terns in hemodynamic data indicative of impending hypotension in

intensive care

In the intensive care unit (ICU), clinical staff must stay vigilant to promptly detect and treat
hypotensive episodes (HEs). Given the stressful context of busy ICUs, an automated hypotensive
risk stratifier can help ICU clinicians focus care and resources by prospectively identifying
patients at increased risk of impending HEs. The objective of this study was to investigate the
possible existence of discriminatory patterns in hemodynamic data that can be indicative of
future hypotensive risk.

Disadvantages

1) Cannot achieve better performance and accuracy

Similarity-Based Searching in Multi-Parameter Time Series Databases

It presents a similarity-based searching and pattern matching algorithm that identifies time series
data with similar temporal dynamics in large-scale, multi-parameter databases. We represent time
series segments by feature vectors that reflect the dynamical patterns of single and multidimensional physiological time series. Features include regression slopes at varying time scales,
maximum transient changes, auto-correlation coefficients of individual signals, and cross
correlations among multiple signals. We model the dynamical patterns with a Gaussian mixture
model (GMM) learned with the Expectation Maximization algorithm, and compute similarity
between segments as Mahalanobis distances. We evaluate the use of our algorithm in three
applications: search-by-example based data retrieval, event classification, and forecasting, using
synthetic and real physiologic time series from a variety of sources.

Disadvantages

1) Lack in missing value Estimations

Modules
Dataset Selection

MIMIC II is a publicly available database, as part of Physionet , containing physiological signals

and clinical data of more than 2300 ICU patients. Vital signs of most of these ICU patients have
been recorded, sampled every minute. We consider only the Mean Arterial Blood Pressure
(MAP) measurements for our study. Each patient record consists of a time series signal xt; t =
1; : : : ;N where N is the total number of MAP measurements for the patient, and xt is the MAP
measurement at time t, time being reckoned from the start of the measurements in the ICU at an
interval of 1 minute. An Acute Hypotensive Event (AHE) is defined as an interval in the record,
[xi; xi+30], in which at least 27 of the measurements are no greater than 60.

Parameters of Algorithms
Prediction Window: Duration of time (in minutes) before which predictions are made, i.e., we
predict whether or not a patient will have AHE w minutes in the future; we call w the prediction
window.
Observation Window: Duration of time (in min-utes) before the prediction window during which
MAP measurements are considered for training the classifier.
Test Window: Duration of time (in minutes) before the prediction window during which MAP
measure- ments are considered for predicting AHE in a patient

The training set

This module training set of 60 cases from the set of MIMIC II records meeting the initial
selection criteria and chose a time interval T0 for each case. The training set consisted of 15
records from each of groups H1, H2, C1, and C2. The classification of each case, and the data
before and after T0, were available for study.

Testing and Prediction

From the newly available records meeting the initial selection criteria, we selected 50 for the
Challenge test sets, and chose a T0 for each case. Each selected record was divided into an a
segment including all data available more than 10 hours before T0, typically beginning at or
shortly following the patients admission to the ICU; a 10-hour b segment beginning at the end
of the a segment and ending at T0; and a final c segment beginning at T0 and ending at the
patients discharge from the ICU

Our classification algorithm can easily be adapted to run in batch/online mode. As new training
data arrives are updated. The mean and standard deviation values can be updated using online
algorithms. In the data we observe patients whose MAP measurements actuate considerably in
the observation window. For ex- ample, in some cases, the BP remains in the normal range
Initially and becomes low and again comes back to the nor- mal range after some time. In such
cases, the mean is often close to the decision boundary and a single boundary clas- si_er tends to

misclassify them. These cases are captured in the uncertainty region in our classifier and the
distance based rule classifies them well in practice.

Performance Evaluation:

In this module the performance and accuracy of the proposed is analyzed and compared with
existing Methods.

CONCLUSION
It Describes a new classification method that can be used to identify ICU patients who are at risk
for Acute Hypotensive Episodes (AHE). Our extensive experi-ments by varying the observation,
test and prediction win-dows for the tested algorithms have not been conducted in previous
studies. Thus we thoroughly evaluate our method comparing it with mean-based and neural
network based predictors, which have been the best predictors until now, as well as some other
methods that were previously untested. These experiments demonstrate that our method outperforms existing methods for predicting AHE in a variety of experimental conditions. Our
algorithm can easily be inte-grated into patient bedside monitors to provide alerts and
noti_cations to ICU sta_ regarding potential episodes of acute hypotension in patients.Our
method has both high sensitivity and speci_city which is important in ICU conditions. Methods
with low sensitiv-ity will not be able to predict AHE in patients well and are clearly unusable.

Future Enhancements

Moreover methods with high sensitivity and low specificity (such as previous SVM based
approaches) are also practically undesirable since their implementations in real ICU monitors
will raise a number of false alarms (by predicting AHE for even those patients who do not
develop AHE).In future work we have to apply it in real time and take necessary steps to achieve
very high predictive accuracy.

References
[1] X Chen, D Xu, G Zhang, and R Mukkamala. Fore- casting acute hypotensive episodes in
intensive care
patients based on a peripheral arterial blood pressure waveform. In Computers in Cardiology,
2009, pages 545{548. IEEE, 2009. [2] Cristina Crespo et al. Precursors in the arterial blood
pressure signal to episodes of acute hypoten- sion in sepsis. In Proceedings of the 16th
International EURASIP Conference BIOSIGNAL, volume 16, pages 206{8, 2002.
[3] A. Gha_ari, M. R. Homaeinezhad, M. Atarod, and M. Akraminia. A methodology for
prediction of acute hypotensive episodes in ICU via a risk scoring model including analysis of
ST-segment variations. Cardio- vascular Engineering, 10(1):121{29, 2010.
[4] Marzyeh Ghassemi. Methods and models for acute hy- potensive episode prediction. In MSc
Thesis, pages 541{ 544. University of Oxford, 2011.
[5] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdor_, P. Ch. Ivanov, R. G. Mark, J. E.
Mietus,
G. B. Moody, C.-K. Peng, and H. E. Stanley. Phys- ioBank, PhysioToolkit, and PhysioNet:
Components of
a new research resource for complex physiologic signals. Circulation, 101(23):e215{e220, 2000
(June 13).
[6] JH Henriques and TR Rocha. Prediction of acute hy- potensive episodes using neural network
multi-models. In Computers in Cardiology, 2009, pages 549{552. IEEE, 2009.
[7] J Hope Kilgannon, Brian W Roberts, Lisa R Reihl, Michael E Chansky, Alan E Jones, R
Phillip Dellinger,

Joseph E Parrillo, and Stephen Trzeciak. Early arterial hypotension is common in the postcardiac arrest syn- drome and associated with increased in-hospital mor- tality. Resuscitation,
79(3):410{416, 2008.
[8]
Richard
E.
Klabunde.
http://www.cvphysiology.com/Blood%

Cardiovascular

physiol-

ogy

concepts.

20Pressure/BP006.htm.

[9] Donald E Knuth. The art of programming, vol. 2, semi- numerical algorithms, 1981.
[10] Anand Kumar, Daniel Roberts, Kenneth E Wood, Bruce Light, Joseph E Parrillo, Satendra
Sharma,
Robert Suppes, Daniel Feinstein, Sergio Zanotti, Leo Taiberg, et al. Duration of hypotension
before initiation of e_ective antimicrobial therapy is the critical determinant of survival in human septic
shock. Critical Care Medicine, 34(6):1589{1596, 2006.
[11] Joon Lee and Roger G Mark. An investigation of pat- terns in hemodynamic data indicative
of impending hy- potension in intensive care. Biomedical Engineering Online, 9(1):62, 2010.
[12] Kerry L Lee, Lynn H Woodlief, Eric J Topol, W Dou-glas Weaver, Amadeo Betriu, Jacques
Col, Maarten Simoons, Phil Aylward, Frans Van de Werf, Robert M Cali_, et al. Predictors of 30Day Mortality in the Era of Reperfusion for Acute Myocardial Infarction Results from an
International Trial of 41021 patients. Circula-tion, 91(6):1659{1668, 1995.
[13] LH Lehman, M Saeed, GB Moody, and RG Mark.Similarity-based searching in multiparameter time se-ries databases. In Computers in Cardiology, 2008, pages 653{656. IEEE,
2008.
[14] Greg S Martin, David M Mannino, Stephanie Eaton,and Marc Moss. The epidemiology of
sepsis in the United States from 1979 through 2000. New England Journal of Medicine,
348(16):1546{1554, 2003.
[15] MAMneimneh and RJ Povinelli. A rule-based approach for the prediction of acute
hypotensive episodes. In Computers in Cardiology, 2009, pages 557{560. IEEE,2009.