VITAL-SIGN DATA FUSION MODELS FOR POST-OPERATIVE

PATIENTS

Marco A. F. Pimentel

, David A. Clifton

, Lei Clifton

, Peter J. Watkinson

and Lionel Tarassenko

Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford, Oxford, U.K.

Nufﬁeld Department of Anaesthetics, University of Oxford, Oxford, U.K.

Keywords:

Patient Monitoring, Early Warning Scores, Novelty Detection.

Abstract:

Deterioration in Patients who undergo upper-gastrointestinal surgery may be evident in the vital signs prior

to adverse events. A dataset comprising observational vital-sign data from 128 post-operative patients was

used to explore the trajectory of patients vital-sign changes during their stay in the post-operative ward. A

model of normality based on pre-discharge data from patients who had a “normal” recovery was constructed

using kernel density estimates, and tested with “abnormal” data from patients who deteriorate sufﬁciently to

be re-admitted to the Intensive Care Unit. The results suggest that the criticality of post-operative patients can

be evaluated by assessment of the distributions of their vital signs after their admission to the post-operative

ward.

1 INTRODUCTION

A large number of preventable and avoidable deaths

occur in hospitals each year in the UK, due to adverse

events such as cardiac arrest and unplanned admis-

sion into the ICU from other hospital wards (Bardell

et al., 2003; Kause et al., 2004). Patients who undergo

surgery for the removal of upper- gastrointestinal (GI)

cancer have a high incidence of post-operative com-

plications, often resulting in readmission to the In-

tensive Care Unit (ICU) several days after surgery

(Thompson et al., 2003; Veltkamp et al., 2002). Fail-

ure to identify such deteriorations in a timely manner

has led to the design of a two-phase clinical trial at

the Oxford Cancer Hospital, the “CALMS2” trial, in

which ambulatory post-operative cancer patients are

monitored using wearable sensors that are connected

by the hospital wireless network to patient monitors

and nursing display stations. Vital-sign measure-

ments are also made periodically by the nurses on the

ward as part of the usual post-operative care.

Our approach to monitoring in-hospital patients

relies on constructing models of normality based on

the vital signs acquired from a large population of

acutely-ill patients (Tarassenko et al., 2006; Hann,

2008). We propose to improve upon existing models

by investigating models tuned to speciﬁc patient pop-

ulations. Post-operative cancer patients are recover-

ing from surgery, so that they start in their most acute

state and then gradually stabilise. In this work, we

therefore adopt a novel approach in which the aim

is to learn the vital-sign trajectories associated with

“normal” recovery of these patients, allowing “abnor-

mal” trajectories to be identiﬁed in subsequent pa-

tients.

We present preliminary results of the analysis of

data acquired during Phase I of the CALMS2 clinical

trial. A multivariate model of the distribution of vital-

sign data from “normal” patients, which describe the

normal physiological trajectory, is constructed using

a non-parametric statistical approach, a kernel den-

sity estimate. This model is then tested on “abnor-

mal” data from patients who deteriorate sufﬁciently

after surgery to be re-admitted to the ICU.

2 DATASET

The dataset used for the work described by this paper

comprises measurements of heart rate (HR), respira-

tory rate (RR), peripheral arterial oxygen saturation

(SpO

), systolic blood pressure (SBP) and tempera-

ture acquired by ward staff every hour or every two

hours in the ﬁrst days after patient admission (de-

pending on the patient’s condition), and every four

hours in the last days of the patient’s stay on the ward.

A set of 15, 029 observations X ∈ R

was obtained

from 169 patients. These patients were admitted to

410

A. F. Pimentel M., A. Clifton D., Clifton L., J. Watkinson P. and Tarassenko L..

VITAL-SIGN DATA FUSION MODELS FOR POST-OPERATIVE PATIENTS.

DOI: 10.5220/0003789104100413

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2012), pages 410-413

ISBN: 978-989-8425-89-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

the Upper GI ward and the median length of stay on

the ward was 8 days (IQR 6-11). In order to select

a subgroup of patients with “normal” recovery from

surgery (class C

), we chose those patients who stayed

more than 4 days and less than 25 days on the ward,

and who were not (re-)admitted to the ICU. This re-

sulted in 128 patients being included in class C

It is assumed, a priori, that each of the D = 5 pa-

rameters has equal importance in the patient model.

Each parameter was therefore scaled to have the

same dynamic range using a zero-mean, unit-variance

transformation.

The vital-sign trajectory throughout patient stays

on the ward was evaluated by examining the 5 sub-

groups of data:

• G

: comprises the ﬁrst clinical observation per-

formed when a patient was admitted to the ward;

• G

: is the average of all observations performed

on the ﬁrst day of a patient’s stay on the ward;

• G

: comprises the average of all observations per-

formed on the day that corresponds to half of the

length of the patient’s stay on the ward (if the pa-

tient stayed 5 days on the ward prior to discharge,

this day would correspond to day three);

• G

: contains the average of all observations per-

formed on the day that corresponds to 75% of the

length of the patient’s stay on the ward;

• G

: contains the average of all observations per-

formed on the last day of the patient’s stay on the

ward.

These ﬁve subgroups were deﬁned in this way be-

cause of the different lengths of patient stay on the

ward (from 4 to 25 days).

3 DATA VISUALISATION

3.1 Methodology

The ﬁrst stage in constructing a model of normal-

ity for novelty detection usually consists in obtain-

ing more insight into the structure of the data. Pro-

cedures for visualisation of the data in their original

high-dimensional space are required. The relations of

proximity of the vital-sign data in their original space,

with D > 3, can be visualised through a non-linear

projection from R

to R

or R

A Sammon-mapping algorithm (Sammon, 1969)

was implemented and applied in the 640 mean vec-

tors contained in the 5 subgroups (G

, G

, with 128 data in each) from patients belonging to

class C

3.2 Results

Represented in each map of Figure 1 are the projected

data from G

, G

superimposed on the pro-

jected data points from G

(data from the last day on

the ward for each patient).

It may be seen from the plots that the groups form

Figure 1: Sammon maps obtained for the group dataset: (a) G

and G

, (b) G

and G

, (c) G

and G

, and (d) G

and G

; in

(a)-(d) G

, G

and G

are represented by red {} and G

by blue {o}.

VITAL-SIGN DATA FUSION MODELS FOR POST-OPERATIVE PATIENTS

411

clusters with some overlap between them. If we con-

sider projected data from groups G

and G

(shown in

red and blue, respectively in the top-left plot), it may

be seen that the former are dispersed, while the latter

are more concentrated. It may also be seen that the

distributions of the data for G

(data that correspond

to the midpoint of patient’s stay) and G

groups are

very similar and evenly concentrated. This suggests

that for the 128 patients with a normal recovery pro-

cess from surgery, there are no large changes in the

vital-sign distributions after the midway point during

their stay on the ward. These results suggest the pos-

sibility that patients belonging to class C

could have

been discharged early, or provided with a lower level

of care from the halfway point of their stay, as they

had already been stabilised and no signiﬁcant changes

in their vital-sign distributions subsequently occured.

4 MODEL OF NORMALITY

4.1 Methodology

For the construction of the model of normality, we

considered the G

subgroup of measurements as the

“normal” dataset. This contains the most stable phys-

iological variables (because these data are acquired

just before discharge, when the patient is most “nor-

mal” ). At discharge the patient is stable and deemed

well enough to go home. Therefore, the G

dataset

of N = 128 prototype vectors, X ∈ R

, was used for

construction of the model of normality.

A kernel density estimate (Bishop, 2007) is a tech-

nique that allows the underlying 5-D vital-sign proba-

bilty density function (pdf) to be estimated from train-

ing data. While other methods, such as Gaussian mix-

ture models, were considered, a kernel density esti-

mate was chosen as it has the advantage of being a

fully non-parametric method, so no a priori assump-

tions are made about the form of the probability dis-

tribution.

This method was initially used to estimate the pdf

of the 128 prototype vectors, x

,...,x

p(x) =

N(2π)

D/2

∑

i=1

−

|x−x

2σ

(1)

which is a weighted sum of Gaussian kernels centred

on the 128 prototype vectors, x

, and where each ker-

nel is isotropic with variance σ

. The variance was

determined using the nearest-neighbour method pro-

posed by Bishop (Bishop, 2007).

In the ﬁnal model, only those patients who were

discharged within 5 to 9 days were considered for

building the estimate of the pdf (N = 78). The ratio-

nale for this decision was to focus on the most “nor-

mal” patients from class C

(i.e. the patients who re-

cover entirely as expected after surgery). To identify

those patients, the mode of the lengths of stay on the

ward was calculated, and only the patients whose stay

on the ward was within the mode ± 2 days were taken

into account. This corresponds to the range [5 9] days.

The averages of the observations from each day for

each of these patients were calculated and the pdfs

p(x|θ), where θ corresponds to the model parameters,

were then determined from Equation (1) with N = 78.

In order to estimate the abnormality of a

previously-unseen test pattern, the departure from

normality is usually quantiﬁed using novelty deﬁned

as z(x) = − log p(x|θ), where z(x) is the novelty

score. This process leads to “normal” data, which

have higher likelihoods p(x|θ), generating low nov-

elty scores, and “abnormal” data, which have lower

likelihoods, generating high novelty scores.

4.2 Results

The novelty scores z(x) from the most “normal” 78

patients using the 5-D pre-discharge data (G

) to con-

struct the model of normality are shown in Figure 2,

averaged for each patient for each day. These scores

conﬁrm the results given by the Sammon maps. From

the trajectory of the novelty z(x) shown in Figure 2 we

can conclude that there is a signiﬁcant decrease of the

value of z(x) during the ﬁrst 3 days. From day 4 on-

wards, z(x) remains approximately constant. The ﬁrst

3 days could be associated with the patient’s recov-

ery in the ﬁrst post-operative days (Thompson et al.,

2003), whereas after day 4, the majority of patients

Figure 2: Representation of average of the novelty scores

z(x) = −log p(x|θ) with time for “normal” patients, where

the data probability density was constructed using pre-

discharge data from the same group of patients. Error bars

denoted one standard error of the group mean.

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

412

included in this study appear to have fully recovered

from surgery and to be stable. They are well enough

to either be discharged or be provided with a lower

level of care for their remaining time on the ward.

The multivariate model of normality was also

tested on “abnormal” vital-sign data from two patients

who deteriorated sufﬁciently after surgery to be re-

admitted to the ICU. The model was able to identify

patient deterioration up to 12 hours before the pa-

tient’s readmission to the ICU. These results are omit-

ted for brevity.

5 CONCLUSIONS AND FUTURE

WORK

We have presented the preliminary analysis of data

acquired from patients who were admitted to the

post-operative Upper GI ward after cancer surgery.

We studied the vital-sign distributions at the time of

the patient’s admission to the ward, halfway through

their stay, and near the time of discharge. Although

changes in vital-sign distributions from “normal” pa-

tients between admission to the ward and subsequent

discharge were found, no signiﬁcant changes in these

distributions were observed from halfway through

their stay to the time of discharge, which suggests that

these patients could have been discharged earlier or

provided with a lower level of care from the halfway

point of their stay on the ward.

A multivariate model of the distribution of vital-

sign data from “normal” patients, which describe the

normal trajectories, was constructed using a kernel

density estimate. The model conﬁrmed the results ob-

tained with the Sammon maps.

Future work will concentrate on the reﬁnement of

existing techniques for the target population group,

and on the improvement of model construction using

more complex dynamical modelling methods. How-

ever, there are a number of limitations that must be

overcome in future analysis. The dataset used in the

analysis described by this paper consisted of measure-

ments of vital signs acquired periodically (every 2 or

4 hours) by ward staff during Phase I of the clinical

trial. These infrequent patient observations combined

with lower nurse:patient ratios (typically, 1:4 to 1:10

in step-down wards), can lead to unnoticed clinical

deterioration and, consequently, to increased numbers

of adverse events. We are developing a continuous

vital-sign monitoring system to provide early warn-

ing of patient deterioration in a robust manner with

low numbers of false alarms.

ACKNOWLEDGEMENTS

The work described in this paper was funded by the

NIHR Biomedical Research Centre Programme, Ox-

ford. MAFP was supported by the RCUK Digital

Economy Programme grant number EP/G036861/1

(Oxford Centre for Doctoral Training in Healthcare

Innovation), and Dr David Clifton was supported by

the Wellcome Trust and the EPSRC under grant num-

ber WT 088877/Z/09/Z. The authors also wish to

thank the support of all clinical staff involved in the

collection of the data used in this investigation.

REFERENCES

Bardell, T., Legare, J. F., Buth, K. J., Hirsch, G. M.,

and Ali, I. S. (2003). ICU readmission after car-

diac surgery. European Journal of Cardio-Thoracic

Surgery, 23(3):354–359.

Bishop, C. M. (2007). Pattern Recognition and Machine

Learning. Springer.

Hann, A. (2008). Multi-parameter monitoring for early

warning of patient deterioration. Ph.D. thesis, Uni-

versity of Oxford.

Kause, J., Smith, G., Prytherch, D., Parr, M., Flabouris,

A., and Hillman, K. (2004). A comparison of an-

tecedents to cardiac arrests, deaths and EMergency in-

tensive care admissions in australia and new zealand,

and the united kingdom–the ACADEMIA study. Re-

suscitation, 62(3):275–282.

Sammon, J. W. (1969). A nonlinear mapping for data struc-

ture analysis. IEEE Transactions on Computers, C-

18(5):401– 409.

Tarassenko, L., Hann, A., and Young, D. (2006). Integrated

monitoring and analysis for early warning of patient

deterioration. British Journal of Anaesthesia, 97(1):64

–68.

Thompson, J. S., Baxter, B. T., Allison, J. G., Johnson,

F. E., Lee, K. K., and Park, W. Y. (2003). Temporal

patterns of postoperative complications. Arch Surg,

138(6):596–603.

Veltkamp, S. C., Kemmeren, J. M., van der Graaf, Y.,

Edlinger, M., and van der Werken, C. (2002). Pre-

diction of serious complications in patients admitted

to a surgical ward. The British Journal of Surgery,

89(1):94–102.

VITAL-SIGN DATA FUSION MODELS FOR POST-OPERATIVE PATIENTS

413