Vegetative State: Early Prediction of

Clinical Outcome by Artificial Neural Network

L. Pignolo

F. Riganello

, A. Candelieri

and V. Lagani

S. Anna Institute, RAN – Research on Advanced Neuro-rehabilitation, Crotone, Italy

Laboratory for Decision Engineering and Health Care Delivery

Department of Electronic Informatics and Systemistics

University of Calabria, Cosenza, Italy

Abstract. Residual brain function has been documented in vegetative state

patients, yet early prognosis remains difficult. Purpose of this study was to

identify by artificial Neural Network procedures the significant neurological

signs correlated to, and predictive of outcome. The best networks test set

accuracy was 70%, 72% and 70% for the entire patients’ group and the

posttraumatic and non-posttraumatic subgroups, respectively. The method

accuracy does not reflect a perfect classification, but is significantly far from

the random or educated guess and is in accordance with the results of previous

clinical studies.

1 Introduction

The Vegetative State (VS) is a clinical condition characterized by the absence of

awareness (of self and environment), voluntary or purposeful behavioral responses to

external stimuli, and communication in the severely brain damaged. Subjects in VS

are otherwise awake, often with wakefulness-sleep cycles [1, 2, 3, 4, 5, 6]. Recovery

(with varying residual disabilities) occurs only in a portion of patients; resources,

staff, logistics and costs requirements for the care of these subjects are imposing

irrespective of outcome. Purpose of this study was to identify by artificial intelligence

procedures a significant model supporting decision in the early prognosis of VS

subjects [7, 8]. It should be noted in this regard that evidence-based neurology

indicates significant neurological signs correlated to, and predictive of outcome.

Prognosis can be modeled as a regression, classification or survival analysis problem

by traditional statistics or machine learning techniques [9]. This study is purposed to

demonstrate that reliable classification models predictive of the vegetative state

outcome prognosis can be obtained by Artificial Neural Networks (ANN) techniques.

Section 2 of this paper outlines dataset and pre- processing; Section 3 describes the

experimentation protocol for the training of classification models; Sections 4 and 5

summarize and comment the results.

Pignolo L., Riganello F., Candelieri A. and Lagani V. (2009).

Vegetative State: Early Prediction of Clinical Outcome by Artiﬁcial Neural Network.

In Proceedings of the 5th International Workshop on Artiﬁcial Neural Networks and Intelligent Information Processing, pages 91-96

DOI: 10.5220/0002264300910096

 SciTePress

2 Data Collection and Pre-processing

2.1 Data Collection

Three hundred and thirty three subjects in VS consecutively admitted to the dedicated

semi-intensive care unit of the S. Anna – RAN Institute (Crotone, Italy) over a 9-year

period (April 1998–March 2006) were considered retrospectively. The VS was

clinically defined in all subjects compliant to the criteria suggested by the Multi-

Society Task Force and the guidelines of the London Consensus Conference (Multi-

Society Task Force, 1994).

For each patient, were entered in the dataset: age, sex, etiology of brain injury

(posttraumatic or non-posttraumatic), rating at the Glasgow Coma Scale (GCS) [10]

at admission, and twenty-two neurological signs of established relevance in coma and

VS [11] (Tables 1) assessed by the attending physician at two-week intervals

following procedures and criteria predefined as intrinsic to the UNI ENI ISO

9001:2000 quality standards. Each sign was present or absent (binary attribute). The

subjects’ condition at discharge was measured by the Glascow Outcome Scale (GOS):

GOS

=death; GOS

=vegetative state exceeding 1 year in duration; GOS

=recovery,

with severe disabilities; GOS

=recovery, with mild disabilities; and GOS

=full

recovery or recovery with minimal disabilities not interfering with the everyday life

[12]. The GOS is widely used in the evaluation of the VS outcome, but the subject’s

assignment to any GOS class is subjected to misclassification [13] which could affect

the training of classification models. Therefore, the first two classes and the latest two

classes of GOS were combined into the GOS

1-2

and GOS

4-5

classes respectively, with

a resulting sharper separation among classes. The prediction of outcome was

estimated at admission and after 50, 100 and 180 days after admission.

Table 1. Clinical signs assessed at two-week intervals and entered into the artificial neural

network processing as potential prognostic factors.

Decerebration

Decortication

Conjugated gaze deviation

Skew eye deviation

Blink reflex

Cilio-spinal Reflex

Tactile-oral Reflex

Optic-oral Reflex

Bulldog Reflex

Grasping reflex

Corneal Reflex

Corneal-mandibular reflex

Threat reflex

Myotactic-cervical reflex

Chewing reflex

Sucking Reflex

Oculo-cephalic reflex (with disappearance of the doll’s head phenomenon)

Absence of spontaneous motility

Eye tracking

Snout Rabbit sign

Half-moon pucker sign

Klippel sign

2.2 Pre-processing

Etiology of brain injury and the pathophysiology underlying VS are known to

influence the outcome. The dataset inclusive of all patients and two data sub-sets of

the posttraumatic (n=213) and non-posttraumatic patients were considered.

Continuous numerical attributes (such as age and GCS level) were normalized in the

interval [0;1] for each dataset; remaining attributes were binary and did not require

pre-processing.

3 Experimentation

3.1 Parameter Configuration

The classification models were structured as classical feed-forward ANN , with one or

two hidden layers and sigmoid function activation [14, 15]. The number of neurons

was varied among 1, 2, 4, 6, 10, 15, 20, 25, 30 and 40 for both the first and second

hidden layer. The Stuttgart Neural Network Simulator (SNSS) was used for all the

experimentations [16]

The training of the ANN was performed by using the standard Back Propagation

algorithm and the “Enhanced Back Propagation” algorithm. The latter introduces the

previous arc weight change as a parameter for computing the new arc weight change.

SNNS implements both algorithms with the Std_Backpropagation and

BackpropMomentum functions. In particular, the Std_Backpropagation function

requires the specification of the parameter η (learning rate) and d

max

(maximal

difference between expected and calculated output for each neuron). Besidea η, the

BackpropMomentum function needs the momentum μ measuring the influence of the

previous arc weight change on the current weight calculation. Table 2 shows the

parameters configuration used for the training algorithms.

Table 2. Training algorithms parameters configurations.

Std_Backpropagation BackpropMomentum

η d

max

η Μ

0.1 0.1 0.1 0.2

0.3 0.2 0.3 0.8

0.5 0.5

0.7 0.7

0.9 0.9

3.2 Experimentation Protocol

We used a Training–Validation–Test (TVT) procedure to select the best parameter

configuration regulating both the network structure and the training algorithm

operation. In particular, for each dataset the following steps were applied:

1. creation of training, validation and test set (see Table 3);

2. for each combination of network and training algorithm parameters:

a. execution of 200 training cycles;

b. evaluation of network accuracy on the validation set;

c. if the total number of training cycles is 20000, then stop; otherwise,

return to step a;

3. selection of the network with the best accuracy on the validation set;

4. evaluation of accuracy on the test set.

At the end of the TVT procedure, we obtained three trained ANN (one for each

dataset) with their respective accuracy on the test set.

Table 3. Subdivision of instance among training, validation and test sets.

Dataset Training Validation Test

NPT Dataset 80 20 20

PT Dataset 133 30 50

Entire Dataset 200 53 80

4 Results

The best networks test set accuracy was 70%, 72% and 70% for the entire patients’

group and the posttraumatic and non-posttraumatic subgroups, respectively. The best

parameter configurations are reported in Table 4.

Table 4. Configurations parameters of the best networks. BP: standard back propagation

algorithm; EBP: enhanced back propagation; na: not applicable.

Dataset

Entire

dataset

NPT

dataset

hidden layer 30 1 6

hidden layer 30 N.A. N.A.

Training algorithm BP EBP BP

η 0.7 0.1 0.7

N.A. N.A. 0.2

μ 0.2 0.8 N.A.

A better understanding of the classificatory performance can be obtained through

the analysis of the confusion matrices (see Tables 5 and 6) indicating misclassified

elements. We decided to assign instances with unclear evaluation to the

“misclassified” class (e.g. the same instance was assigned to two classes at the same

time with similar probability).

Table 5. Entire dataset confusion matrix.

predicted class

Misclassified

real

class

1_2 3 4_5

1_2 16 1 3 2

3 5 2 11 0

4_5 1 0 38 0

Table 6. Posttraumatic dataset confusion matrix.

predicted class

Misclassified

real

class

1_2 3 4_5

1_2 6 0 0 3

3 0 0 7 2

4_5 0 0 30 2

Table 7. Non-posttraumatic dataset confusion matrix.

predicted class

Misclassified

real

class

1_2 3 4_5

1_2 8 2 0 0

3 0 5 0 1

4_5 1 1 1 1

5 Comment

The method accuracy does not reflect a perfect classification, but is significantly far

from the random or educated guess and is in accordance with the results of previous

clinical studies [11]. It should be noted that class GOS

has a larger error estimate

both in the entire dataset and in the posttraumatic sub-set. The higher

misclassification depends on this class taking into account all patients with a severe

motor outcome (e.g. paresis of one or more limbs), impaired consciousness (e.g.

global amnesia) or both. GOS

can therefore be heterogeneous and ANN are unable to

identify a major labeling characteristic. Interestingly, test set patients with GOS

the non-posttraumatic dataset are well classified, while GOS

4-5

subjects of the same

dataset are poorly classified. The limited size of the non-posttraumatic sample does

not allow further investigation of such phenomenon.

References

1. Jennett B, Plum F. Persistent vegetative state after brain damage: a syndrome in search of a

name. Lancet 1972;1:734-6.

2. Dolce G, Sazbon L. The posttraumatic vegetative state. Stuttgart, Thiene, 2002.

3. Laureys S. The neural correlate of (un)awareness: lessons from the vegetative state. Trends

Cogn Sci 2005;9:556-9.

4. Jennett B. The vegetative state. Cambridge, UK, University Press, 2002.

5. Multi-Society Task Force on PVS. Statement on medical aspects of the persistent

vegetative state. N Eng J Med 1994; 330:. 1499-1508.

6. Zeman A. Consciousness. Brain 2001;124:1263-89 (review).

7. Braakman R, Jennett WB, Minderhoud JM. Prognosis of the post traumatic vegetative state.

Acta Neurochirurgica 1988;95): 49-52.

8. Schmutzard E, Kampfl A, Franz G, Pfausler B, Haring HP, Ulmer H, Felber S,

Golaszewski S, Aichner F. Prediction of recovery from post traumatic vegetative state with

cerebral magnetic-resonance imaging. Lancet 1998;351:1763-67.

9. Rovlias A, Kotsou S.Classification and regression tree for prediction of outcome after

severe head injury using simple clinical and laboratory variables. J Neurotrauma.

2004;21:886-93.

10. Teasdale G, Jennet B. Assessment of coma and impaired consciousness: a practical scale.

Lancet 1974; 2:81-84.

11. Dolce G., Quintieri M., Serra S., Lagani V., Pignolo L. Clinical signs and early prognosis: a

decisional tree, data mining study. Brain Injury. 2008, 22:7, 617 — 623.

12. Jennet B., Bond M. Assessment outcome after severe brain damage: a practical scale.

Lancet 1976; 1: 480-484.

13. Pignolo, L., Quintieri M., Sannita, W. G. 'The Glasgow outcome scale in vegetative state: A

possible source of bias, Brain Injury, 2009, 23:1,1 — 2.

14. Holte, R.C. (1993). Very simple classification rules perform well on most commonly used

datasets. Machine Learning, 11, 63–90.

15. Van Bemmel, J.H.,&Munsen, M.A. (1997). Handbook of medical informatics. Berlin:

Springer-Verlag.

16. Zell, A., Mache, N., Hubner, R., Schmalzl, M., Sommer, T., Korb, T. SNNS Stuttgart

Neural Network Simulator Users Manual, Version 2.0, report (1992),No. 3/92, IPVR,

Universität Stuttgart.