FUNCTIONAL STATUS AND THE EYE-TRACKING RESPONSE

A Data Mining Classification Study in the Vegetative

and Minimaly Conscious States

A. Candelieri

S. Anna Institute and RAN - Research in Advanced Neurorehabilitation, Crotone, Italy

Laboratory of Decision Engineering for Healthcare Delivery, Department of Electronics Informatics and Systems

University of Calabria, Cosenza, Italy

F. Riganello, D. Cortese

S. Anna Institute and RAN - Research in Advanced Neurorehabilitation, Crotone, Italy

W. G. Sannita

Department of Neuroscience, Ophthalmology and Genetics, University of Genova, Genova, Italy

Department of Psychiatry, State University of New York, Stony Brook, NY, USA

Keywords: Vegetative State, Minimally Conscious State, Eye-tracking, Data Mining.

Abstract: Eye-tracking is defined as the “pursuit eye movement or sustained fixation that occurs in direct response to

moving or salient stimuli”; it is a key descriptor of the evolution from the vegetative (VS) to the minimally

conscious (MCS) state and predicts better outcome. In this study, several physiological parameters (such as

heart beat, Galvanic Skin Response [GSR], Blood Volume Pulse [BVP], respiratory rate and amplitude)

were recorded while a medical examiner searched for eye-tracking by slowly moving a visual stimulus

horizontally and vertically in front of the subject. Seven patients in VS and 8 in MCS were studied. The

Heart Rate Variability (HRV) was analyzed to obtain time and frequency descriptors. Different

classification methods were adopted to search for a plausible relationship between the subject psycho-

physiological state and observable eye-tracking to stimuli. The performance of different classifiers was

computed as Balanced Classification Accuracy (BCA) and evaluated through suitable validation technique.

A Support Vector Machine (SVM) classifier provided the most reliable relationship: BCA mean was about

84% on fold cross validation and about 75% on an independent test set of 6 patients (3 VS and 3 MCS).

1 BACKGROUND & RATIONALE

Eye-tracking, the pursuit eye movement or sustained

fixation that occurs in direct response to moving or

salient stimuli (Vanhaudenhuyse, Schnakers, Brédart

and Laureys, 2008), it is usually observed in 20%

and 82% of subjects in the vegetative (VS) and

minimal conscious (MCS) states, respectively

(Giacino, Zasler, Katz, Kelly, Rosenberg, and Filley,

1997; Royal College of Physicians, 1996;

Schnakers, Vanhaudenhuyse, Giacino, Boly,

Majerus, Moonen and Laureys, 2009).). It is a key

descriptor of the evolution from VS to MCS.

We retrospectively observed eye-tracking in 73%

of 395 patients in a vegetative state, referred to the

S. Anna - RAN Institute from intensive care,

neurological or neurosurgery units in the years 1998-

2008. These 395 patients could be clustered by

etiology of brain damage in 3 different groups:

posttraumatic (n=248), vascular (n=119) or anoxic-

hypoxic (n=28). Eye-tracking was already observed

within 50 days from brain injury in about 50% of

posttraumatic and vascular subjects and in 20% of

anoxic-hypozxic patients. After 230 days, eye-

tracking had re-appeared in about 90% of

posttraumatic and vascular subjects and in 67% of

anoxic patients. Subjects with early recovered eye-

tracking had a better outcome at discharge or at the

138

Candelieri A., Riganello F., Cortese D. and G. Sannita W..

FUNCTIONAL STATUS AND THE EYE-TRACKING RESPONSE - A Data Mining Classiﬁcation Study in the Vegetative and Minimaly Conscious States.

DOI: 10.5220/0003128201380141

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2011), pages 138-141

ISBN: 978-989-8425-34-8

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

end of follow-up (Spearman nonlinear correlation

coefficient=-0.365, p-value<0.001). In this respect,

eye-tracking proved an efficient predictor of

outcome also in a study assessing at regular time

intervals the presence/absence of 21 neurological

signs. A data mining decision tree model identified

eye-tracking as the best predictor of favorable

outcome in the vegetative state (Dolce, Quintieri,

Serra, Lagani and Pignolo, 2008; Pignolo, Riganello,

Candelieri and Lagani, 2009).

We also searched for eye-tracking at different

times over the day (3 observations in the morning

and 3 in the afternoon) in subjects in VS (n=9) or

MCS (n=13). Eye-tracking was observed at any time

during the day in 62% of MCS subjects, and never in

67% of VS patients. About 33% of subjects in VS

presented eye-tracking at least once in the day, while

38% of subjects in MCS never showed it. These

percentages are consistent with the reported rate of

misdiagnosis between VS and MCS and suggest that

eye-tracking may depend on the subject’s psycho-

physiological condition to occur (submitted).

We decided to test the relationship between eye-

tracking and the physiological condition as

characterized by Heart Rate Variability (HRV)

analyses. HRV is an emerging objective measure of

the continuous interplay between sympathetic and

parasympathetic subsystems (Task Force of

European Society of Cardiology and North

American Society of Pacing and Electrophysiology

of Circulation, 1996) and provides information on

complex brain activation as well (Dolce, Riganello,

Quintieri, Candelieri and Conforti, 2008; Riganello,

Quintieri, Candelieri, Conforti and Dolce, 2008;

Appealhans and Luecken, 2006, Friedman, (2007)

Kreibig, 2010). In previous studies on controls, brain

injured conscious patients and subjects in a

vegetative state we obtained evidence f a correlation

between the response to external stimuli and HRV

(mainly expressed by the normalized low-frequency

[0.04-0.15 Hz] band power descriptor nuLF)

(Riganello, Pignolo, Lagani and Candelieri, 2009;

Riganello and Candelieri, 2010; Riganello,

Candelieri, Quintieri, Conforti and Dolce, 2010).

In this respect, our working hypothesis is that a

subject’s physiological status can be (partially)

described by the HRV parameters and that a

consistent response, in our case an eye-tracking, may

depend on its variations.

2 MATERIALS & METHODS

Eye-tracking was searched for in 9 and 13 patients in

VS and MCS, respectively. Three different visual

stimuli, namely a mirror, a green light and a bright

red ball were used. The test was repeated several

times for each subject in the absence of indications

of sleepiness, stress, pain or discomfort. During this

procedures, several physiological parameters were

recorded (Nexus-10 device, Mind Media BW,

Roermond-Herten, NL): Galvanic Skin Response

(GSR), respiratory rate and amplitude, heart rate,

coherence between the heart and respiratory rates,

blood volume pulse (BVP), and the heart rate

variability normalized band power and peak

frequency in the low frequency interval (nuLF and

peakLF). A dataset including 220 test conditions

(also including the stimulus used to elicit an eye-

tracking response and clinical condition [VS or

MCS]) was built for the data mining classification

task.

Data Mining classification approaches were to

identify the fuctional condition that best correlated

with the observation of eye-tracking. The established

updated data mining techniques provided by WEKA

(Waikato Environment for Knowledge Analysis)

open-source software were used in the classification

task (Witten and Eibe, 2005). Decision Trees, Rule-

based Learning algorithm (OneR, Ridor and JRip)

and Support Vector Machines (SVM) were used.

A Chi-Squared Feature Selection method was

used to rank the study variables based on their

correlation with the class value (presence and

absence of eye-tracking). Such an approach usually

improves the classifiers performance and should

provide medical experts with information on the

physiological parameters facilitating eye-tracking.

The classifiers’ reliability was evaluated by the

Balanced Classification Accuracy (BCA) computed

as the mean of correct classifications among classes.

In addition, all instances related to a single patient

were entered into a separate fold; the training was

performed on all remaining folds and the extracted

model was tested on the fold left apart. Such a cross

validation procedure (repeated for each patient and

comparable to the leave-one-out validation) avoids

over-fitting, dependency by the patient related

information, circularity in the analysis (double

dipping) (Kriegeskorte, Simmons, Bellgowan and

Baker, 2009), and estimates the reliability of the

extracted criteria.

3 RESULTS

The Chi-Squared feature selection separately

performed on each fold and the anaysis of the

FUNCTIONAL STATUS AND THE EYE-TRACKING RESPONSE - A Data Mining Classification Study in the

Vegetative and Minimaly Conscious States

139

resulting ranked list selected three parameters best

correlating with eye tracking: nuLF, peakLF and

clinical condition (Table 1).

Table 1: First, second and third ranked features on 15

folds (each fold consists in several observations of a single

patient).

Chi-Squared-based Features Ranking

First Second Third

Fold 1 nuLF peakLF condition

Fold 2 condition nuLF peakLF

Fold 3 nuLF peakLF condition

Fold 4 nuLF peakLF condition

Fold 5 stimulus condition Resp.Rate

Fold 6 nuLF peakLF condition

Fold 7 nuLF peakLF condition

Fold 8 nuLF peakLF condition

Fold 9 nuLF peakLF condition

Fold 10 stimulus condition Resp.Rate

Fold 11 nuLF peakLF condition

Fold 12 nuLF peakLF BVP

Fold 13 condition nuLF peakLF

Fold 14 nuLF peakLF condition

Fold 15 nuLF peakLF condition

nuLF ranked as the best correlating parameter,

peakLF ranked second and clinical condition third.

No consistent trend of correlation was observed for

other parameters.

In a first classification task, J48, Ridor, JRip,

OneR and SVM were used to define a reliable

classification model explaining the relationship

between the three variables and eye-tracking. The

best model was provided by OneR algorithm (with

bucket size=3) which only used nuLF value to

predict eye-tracking. This decision model provided

BCA=81.80% averaged on the 15 folds (standard

deviation=14.92%) and BCA=82.66% on overall

validation phase.

All the other classification algorithms presented

averaged BCA lower than 80%.

Consistent with the study purpose, etiology was

excluded and analysis focused on nuLF and peakLF.

The operation did not change the OneR’s

performance, but increased BCA for most of the

other approaches (between 0.02% and 1.34% for

J48, JRip and Ridor), in particular for SVM (with

radial basis function kernel and gamma=25) which

resulted the most reliable model of the entire

analysis, with overall BCA=85% and BCA=84.10%

averaged on folds (standard deviation=16.56%).

When only 2 parameters (nuLF and peakLF)

were used, the non-linear relationship learned

through SVM (figure 1) proved able to convey new

information of possible medical use.

The actual reliability of the learned criterion was

estimated by applying without any retraining the

SVM model to a dataset from 6 patients (3 in VS

Figure 1: Presence (green) or absence (red) of eye tracking

(green for present and red for absent) against decision

function predictions (white regions as high probability for

eye tracking, black regions as high probability for no eye

tracking).

and 3 in MCS, 167 and 216 eye-tracking

observations respectively). The assessment

performed by the model based on the physiological

parameters value was compared to that by a medical

examiner with no information on the subject’s

physiological parameters. The model proved

reliable, with an overall BCA of 74.92%, (84.54%

and 73.45% for VS and MCS, respectively). The

worse performance on MCS subjects depends on a

lower sensitivity, due to eye-tracking detected by the

medical examiner when the decision model

indicated a non-optimal status. This kind of error

suggests that patients in MCS may be able to

provide eye-tracking even if their physiological

condition is evaluated as border-line by the SVM-

based criterion. A plausible reason may concern a

recovery of more complex consciousness patters

requiring less strong constraints on physiological

condition, respect to VS (Bosco, Lancioni, Olivetti

Belardinelli, Singh, O’Reilly and Sigafoos, 2010;

Andrews, Murphy, Munday and Littlewood, 1996).

4 CONCLUSIONS

A relationship between the physiological status of

subjects in the VS or MCS and their eye-tracking

response to a visual stimulus would be crucial to a

better understanding of the evolution from the

former to the latter clinical condition. The

relationship among a number of physiological

parameters and eye-tracking presence/absence (as

assessed by a medical examiner) was studied by

several data mining classification approaches

provided by the WEKA open-source tool. Two

parameters obtained by HRV Analysis (nuLF and

peakLF) proved highly correlated to eye-tracking. A

SVM classifier provided a reliable criterion to

HEALTHINF 2011 - International Conference on Health Informatics

140

predict eye-tracking simply by evaluating these two

HRV spectral parameters. Reliability, computed as

Balanced Classification Accuracy (BCA), proved

remakably high for this SVM model.

The training set size (220 instances) warrants

caution and additional research on large datasets is

advisable. Although preliminary, our results are

quite interesting and encouraging because the

reliability obtained on an independent data set (383

instances).

The correlation between the physiological status

(as indicated by the HRV descriptors) and eye-

tracking nevertheless appears applicable to mark

with better precision the evolution from the

vegetative to the miniman conscious state and

reduce misdiagnosis.

Further research is planned to assess the criterion

diagnostic reliability and the suitability of extended

application to calibrate type and timing of (visual)

stimulation paradigms potentially supporting

recovery of consciousness in VS and MCS patients.

REFERENCES

Andrews, K., Murphy, L., Munday, R., Littlewood, C.

(1996). Misdiagnosis of the vegetative state:

retrospective study in a rehabilitation unit. British

Medical Journal, 313, 13–16.

Appelhans, B. M., Luecken, L. J. (2006). Heart rate

variability as an index of regulated emotional

responding. Review of General Psychology, 10, 229–

240.

Bosco, A., Lancioni, G. E., Olivetti Belardinelli, M.,

Singh, N. N., O’Reilly, M. F., Sigafoos, J. (2010).

Vegetative state: efforts to curb misdiagnosis.

Cognitive Processing, 11(1), 87–90.

Dolce, G., Quintieri, M., Serra, S., Lagani, V., Pignolo, L.

(2008). Clinical signs and early prognosis in

vegetative state: A decisional tree, data-mining study.

Brain Injury, 22(7), 617–623.

Dolce, G., Riganello, F., Quintieri, M., Candelieri, A.,

Conforti, D. (2008). Personal interaction in the

vegetative state. A data mining study. Journal of

Psychophysiology, 22(3), 150–156.

Friedman, B. H. (2007). An autonomic

flexibility-neurovisceral integration model of anxiety

and cardiac vagal tone. Biological Psychology, 74,

185–199.

Giacino, J. T., Zasler, N. D., Katz, D. I., Kelly, J. P.,

Rosenberg, J. H., Filley, C. M. (1997). Development

of practice guidelines for assessment and management

of the vegetative and minimally conscious states.

Journal of Head Trauma Rehabilitation, 12(4), 79–89.

Kreibig, S .D. (2010). Autonomic nervous system activity

in emotion: A review. Biological Psychology, 84,

394–421.

Kriegeskorte, N., Simmons, W. K., Bellgowan,

P. S. F.,

Baker, C. I. (2009). Circular analysis in systems

neuroscience: the dangers of double dipping. Nature

Neuroscience, 12, 535– 40.

Pignolo, L., Riganello, F., Candelieri, A., Lagani, V.,

(2009). Vegetative State: early prediction of clinical

outcome by artificial neural network. In Proceedings

of 5th International Workshop on Artificial Neural

Networks and Intelligent Information Processing –

ANNIIP 2009, 91–96.

Riganello, F., Quintieri, M., Candelieri, A., Conforti, D.,

Dolce, G. (2008). Heart rate response to music. An

artificial intelligence study on healthy and traumatic

brain injured subjects. Journal of Psychophysiology,

22(4), 166–174.

Riganello, F., Pignolo, L., Lagani, V., Candelieri, A.

(2009). Data mining approaches for the study of

emotional responses in healthy controls and traumatic

brain injurd patients: comparative analysis and

validation. In Proceedings of 5th International

Workshop on Artificial Neural Networks and

Intelligent Information Processing – ANNIIP 2009,

125–133.

Riganello, F., Candelieri, A. (2010). Data mining and the

functional relathionship between heart rate variability

and emotional processing. Comparative analyses,

validation and application. In Proceedings o Healthinf

2010, 3rd International Conference on Health

Informatics, 159–165.

Riganello, F., Candelieri, A., Quintieri, M., Conforti, D.,

Dolce, G. (2010). Heart rate variability: an index of

brain processing in vegetative state? An artificial

intelligence data mining study. Clinical

Neurophysiology, doi:10.1016/j.clinph.2010.05.010.

Royal College of Physicians. (1996). Guidance on

diagnosis and management: Report of a working party

of the Royal College of Physicians, London, Royal

College of Physicians.

Schnakers, C., Vanhaudenhuyse, A., Giacino J. T., Boly,

M., Majerus, S., Moonen, G., Laureys, S. (2009).

Diagnostic accuracy of the vegetative and minimally

conscious state: Clinical consensus versus

standardized neurobehavioral assessment.

BMC Neurology, 9, 35.

Task Force of European Society of Cardiology and the

North American Society of Pacing and Electro-

physiology of Circulation. (1996). Heart rate

variability: standard of measurement, physiological

interpretation, and clinical use. Circulation, 93, 1043–

1065.

Vanhaudenhuyse, A., Schnakers, C., Brédart, S., Laureys,

S. (2008). Assessment of visual pursuit in

postcomatose states: use a mirror. J Neurol Neurosurg

Psychiatry, 79(223), doi:10.1136/jnnp.2007.121624

Witten, H. W., Eibe, F. (2005). Data Mining – Practical

machine learning tools and techniques with Java

implementations. San Francisco, CA. Morgan

Kaufman.

FUNCTIONAL STATUS AND THE EYE-TRACKING RESPONSE - A Data Mining Classification Study in the

Vegetative and Minimaly Conscious States

141