See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/239932775

Diagnosis of VisuoSpatial Neglect Using Dynamic Sequence Features from a

Cancellation Task

Article  in  Pattern Analysis and Applications · August 2002

DOI: 10.1007/s100440200023 · Source: DBLP

CITATIONS READS
15 1,148

3 authors, including:

Richard Guest
University of Kent
123 PUBLICATIONS   561 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The SuperIdentity View project

Private IDentifiaction as a Service View project

All content following this page was uploaded by Richard Guest on 04 August 2014.

The user has requested enhancement of the downloaded file.

Pattern Analysis & Applications (2002)5:261–270
Ownership and Copyright
 2002 Springer-Verlag London Limited

Diagnosis of Visuo-Spatial Neglect Using Dynamic

Sequence Features from a Cancellation Task
R. M. Guest1, M. C. Fairhurst1 and J. M. Potter2
1
Department of Electronics, University of Kent, Canterbury, UK; 2Nunnery Fields Hospital, Canterbury, Kent, UK

Abstract: Visuo-spatial neglect is recognised as a major barrier to recovery following a stroke or head injury. A standard clinical assessment
technique to assess the condition is a pencil-and-paper based cancellation task. Traditional static analysis of this task involves counting the
number of targets correctly cancelled on the test sheet. Using a computer-based test capture system, this paper presents the novel application
of using a series of standard pattern recognition techniques to examine the diagnostic capability of a number of dynamic features relating
to the sequence in which the targets were cancelled. While none of the individual dynamic features is as sensitive to neglect as the
conventional static analysis, a series of standard multi-dimensional feature analysis techniques are shown to improve the classification
accuracy of the dynamic properties of task execution, and hence the sensitivity to the detection of neglect and the validity of this novel
application. Combining the outcome of the dynamic sequence-based features with the conventional static analysis further improves the
overall sensitivity of the two cancellation tasks included in this study. The algorithmic nature of the methodology for feature extraction
objectively and consistently assesses patients, thereby improving the repeatability of the task.

Keywords: Cancellation task; Clinical evaluation; Drawing dynamics; Multi-feature analysis; Stroke; Visuo-spatial neglect

1. INTRODUCTION containing a series of targets which must be located and

cancelled (marking/crossing out of each target) using a pen.
Visuo-spatial neglect is a condition that can occur following Conventional assessment of the task involves counting the
a stroke or head injury causing a patient to fail to react to number of targets marked. Typical right CVA neglect
stimuli positioned within the opposite side of the visual field performance results in failure to cancel the targets on the
from the location of the lesion [1]. This can lead to problems left-hand side of the overlay, so the number of targets left
in performing everyday tasks such as washing, dressing and uncancelled provides a quantitative assessment of the severity
eating [2]. The effects are more prevalent in patients with of neglect [5]. Practically, diagnosis of neglect would typically
a lesion in the right hemisphere of the brain (Right Cerebro- be made using such cancellation tasks alongside other neur-
Vascular Accident – RCVA), leading to a deficit in the left opsychological assessments and clinical evaluation/judgement
of the visual field. Visuo-spatial neglect is recognised as a of patients.
barrier to recovery following a stroke, and therefore accurate A cancellation task widely regarded as a standard for
assessment of the condition is critical to the selection of assessing neglect is the Albert’s Cancellation Task [6]. In
effective schemes of rehabilitation [3]. Neglect is not an ‘all- this assessment, a test subject must cancel 40 lines on a
or-nothing’ condition [4]; the severity of neglect is depen- single sheet of paper (Fig. 1). A variation on the basic
dent on the location and volume of the lesion, and therefore cancellation task introduces distractor targets on the overlay
a diagnostic test of neglect must be sensitive to a wide range which must not be cancelled. These tasks require extra
of patient performance. attentional processing, which is able to increase the sensitivity
A standard method for the detection and severity assess- to neglect characteristics [7]. Tasks such as the star cancel-
ment of neglect is the ‘pencil and paper’ based cancellation lation task [8] use letters and large stars as distractors
task. The test subject is presented with a sheet of paper amongst the cancellation targets (smaller stars). Increasing
the distractor to target ratio on a single overlay has also
Received: 16 April 2001 been found to increase the sensitivity to neglect [9].
Received in revised form: 28 May 2001 Several studies have investigated the cancellation perform-
Accepted: 31 May 2001
ance of neglect patients by dividing the cancellation overlay
262 R. M. Guest et al.

resolution static data (the outcome of the drawing process)

and dynamic data (time-based constructional aspects) can be
investigated with accuracy and consistency. This work is part
of a wider study which uses the computer-based system to
analyse the responses from a series of pencil-and-paper based
tasks, and extract the dynamics of drawing performance
within neglect, stroke control and age-matched populations
[14,15]. Any performance differences noted between the
test groups from previously unmeasurable novel features
constitute a potential increase in test sensitivity which, com-
bined with the enhanced accuracy and objectivity of the test
infrastructure, provides a more stable test platform for the
detection of neglect.
In the trial we are deliberately using standard multiple
classification techniques [16,17] such as majority voting,
discriminant analysis [18] and Behaviour Knowledge Space
[19], with the outcome of each of the features extracted
from the test response. These techniques have been widely
Fig. 1. Albert’s cancellation task [6]. used to improve the classification rates of pattern recognition
applications [20–22], including handwriting recognition
[23,24] and medical image processing assessment [25–27].
into quadrants rather than by traditional left/right visual In this way, we hope to demonstrate the effectiveness of the
field analysis. Using a standard Albert’s task with right CVA task application and novel feature measurement without the
neglect patients, both Halligan and Marshall [10] and Mark necessity of introducing new classification algorithms. Rah-
[11] showed that the greatest number of omissions occurred man and Fairhurst [23] showed that by combining four
in the bottom left-hand quadrant of the overlay. From these classifiers to analyse handwritten characters, the correct rec-
omission patterns, it is possible to hypothesise that the ognition rate was improved from 81.05% (average recog-
sequential aspects of test performance, such as the order in nition rate of individual classifiers) to 88.41%. Combining
which the targets were cancelled and the starting target these proven methodologies for classification with the estab-
within the sequence, are also affected by the neglect con- lished clinical basis of the cancellation task, the objectives
dition. The timing and executional aspects of cancellation of the experiments presented in the paper are twofold: first,
task completion have previously been investigated both by to assess initially whether any of the dynamic features are
direct visual observation and analysis of videoed evidence. able to attain the same classification ability as the existing
Search patterns and cancellation strategies have been analysed static standard; and secondly, and more importantly, to
by Chatterjee et al [12]. While expected target omissions on ascertain whether, using standard multi-classification pattern
the left-hand side of the overlay were reported, regular recognition techniques, the clinical sensitivity of the tasks in
patterns of cancellation with movement predominantly in improved. The emphasis, therefore, is on the application of
the vertical plane were made by the single neglect patient standard pattern recognition techniques to establish a novel
used as a case study for the trial. An investigation using a solution to aid a clinical diagnosis. Future work may then
larger group of both neglect and stroke control test subjects focus on designing novel classification strategies to improve
(stroke subjects without neglect) may produce sequencing or refine diagnostic capability.
results correlating with the severity of the neglect or other
features extracted from the sequence of cancellation.
The aim of the research reported here is to investigate
the diagnostic properties of the sequential aspects of task 2. METHOD
execution, and to ascertain whether the sensitivity of a
cancellation task can be increased through this sequential Data for the evaluation trial was collected from two cancel-
performance analysis. In addition, standard pattern recog- lation tasks. The first was a standard Albert’s Cancellation
nition techniques will be used to evaluate the results of the Test containing 40 target lines of length 25 mm arranged at
novel dynamic performance features extracted from the data, random orientations and in a pseudo-random linear grid
to assess if task sensitivity can be increased further. formation measuring 222 ⫻ 181 mm (Fig. 1). The test
Algorithmically extracting features from responses enables subject was required to cancel all of the lines on the overlay.
an objective assessment of the task execution (not reliant The second task contained a series of 12 targets (‘O’
on a marker’s experience and judgement), which in turn characters) and 12 distractors (‘X’ characters), both of
allows a standardised exploration of diagnostic sensitivity dimensions 7 ⫻ 7 mm arranged in a pseudo-random linear
across test subjects. Using a novel computer-based testing formation of dimension 198 ⫻ 107 mm (Fig. 2). The test
system [13], whereby test responses are captured using a subject was required to cancel all the ‘O’ characters on the
graphics digitisation tablet interfaced to a standard PC which overlay, ignoring the distractors. The overlay was designed
can time-stamp the positional data from the pen, both high so that two ‘O’ and two ‘X’ characters appeared in each
Diagnosis of visuo-spatial neglect using dynamic sequence features 263

used throughout the experimentation to assess the accuracy

of features and methods.
Responses were assessed on the basis of a single static
and seven sequence-based dynamic features extracted from
the captured task execution data:
Neglect detection thresholds (a score obtained using a single
feature below which neglect is defined) were established
individually for features returning quantifiable data (for
example, number of cancellations made). The threshold was
set as the median response obtained by members of the age
matched and stroke control groups within the training set
of data. A median value was found to give the best separation
between groups given the distribution of feature results. It
also prevents outlying feature values from distorting a thres-
Fig. 2. OX cancellation task.
hold.
For features returning assigned identifiers (for example,
starting quadrant), the modal response was used as a normal
vertical column, thereby forcing a normal cancellation strat-
reference value. In analysis, any response that did not con-
egy into predominately horizontal or vertical movements.
form to the modal control response was classified as a
The test overlays were positioned separately on a graphics
neglect (fail).
digitisation tablet (a Wacom WD1212 with spatial accuracy
These thresholds and reference values are shown in Tables
of 6.25 lines/mm) connected to a standard portable com-
2 and 5. The ‘pass/fail’ system is used to provide an instant
puter. The test subject was seated directly in front of the
indication of neglect behaviour for use in a clinical environ-
tablet. As a test subject moved the pen and marked the
ment. This system is analogous to the thresholded system
overlay using a cordless and batteryless pen (of comparable
employed within the Rivermead BIT [8]. While analysis of
dimensions to a normal biro pen), the positional data was
continuous data will provide a performance metric, the
sampled at a rate of 100 Hz by the computer. The data was
thresholds/reference values enable an instant performance
stored sequentially and time-stamped as received. Perform-
judgement, which can be combined with other clinical evalu-
ance related features can then be extracted from this stored
ations to reach a diagnosis. These thresholds/reference values
data stream.
were also used in research to identify the classification rate
Data from three test groups were analysed in this trial:
of correctly identified subjects from the test sets for each of
쐌 28 right CVA stroke subjects. 16 male, 12 female, mean the features included in the analysis. Comparing the per-
age = 74.8, age SD = 7.43, age range 90–62, diagnosed formance from the existing static feature (detailing the num-
as exhibiting neglect using a recognised test battery ber of cancellations made) against the novel dynamic fea-
(Rivermead Behavioural Inattention Test – BIT [8]) and tures, an instant benchmark is formed establishing whether
clinical observations. Mean BIT score for neglect group = superior performance is observed.
89.21, SD = 25.34, range 129–42.
쐌 51 right CVA stroke subjects without neglect. 28 male, 2.1 Static Features
23 female, mean age = 74.29, age SD = 8.17, age range
92–57. Mean BIT score for stroke control group = 142.67, A conventional assessment of counting the number of correct
SD = 4.96, range 146–130. cancellations made during the task provided the static feature.
The detection rate of neglect from this standard feature was
쐌 13 age matched control subjects with no history of cereb-
used as a benchmark against which the sensitivity of the
rovascular or cardiovascular disease. 5 male, 8 female,
previously unobtainable dynamic features could be measured.
mean age = 72.77, age SD = 4.07, age range 79–63. A
In assessing the Albert’s task, the centre vertical line of four
BIT test was not performed by the age matched group as
targets were not counted, as these were used to demonstrate
it was assumed that no neglect was present.
the cancellation process (as defined in the Albert’s task
For the purposes of the trial, the age matched and stroke marking instructions [6]). This resulted in a maximum score
control groups were combined into a single group of 64 of 36. The maximum score obtainable from the OX overlay
test subjects. In this way, differences between a neglect and was 12.
non-neglect population can be assessed.
For the analysis of features extracted from test responses, 2.2. Dynamic Features
the test groups were divided equally to form disjoint training
and testing data sets, with 32 control and 14 neglect subjects Seven dynamic features assessing both the sequence in which
in each. A proportionate number of subjects were taken the targets were visited and the pen movement trajectory
from the age matched and stroke control groups to form followed during execution of cancellation task, are defined
the single control group. Having separate training and testing below. Features directly assessing the order in which the
data facilitates the use of supervised learning classification targets were cancelled have been defined in detail by the
algorithms. To attain consistency, the testing set of data was authors [28], so only a brief overview is included here.
264 R. M. Guest et al.

Table 1. Two-dimensional behaviour knowledge space

Expert 1

Pass Fail
Fig. 3. (i) Raster, and (ii) snake cancellation strategies.
Expert 2
Pass BKS BKS
(Pass,Pass) (Fail,Pass) duplications and, in the case of the OX task, all ‘X’ character
Fail BKS BKS (distractor) identification references.
(Pass,Fail) (Fail,Fail) A direct correctness-of-match between the each model
and observed sequence is calculated which assesses the match
in each of the sequence positions. The number of correct
matches is recorded. The observed sequence is then barrel
Assessment is made about how the observed sequence drawn shifted one position to the right (the location identifier in
by the test subject differs for a series of standard predefined the extreme right-hand position is moved to the start of the
model sequences for each overlay. Cancellation targets are sequence), and again, a correctness of match score is calcu-
assigned notional numerical identifiers which are used to lated. This procedure is repeated until the observed sequence
define and represent sequences. Sixteen model sequences ‘returns’ to its original position. Figure 4 shows this process.
have been defined for each cancellation task encompassing: As can be seen, the number of matches varies as the pattern
is shifted.
(i) A start position in each corner of the overlay. From this process, four sequence features are obtained:
(ii) A cancellation sequence predominantly in both the
쐌 The highest number of matches, measuring the conformity
horizontal (left to right) and vertical (top to
bottom) planes. of the completion sequence to one of the model sequences.
(iii) A cancellation sequence in a raster (movement to end 쐌 The number of shifts required to obtain the highest match.
This result also represents the position within the model
of row or column and then ‘jump back’ to start of
sequence where the observed sequence started to conform
next row or column) or snake (movement to end of
row or column, and then continue cancellation in to a predefined sequence.
opposite direction) movement. These movements 쐌 The movement plane of the pre-defined sequence produc-
(made with the pen off the paper) between targets ing the highest match (either horizontal or vertical).
are shown in Fig. 3. 쐌 The method of traversal of the pre-defined sequence pro-
ducing the highest match (either snake or raster).
The observed sequence is classified by finding the best match
against the model sequences. Prior to classification, the Three other novel sequence-based features are extracted from
observed sequence is pre-processed to remove sequence the observed sequence:
쐌 Starting Quadrant – indicates the quadrant of the overlay
within which the initial target of the observed cancellation
Table 2. Albert’s cancellation task – Neglect Detection sequence is located. The sequence shown in Fig. 5, for
Thresholds/Reference Values and Correct Detection example, starts in the top right-hand quadrant. The
Percentages expected execution pattern from a normal Western langu-
age response would be to start the cancellation sequence
Feature Neglect Detection Correct at the left of the overlay [29]. With RCVA, neglect subjects
Threshold/ Detection typically failing to cancel the left-hand side targets on the
Reference (%) overlay, the sequence starting quadrant will be affected
by a neglect response.
Static Feature 쐌 Duplications – indicates the number of duplicate cancel-
Number of Cancellations 36 71.7
Dynamic Feature
Method of Traversal Raster 60.9
Movement Plane Vertical 58.7
Number of Shifts 1 56.5
Number of Matches 10 63.0
Intersection 1 58.7
Starting Quadrant Top Left 65.2
Duplication 0 69.6 Fig. 4. Barrel shift correctness of match operation. The figure in the
parenthesis is the correctness-of-match score.
Diagnosis of visuo-spatial neglect using dynamic sequence features 265

identifier data) or below the performance (in the case of

quantifiable data) of the normal population.
By simply analysing the percentage of correct classifications
from each dynamic feature against the classification rate of
the static number of cancellations feature, it is possible to
ascertain if an increase in neglect detection sensitivity is
obtained. This ‘pass/fail’ method enables a rapid and simple
comparison of performance.

2.4. Multiple Feature Analysis

The binary ‘pass/fail’ outcome from each feature can be

combined using a series of decision rules facilitating a multi-
dimensional analysis. This may increase task sensitivity by
establishing neglect performance interaction between two or
Fig. 5. OX cancellation task showing a double path crossing and more of the sequence-based dynamic features. The perform-
sequence duplication (numbers refer to the movement order ance of three multi-dimensional analysis techniques were
between targets). considered: threshold voting, feature weighted discriminant
analysis [18] and Behaviour Knowledge Space [19]. Two
separate experiments for each of the cancellation tasks were
lations that were made within a sequence (i.e. the number undertaken using these methods:
of multiple cancellations of single targets). Figure 5 shows
a cancellation sequence with a single target duplication. 1. Assessment of multi-dimensional analysis using the seven
While duplications may be assessed on a static basis dynamic features.
by examination of the completed cancellation overlay, 2. Assessment of multi-dimensional analysis using both the
examination of the sequence data from the computer static and dynamic features to determine whether this
capture system will reveal a pen placement on targets that can further increase task sensitivity.
are obscured or drawn with insufficient pressure to enable
By using a simple pass/fail mark as an input vector to our
a visual analysis. Duplications within a sequence indicate
‘tried and tested’ classification configurations, we present a
a performance deficit caused by the subject being unable
simplified multi-feature analysis. The purpose of this simpli-
to visualise whether a particular target has been cancelled.
fication is to establish whether the general principles of
This process is usually made when making a second pass
multi-feature analysis will result in enhanced classification
over the overlay to establish that all targets have been
ability compared to the individual features. The aim is to
marked. Expected normal performance would be for a
demonstrate that our novel features are able to increase the
single cancellation to be made on each target with no
accuracy of neglect detection over the existing methods
duplications.
of neglect assessment (statically counting the number of
쐌 Path Crossings – indicates an irregularity in cancellation cancellations made). Future work will concentrate on the
sequence by analysing the number of times the pen tra- analysis of the normalised (or otherwise) ‘raw’ feature values
jectory crosses [28]. In Fig. 5, two crossing incidents are within a multi-feature assessment, and their application in
detected. The ‘paths’ have been superimposed on the the context of novel classifier combinations.
image to represent the pen movement trajectory between
cancellation targets. Consecutive pairs of cancellation 2.4.1. Classifier 1: Threshold Voting. A response score,
points within the sequence from virtual lines indicating between 0 and 7, was obtained by summing the individual
the sequence path. Each virtual line is compared against binary outcomes from the dynamic features. A response score
others in the sequence to test for crossings using a stan- fail threshold can be established, whereby all subjects who
dard line intersection routine [30]. The case where one score below this threshold are categorised as neglect subjects,
line starts and another ends at the same point (i.e. the such that:
next line in sequence) is ignored.

冉冘 冊
冦
N
2.3. Individual Feature Analysis Neglect if r(i) ⬎ ␣
classification = i=1 (1)
A binary ‘pass/fail’ output was produced by comparing each
feature against the neglect detection threshold and reference Control otherwise
values. For the quantifiable data items a binary 0 (or ‘fail’)
was awarded if the result was below the median threshold. where N = number of features, r(i) = ‘pass/fail’ binary result
For assigned identifier data items, a fail was awarded if the from the ith feature, and ␣ = response score fail threshold.
observed result was not equal to the modal result (reference In implementing this system, a test subject has to ‘pass’
value) obtained from the control group. A ‘fail’ indicates a with respect to a particular number of features for an overall
response which is not-identical to (in the case of assigned response to be classified as normal. By varying the fail
266 R. M. Guest et al.

Table 3. Albert’s Cancellation Task – Decision Rule Classi- the discriminant score. In the trial this is varied initially
fication Percentages (dynamic) from 2–7 features (the results from the seven dynamic
features), and in the second experiment 2–8 features
Response score Correct classification (%) when including the static result in the analysis. It is
fail threshold possible to obtain optimum coefficients for each feature
which, when summed, maximise the degree of dis-
7 39.1 tinguishability between neglect and control subjects
6 52.2 [31,32].
5 63.0 Stage 3: Bayes’ Classification – from each discriminant
4 71.7 score, it is possible to use a standard Bayes’ classification
3 73.9 rule [33] to classify a score into one of the two subject
groups.
2 76.1
1 69.6
2.4.3. Classifier 3: Behaviour Knowledge Space. The third
classification methodology used the Behaviour Knowledge
Space (BKS) method [19] to perform a multi-dimensional
analysis using the binary ‘pass/fail’ outcome from each of
threshold, the optimum response score correctly classifying the individual features. Traditionally, a BKS classifier takes
subjects can be established. Analysis of correct classification the outcome from a series of classifiers (experts), all assessing
rates compared against a particular fail threshold was perfor- a common feature set to form an overall judgement on a
med using the test set of data for each task. given system. Here, analysing the results from the cancel-
lation tasks, each expert is an individual feature. A BKS
2.4.2. Classifier 2: Feature Weighted Discriminant classifier uses an n-dimensional space defined by the number
Analysis. Using a standard discriminant analysis [18], a of experts providing a unique method of classification on
series of coefficients were calculated, providing a weight for an individual or set of features. In this explanation, we shall
each feature score to produce the best classification separ- consider a two-dimensional BKS (two experts), where each
ation between the control and neglect groups. There are feature classifies an input into one of two classes (pass/fail).
three principal stages involved in this process: The four ‘units’ of the BKS represent the possible input
Stage 1: Feature Correlation – assessment of the correlation combinations of the experts. This can be visualised in Table
(relationship) between pairs of features to form a 1. The active unit is selected based on the decisions of each
correlation matrix. A pooled within-group correlation of the experts.
matrix is calculated separately for each group (neglect and For the example two-dimensional BKS, within each BKS
controls), and then averaged to form an overall matrix. unit is stored three items based on a set of training data
supplied to the BKS:
Stage 2: Discriminant Score – this stage algorithmically
describes the relationships identified in the correlation 쐌 N(m) – the total number of incoming samples belonging
matrix, by finding coefficients for the equation: to class m (either neglect or control). As there are two
categories in the example, m = 2.
D = B0 + B1X1 + B2X2 + % + BpXp (2) 쐌 Te – total number of incoming samples within the BKS
unit:
where D is the discriminant score B0. .P are the individual

冘
feature coefficients, and X1. .P are the individual results m
from each of the features extracted from a single cancel- Te = N(m) (3)
lation response. P is the number of features included in 1

Table 4. Optimum feature combinations for BKS classification of Albert’s data

Classification Number of Features Number of

Rate (%) Features Unknown
Classifications

73.9 4 Intersection, Number of Sequence Shifts, Movement Plane and Sequence 0

Starting Quadrant
74.4 3 Number of Sequence Matches, Number of Sequence Duplications and 3
Starting Quadrant
Diagnosis of visuo-spatial neglect using dynamic sequence features 267

Table 5. OX cancellation task – Neglect Detection

Thresholds/Reference Values and Correct Detection Percentages

Feature Neglect Detection Correct

Threshold/ Detection
Reference (%)

Static Feature
Number of Cancellations 12 76.1
Dynamic Feature
Method of Traversal Raster 58.7
Movement Plane Vertical 63.0
Number of Shifts 0 73.9
Number of Matches 8 65.2
Intersection 0 60.9
Starting Quadrant Top Left 67.4
Fig. 6. Average BKS classification error for known responses from
Duplication 0 71.7 the Albert’s cancellation test.

number of experts, and hence the dimensionality of the

쐌 Re – the best representative class of the BKS unit.
BKS, between 2 and 7 (2–8 when including the static feature)
Separate values of N(m), Te and Re are calculated for each to find the optimum classification ability.
BKS unit.
Once a BKS has been trained, it can be used to determine
the best representative class outcome using a given series of 3. RESULTS
expert results. For example, when presented with a test
response, the outcome of expert 1 was a pass, whereas expert The results are first presented for the Albert’s Task, and
2 failed the response. These outcomes are then presented to then for the OX Cancellation task using the dynamic features
a trained BKS, where the unit of interest would therefore in a multi-dimensional analysis. The results from combining
be BKS(Pass, Fail). This select unit’s Re value provides the the static and dynamic features are then presented at the
largest represented class, in terms of number of represented end of this section.
samples as a proportion of overall class size, given the
specific combination of expert outcomes, and hence the 3.1. Albert’s Individual Feature Analysis
most likely output of the classification process.
A separate BKS was created for both of the tasks using The correct classification percentages for each of the features
the training set of data. The optimum number of experts (i.e. test subjects classified in their correct clinical groupings)
(features) and, secondly, which features should be selected from the Albert’s task testing data set are show in Table 2.
to form this optimum configuration, was explored. In the As can be seen, none of the dynamic sequence-based features
work reported in this paper, we experimentally vary the is able to detect neglect as accurately the static number of
cancellations feature (71.1%). Also shown in Table 2 are
the control group responses that form neglect detection
Table 6. OX cancellation task – Decision Rule Classification
thresholds/reference values for each feature within the
Percentages (Dynamic)
Albert’s task. A median value is used for the quantifiable
data (number of cancellations, shifts, matches, intersections and
Response score Correct classification (%)
duplications). A modal reference value is used for the assigned
fail threshold
identifier data (method of traversal, movement plane and
starting quadrants).
7 45.7 3.2. Albert’s Dynamic Feature Threshold Voting
6 56.5
5 63.0 Table 3 shows the classification accuracy from the Albert’s
4 69.6 task applying different response score fail thresholds using the
3 78.3 binary outcomes from the dynamic features. A response score
fail threshold requiring two or more feature passes to classify
2 78.3
the response as ‘normal’ produces the best overall classi-
1 69.6 fication. The detection rate at this threshold (76.1%) pro-
duces an improvement on the most accurate detection rate
268 R. M. Guest et al.

of the static number of cancellation feature. The number of

sequence shifts feature provides the best detection rate. The
neglect detection thresholds/reference values derived from
the training set are also displayed in Table 5.

3.6. OX Dynamic Feature Threshold Voting

Applying different response score fail thresholds to the sum

of the dynamic feature pass/fail values, it can be observed
that thresholds requiring two or more feature passes for the
response to be classified as normal provides the best classi-
fication rate, producing an improvement over all of the
singular dynamic feature classification rates (Table 6) and
the static number of cancellations feature.

3.7. OX Dynamic Feature Discriminant Analysis

Fig. 7. Average BKS classification error for known responses from Applying a discriminant weighted analysis using the dynamic
the OX cancellation test. features from the OX task produced an identical classification
performance as the single static feature (76.1%). This, how-
ever, is an improvement on the classification ability of any
of a singular dynamic feature (69.6%) and, indeed, the static of the individual dynamic features.
number of cancellations feature (71.7%).
3.8. OX Dynamic Feature BKS Analysis
3.3. Albert’s Dynamic Feature Discriminant Analysis. A
discriminant weighted analysis using the dynamic features The best classification rate of OX features using the BKS
produced a correct classification percentage of 76.1%. This was 75.7%. Six experts were used to achieve this (Number
method, having identical performance to the voting threshold of Matches, Intersections, Number of Sequence Shifts, Method
classifier, is also an improvement on the conventional static of Traversal, Movement Plane and Sequence Starting
feature for this task. Quadrant). Nine of the 46 classification were unknown. The
best classification rate when none of the responses was
3.4. Albert’s Dynamic Feature BKS Analysis unknown was 73.9%, using three features (Intersections,
Number of Sequence Shifts and Starting Quadrant). Using the
The number of experts used to train and test the BKS was OX data, a BKS classifier is unable to obtain the classification
varied (between 2 and 7 – representing the seven dynamic rate set by the benchmark number of cancellations feature
features), as was the selection of individual experts (features) on this task.
that were presented to a BKS. In this way, it was possible
to assess the optimum number and combination of experts 3.9. Combined Static and Dynamic Analysis
to provide maximum accuracy of classification. It is possible
that, when training the BKS classifier, not every combination The results reported here have shown that improved detec-
of expert ‘pass/fail’ classifications has an entry within the tion of neglect can be obtained using a multi-dimensional
knowledge space. For example, the situation may never arise analysis of the dynamic features extracted from the test
within the training data that all experts ‘pass’ a response. If response. Table 7 shows the results of including the static
this situation arises in the testing data, then an unknown number of cancellation made feature ‘pass/fail’ mark in with
pattern response is returned (i.e. no conclusion can be the multi-dimensional analysis to assess whether the detec-
reached about the classification probability). The number tion rate can be improved further.
of unknown patterns returned by a particular BKS was As can be seen, all classifiers improve on their score
also established. except Threshold Voting, which retains the same classi-
The best individual classification rate achieved from the fication ability. To further explore the results contained in
Albert’s data is shown in Table 4. These results show that, Table 7, Figs 6 and 7 show the error curves from applying
as with the other multi-dimensional classification methods, the BKS to the Albert’s and OX data, respectively. These
the BKS accuracy is an improvement on the static number curves show the mean error for each number of features
of cancellations feature recognition rate. presented to the classifier for known classification patterns.

3.5. OX Individual Feature Analysis

4. CONCLUSION
The classification ability of features from the OX task are
displayed in Table 5. As with the Albert’s task, none of the Using the conventional static assessment criterion as a bench-
dynamic features is individually able to achieve the accuracy mark against which the classification performance of the
Diagnosis of visuo-spatial neglect using dynamic sequence features 269

Table 7. Classification performance combining static and dynamic features

Task Classification Method Optimum Features Improvement

Classification Rate (%)

Albert’s Threshold Vote 76.1 All ✗

Discriminant 80.4 All ✓
BKS (Known) 78.3 a
✓
BKS (Unknown) 84.4 b
✓
OX Threshold Vote 78.3 All ✗
Discriminant 80.4 All ✓
BKS (Known) 76.1 c
✓
BKS (Unknown) 85.0 d
✓
a
Number of Cancellations, Intersections, Start Quadrant (3 Features)
b
Number of Cancellations, Sequence Matches, Intersections, Sequence Shifts, Movement Plane, Start Quadrant (6 Features)
c
Number of Cancellations, Intersections, Movement Plane, Duplications (4 Features)
d
Number of Cancellations, Sequence Matches, Sequence Shifts, Duplications (4 Features)

devised novel dynamic features can be assessed, none of the Acknowledgement

dynamic features considered individually was as sensitive to
the detection of neglect. It has been shown, however, that The authors acknowledge the support of the South Thames
by adopting a multi-dimensional approach to the assessment NHS R&D Project Fund.
of features, the classification rate can be improved. On both
tasks the use of threshold voting provides the best classi-
fication rate of the three methodologies used. Including the
static feature within the multi-dimensional analysis further
References
increases the accuracy of the tasks. In general, the results
1. Halligan PW, Marshall JC, Wade DT. Visuospatial neglect:
across all experimentation within this study have shown that Underlying factors and test sensitivity. The Lancet 1989; ii:
the OX task is more sensitive to the detection of neglect, 908–910
confirming the positive effect of distractor characters on 2. Edmans JA, Lincoln N. The relationship between perceptual
test sensitivity. deficits after stroke and independence in activities of daily living.
Importantly, using this novel application of pattern recog- British Journal of Occupational Therapy 1990; 53: 139–142
nition techniques, it has been shown that dynamic construc- 3. Robertson IH. The rehabilitation of attentional and hemi-atten-
tional aspects of task performance are able to classify and tional disorders. In: Riddoch MJ, Humphreys GW (eds) Cogni-
tive Neuropsychology and Cognitive Rehabilitation. Lawrence
detect neglect from subject responses. The use of these novel Erlbaum, 1994; 173–186
features leads to a further understanding of the condition 4. Behrmann M, Moscovitch M, Black S, Mozer M. Perceptual
of neglect, providing information concerning task execution and conceptual mechanisms in neglect dyslexia. Brain 1990; 113:
to clinicians and neuropsychologists which was previously 1163–1183
unobtainable. The use of these cancellation tests would most 5. Heilman KM, Watson RT, Valenstein E. Localization of lesions
likely be to provide a diagnostic aid alongside other clinical in neglect and related disorders. In: Kertesz A (ed), Localization
and Neuroimaging in Neuropsychology. Academic Press, 1994;
and functional assessments of neglect. Improving the sensi- 495–523
tivity of the task provides a more accurate assessment tool 6. Albert ML. A simple test of visual neglect. Neurology 1973; 23:
to support clinical investigation. While the classification rates 658–664
obtained may not be suitable in isolation for clinical diag- 7. Kaplan RF, Verfaellie M, Meadows ME, Caplan LR, Pessin MS,
nosis, comparison against the existing standard (number of DeWitt LD. Changing attentional demands in left hemispatial
cancellations made) indicates the enhancement in sensitivity neglect. Archives of Neurology 1991; 48: 1263–1266
provided by this method. Improving the detection rate for 8. Wilson B, Cockburn J, Halligan P. Development of a behavioral
test of visuospatial neglect. Archives of Physical Medicine and
neglect and the extent to which assessments can be linked Rehabilitation 1987; 68: 98–102
quantitatively to particular aspects of performance enables 9. Geldmacher DS. Effects of stimulus number and target-to-
an increase in accuracy of rehabilitation programme selec- distractor ratio on the performance of random array letter
tion, and allows the standardisation of patient progress cancellation tasks. Brain and Cognition 1996; 32: 405–415
monitoring. 10. Halligan PW, Marshall JC. Is neglect (only) lateral? A quadrant
270 R. M. Guest et al.

analysis of line cancellation. Journal of Clinical and Experimental 2000. He was employed from 1995 to 2000 as a Research Assistant at the
Neuropsychology 1989; 11: 793–798 University of Kent in the field of image analysis and processing, specifically the
analysis of drawing dynamics for the clinical evaluation of patients. He has
11. Mark VW, Heilman KM. Diagonal spatial neglect. Journal of recently been appointed to a Lectureship within the Department of Electronics at
Neurology, Neurosurgery and Psychiatry 1998; 65: 348–352 Kent. Dr Guest’s research interests include medical diagnosis using computer
12. Chatterjee A, Mennemeier M, Heilman KM. Search patterns and analysis of hand-drawn and other biometric data, and multi-modal biometric
neglect: A case study. Neuropsychologia 1992; 30: 657–672 analysis for security.
13. Fairhurst MC, Guest RM. Analysing visuo-spatial neglect through
figure-copying tasks. Proceedings 3rd IEE European Workshop
on Handwriting Analysis and Recognition, IEE, 1998; 2.1–2.5
14. Potter J, Deighton A, Patel M, Fairhurst M, Guest R, Donnelly Michael Fairhurst has been on the academic staff of the Department of Electronics
N. Computer recording of standard test of visual neglect in at the University of Kent since 1972. He has wide research interests in image
stroke patients. Clinical Rehabilitation 2000; 14: 441–446 analysis and computer vision, with a particular focus on computational architec-
tures for image analysis, and the implementation of high performance classification
15. Guest RM, Fairhurst MC. Analysing constructional aspects of algorithms. Application areas of principal concern include handwritten text reading
figure completion for the diagnosis of visuospatial neglect. Pro- and document processing, security and biometrics, and medical image analysis.
ceedings of ICPR2000, IEEE Press 2000; 4: 316–319 Professor Fairhurst is a past chairman of the IEE Professional Group E4 on Image
16. Xu L, Krzyzak A, Suen CY. Methods for combining multiple Processing and Vision, and has been chairman of several of the IEE International
Series of Conferences on Image Processing and Applications. He has been a
classifiers and their application to handwriting recognition. IEEE member of many Conference Organising/Programme Committees, and is an
Trans System, Man and Cybernetics 1992; 22: 418–435 Editorial Board member for several international journals in the image analysis
17. Kittler J, Hatef M, Duin RPW, Matas J. On combining classifiers. field. He has published more than 250 papers in the technical literature, and has
IEEE Trans PAMI 1998; 20: 226–239 authored an undergraduate textbook on computer vision.
18. Lachenbruch PA. Discriminant Analysis. Hafner Press, New
York, 1975
19. Huang YS, Suen CY. A method of combining multiple experts
for the recognition of unconstrained handwritten numerals. IEEE Jonathan Potter DM, FRCP is a Consultant Geriatrician at the Kent & Canterbury
Trans PAMI 1995; 17: 90–95 Hospital, Canterbury. His clinical responsibilities include the running of a 16-
bedded Stroke Unit in Canterbury. The unit has a long running commitment to
20. Kittler J. Combining classifiers: A theoretical framework. Pattern research, and has undertaken work evaluating of methods of rehabilitation, and
Analysis and Applications 1998; 1(1): 18–27 the comparison of outcomes between units. Dr Potter’s academic commitments
21. Lam L, Suen CY. Application of majority voting to pattern include: Honorary Senior lecturer, the Centre for Health Services Studies (CHSS)
recognition: An analysis of its behavior and performance. IEEE at the University of Kent at Canterbury. Work with the CHSS focuses on
Trans System, Man and Cybernetics 1997; 27(5): 553–568 evaluation of standardised assessment for older people; Honorary Senior Lecturer,
the Kent Institute of Health and Medicine (KIHMS) at the University of Kent at
22. Martinez AM, Kak AC. PCA versus LDA. IEEE Trans PAMI Canterbury, developing research and training within East Kent; and Associate
2001; 23(2): 228–233 Director of the Royal College of Physicians of London Clinical Effectiveness and
23. Rahman AFR, Fairhurst MC. Serial combination of multiple Evaluation Unit (CEEU), responsible for the Health Care of Older People pro-
experts: A unified evaluation. Pattern Analysis and Applications, gramme. The unit works to develop guidelines, audit tools and indicators of high
1999; 2(4): 292–311 quality care.
24. Impedovo S, Salzo A. A new evaluation method for expert
combination in multi-expert system designing. Lecture Notes in
Computer Science 2000; 1857: 230–239
25. Steuer M, Caleb P, Drummond GB, Black AMS. Visualisation Originality and Contribution
and categorisation of respiratory mechanism using self organising
maps. IEE Proc. Science, Measurement and Technology 2000; This paper reports work carried out to further the development of a clinical tool
147(6): 339–344 to support the diagnosis and assessment of visuo-spatial neglect in stroke patients
through the automatic processing of responses from a standard neuropsychological
26. Blanzieri E, Eccher C, Forti S, Sboner A. Exploiting classifier cancellation task. Although considerable work has been reported in the
combination for early melanoma diagnosis support. Lecture psychological/medical literature which investigates this condition, surprisingly little
Notes on Artificial Intelligence 2000; 1810: 55–62 work has been undertaken to use automated software tools to investigate the
27. Constantinidis AS, Fairhurst MC, Rahman AFR. Detection of sequential/constructional aspects of test performance.
circumscribed masses in digital mammograms using behaviour- The application of pattern recognition reported here is novel since it has a
focus on the extraction of dynamic features to characterise patient performance,
knowledge space method. Electronic Letters 2000; 36(4): 302–303 thereby analysing patient responses in a way which is entirely unavailable to more
28. Donnelly N, Guest R, Fairhurst M, Potter J, Deighton A, Patel conventional testing tools. In particular, we investigate the relative merits of static
M. Developing algorithms to enhance the sensitivity of cancel- vs dynamic features in relation to their diagnostic power, and show how the
lation test of visuo-spatial neglect. Behavior Research Methods, procedures we propose can combine these diverse sources of evidence to offer a
Instruments and Computers 1999; 31(4): 668–673 technique which allows a more sensitive evaluation of this debilitating condition.
The contribution to the field of pattern analysis lies principally in the novel
29. Chedru F, Leblanc M, Lhermitte F. Visual searching in normal way in which standard techniques, which are central to the field, can be exploited
and brain-damaged subjects. Cortex 1973; 9: 94–111 in a powerful way to address an application domain which has been largely
30. Avro J, Graphic Gems II. Academic Press, 1991 unexplored hitherto. Additionally, the principles involved in seeking to integrate
diverse sources of information to enhance robustness and reliability allow us to
31. Morrison DF. Multivariate Statistical Methods. McGraw-Hill, evaluate directly several alternative strategies for improving processing capabilities
New York, 1967 and dignostic resolution. Finally, the notion of identifying and computing descrip-
32. Tatsuoka MM. Multivariate Analysis. Wiley, New York, 1971 tors of dynamic sequences of actions to characterise performance represents an
33. Tou JT. Pattern Recognition Principles. Addison-Wesley, important principle which could be applicable in a variety of other task domains.
Reading, MA, 1974

Correspondence and offprint requests to: R. M. Guest, Department of Elec-

Richard Guest obtained a BEng in computer science from the University of York tronics, University of Kent, Canterbury, Kent CT2 7NT, UK. Email:
in 1995 and a PhD in electronic engineering from the University of Kent in rmg얀ukc.ac.uk

View publication stats