Stroke is one of the major diseases that can cause human deaths. However, despite the frequency and importance of stroke, there are only a limited number of evidence-based acute treatment options currently available. Recent clinical research has indicated that early changes in common physiological variables represent a potential therapeutic target, thus the manipulation of these variables may eventually yield an e ective way to optimise stroke recovery. Nevertheless the accuracy of prediction methods based on statistical characteristics of certain physiological variables, such as blood pressure, glucose, is still far from satisfactory due to vague understandings of e ects and function domain of those physiological determinants. Therefore, developing a relatively accurate prediction method of stroke outcome based on justi able determinants becomes more and more important to the decision of the medical treatment at the very beginning of the stroke. In this work, we utilize machine learning techniques to nd correlations between physiological parameters of stroke patient during 48 hours after stroke, and their stroke outcomes after three months. Our prediction method not only incorporates statistical characteristics of physiological parameters, but also considers physiological time series patterns as key features. Experiment results on real stroke patients' data indicate that our method can greatly improve prediction accuracy to a high precision rate of 94%, as well as a high recall rate of 90%.
Stroke is a common cause of human death and is a major cause of death after
ischemic heart disease [
Acute Ischemic Stroke Prediction as blood pressure, temperature and blood glucose levels, may eventually yield an e ective and potentially widely applicable range of therapies for optimising stroke recovery, such as abbreviating the duration of ischaemia, preventing further stroke, or preventing deterioration due to post-stroke complications.
The correlations between blood pressure and stroke outcomes have been
widely studied in the literature. It is stated in current guidelines that a
signi cant decrease of BP during the rst hours after admission should be avoided,
as it correlates with poor outcomes, measured by Canadian Stroke Scale or
modi ed Rankin Score (mRS), at 3 months [
Research also shows associations between other physiological variables and stroke outcomes. Abnormalities of blood glucose, heart rate variability, ECG and temperature may be predictors of 3-month stroke outcome.
Most of the above analyses are based on periodically recorded physiological
parameters, hourly or daily, up to 3 months. Whether continuous data patterns,
such as data trends, have a similar predictive role is still uncertain. Although it is
clear that the after stroke elevated 24-hours blood pressure levels predict a poor
outcome, few studies have investigated the predictive ability of more
sophisticate trends, e.g. combined trends of several physiological parameters. Yet this
could be an e ective way to readily obtain important prognostic information for
acute ischemic stroke patients. Dawson et al did pioneering works on associating
shorter length (around 10 minutes) beat-to-beat BP with acute ischemic stroke
outcomes [
Obviously mining physiological data patterns can be easily aligned with time series data classi cation, which is a traditional topic and has attracted intensive studies. Although there exist many sophisticate time series data mining techniques, we nd that most of them, if not all, are not applicable to our application scenario, due to the always incomplete, non-isometric physiological data collected from patients. Therefore, in this paper, we incorporate a simple yet powerful time series data pattern analysing method, trend analyses, into our prediction method. By utilising those trend features, together with values of traditional physiological variables, we design an e cient algorithm that can predict 3-month stroke outcome with high accuracy.
In summary, we list our contributions in this paper: { We propose using trend patterns of physiological time series data as a new set of stroke outcome prediction features, { We design a novel prediction algorithm which can accurately predict 3months stroke outcomes with high precision and recall rate, when tested against a real data set.
The rest of this paper is organised as follows. Section 2 introduces works related to stroke outcome predictions. Section 3 presents our prediction methods. Section 4 reports empirical study results. And section 5 concludes this paper with possible future studies. 2
The relationship between beat-to-beat blood pressure (BP) and the early
outcome after acute ischemic stroke was rstly described in [
A further investigation on BP was done in [
Relationships between other physiological variables and stroke outcome have
also been studied in literature. Abnormalities of serum osmolarity, temperature,
blood glucose, SPO2 may be predictors of stroke outcomes. More speci cally,
heart rate and ECG, can be correlated to stroke outcomes at 3-months:
{ Heart Rate Variability: Gujjar et al. reported that heart rate variability is
e cient in predicting stroke outcome. Speci cally they studied continuous
echocardiogram of 25 patients with acute stroke and concluded that the
eyeopening score of Glasgow Coma Scale and low-frequency spectral power were
factors that were independently predictive of mortality [
Acute Ischemic Stroke Prediction
Our prediction method adopts statistical values of physiological parameters and
also incorporates the descriptive ability of the physiological patterns as features
to predict 3-months stroke outcomes. Particularly, we use the trend pattern of
time series data as new add-on features to form an initial feature set. Then we
apply the logistic regression method to classify stroke patient outcomes into two
groups: good vs. bad. Note that there exist di erent clinical criteria in de ning
good/bad outcomes. We will report empirical study results on all criteria in the
next section. Cross validation is also adopted to obtain an unbiased assessment of
classi er performance, by which the physiological determinants can be accurately
identi ed in the last stage. Finally, we select a subset of features that can most
accurately predict 3-months stroke outcomes. Figure 1 presents logic ows of our
method. We use Rankin Scale to represent various outcomes at 3 months after
stroke (RS3) [
Moreover, for each physiological parameter, we compute trends through
partitioning the time series data into non-overlapping, continuous blocks. Although
there exists many trend and shape detection methods in the literature, such as
[
In each partition, accordingly we generate 6 new features, as shown below, to represent the trend pattern: 1. yChange: the di erence between the value at the end of a trend and the value at the start of a trend yChange = y(end of trend) y(start of trend) 2. absYChange: the absolute value of the yChange 3. slope: the slope of the trend 4. sign: the direction of the trend 5. NumofMeasure: the number of values in a partition 6. FreqofMeasure: the average time interval between measurements, i.e.
F reqof M easure =
T rend Length
N umof M easure
The initial feature set comprised physiological values and their trend patterns. We apply the logistical regression method to classify the good/bad stroke outcomes based on this initial feature set. 3.2
In statistics, logistic regression is a type of regression analysis used for predicting the outcome of a binary dependent variable (a variable which can take only two possible outcomes, e.g. \yes" vs. \no" or \success" vs. \failure") based on one or more predictor variables. Like other forms of regression analysis, logistic regression makes use of one or more predictor variables that may be either continuous or categorical. Unlike ordinary linear regression, however, logistic regression is used for predicting binary outcomes rather than continuous outcomes. Logistic regression adopted here is a type of regression analysis used for predicting the outcome of stroke (\good" vs. \bad") based on features in our initial feature set.
To obtain an unbiased assessment of classi er performance, the Leave-OneOut Cross validation technique is adopted. Suppose N folds are employed, this
Acute Ischemic Stroke Prediction technique withholds a subject from the training set for each run to later test with. Once a record has been withheld for testing, the classi er is trained using the remaining N-1 subjects. The withheld subject is then reintroduced for classi cation. 3.3
We use two greedy search strategies to nd the best feature subset that can achieve highest prediction accuracy. Speci cally, we use backward search and forward search: backward search : A greedy backward search is performed to identify a near optimum subset of features. Starting with all features, in sequence, the feature which improves prediction accuracy the most (or decreases it the least) is removed from the current set of features and retained as an intermediate feature subset. This is repeated until all features have been removed. The intermediate feature subset which provides the maximum performance, compared to all other subset evaluated, is selected as the nal feature set. forward Search A sequential forward oating search algorithm is used for feature selection, in an attempt to discover the optimal subset of features from the pool of available candidate features. This strategy begins with a forward-selection process, selecting a single feature from the pool of available features, which improves the prediction accuracy most. After this selection, removal of a feature from the set of selected features is considered. The process of possible feature addition, followed by possible feature removal, is iterated until the selected feature set converges. 4
In this section, we report experiment results through testing our prediction method on a real data set of stroke patients. Firstly, we introduce the physiological data sets of stroke patients and the good/bad criteria used in our study. Then we report prediction accuracy based on various combination of feature sets. Our study was approved by a ethics committee of the related institution. 4.1
A cohort of 157 patients with acute ischaemic stroke were recruited. Patients presenting to the Emergency Department of the Royal Brisbane and Women's Hospital, an Australian tertiary referral teaching hospital, within 48 hours of stroke or existing inpatients with an intercurrent stroke were enrolled prospectively. Important physiological parameters, such as blood pressure, were recorded at least every 4 hours from the time of admission until 48 hours after the stroke. These values were used as the outcome variable in the analyses. The measurements from patients who died during these rst 48 hours were also included in the analyses. Furthermore, some demographic and other stroke-related data were also collected such as the age and gender. The age range of these 157 patients was 16 to 92 years with median age 75 years. The patient distribution based on di erent values of RS3 is showed in Figure 2. As shown in Figure 2, RS3 score varies between 0 and 6. Patients with RS3 = 6 means the subject is dead after three months and RS3 = 0 means the subject recovers quite well after three months. Based on RS3 values, patient outcomes can be divided into good/bad groups basing on di erent grouping criteria. Figure 3 illustrates patient distributions under three type grouping criteria. 4.3
Applying techniques described in Section 3, we run experiments on various grouping criteria to test our stroke outcome prediction algorithm. We always notice that `backward search' generates more accurate prediction results, which will thus be used as our default feature set search strategy. Figure 4 shows prediction accuracy comparisons under all three types of grouping criteria. In Figure 5, we also evaluate the e ciency of including trend pattern as prediction
Acute Ischemic Stroke Prediction features. Experiment shows that by adding those simple trend features, the prediction accuracy on all three grouping types is unanimously boosted from 71% to 89 91%. 5
In this paper, we describe novel algorithms to predict three months stroke outcomes. We have quanti ed the great improvements brought by including physiological data trend patterns as features of a classi er. We believe that these trends play important roles on three months outcomes of stroke patients. The e ciency and accuracy of our algorithm have also been demonstrated through our experiments.
In our future work, we will rst try to locate the most important trend pattens for stroke outcome predictions. Then we will work with healthcare professionals to nd clinical ground truth beneath those physiological trend patterns of stroke patients. This will greatly bene t clinical treatments of acute ischemic stroke. We also plan to run clinical trials to validate our prediction methods on other real data sets of stoke patients.
Acute Ischemic Stroke Prediction