KEDRI - Stroke

KEDRI
AUT

KEDRI
AUT Main Nav
Centre Banner
Main Content

Stroke

Stroke Risk Prediction Using Spiking Neural Network Methods

By Othman, M. (2015) [1]

 

Background and Motivation of Study

In New Zealand, stroke has become the third leading cause of death after cancer and heart disease, and it is the greatest cause of disability in older people [2]. Each year it is predicted that over 7,000 New Zealanders will experience a stroke event, and of this population at least three-quarters will die or be dependent on others for care one year after the event [3]. Health costs associated with stroke both to individuals and the health system is high and increasing. When severe, stroke has a substantial impact on the psychological and physical well-being of both patients and their families [4].
It is often observed that individuals with similar medical risk factors suffer a stroke at different time points to each other. It has been suggested that the external environment plays a significant role as a trigger of stroke. Moving toward personalized preventive measures, in this study we apply an individualized approach to stroke risk assessment using NeuCube SNN architecture in Figure 1. The first objective of this study is to verify that the increment of stroke risk in an individual with modifiable risk factors for stroke is triggered by a combination of environmental factors. Secondly, to understand the association of combined environmental variables toward a stroke event on individual level of risk. Thirdly, exploring the influence of prolonged exposure to inclement environmental conditions several days before a stroke event that can help us determine the earliest time point at which we can best predict the risk of stroke incident in individuals.

NeuCubeFigure 1. A functional diagram of the NeuCube SNN architecture, consisting of: Input data encoding module; 3D SNN Neurogenetic cube using STDP learning rule; Output module for classification or regression using RO and STDP learning rules; Gene regulatory network (GRN) module (optional) [5]

Dataset

The dataset consists of 2805 samples (all with first-ever occurrence of stroke) taken from the Auckland Regional Community Stroke Study (ARCOS) population. This data set has been published in several studies where they apply different methods to the case study, for example statistical methods [6], FaLK-SVM (global method) [7], and a combination of evolving personalized method (evoPM) with gravitational search algorithm and recurrent reservoir [8]. In exploring the feasibility of NeuCube M1 (system) for predictive modelling of SSTD this study has been published recently in 2014 International Joint Conference on Neural Network (IJCNN) [9]. The third ARCOS study was chosen which is between 2002 and 2003, consisting of 1207 subjects. These subjects were then divided into several subgroups stratified by season (summer, autumn, winter, spring), history of hypertension, smoking status and age to explore the susceptibility of groups to the influence of environmental changes (see Table 1). Each subject is described by 12 environmental variables (wind speed, wind chills, temperature dry, temperature wet, temperature max, temperature min, humidity, atmospheric pressure, nitrogen dioxide, sulphur dioxide, ozone gas, solar radiation) measured daily within a time window before the stroke event, we created personalised models of 108 individuals using the NeuCube SNN architecture from Figure 1.

 Stroke Occurrences Dataset

Table 1: Stroke Occurrences Dataset.

Since the data consists only of stroke subjects, the time window between days 60 and 40 before the event was used as ‘low risk’ and the days between 20 and 1 days before the event as high risk (see Figure 2). 

Time windows to discriminate between ‘low risk’ and ‘high risk’ stroke class

Figure 2: Time windows to discriminate between ‘low risk’ and ‘high risk’ stroke class.

Revealing dynamic triggering patterns

A personalised SNN model can be used to reveal dynamic patterns of consecutive changes in the environmental variables over time that would represent meaningful, predictive triggering patterns for an individual (see Figure 3). Figure 3 suggests that consecutive changes of input variables, represented as trajectories of spiking activities of neurons in the SNN cube, would have an impact on a stroke risk.

Revealing dynamic temporal patterns as trajectories of major consecutive changes in the environmental variables presented as a temporal order from 1 to 10, related to Low Risk and High Risk of stroke prediction for two selected individuals: Subject 20 in the summer season - (a) low risk trajectory pattern, (b) high risk trajectory pattern; Subject 1 in the spring season - (c) low risk trajectory pattern, (d) high risk trajectory pattern. 

Figure 3. Revealing dynamic temporal patterns as trajectories of major consecutive changes in the environmental variables presented as a temporal order from 1 to 10, related to Low Risk and High Risk of stroke prediction for two selected individuals: Subject 20 in the summer season - (a) low risk trajectory pattern, (b) high risk trajectory pattern; Subject 1 in the spring season - (c) low risk trajectory pattern, (d) high risk trajectory pattern.

Seasonal Variation Analysis [1]

It could be argued that each season forms a different population grouping due to its different environmental conditions. To determine whether any differences were significant or not a two sample t-test with unequal variances was carried out. Using this t-test with all variant couplings of seasons, it was found that it is possible to group the seasons into two separate populations, see Table 2. As can be seen from Table 3, seasonal comparisons outside these groupings did show a statistically significant difference in population means, with an alpha value of 0.05.

When comparing the standard deviation of the environmental data within each season autumn showed the largest value at 4.19, see Table 4. This indicates that there is a greater level of variation within the data in this season than in any of the others, which may in turn affect the model built upon this data and its accuracy of prediction.

Results of t-test for significant difference between seasonal groups  

Table 2: Results of t-test for significant difference between seasonal groups.

Two sample t-test with unequal population variances 

 Table 3: Two sample t-test with unequal population variances.

Mean and standard deviation for environmental data within each season 

 Table 4: Mean and standard deviation for environmental data within each season.

Conclusion

The novelty of this research is in proving there are hidden associations amongst the environmental features that increase the risk of stroke for an individual. This research has successfully achieved the objectives by predicting accurately the risk of stroke occurrences at the earliest time point at 6 days prior, and demonstrates that analysing all the features collectively can accurately predict stroke risk. We also confirmed that there is a cascading effect unique for each individual depending on their exposure to certain environmental factors within a specific time window.

References: 

  1. Othman, M. Spatial-temporal data modelling and processing for personalised decision support (Doctoral dissertation, Auckland University of Technology). 2015. (http://hdl.handle.net/10292/9079 ) 
  2. Martin Tobias, Jit Cheung and Harry McNaughton. Modelling stroke: a multi-state life table model. Ministry of Health, 2002. 
  3. John Gommans, Alan Barber, Harry McNaughton, Carl Hanger, Patricia Bennett, David Spriggs and Jonathan Baskett. Stroke rehabilitation services in New Zealand. The New Zealand Medical Journal, vol. 116, no. 1174, pages U435–U435, 2003.
  4. Maree L Hackett, John R Duncan, Craig S Anderson, Joanna B Broad and Ruth Bonita. Health-related quality of life among long-term survivors of stroke results from the Auckland stroke study, 1991–1992. Stroke, vol. 31, no. 2, pages 440–447, 2000.
  5. Nikola K Kasabov. NeuCube: A spiking neural network architecture for mapping, learning and understanding of spatio-temporal brain data. Neural Networks, vol. 52, pages 62–76, 2014.
  6. Craig S Anderson, Kristie N Carter, Maree L Hackett, Valery Feigin, P Alan Barber, Joanna B Broad, Ruth Bonitaet al. Trends in stroke incidence in Auckland, New Zealand, during 1981 to 2003. Stroke, vol. 36, no. 10, pages 2087–2093, 2005.
  7. Wen Liang, Yingjie Hu, Nikola Kasabov and Valery Feigin. Exploring associations between changes in ambient temperature and stroke occurrence: comparative analysis using global and personalised modelling approaches. In Neural Information Processing, pages 129–137. Springer, 2011.
  8. Nikola Kasabov, Valery Feigin, Zeng-Guang Hou, Yixiong Chen, Linda Liang, Rita Krishnamurthi, Muhaini Othman and Priya Parmar.Evolving spiking neural networks for personalised modelling, classification and prediction of spatio-temporal patterns with a case study on stroke. Neurocomputing, vol. 134, pages 269–279, 2014.
  9. Muhaini Othman, Nikola Kasabov, Enmei Tu, Valery Feigin, Rita Krishnamurthi, Zhengguang Hou, Yixiong Chen and Jin Hu. Improved predictive personalized modelling with the use of Spiking Neural Network system and a case study on stroke occurrences data. In Neural Networks (IJCNN), 2014 International Joint Conference on, pages 3197–3204. IEEE, 2014. (http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6889709&tag=1 )

Supplementary material

 



Last updated: 21 Oct 2015 1:15pm

Auckland University of Technology, New Zealand | Copyright © | Privacy | Site map