Title: Presentation of: Microdialysis Monitoring as a Mean to Detect Cerebral Ischemia by use of Data Minin
1Presentation ofMicrodialysis Monitoring as a
Mean to Detect Cerebral Ischemia by use of Data
Mining
Project group 859a
- Tuesday the 29th of June 2004
- 9.00
2Agenda
- Introduction (Nicolaj) E
- Data Mining Overview (Nicolaj)
- Results Clustering (Svend)
- Results Linear Regression (Svend)
- Results Neural Networks (Tue)
- Conclusion (Michael)
- Future Work (Michael)
3Introduction (1/6)
- aSAH is a bleeding from an aneurysm in the
subarachnoid space.
Aneurysm
4Introduction (2/6)
- Facts about aSAH
- Incidence 15 per 100000 per year.
- aSAH is combined with a high rate of mortality
and morbidity. - 15 die before they reach the hospital.
- a total of 40 die in the acute phase.
- More than half of the survivors must live with
weak or severe morbidity.
5Introduction (3/6)
- Many of the complications are directly or
indirectly related to cerebral energy metabolism. - It is ususally caused by the impaired blood flow
(substrate supply). - Cerebral ischemia a local lack of blood in the
brain.
6Introduction (4/6)
- Microdialysis is a way to measure the
concentration of energy related metabolites.
7Introduction (5/6)
- Energy related metabolites
- Glucose
- Pyruvate
- Lactate
- Glycerol
8Introduction (6/6)
- Is it possible to predict complications related
to metabolism? - Microdialysis can provide the needed information
about the current cerebral metabolism. - Data mining can be used to analyze the collected
data in order to produce knowledge from the data.
9Agenda
- Introduction (Nicolaj)
- Data Mining Overview (Nicolaj) E
- Results Clustering (Svend)
- Results Linear Regression (Svend)
- Results Neural Networks (Tue)
- Conclusion (Michael)
- Future Work (Michael)
10Data Mining Overview (1/10)
- Data mining is a process that turns data into
knowledge. - Methodologies like CRISP-DM add stability to the
project.
11Data Mining Overview (2/10)
- Data Preprocessing steps are very time consuming.
- Cleaning - garbage in, garbage out.
- Includes 4 mis'es
- Missing values.
- Mis-measurements.
- Mis-placements.
- Mis-spellings.
- Transformations, DB- technology saves time.
12Data Mining Overview (3/10)
- How to deal with mis-measurements?
- Remove outlier values.
- Slide window of size 5 over the data.
- Calculate mean and std.dev.
- Remove outliers, i.e. values greater than /-
2std.dev. - Replace with the mean value.
- Noise Reduction.
- Replace each value with a moving average.
13Data Mining Overview (4/10)
- Data Warehouse Technology can help save time for
producing Transformations. - In a DW data are stored in a summarized manner.
- It allows instant answers to rapid questions.
- OnLine Analytical Processing (OLAP).
14Data Mining Overview (5/10)
- Data are stored in Cubes (metaphore).
- A cube has multiple dimensions.
- The content of the cube is measures.
- In our case
- dimensions are
- clinical outcome, age, time, sex, events etc.
- measures are
- glucose, pyruvate, lactate, glycerol, L/P, L/G.
15Data Mining Overview (6/10)
16Data Mining Overview (7/10)
- OLAP Example What was the average metabolism for
patients who died, had ischemia, had normal
outcome over time? - This question has been answered using a DW tool.
17Data Mining Overview (8/10)
18Data Mining Overview (9/10)
19Data Mining Overview (10/10)
20Agenda
- Introduction (Nicolaj)
- Data Mining Overview (Nicolaj)
- Results Clustering (Svend) E
- Results Linear Regression (Svend)
- Results Neural Networks (Tue)
- Conclusion (Michael)
- Future Work (Michael)
21Clustering
- Want to find homogeneous groups of data
- Averge distance to the centroid
- Cluster at global normalized data
- Glucose
- Glycerol
- Lactate
- Pyruvate
- L/G ratio
- L/P - ratio
22Clustering with 5 clusters (1/4)
6.73 1.09
0.54 0.15
354 88.5
84.2 35.8
Severe ischemia found by Schulz,2000
23Clustering with 5 clusters (2/4)
97.8 32.2
16.7 4.70
Severe ischemia found by Schulz,2000
24Clustering with 5 clusters (3/4)
25Clustering with 5 clusters (4/4)
- Cluster 5 is a ischemic pattern cluster, but with
only 4 patients, who had a ischemic pattern. - Patient js085 two instance are measured under
surgery and has non ischemic pattern described.
26Agenda
- Introduction (Nicolaj)
- Data Mining Overview (Nicolaj)
- Results Clustering (Svend)
- Results Linear Regression (Svend)E
- Results Neural Networks (Tue)
- Conclusion (Michael)
- Future Work (Michael)
27Linear regression (1/2)
- Why linear regression?
- High correlation between lactate and pyruvate
- Will there be differents between gorups of
patients? - Normal pattern
- Ischemic pattern
- Dead patients
28Linear regression (2/2)
- Are there a linear regression in data?
- Simple linear regrsssion model
- y ß0 ß1x e gt E(y) ß0 ß1x
- Estimated simple linear regrsssion equation
- y b0 b1x
- Testing for significance H0ß1 0 Ha ß1 ? 0
- t-Test F-Test
- Can it be used to define ischemia pattern?
29Normal group
- 20 patients
- 1892 instances
- Lac0,5016,70Pyr
- Hoß10 Reject , with a 0.01
99 predicted
30Ischemia group
- 9 patients
- 791 instances
- Lac0,4024,38Pyr
- Hoß10 Reject, with a 0.01
99 predicted
31Dead group
- 6 patients
- 410 instances
- Lac1,8318,29Pyr
- Hoß10 Reject, with a 0.01
99 predicted
32The 3 groups compared
- It is difficult to make decision about ischemic
pattern from the linear regression
99 predicted normal
33Patient variation (1/2)
Patients in the Ischemic pattern group
Lac 6.73 1.09
L/P 97,8 32.2
Pyr 0.084 0.036
Severe ischemia found by Schulz,2000
34Patient variation (2/2)
Patients in the Dead group
Lac 6.73 1.09
L/P 97,8 32.2
Pyr 0.084 0.036
Severe ischemia found by Schulz,2000
35Trend for patient
js059
js106
js071
36Agenda
- Introduction (Nicolaj)
- Data Mining Overview (Nicolaj)
- Results Clustering (Svend)
- Results Linear Regression (Svend)
- Results Neural Networks (Tue) E
- Conclusion (Michael)
- Future Work (Michael)
37Neural network introduction
38Modelling with NN
- Different approaches
- With time and without time
- Different sizes of the hidden layer
- Different combinations of input
39Fundamentals of each neural network
- Each neural network consists of an input layer,
hidden layer, and output layer - Each network has been trained using 15 neurons in
the hidden layer - Each network follows the principles of the back
propagation technique - A sigmoid tranfer function has been used in
determining output by the neurons
40First models(1/3)
The network is used to predict 40 values from the
40 actual values
41First models(2/3)
The network is used to predict 40 values from the
40 actual values plus the 10 timestamps
42First models(3/3)
And so on toward the 10th forecasted value...
The second neural network in the series
The third neural network in the series
The first neural network in the series
43Final model
And so on toward the 10th forecasted value...
The second neural network in the series
The third neural network in the series
The first neural network in the series
44Results by visual inspection
45Result by tests(1/3)
- MSE and Pearsons correlation coefficient
- Value of error and linarity
- Range of PCC at 10 0.53-0.79
- Range of PCC at 7 0.70-0.83
- MSE ranges between 0.001-0.039
- Significance test is needed
46Result by test (2/3)
- T-test
- Assumes normal distribution
- Visual inspection
47Result by test (3/3)
- Glucose H0 rejected from 7 to 10
- Glycerol H0 rejected at 1, 2, 9 and 10
- Lactat H0 rejected at 10
- Pyruvate H0 rejected at 10
48Conclusion of the neural network
- It is sensible to up to the 4th predicted value
of each metabolite. - It is not possible to use time
- if the data was collected at with equal time
intervals the accuracy might increase
, but
49Agenda
- Introduction (Nicolaj)
- Data Mining Overview (Nicolaj)
- Results Clustering (Svend)
- Results Linear Regression (Svend)
- Results Neural Networks (Tue)
- Conclusion (Michael) E
- Future Work (Michael)
50The thought application
- Basic functionality
- Visualizing the past 10 values
- And the 4 predicted values
- Indicator lines based on likelihood of a given
event - Individual settings for the patients (Alarm
settings)
51Deriving this solution (1/2)
- Cleaning for outliers and such
- Better cleaning principles has been suggested
- Neural network
- Gave information on future values
- Estimated a 3-4 hour future forecast
- Clustering
- Resulted in information that is interesting, and
relevant, but difficult to implement
52Deriving this solution (2/2)
- Use Data Warehouse principles by SQL statements
for store, fetch and visualize data - Finally, suggestions for clinical use of the
system
53Solving the problemFormulation of the problem
- An analysis of the potential in microdialysis
monitoring is wanted - Can microdialysis monitoring be used to improve
the outcome for aSAH patients and can this be
done by the use of data mining techniques?
54Is the problem solved?
- To some extend, yes...
- This project has investigated the potential of
microdialysis - Positive results for neural network, and
interesting information on microdialysis
substrates - All of which indicating significant possibilities
- A complete implementation of results are
impossible before real knowledge of its
influence has been investigated
55Agenda
- Introduction (Nicolaj)
- Data Mining Overview (Nicolaj)
- Results Clustering (Svend)
- Results Linear Regression (Svend)
- Results Neural Networks (Tue)
- Conclusion (Michael)
- Future Work (Michael) E
56Follow-up assignments
- The possibilities by the use of EEG in
conjunction with microdialysis - Improvements in the CMA equipment
- Optimize ways of collecting data
- Correction of modelling methods
57EEG as supplement to micordialysis (1/3)
- EEG for clinical use by surgecal procedures
- Quantitative EEG
- Power Spectral Analysis to gain information on
frequency - Case shown by next slice
- Reliability has been shown for use of EEG as an
indicator for Cerebral Ischemia
The reliability of Quantitative
Electroencephalography as an indicator of
Cerebral Ischemia Adams D.C. et al.
58EEG as supplement to micordialysis (2/3)
- 74 year old patient, having a left posterior MCA
infact presented by dysphasia and right
hemiparesis. - The EEG shows continuous left hemisphere slowing.
- A cortical infact of this size is typically
continuous.
59EEG as supplement to micordialysis (3/3)
These two blocks can supplement each other
60Better registration by automatisation (1/2)
61Better registration by automatisation (2/2)
- A solution is not eminent.
- Re-evaluate design and collection of data.
- The advantage Less strain on clinical staff and
more secure measurements (Time variance)
62...Better registration though is not just
automatisation
- Procedures for collecting data
- Time intervals between measurements
- Incoorperation these in existing rutines
- Standardization of registration (Same way of
representing data) - Could easily be accomplished by implementing a
system like the developed - Remove the clinical way of collecting data
- Trouble with existing data from collection bias
- Scratch my back, Ill scratch yours
63Improvements in modelling
- First of Substantual amount of data (Not enough
data is accessable) - Perhaps different types of modelling or different
use of allready used techniques - E.G. Neural Network Radical Basis Functions
- Different modelling tools, Clementine has its
limitations - Hugin may be the solution in cases of Baysian
Network proposals - Investigation of inter-cluster likelihood and so
on...
64This was the presentation by project-group
859a Thank you for listening
...By further investigation that is