Menu
Index

Sample run 1

 
In this sample run, a sample application was performed to illustrate PSGMiner. Although it is based on real data, the purpose of this example is simply to explain how PSGMiner could be used for testing a hypothesis, and it is not meant to suggest a clinical procedure for experiments of  PLM (Periodic Leg Momement). In this sample study it is aimed to investigate differences which occur in the PSG, features extracted from All channel (without leg channel) during 1 seconds before PLM Left occurrence and during 1 seconds following the onset of the PLM Left. In other words, this was done to see whether the before PLM Left period and onset of PLM Left could be classified using signals from other PSG channels. The detailed  statistical verification of the study results for the sample run is shown in the user interface of Machine Learning Module and below. According to the results for the sample run, the accuracy rate is 51 %. A careful study of the results indicates that there is no important distinction between before PLM period and onset of PLM. However, this is by no means a demanding study, and it was only carried out to illustrate how our software can be used in clinical events. The accuracy rate can be increased by using different algorithms or parameters.
PSG data
PSG data that are used at the designing stage of the study and test run 1 were taken from 153 patients (M/F, 112/41; mean age, 61,48±11.54 years, age range, 32-94).  These patients were actual patients from the archive of a sleep laboratory. Recordings were used in retrospective manner. All patients were diagnosed with periodic leg movement disorder and underwent a one-night PSG for diagnosis.These PSG records are recorded with 44-channel polygraph (Compumedics 44E series, Avustralia) and these archive records are obtained through the use of complete PSG techniques in the Sleep Laboratory of the Faculty of Medicine, Trakya University. PSG assembly involves 6-channel EEG (F3-A2, F4-A1, O1-A2, O2-A1, C3-A2 and C4-A1), right and left electrooculography (EOG) (LOC-A2 and ROC-A1), leg electromyography (EMG), chin EMG, electrocardiography (ECG), oxygen saturation with fingertip pulse oximeter (SpO2), thermistor (for upper airway signals), thorax and abdomen, snoring (microphone), and measurement of body positions. EEG electrodes are placed according to the internationally accepted 10-20 system. It is determined that the sampling rate used during the recording of lung and upper airway signals is 256 Hz, the sampling rate for thoracic and abdominal respiration signals is 128 Hz, the lower and upper frequencies of the filtering done is 0.3 Hz and 30 Hz respectively. PSG data are manually scored by a physiologist, according to the International Classification of Sleep Disorders (ICSD), which was written by AASM.The study protocol was approved by local ethics committee.
 
General characteristics of the study group.
 
Age
Weight (Kg)
Length (cm)
PLM Index
Minimum
32
47
150
28
Maximum
94
195
192
100
Average
61.48
93.92
170.9
66.86
Standard deviation
11.54
20.19
8.57
16.26
 
Feature selection
Channels and features: All available channels and features in PSGMiner
Evaluator: InfoGain
Search method: Ranker
 
After feature selection, selected channels and features for classification.
No         
Channel
Feature
1
LOC
Minimum value
2
C4-A1
Slope
3
F3-A2
Kurtosis
4
F3-A2
Intensity weighted mean frequency
5
F3-A2
Line Length
6
LEGL
class {0,1}
class 0:Not PLM Left
class 1:PLM Left
 
Classification Results (PSGMiner and WEKA)
 
PSGMiner
 
K-Nearest Neighbor
 
K : 1
 
---------------
 
Instances :960
Attributes :6
   
--Summary-- 
   
Accuracy :%51.354
Root mean squared error :0.697
Kappa Value :0.027
0==>
TP :250 TN :243 FP :242 FN :225
Sensitivity :0.526
Specificity :0.501
Precision :0.508
Negative Predictive Value :0.519230769230769
False Negative Rate :0.474
F1 score :0.517
Matthews correlation coefficient :0.162361511508771
ROC Area :0.513673358654368
1==>
TP :243 TN :250 FP :225 FN :242
Sensitivity :0.501
Specificity :0.526
Precision :0.519
Negative Predictive Value :0.508130081300813
False Negative Rate :0.499
F1 score :0.51
Matthews correlation coefficient :0.162361511508771
ROC Area :0.513673358654368
Average==>
Sensitivity :0.513541666666667
Specificity :0.513805050642069
Precision :0.51373824134876
Negative Predictive Value :0.513622609182823
False Negative Rate :0.486458333333333
F1 score :0.513478850239072
Matthews correlation coefficient :0.162361511508771
ROC Area :0.513673358654368
Confusion Matrix 0 (Classified as)  1 (Classified as)
            0                  250                  225
            1                  242                  243
____________________________________________
 
Neural Network (Rprop)
 
Neuron Type : Sigmoid
Hidenlayer Count : 1
Hidenlayer 0 Neuron Count : 4
Learning Rate : 0.3
Maximum Iteration : 500
Epsilon : 1E-8
Weights Initial : 0.1
Weights Max : 50
 
---------------
 
Instances :960
Attributes :6
   
--Summary-- 
   
Accuracy :%51.25
Root mean squared error :0.517
Kappa Value :0.022
0==>
TP :170 TN :322 FP :163 FN :305
Sensitivity :0.358
Specificity :0.664
Precision :0.511
Negative Predictive Value :0.513556618819777
False Negative Rate :0.642
F1 score :0.421
Matthews correlation coefficient :0.171792475426403
ROC Area :0.510906131307651
1==>
TP :322 TN :170 FP :305 FN :163
Sensitivity :0.664
Specificity :0.358
Precision :0.514
Negative Predictive Value :0.510510510510511
False Negative Rate :0.336
F1 score :0.579
Matthews correlation coefficient :0.171792475426403
ROC Area :0.510906131307651
Average==>
Sensitivity :0.5125
Specificity :0.509312262615301
Precision :0.512049429812588
Negative Predictive Value :0.512017699517699
False Negative Rate :0.4875
F1 score :0.500789096445616
Matthews correlation coefficient :0.171792475426403
ROC Area :0.510906131307651
Confusion Matrix 0 (Classified as)  1 (Classified as)
            0             170             305
            1             163             322
 
____________________________________________
 
WEKA Results
 
Neural Network (MultilayerPerceptron )
 
Options: -L 0.3 -M 0.2 -N 500 -V 0 -S 0 -E 20 -H a
 
 
Correctly Classified Instances         487               50.1545 %
Incorrectly Classified Instances       484               49.8455 %
Kappa statistic                          0.0024
Mean absolute error                      0.4973
Root mean squared error                  0.5032
Relative absolute error                 99.4626 %
Root relative squared error            100.6445 %
Total Number of Instances              971    
 
 
=== Detailed Accuracy By Class ===
 
              TP Rate   FP Rate   Precision   Recall  F-Measure   ROC Area  Class
                0.464     0.461      0.497     0.464     0.48       0.502    0
                0.539     0.536      0.506     0.539     0.522      0.502    1
Weighted Avg.    0.502     0.499      0.501     0.502     0.501      0.502
 
 
=== Confusion Matrix ===
 
  a   b   <-- classified as
223 258 |   a = 0
226 264 |   b = 1
____________________________________________
 
K-Nearest Neighbor
 
Options: -K 1 -W 0 -A "weka.core.neighboursearch.LinearNNSearch -A "weka.core.EuclideanDistance -R first-last""
 
IB1 instance-based classifier
using 1 nearest neighbour(s) for classification
 
 
Correctly Classified Instances         526               54.171  %
Incorrectly Classified Instances       445               45.829  %
Kappa statistic                          0.0835
Mean absolute error                      0.4584
Root mean squared error                  0.6762
Relative absolute error                 91.6851 %
Root relative squared error            135.2472 %
Total Number of Instances              971    
 
 
=== Detailed Accuracy By Class ===
 
              TP Rate   FP Rate   Precision   Recall  F-Measure   ROC Area  Class
                0.545     0.461      0.537     0.545     0.541      0.54     0
                0.539     0.455      0.547     0.539     0.543      0.54     1
Weighted Avg.    0.542     0.458      0.542     0.542     0.542      0.54
 
 
=== Confusion Matrix ===
 
  a   b   <-- classified as
262 219 |   a = 0
226 264 |   b = 1
 
____________________________________________
 
Naive Bayes
 
Correctly Classified Instances         496               51.0814 %
Incorrectly Classified Instances       475               48.9186 %
Kappa statistic                          0.0277
Mean absolute error                      0.4954
Root mean squared error                  0.5234
Relative absolute error                 99.086  %
Root relative squared error            104.683  %
Total Number of Instances              971    
 
 
=== Detailed Accuracy By Class ===
 
              TP Rate   FP Rate   Precision   Recall  F-Measure   ROC Area  Class
                0.848     0.82       0.504     0.848     0.632      0.532    0
                0.18      0.152      0.547     0.18      0.27       0.532    1
Weighted Avg.    0.511     0.483      0.525     0.511     0.45       0.532
 
 
=== Confusion Matrix ===
 
  a   b   <-- classified as
408  73 |   a = 0
402  88 |   b = 1
 
____________________________________________
 
Decision Trees
 
 
Options: -C 0.25 -M 2
 
J48 pruned tree
------------------
 
Correctly Classified Instances         528               54.3769 %
Incorrectly Classified Instances       443               45.6231 %
Kappa statistic                          0.0909
Mean absolute error                      0.4922
Root mean squared error                  0.4985
Relative absolute error                 98.4464 %
Root relative squared error             99.6974 %
Total Number of Instances              971    
 
 
=== Detailed Accuracy By Class ===
 
              TP Rate   FP Rate   Precision   Recall  F-Measure   ROC Area  Class
                0.748     0.657      0.528     0.748     0.619      0.542    0
                0.343     0.252      0.581     0.343     0.431      0.542    1
Weighted Avg.    0.544     0.452      0.555     0.544     0.524      0.542
 
 
=== Confusion Matrix ===
 
  a   b   <-- classified as
360 121 |   a = 0
322 168 |   b = 1
 
 
 
The online help was made with Dr.Explain