Presentation is loading. Please wait.

Presentation is loading. Please wait.

Final Project Classification of Sleep data Akane Sano Affective Computing Group Media Lab.

Similar presentations


Presentation on theme: "Final Project Classification of Sleep data Akane Sano Affective Computing Group Media Lab."— Presentation transcript:

1 Final Project Classification of Sleep data Akane Sano akanes@mit.edu Affective Computing Group Media Lab

2 Polysomnography Multi-parametric test to evaluate sleep

3 Data and Labels Data: – Healthy students (N=7) – Electroencephalogram(EEG) 100Hz – Electro dermal activity(EDA) 32Hz – Motion 32Hz Labels : sleep stages (every 30s) (Wake, REM, NonREM1-4, Movements/Noise) -> Wake, REM, Non-REM, Movements/Noise

4 Sleep Stage NREM

5 Sleep Stage EEG Motion EDA [Hz]

6 Questions Which features are the best to estimate sleep stage? How accurate can we estimate sleep stage from EDA and motion?

7 data SubjectWAKEREMNREMMOVEMENT Q3171245270 R151417193 S461847149 T101896300 U512097093 V382037035 W4917973532 %8.018.872.40.8

8 Features Recorded signals were segmented into 30s window Computed with MATLAB EEG : – Frequency Energy δ:0.5-4,Hz θ:4-8 α:8-13Hz β:13-40Hz γ:40-50Hz) Motion: – Amplitude – Standard Deviation – Zero-crossing – Frequency Energy EDA : Amplitude (Normalized) Standard Deviation # of peaks Gradient Frequency Energy

9 Red Non-REM, blue Wake pink REM black M

10 Methods Using Matlab k Nearest Neighbors (k=1-199) Support Vector Machine (libSVM) – Linear – Polynomial – Radial Basis Function – Dynamic features (still working) Neural Network (n=2, 4, 6, 8, 10, 20) Hidden Marcov Model / Baysian Network (still working) – Gaussian Mixture Model (HMM toolbox Errors, Bayes Net toolbox : Errors) – Discrete Model (HMM toolbox Errors)

11 Methods (cont.) Leave one subject out Compare Accuracy – Classification Methods – Features

12 Results – Accuracy [%] (SVM) SVM FeaturesLinear PolyRGB ALL73.272.471.5 EEG71.372.669.9 EDA72.6 MOTION72.6 70.4 W30.7 REM32.1 NREM91.2 MOVE0 misclassified to NREM W0 REM0 NREM100 MOVE0 predicted WRNM true W0.30.10.60.0 R 0.30.70.0 NR0.00.10.90.0 M0.30.10.50.0

13 Results – Accuracy [%] (kNN) featureskNN ALL76.0n=137 EEG76.7n=9 EDA70.1k=199 MOTIO N 70.3n=197 Improved For ALL, EEG, but decreased for EDA and MOTION

14 Results – Accuracy [%] Neural Network Node # Features ALLEEGEDA MOTIO N 2 82.885.272.672.5 4 84.985.672.573.1 6 84.986.172.973.1 8 85.186.873.073.3 10 85.586.573.773.4 20 86.088.274.073.3 Improved!

15 Another Better Feature? Added Elapsed time (0:Record Start-1:End) SVM kNN LinearPolyRGB ALL73.572.171.7ALL76.1 EEG72.272.670.9EEG75.5 EDA72.671.771.9EDA68.6 MOTIO N 72.6 MOTION70.0 0.3% UP! Improved For ALL, EEG, but decreased for EDA and MOTION

16 Neural Network (Elapsed Time added) Node # Features ALLEEGEDAMOTION 284.383.173.073.8 486.685.772.473.6 686.785.073.974.3 887.485.774.474.0 1087.586.474.274.3 20 87.689.274.8 Improved!

17 Best classifier Neural Network(n=20) EEG with elapsed time Precited WRNM True W0.70.10.30.0 R 0.80.20.0 NR0.00.10.90.0 M0.30.00.20.5

18 Summary Node # Features ALLEEGEDA MOTIO N SVM73.272.6 kNN 76.076.770.170.3 Neural 86.088.274.0 73.4 SVM with time 73.572.6 kNN with time 76.1 75.568.670.0 Neural with time 87.689.274.8

19 Conclusion EDA and Motion showed less accuracy than EEG and ALL Wake, REM, Movement were misclassified to N- REM Neural Network showed the best accuracy Elapsed Time might be effective Future Work – Temporal model – Features


Download ppt "Final Project Classification of Sleep data Akane Sano Affective Computing Group Media Lab."

Similar presentations


Ads by Google