1. Automatic detection of spiking events in EMFi sheet during sleep
Jarmo Alametsä, Esa Rauhala, Eero Huupponen, Antti Saastamoinen, Alpo Värri, Atte Joutsen, Joel Hasan, Sari-Leena Himanen 
Medical Engineering and Physics 
Volume 28, Issue 3, Pages (April 2006)
DOI: /j.medengphy
Copyright © 2005 IPEM Terms and Conditions

2. Fig. 1
An example of ECG and corresponding (5–30Hz filtered) EMFi signal segments of 6s. Characteristic components of BCG can be seen in the EMFi signal so that the G, H, I, J and K denote the systolic deflections and L, M and N the diastolic deflections of the BCG complex [20]. The QRS complex of the ECG can be seen preceding the BCG complex during each heart cycle.
Medical Engineering and Physics  , DOI: ( /j.medengphy )
Copyright © 2005 IPEM Terms and Conditions

3. Fig. 2
Block diagram of the developed spiking detector. Letter k indicates time in seconds, A[k] is the amplitude level feature and Ac[k] is the change feature, formed on the basis of history of 60s. The detector output is obtained with 1-s time resolution.
Medical Engineering and Physics  , DOI: ( /j.medengphy )
Copyright © 2005 IPEM Terms and Conditions

4. Fig. 3
An example of ECG, nasal flow pressure and (6–16Hz filtered) EMFi signal segments of 20s with scored spiking events (a). Occurrence of spiking seems to correlate with the flat flow pressure signal with high-frequency activity. Corresponding change feature Ac[k] values of methods 6–10 are seen in (b). Dashed vertical lines indicate the time instants of spiking events. Method 6 is indicated with triangles, method 7 with circles, method 8 with squares, method 9 with diamonds, method 10 with stars. All these versions of change features seem to show large increases at the times of spiking events. Different versions are seen to act with different dynamics, as intended.
Medical Engineering and Physics  , DOI: ( /j.medengphy )
Copyright © 2005 IPEM Terms and Conditions

5. Fig. 4
Comparison of spiking detection outcome from all six night recordings obtained with different feature versions, with frequency range 3–16Hz (a) and 6–16Hz (b). ROC curves are labelled so that method 1 is indicated with triangles, method 2 with circles, method 3 with squares, method 4 with diamonds and method 5 with stars (a). Similarly, method 6 with triangles, method 7 with circles, method 8 with squares method 9 with diamonds, method 10 with stars (b).
Medical Engineering and Physics  , DOI: ( /j.medengphy )
Copyright © 2005 IPEM Terms and Conditions

6. Fig. 5
Performance curves of different feature versions from all six night recordings from the frequency range 3–16Hz (a) and 6–16Hz (b); true positive rate on the left picture and false positive rate on the right picture. Method 1 is indicated with triangles, method 2 with circles, method 3 with squares, method 4 with diamonds and method 5 with stars. Similarly, method 6 is indicated with triangles, method 7 with circles, method 8 with squares, method 9 with diamonds, and method 10 with stars. The effect of threshold value λ into the operation of detection algorithm can be seen in these pictures. By example, analogy between pictures can be seen between performance curves 6–16Hz and ROC Curve (Fig. 3; b). True positive rate 80% and false positive rate 19% were produced by the same threshold λ of 25%.
Medical Engineering and Physics  , DOI: ( /j.medengphy )
Copyright © 2005 IPEM Terms and Conditions

7. Fig. 6
Likelihood functions of two versions of change features from all six night recordings, method 10 (a) and method 2 (b). Method 10 provided the best results (Fig. 4) and method 2 the worst, which can be seen as smaller overlap (dotted area) between spiking and non-spiking seconds in method 10 than in method 2 (Fig. 3). In method 10 the relative area of dotted are was 36.5% and corresponding area in method 2 was 42.6%.
Medical Engineering and Physics  , DOI: ( /j.medengphy )
Copyright © 2005 IPEM Terms and Conditions