1. Politecnico di Milano
Dipartimento di Elettronica, Informazione e Bioingegneria (DEIB)
Biomed Meeting
Sara Bridio
sara.bridio@mail.polimi.it
Thursday, May 12, 2016
Giulia Core
giulia.core@mail.polimi.it
Implementation methods
4. 4
Preprocessing
fir1( #coefficients , f_cutoff_norm )
filtfilt( filter, 1 , signal )
Filtering
FIR filter
f_cutoff = 8 Hz
Signal frequency
0,001-2 Hz
Low pass filter
5. 5
Preprocessing
Peak detection algorithm AMPD [1]
Segmentation
256 samples for
each segment
Resample
[1] F. Scholkmann, J. Boss and M. Wolf, “An Efficient Algorithm for Automatic Peak Detection in Noisy Periodic and Quasi-Periodic
Signals“, 2012
Drift elimination
Segments of PPG signal from subject 1
Segments of PPG signal from subject 1, after drift elimination
7. 7
Preprocessing
Segments Matrix for each subject
Segm
Samples
1 2 … 256
Segm 1
Segm n
…
rows: segments belonging to a subject
columns: samples
8. 8
Features extraction
Template creation
PPG signal template
Template = mean(segm_mat)
Segments matrix
Template of PPG signal from subject 1
Template of PPG signal from subject 2
10. 10
Features extraction
1st derivative template 2nd derivative template
Template of 1st derivative of PPG signals from subject 1
Template of 1st derivative of PPG signals from subject 2 Template of 2nd derivative of PPG signals from subject 2
Template of 2nd derivative of PPG signals from subject 1
11. 11
Features extraction
Template matching
3 matrixes
Euclidean distance
3 matrixes
with distances
One for each type of template
PPG signal template
1st derivative template
2nd derivative template
Sum of distances dij
One for each type of template
i: template subject i
j: template subject j
12. 12
Features extraction
3 matrixes with distances
T1 T2
T1
T2
…
…
…
…
Tn
Tn
0
0
0
0
0
d12
d12
d1n
d1n
d2n
d2n
…
…
… …
… … …
……
…
…
… …
…
Ti: Template belonging to subject i
dij: sum of euclidean distances point to point between Ti and Tj
14. 14
Evaluation
Classificator k-nearest neighbors
• Subject 1
Subject 2
Subject 3
Template to assign
• k arbitrarily determined
• Euclidean distance calculated between and stored data points
• Majority ranking on Euclidean distance: the template is assigned to
the class with the majority among the k closest templates
Da mettere titoli piccoli
+
Aumenta lunghezza segnale
Fasi di implementazione
Guardare a cosa servono i coefficienti!!!
Da dire che #coefficient=127?Se coeff. Aumenta aumenta la latenza, ma aumenta la precisione del filtraggio.
Per risolvere il problema della latenza usiamo filtfilt?
Filtfilt applica segnale avanti e indietro così toglie la latenza
Tagliare a metà immagine
Peak detection noisy periodic-semiperiodic per tutti i segnali
Segmentation semplice funzione per selezionare i segmenti utili
Resample
Normalization
Peak detection noisy periodic-semiperiodic per tutti i segnali
Segmentation semplice funzione per selezionare i segmenti utili
Resample
Normalization
Matrice segmenti Per ogni soggetto
Aggiungere i vantaggi…Eventualmente dire che noi utlizziamo l’ordine di accuratezza 6?Quali sono i coefficienti?
3 matrixes righe=segmenti, colonne=campioni
Spiegare cos’è dij
Simmetria
Diagonale con zeri
Da aggiungere: piccola legenda
k value arbitrarily determined
Euclidean distance between data points referred to different periods
Distances ranking: the closest k number of stored templates with the smallest distance are ascertained
Among the k number of class labels ascertained, the one with the majority is determined. This determined stored template is assigned as the outcome of the unknown template
Determinare k perché la valutazione si fa con i ktemplate più vicini e quindi più simili al nostro template appena acquisito