Premature Ventricular Contractions’ Detection Based on Active Learning

Table 3

Overall algorithm flow of PVC detection.

---

Algorithm ALPVCsD

---

Input: Unlabeled data set Z = (x_i, y_i), i = 1, …, N,

   Number of initial selected samples: K,

   Number of samples selected during iteration: i.

Output: Labeled data set L = (a_j, b_j), j = 1, …, M (0 < M < N),

   Trained classifier model: random forest

Initialize: centroids = [one random selected points]

   Z_len ⟵ length (Z)

   L_len ⟵ length (L)

   Feature extraction from data by self convolution encoder

For j from 1 to K − 1

  Distance ⟵ { }

  For t from 1 to length (centroids)

   compute distance of “point” in Z to centroids and store the point P with minimum distance

  End centroids = centroids + P

End

L ⟵ L ∪ centroids (annotated the K data by Oracal)

Z ⟵ Z\centroids

Z_len ⟵ Z_len – K

L_len ⟵ L_len + K

Repeat:

  Training random forest with centroids.

  If the number of unlabeled data <1/2 N:

   Break

 For t in Z

   calculate the minimum confidence using equation (3) and store it in Q

  End

  P_data ⟵ Sort Q and select the smallest 50 data

  L ⟵ L ∪ P_data

  Z ⟵ Z\ P_data train random forest use the 50 labeled data

  Z_len ⟵ Z_len – 50

  L_len ⟵ L_len + 50

End

---