4. Clinical Prediction Models

Question 1

what is a perceptron

Question 2

what is an svm

Answer 2

f is a sigmoid function and output is consistent with binary labels

Question 3

what is a multi-layer perceptron

Answer 3

f is a sigmoid function or other non-linear function, that can form a stack of many layers

Question 4

what is deep learning

Question 5

what are some famous NN

Answer 5

googleNet
alexNet

Question 6

what should be considered for knn

Answer 6

• value of k (number of nearest neighbours to retrieve), rule is that k = sqrt(number training points)
• what distance metric to use
• computation (data size, dim size)

Question 7

what happens if k is too small/ too large

Answer 7

• too small = not general, sensitive to noise
• too large = may include points from other classes

Question 8

what are some distance measures for knn

Answer 8

• euclidean = straight line distance between points
• manhattan = measures distance if a grid like path is followed
• cosine = angle between two vector representations of data points
• hamming = the number of points that differ between two collections of values

Question 9

draw cosine distance

Question 10

draw manhattan distance

Question 11

draw euclidean distance

Question 12

how do we deal with the computational complexity of knn

Answer 12

• compute only an approximate distance
• remove redundant data
• pre-sort (index)

Question 13

what are some evaluation metrics

Answer 13

confusion matrix:
- tp = positive values correctly identified
- fp = negative values identified as positive

recall = ratio of actual positives to those correctly identified

precision = ratio true positives actually classified as positives

f = harmonic mean of precision & recall