3. Visual Healthcare Analytics

Question 1

what is parallel coordinate

Answer 1

• visual of multidim data
• given NxM table with N patients and M clinical var, a line chart is generated by displaying M equally spaced vertical axes with individual ranges
• can then filter and explore correlations between vars for patients

Question 2

draw a parallel coordinate graph

Question 3

what is a chord visualisation plot

Answer 3

a diagram illustrating the connection between different variables

Question 4

draw a chord visualisation plot

Question 5

what are 2 dim reduction algos

Answer 5

• pca
• t-SNE

Question 6

what is t-SNE

Answer 6

an algorithm that calculates similarity via the high dim and low dim space. it computes the distance between instances in both spaces and tries to optimise these similarity measures using a cost function

Question 7

what is pca

Answer 7

• unsupervised linear dimensionality reduction
• visualisation technique

Question 8

how is pca calculated at a high level (variance)

Answer 8

the greatest variance by some scalar projection becomes the first PC (coordinate), and the second greatest becomes the second PC etc etc.

Question 9

how can we determine the amount of preserved information after pca is applied

Answer 9

variance - compare that of all original dimensions, with the variance of the reduced dimensions

Question 10

after performing pca, two variables have a variance of 1.46 (A), and 0.2 (B). the entire dataset itself has a variance of 2.06, what can be inferred by this

Answer 10

we can infer that variable A alone can explain most of the information of the output predicted by the two variables

Question 11

how is pca calculated

Question 12

how are the number of pca dimensions measured

Answer 12

• consider using a scree plot. it shows the variance explained by each PC, based on the number of PCs used

Question 13

draw a scree plot

Question 14

when should we not use PCA

Answer 14

pca is linear because when calculated, it is projected as a linear vector [finish this]

Question 15

what to use instead of pca

Answer 15

if it’s not a linear transformation, consider distributed stochastic neighbourhood embedding (t-SNE)

Question 16

how is t-sne calculated

Answer 16

1. for every point in 2D space, centre a Gaussian distribution over that point (this denotes the similarity by probability, of the two points)
2. repeat step 1, this time with a Cauchy/student t-distbn (these are the set of prob. for the low dim space)
3. now map the high dim and low dim using a loss function to minimise the distance using Kullback-Liebler divergence (KL)
4. finally, use gradient descent to minimise the KL cost function

Question 17

what is the advantage of the student t over Gaussian

Answer 17

student-t has heavier tails than normal, so it can account for less similar points (further distances)

Question 18

compare t-SNE & PCA

Answer 18

• linearity
• t-SNE doesn’t keep the original data structure while PCA does
• t-SNE produces well separated clusters
• t is computationally expensive for large datasets
• PCA is consistent, t-SNE is stochastic and can produce varying embeddings
• PCA requires no hyperparam optimisation