Table 2.

The detailed algorithm for PUNCH.

Given a sample of n individuals with their corresponding score sets X ≡ {x¹, x²,…, xⁿ}, each having m different scores, x ≡ {x₁,x₂,…, x_m}, their diagnosis labels y ∈ {0,1}, and noise variances

σ_{j}^{2}

of scores,

1. For each score s_j

Sample noise n_j from $N (0, σ_{j}^{2})$
Calculate ${\hat{x}}_{j}$ , by ${\hat{x}}_{j} = x_{j} + n_{j}$
Calculate w_j, by using ${\hat{x}}_{j}$ (Equation 4)
Calculate $p (y = 1 | {\hat{x}}_{j}, s_{j})$ (Equations 2 and 3)

2. Calculate probabilities p(s_j) by normalizing weights w_j

3. Fuse individual severity beliefs to get

p (y = 1 | \hat{x})

(Equation 1)

4. Repeat 1–3 to get a distribution over

p (y = 1 | \hat{x})

5. Calculate expected value

E [p (y = 1 | \hat{x})]

by averaging