Sigma tuning of gaussian kernels: detection of ischemia from magnetocardiograms

An apparatus for improving the detection of ischemia from magnetocardiographs, comprising computerized storage and processing capacity for:

• associating each attribute of a Gaussian or Radial Basis Function data kernel with its own Parzen window sigma, wherein said data kernel is derived from a magnetocardiograph;
• forming each of said sigmas into an initial vector σ comprising i scalars σi;
• starting with an initial guess σ0 and calculating an initial error metric E(σ0)=Q02 from a leave-one-out (or leave-several-out) K-PLS model, with scalar parameter λ=1;
• a) for each said scalar σi, calculating a corresponding element ΔE in E(σ) by perturbation;
• b) solving the equation (H+λI)Δσ=−∇E(σ0) for Δσ, where H is a Hessian matrix;
• c) if E(σ)=Q2 has become smaller by virtue of said solution b), updating said σ using said E(σ) and reducing said λ→βλ where β<1, otherwise making no change in said σ and increasing said λ→βλ where β>1;
• d) if said λ>1, capping λ to λ=1;
• e) iterating said a) through d) until either said E(σ) can no longer be improved or until a predetermined iteration limit has been reached;
• and
• f) replacing said initial vector σ with the updated σ resulting from said e).

A non-transitory computer-readable medium for improving the detection of ischemia from magnetocardiographs, comprising a set of instructions executable by a computerized system for:

• associating each attribute of a Gaussian or Radial Basis Function data kernel with its own Parzen window sigma, wherein said data kernel is derived from a magnetocardiograph;
• forming each of said sigmas into an initial vector σ comprising i scalars σi;
• starting with an initial guess σ0 and calculating an initial error metric E(σ0)=Q02 from a leave-one-out (or leave-several-out) K-PLS model, with scalar parameter λ=1;
• a) for each said scalar σi, calculating a corresponding element ΔE in E(σ) by perturbation;
• b) solving the equation (H+λI)Δσ=−∇E(σ0) for Δσ, where H is a Hessian matrix;
• c) if E(σ)=Q2 has become smaller by virtue of said solution b), updating said σ using said E(σ) and reducing said λ→βλ where β<1, otherwise making no change in said σ and increasing said λ→βλ where β>1;
• d) if said λ>1, capping λ to λ=1;
• e) iterating said a) through d) until either said E(σ) can no longer be improved or until a predetermined iteration limit has been reached;
• and
• f) replacing said initial vector σ with the updated σ resulting from said e).