Explanatory analysis of spectroscopic data using
machine learning of simple, interpretable rules
Vibrational Spectroscopy, 32 (1):
33--45 (5 August 2003)A collection of Papers Presented at Shedding New Light
on Disease: Optical Diagnostics for the New Millennium
(SPEC 2002) Reims, France 23-27 June 2002.DOI: doi:10.1016/S0924-2031(03)00045-6
Abstract
Whole organism or tissue profiling by vibrational
spectroscopy produces vast amounts of seemingly
unintelligible data. However, the characterisation of
the biological system under scrutiny is generally
possible only in combination with modern supervised
machine learning techniques, such as artificial neural
networks (ANNs). Nevertheless, the interpretation of
the calibration models from ANNs is often very
difficult, and the information in terms of which
vibrational modes in the infrared or Raman spectra are
important is not readily available. ANNs are often
perceived as 'black box' approaches to modelling
spectra, and to allow the deconvolution of complex
hyperspectral data it is necessary to develop a system
that itself produces 'rules' that are readily
comprehensible. Evolutionary computation, and in
particular genetic programming (GP), is an ideal method
to achieve this. An example of how GP can be used for
Fourier transform infrared (FT-IR) image analysis is
presented, and is compared with images produced by
principal components analysis (PCA), discriminant
function analysis (DFA) and partial least squares (PLS)
regression.