3.4.3. Statistical analysis

Partial Least Squares (PLS) regression models were generated using the ParLeS software [90]. This method has been used to acquire predictive cremation temperature models from the FT-MIR spectral intensities in the range 1800–400 cm^-1 (independent variables, 250 data points). The spectral pre-processing treatments consisted of a light scatter and baseline correction by Standard Noise Variate (SNV), de-noising through a median filter, and mean centering [58]. This pre-processing has been widely used [44, 58, 91, 92] to obtain forecasting models of different environmental and physical variables, using different analytical techniques (FT-IR, Py-CSIA, Py-GC/MS and FTICR/MS) on different natural matrices (e.g., soil, sediment, foodstuff). The protocol of this pre-processing is well exposed in [58]. The Root Mean Squared Error (RMSE) and the Akaike’s Information Criterion (AIC) were used to prevent overfitting and determine the best number of factors (latent variables) for each model. Lastly, the diagnostic spectral regions of the FT-MIR spectra were studied by the combination plot of the Variable Importance for Projection (VIP) values in the spectral range under study. The VIP traces can be useful to identify the independent variables (spectral peaks), that may be linked to the temperature of the cremation. Spurious forecast models due to overfitting were discarded after comparing PLS models calculated with the randomised combustion temperature (cross-validation). The cross-validation model was previously used by [44, 58].