Fourier transform infrared analysis (FTIR) was used in combination with partial least squares regression (PLS) to predict the concentration of acetone in milk. FTIR spectra were compared with results of a gas-chromatographic head space method. Principal component analysis of whole spectra (3000 to 1000 cm(-1)) suggested to reduce the spectrum of analysis for acetone to 1450 to 1200 cm(-1). A second derivative was applied to the spectra to remove baseline effects and further enhance the spectral features. Full cross-validation was used to compare the reference with predicted acetone concentrations of samples not included in model development. PLS applied to the full spectral range resulted in a complex 19-factor model with a cross-validation error of 0.22 mM. After reducing the spectrum and taking the second derivative, we obtained a model with seven factors that yielded a cross-validation error of 0.21 mM. This compares favorably with a previously reported model with 20 factors and an error of 0.25 mM. Using PLS predictions to identify cows with subclinical ketosis resulted in 95 to 100% sensitivity and 96 to 100% specificity when the threshold for subclinical ketosis was 0.4 to 1.0 mM. The corresponding positive predictive values were > or = 76% and the negative predictive values > 98% throughout an assumed range of subclinical ketosis prevalence of 10 to 30%.