Calibrated analyte concentration measurements in mixtures

A method for estimating a target analyte concentration from a test spectrum collected from a sample, the method comprising:

• accessing a first spectral data set that includes a plurality of spectra, the first spectral data set including a training data set of training spectra and a test data set of test spectra;
• accessing independently measured target analyte concentrations for the training spectra;
• for different values of a scalar:
• generating a modified first spectral data set by adding the spectrum of the target analyte multiplied by the scalar and further multiplied by a variable which is random with respect to each sample to the first spectral data set;
• generating modified measured target analyte concentrations by adding the identical random variable to the independently measured target analyte concentrations for the training spectra;
• generating, by a computer, a model that predicts target analyte concentration based on a test spectrum, the model generated based at least in part on the modified first spectral data set and on the modified measured target analyte concentrations;
• cross-validating the generated model;
• selecting the model for the value of the scalar that best optimizes the cross-validated predictions;
• and
• applying the selected model to the test spectrum to estimate the target analyte concentration.

A method for estimating a target analyte concentration from a test spectrum collected from a sample, the method comprising:

• accessing a first spectral data set that includes a plurality of spectra collected from different samples, the first spectral data set including a training data set of training spectra and a test data set of test spectra;
• estimating target analyte concentrations for the spectra in the first spectral data set;
• estimating sample-to-sample variations other than variations of the target analyte concentration, for the samples represented in the first spectral data set, wherein estimating sample-to-sample variations comprises:
• removing contributions of a component to the first spectral data set based on projecting out independently measured spectral features of the component, thereby producing a second spectral data set;
• and
• estimating sample-to-sample variations in the second spectral data set;
• generating, by a computer, a model that predicts target analyte concentration based at least in part on a test spectrum and the test spectrum's sample, the model generated based at least in part on the estimated target analyte concentrations, on the estimated residual sample-to-sample variations estimated from the generated second spectral data set, and on independently measured target analyte concentrations for the training spectra;
• and
• applying the model to the test spectra to estimate the target analyte concentrations.