Performance estimation of diagnostic tests for cervical precancer based on fluorescence spectroscopy: effects of tissue type, sample size, population, and signal-to-noise ratio.

Fluorescence spectroscopy may provide a cost-effective tool to improve precancer detection. We describe a method to estimate the diagnostic performance of classifiers based on optical spectra, and to explore the sensitivity of these estimations to factors affecting spectrometer cost. Fluorescence spectra were obtained at three excitation wavelengths in 92 patients with an abnormal Papanicolaou smear and 51 patients with no history of an abnormal smear. Bayesian classification rules were developed and evaluated at multiple misclassification costs. We explored the sensitivity of classifier performance to variations in tissue type, sample size, tested population, signal to noise ratio (SNR), and number of excitation and emission wavelengths. Sensitivity and specificity could be evaluated within +/- 7%. Minimal decrease in diagnostic performance is observed as SNR is reduced to 15, the number of excitation-emission wavelength combinations is reduced to 15 or the number of excitation wavelengths is reduced to one. Diagnostic performance is compromised when ultraviolet excitation is not included. Significant spectrometer cost reduction is possible without compromising diagnostic ability. Decision-analytic methods can be used to rate designs based on incremental cost-effectiveness.