Infrared Spectroscopy of Urine for the Non-Invasive Detection of Endometrial Cancer

Abstract

Current triage for women with post-menopausal bleeding (PMB) to diagnose endometrial cancer rely on specialist referral for intimate tests to sequentially image, visualise and sample the endometrium. A point-of-care non-invasive triage tool with an instant readout could provide immediate reassurance for low-risk symptomatic women, whilst fast-tracking high-risk women for urgent intrauterine investigations. This study assessed the potential for infrared (IR) spectroscopy and attenuated total reflection (ATR) technology coupled with chemometric analysis of the resulting spectra for endometrial cancer detection in urine samples. Standardised urine collection and processing protocols were developed to ensure spectroscopic differences between cases and controls reflected cancer status. Urine spectroscopy distinguished endometrial cancer (n = 109) from benign gynaecological conditions (n = 110) with a sensitivity of 98% and specificity of 97%. If confirmed in subsequent low prevalence studies embedded in PMB clinics, this novel endometrial cancer detection tool could transform clinical practice by accurately selecting women with malignant pathology for urgent diagnostic work up whilst safely reassuring those without.

Keywords: ATR; FTIR; PLS-DA; cancer; endometrial cancer detection; machine learning; vibrational spectroscopy.

Figure 1

PCA-K-means of the spectra within the fingerprint region. Four clusters are identified by using the elbow method and corroborated the gap statistic. The presence of clusters allows to determine whether the data actually have any possible segregated data.

Figure 2

Discriminant analysis plots of all class comparisons. (A) Controls vs. all cancers. (B) Controls vs. endometrioid cancers (EC). (C) Controls vs. non-endometrioid cancers (NEC). (D) Endometrioid cancers vs. non-endometrioid cancers. (E) Controls vs. stage I cancers. (F) Controls vs. stage IA grade 1 cancers. After the application of PLS-DA it was possible to discern between classes, obtaining a p < 0.0001 shows statistical significance on the discrimination analysis.

Figure 3

Receiver operating characteristic (ROC) curves after the application of OPLS-PLS-DA. Figures (A–F) illustrate the ROC curve of OPLS-PLS-DA processed data (red curves) and unprocessed data (blue curves). These figures demonstrate how the discrimination analysis would behave if unprocessed with OPLS; the models are suboptimal in their discriminatory ability when OPLS is not applied as pre-processing method. The subplots from this figure show the following comparisons: (A) Controls (n = 110) vs. all cancers (n = 109). (B) Controls (n = 110) vs. endometrioid cancers (n = 57). (C) Controls (n = 110) vs. non-endometrioid cancers (n = 52). (D) Endometrioid cancers (n = 57) vs. non-endometrioid cancers (n = 52). (E) Controls (n = 110) vs. stage I cancers (n = 75). (F) Controls (n = 110) vs. stage IA grade 1 cancers (n = 22).

Figure 4

Bar plots from the spectral distribution and the mean absorbance, of the most discriminatory peaks found on (A) Controls vs. all cancers, (B) Controls vs. endometrioid cancers (EC), (C) Controls vs. non-endometrioid cancers (NEC), (D) Non-endometrioid cancers vs. Endometrioid cancers, (E) Controls vs. Stage 1 cancers, and (F) Control vs. Stage 1 grade 1A cancer classes.

Figure 5

Average spectra of the studied groups. The figure depicts the fingerprint region (1800–500 cm⁻¹) and the prospective spectral biomarks who are responsible for the discrimination between the analysed groups. The shaded regions indicate where the prospective biomarks are found, marking potential regions for endometrial cancer diagnostic.