Application of machine learning algorithms for the classification of organs in premature infants using broadband optical spectroscopy

Spectroscopy allows for the collection of optical data from the skin's underlying organs. Data were collected and analyzed from 26 premature infants using broadband optical spectrometry (BOS) within a 350-2500 nm wavelength range using a handheld probe in contact with the skin. Patients varied in physical characteristics (such as weight and skin tone), and scans were taken across multiple days allowing for different subject physical conditions (such as illness, hydration, feed regimen, etc). Statistical analysis and deep learning were leveraged to provide proof-of-concept that optical data is sufficient to distinguish the abdomen from the thigh, indicating that intestinal tissue can be detected, and potential for ischemic disease prediction in future study. We utilized feature-based modeling using principal component analysis (PCA) that discovered a panel of markers from spectra with high univariate AUC and low feature correlation. Our model proposed five features that distinguish abdomen from thigh with an AUC of 0.89 using unsupervised PCA and an AUC of 0.92 using supervised linear discriminant analysis (LDA). Neural network (NN) modeling of a signal wavelength, a panel of 12 selected wavelengths, and a whole spectrum yielded respective accuracies of 62%, 92%, and 95% for spectra-wise, and 65%, 100%, and 100% for subject-wise classification for the validation data set. For the test data set, accuracies of 68%, 84%, and 85% in spectra-wise and 83%, 92%, and 92% in subject wise analysis were achieved. We conclude that analysis of human tissue spectra is sufficient to permit noninvasive characterization of specific underlying organs.