An improved machine learning pipeline for urinary volatiles disease detection: Diagnosing diabetes

Motivation The measurement of disease biomarkers in easily–obtained bodily fluids has opened the door to a new type of non–invasive medical diagnostics. New technologies are being developed and fine–tuned in order to make this possibility a reality. One such technology is Field Asymmetric Ion Mobility Spectrometry (FAIMS), which allows the measurement of volatile organic compounds (VOCs) in biological samples such as urine. These VOCs are known to contain a range of information on the relevant person’s metabolism and can in principle be used for disease diagnostic purposes. Key to the effective use of such data are well–developed data processing pipelines, which are necessary to extract the most useful data from the complex underlying biological structure. Results In this study, we present a new data analysis pipeline for FAIMS data, and demonstrate a number of improvements over previously used methods. We evaluate the effect of a series of candidate operational steps during data processing, such as the use of wavelet transforms, principal component analysis (PCA), and classifier ensembles. We also demonstrate the use of FAIMS data in our pipeline to diagnose diabetes on the basis of a simple urine sample using machine learning classifiers. We present results for data generated from a case-control study of 115 urine samples, collected from 72 type II diabetic patients, with 43 healthy volunteers as negative controls. The resulting pipeline combines the steps that resulted in the best classification model performance. These include the use of a two–dimensional discrete wavelet transform, and the Wilcoxon rank–sum test for feature selection. We are able to achieve a best ROC curve AUC of 0.825 (0.747–0.9, 95% CI) for classification of diabetes vs control. We also note that this result is robust to changes in the data pipeline and different analysis runs, with AUC > 0.80 achieved in a range of cases. This is a substantial improvement in performance over previously used data processing methods in this area. Our ability to make strong statements about FAIMS ability to diagnose diabetes is sadly limited, as we found confounding effects from the demographics when including these data in the pipeline. The demographics alone produced a best AUC of 0.87 (0.795–0.94, 95% CI). While the combination of the demographics and FAIMS data resulted in an improvement on the AUC (0.907; 0.848–0.97, 95% CI), it did not prove to be a significant difference. Nevertheless, the pipeline itself shows a significant improvement in performance over more basic methods which have been used with FAIMS data in the past.