E-Nose Urine Analysis for Non-Invasive Diagnosis of Prostate Cancer: Focus on Data Processing for Drift Compensation

Even though the prostate cancer (KP) represents the fifth most common cancer in men worldwide, the current diagnostic protocol is affected by poor accuracy, which results in the un-detection of cancers and a high false positive rate causing patients’ overtreatment and a huge impact on healthcare spending.In recent years, many innovative approaches, proposing the characterization of biological fluids, especially urine, have been proposed in the scientific literature to improve current procedures. Among them, the most promising results have been obtained by means of the analysis of odours emanated from urine samples by electronic noses (EN). Those studies proved the existence of a specific correlation between urine odour alteration and KP presence that can be exploited to develop a novel, non-invasive and accurate diagnostic tool for the KP.However, no study has already led to the introduction of EN into the clinical practice, despite the high diagnostic accuracies reported (i.e., close or above 90%), which are considerably higher than the 58% achieved by current protocols. One of the main limitations is represented by the fact that no study addresses the problem of sensor drift. Drift, consisting in the deviation over time of sensor response under the same conditions, currently represents the primary obstacle to the scaling up of EN from research objects to effective diagnostic devices for large-scale use because of the progressive worsening over time of the classification performance and the consequent need for periodical recalibration of the system.For this reason, the present study, conceived within a research project proposing an EN for the non-invasive detection of KP, focuses on the description of the data processing protocol specifically developed for compensating drift. In particular, we here describe the adoption of the Orthogonal Signal Correction (OSC) algorithm to implement a drift correction model for a dataset acquired over 9 months, comprising urine headspaces from 122 subjects (81 KP patients and 41 control donors). The model proved effective in mitigating drift effects on 1-year old sensors by restoring their diagnostic accuracy from 55% up to about 80%, which was achieved by new sensors not subjected to drift issues. The developed model was validated by means of double-blind verification tests. The KP diagnosis model built on 1-year-old sensors achieved on double-blind samples an accuracy, sensitivity and specificity of 77% (CI 95% 60.7 - 88.9), 82% (CI 95% 59.7 - 94.8) and 71% (CI 95% 44.0 - 89.7), respectively, thereby confirming results achieved on training data.