Raman Analysis of Saliva as Biomarker of COPD

Overview

Chronic Obstructive Pulmonary Disease (COPD) is a debilitating and chronic lung syndrome that causes accelerated lung function decline and death in the 20% of cases. Mostly, the non-adherence to therapy contributes to symptoms increase, mortality, inability and therapies failure, highly influencing the management costs associated to COPD. The existing procedure of diagnosing COPD is effective and fast. The acute treatment and the subsequent disease management, instead, strictly depend on the currently long and complex process of identification of three factors: COPD phenotype, adherence to chosen therapy and probability of exacerbation events. The knowledge of these factors is needed by clinicians to stratify patients and personalise the therapies and rehabilitation procedures, to initiate an effective disease management. The application of Raman spectroscopy on saliva, representing an easy collectable and highly informative biofluid, has been already proposed for different infective, neurological and cancer diseases, with promising results in the diagnostic and monitoring fields. In this project, we propose the use of Deep Learning analysis of Raman spectra collected from COPD patient's saliva to be combined with other clinical data for the development of a system able to provide fast and sensitive information regarding COPD phenotypes, adherence and exacerbation risks. This will support clinicians to personalise COPD therapies and treatments, and to monitor their effectiveness.

The main goal of the project is to create and validate a new method based on the Raman spectroscopy (RS) analysis of saliva for the optimised and personalised management of patients with Chronic Obstructive Pulmonary disease (COPD). The combination of the clinical instrumental data with the RS-approach will increase the quality of the clinical practice through appropriate stratification of patients, i.e., early identification of COPD phenotypes, consequent attribution of precise therapies, assessment of potential exacerbation risk and adherence to therapy. By the integration of instrumental and RS measures with Artificial Intelligence (AI), patients' COPD phenotype will be predicted allowing to efficiently direct the resources of the health-care system. The feasibility of the work is corroborated by the use of a sensitive, fast and miniaturized RS, used by non-specialized personnel and for the creation of a point-of care (POC) on an accessible biofluid. The multidisciplinary approach in pre-clinical, clinical and big data management fields is achieved through collaboration of academy, clinical research and industry. Starting from the unmet clinical need, CORSAI will build a close link between biomedical research, clinical research, data science towards the integration of PM into clinical practice and on ethical, legal, and social implications across the participating countries and beyond. The main objective is the collection of RS signals from the saliva of COPD patients, characterized for severity stages and phenotypes using GERA instruments, and corresponding CTRL and asthma patients (AsP). The creation and correlation of the dataset will lead to the accomplishment of specific objectives: I) Identification of the specific COPD, CTRL and AsP RF; II) Monitoring of therapy adherence through the drug signal in saliva; III) Definition of COPD phenotypes on the base of the RF correlated with instrumental GERA data; IV) Monitoring of the rehabilitation procedures and effects; V) Association of a high exacerbation risk to specific COPD patients; VI) Creation of a classification model from the RS database; VII) Application of high-performance computing for data analysis; VIII) Integration of the portable RS as POC. The novelty of CORSAI relies in the advanced methodology, brought to the bed side thanks to portable instruments. The minimal invasive procedure used for the saliva collection and the velocity for the Raman acquisition represent relevant advantages allowing the continuous monitoring of patients' adherence to therapy, and the contemporary discrimination of COPD phenotypes with high rate of exacerbation. The feasibility of the project is directly related to the biological sample and proposed technology, already tested in the clinical setting19: i)easy collection and storage of saliva fits the clinical scenario; ii) minimal sample preparation and portable device enable POC use by non-specialized personnel, with AI remote decision guidance. SAMPLE COLLECTION: Saliva collection from all the selected subjects will be performed following the Salivette (SARSTEDT) manufacturer's instructions. To limit variability in salivary content not related to COPD, saliva will be obtained from all subjects at a fixed time, after an appropriate lag time from feeding and teeth brushing. Pre-analytical parameters (i.e. storage temperature and time between collection and processing), dietary and smoking habit will be properly recorded. Briefly, the swab will be removed, placed in the mouth and chewed for 60 seconds to stimulate salivation. Then the swab will be centrifuged for 2 minutes at 1,000 g to remove cells fragments and food debris. Collected samples will be stored at -80° C. SAMPLE PROCESSING: For the Raman analysis, a drop of each sample will be casted on an aluminium foil in order to achieve the Surface Enhanced Raman Scattering (SERS). DATA COLLECTION: SERS spectra will be acquired using an Aramis Raman microscope (Horiba Jobin-Yvon, France) equipped with a laser light source operating at 785 nm with laser power ranging from 25-100% (Max power 512 mW). Acquisition time between 10-30 seconds will be used. The instrument will be calibrated before each analysis using the reference band of silicon at 520.7 cm-1. Raman spectra will be collected from 35 points following a line-map from the edge to the centre of the drop. Spectra will be acquired in the region between 400 and 1600 cm-1 using a 50x objective (Olympus, Japan). Spectra resolution is about 1.2 cm-1. The software package LabSpec 6 (Horiba Jobin-Yvon, France) will be used for map design and the acquisition of spectra. DATA PROCESSING: All the acquired spectra will be fit with a fourth-degree polynomial baseline and normalized by unit vector using the dedicated software LabSpec 6. The contribution of the substrate will be removed from each spectra. The statistical analysis to validate the method, will be performed using a multivariate analysis approach. Principal Component analysis (PCA) will be performed in order to reduce data dimensions and to evidence major trends. The first 20 resultant Principal Components (PCs) will be used in a classification model, Linear Discriminant Analysis (LDA), to discriminate the data maximizing the variance between the selected groups. The smallest number of PCs will be selected to prevent data overfitting. Leave-one-out cross-validation and confusion matrix test will be used to evaluate the method sensitivity, precision and accuracy of the LDA model. Mann-Whitney will be performed on PCs scores to verify the differences statistically relevant between the analysed groups. Correlation and partial correlation analysis will be performed using the Spearman's test, assuming as valid correlation only the coefficients with a p-value lower than 0.05. The statistical analysis will be performed using Origin2018 (OriginLab, USA). DEEP LEARNING: The datasets will be analysed and processed using Deep Learning models with the aim to discover significant patterns that can be used to confirm and analyse trends and to develop predictions and decision support about the COPD stratification. Techniques of data augmentation and automatic hyperparameter optimization will be developed in order to enhance classification performances and improve generalization ability. In order to reach a tradeoff between predictive accuracy and interpretability, a class activation mapping (CAM)-based approach will be applied to visualize the active variables in the spectra in order to identify discriminative pattern to extract the most informative spectral features. UNIMIB and GERA will implement an explanation mechanism to identify the active variables in whole spectrum and interpret the internal feature representations and data transformation pipeline of the CNN model. UNIMIB and GERA will integrate the various computational modules in a modular computational pipeline for patient-wise classification.