Towards quantifying biomarkers for respiratory distress in preterm infants: Machine learning on mid infrared spectroscopy of lipid mixtures

• Published in: Talanta (Oxford) 2024-08, Vol.275, p.126062-126062, Article 126062
• Main Authors: Ahmed, Waseem, Veluthandath, Aneesh Vincent, Madsen, Jens, Clark, Howard W., Dushianthan, Ahilanandan, Postle, Anthony D., Wilkinson, James S., Senthil Murugan, Ganapathy
• Format: Article
• Language: English
• Subjects: ATR-FTIR; Biomarkers - analysis; Humans; Infant, Newborn; Infant, Premature; Learning; Lecithins - analysis; Lecithins - chemistry; Lipid; Lipids - analysis; Lipids - chemistry; Machine; Machine Learning; nRDS; PLSR; Pulmonary Surfactants - analysis; Pulmonary Surfactants - chemistry; Respiratory Distress Syndrome, Newborn - diagnosis; SHAP values; Spectrophotometry, Infrared - methods; Sphingomyelins - analysis
• ISSN: 0039-9140; 1873-3573
• DOI: 10.1016/j.talanta.2024.126062

Neonatal respiratory distress syndrome (nRDS) is a challenging condition to diagnose which can lead to delays in receiving appropriate treatment. Mid infrared (IR) spectroscopy is capable of measuring the concentrations of two diagnostic nRDS biomarkers, lecithin (L) and sphingomyelin (S) with the potential for point of care (POC) diagnosis and monitoring. The effects of varying other lipid species present in lung surfactant on the mid IR spectra used to train machine learning models are explored. This study presents a lung lipid model of five lipids present in lung surfactant and varies each in a systematic approach to evaluate the ability of machine learning models to predict the lipid concentrations, the L/S ratio and to quantify the uncertainty in the predictions using the jackknife + -after-bootstrap and variant bootstrap methods. We establish the L/S ratio can be determined with an uncertainty of approximately ±0.3 mol/mol and we further identify the 5 most prominent wavenumbers associated with each machine learning model. [Display omitted] •Comprehensive calibration of PLSR models using physiological concentrations of lung surfactant lipids.•Prediction intervals for quantified uncertainty in PLSR models.•Use of SHAP values to explain strength of AI model features with a view to optimising a spectroscopic point of care platform.