Using the Expectation Maximization Algorithm with Heterogeneous Mixture Components for the Analysis of Spectrometry Data

TL;DR

This paper introduces advanced EM-based methods for processing MCC/IMS spectrometry data, improving denoising, baseline correction, and peak clustering by handling heterogeneous mixture components, thus enhancing automated analysis in medical gas analysis applications.

Contribution

It proposes novel EM algorithms tailored for heterogeneous mixture components in MCC/IMS data, advancing automated peak analysis without manual intervention.

Findings

Methods outperform existing approaches in evaluation

Provides Python software for all three methods

Enhances automated, accurate MCC/IMS data analysis

Abstract

Coupling a multi-capillary column (MCC) with an ion mobility (IM) spectrometer (IMS) opened a multitude of new application areas for gas analysis, especially in a medical context, as volatile organic compounds (VOCs) in exhaled breath can hint at a person's state of health. To obtain a potential diagnosis from a raw MCC/IMS measurement, several computational steps are necessary, which so far have required manual interaction, e.g., human evaluation of discovered peaks. We have recently proposed an automated pipeline for this task that does not require human intervention during the analysis. Nevertheless, there is a need for improved methods for each computational step. In comparison to gas chromatography / mass spectrometry (GC/MS) data, MCC/IMS data is easier and less expensive to obtain, but peaks are more diffuse and there is a higher noise level. MCC/IMS measurements can be described as samples of mixture models (i.e., of convex combinations) of two-dimensional probability distributions. So we use the expectation-maximization (EM) algorithm to deconvolute mixtures in order to develop methods that improve data processing in three computational steps: denoising, baseline correction and peak clustering. A common theme of these methods is that mixture components within one model are not homogeneous (e.g., all Gaussian), but of different types. Evaluation shows that the novel methods outperform the existing ones. We provide Python software implementing all three methods and make our evaluation data available at http://www.rahmannlab.de/research/ims.