MALDI-TOF mass spectrometry coupled with machine learning: an accurate tool to detect toxigenic Clostridioides difficile strains

Abstract

Purpose

Clostridioides difficile infection (CDI) is one of the most common healthcare-associated infections, causing significant morbidity and mortality. A CDI diagnosis involves a clinical assessment and laboratory testing. This study aimed to evaluate the effectiveness of MALDI-TOF mass spectrometry in identifying toxigenic C. difficile strains.

Methods

From May 2019 to March 2024, all clinical samples received for C. difficile and toxins testing by PCR were also tested by culture under anaerobic conditions. At least one MALDI-TOF spectrum was used to identify all the C. difficile strains that were isolated. Whole genome sequencing was then performed to characterise the C. difficile isolates by searching for toxin genes. Machine learning and deep learning algorithms were used to distinguish between toxigenic and non-toxigenic C. difficile strains.

Results

A total of 389 C. difficile strains, isolated from the stools of 315 patients, were analyzed. Of these, 249 were toxigenic (225 ToxA + B + CDT−, 22 ToxA + B + CDT+, and 2 ToxA−B + CDT+), and 140 were non-toxigenic (ToxA−B−CDT−). The support vector machine algorithm produced the most accurate results (91.5 % accuracy and a κ value of 0.831) with high sensitivity (93.1 %) and specificity (90 %), as well as high positive and negative predictive values (PPV of 90.3 % and NPV of 92.3 %, respectively). The random forest algorithm demonstrated accuracies as high as 87.7 %.

Conclusions

Supervised machine learning paves the way for inexpensive and easy-to-use alternative methods that enable the use of MALDI-TOF MS to recognise toxigenic and non-toxigenic C. difficile strains.