Abstract
Understanding how intestinal microbes contribute to and identifying risk factors predictive of clinical outcomes is necessary to improve patient care. However, human-microbe interactions are inherently complex, with numerous host and microbial variables that can be difficult to disentangle using traditional approaches. To address this complexity, machine learning, which adapts to and infers from data without having explicit human instructions on how to do so, offers a powerful set of tools capable of analyzing large, multidimensional datasets and uncovering patterns and relationships that might otherwise remain obscured. Machine learning algorithms vary in sophistication and interpretability with more simple and easily interpretable models like linear regression, random forest which can capture more complexity but is harder to understand the biological implications of, and sophisticated but “black box” models like neural networks. The objective of this work is to further the understanding of enteric microbes in two clinical settings using machine learning.
We first sought to validate available tools for predicting recurrent Clostridioides difficile infection (CDI). C. difficile is an anaerobic bacterium which can cause diarrheal disease, especially in patients exposed to antibiotics. While there are treatments for CDI, there is an approximately 20% recurrence rate, with risk increasing for each subsequent recurrence. Treatment and prevention strategies for recurrence are often invasive or cost prohibitive. A clinical tool to identify patients at risk of recurrence would greatly aid in clinicians’ abilities to identify patients most in need of intervention. In this work, recurrence risk scores (by Larrainzar-Coghen, Reveles, D’Agostino, Cobo, and Eyre et al.), consensus guidelines risk criteria, the leading severity score (ATLAS), and PCR cycle threshold (as marker of fecal organism burden) were validated using electronic medical records from hospitalized adult patients with CDI at University of Virginia Medical Center. All tools performed poorly at predicting recurrent C. difficile infection. The models were especially poor at predicting recurrence in patients with a prior history of infection. These findings suggest that the current clinical markers and typical models used are inadequate for successful prediction of recurrent CDI. Novel biomarkers and higher-dimensional models are needed to better capture risk factors and provide crucial information for clinicians.
Next, we investigated the relationship between the first-trimester microbiome, inflammation, and small vulnerable newborns (SVN). It was previously known that children who are born preterm, low birth weight, or small for gestational age are at higher risks for neonatal death, developmental delays, and non-infectious and infectious disease later in life. This is especially concerning in low- and middle-income countries which have higher rates of adverse outcomes. To investigate the potential role of the first-trimester maternal gut microbiome in SVN, 213 pregnant Bangladeshi women were enrolled during the first trimester of pregnancy. Blood and stool samples were collected at enrollment and birth outcomes within 72 hours of delivery. Using random forest machine learning models, the ability of first-trimester maternal microbiome to predict SVN was evaluated. SVN could be discriminated from non-SVN (AUC = 0.653, 95% CI 0.518-0.789) supporting that the microbiome may influence gestation and fetal development. The microbiome also predicted the inflammatory markers plasma alpha-1-acid glycoprotein and fecal calprotectin (AUC = 0.742, 95% CI: 0.616-0.868 and AUC = 0.793, 95% CI: 0.680-0.905, respectively) and α-1-acid glycoprotein was significantly associated with SVN (logistic regression: aOR = 1.98; 95% CI: 1.04-3.85). These findings support that the first-trimester maternal gut microbiome may contribute to SVN via induction of chronic systemic inflammation.
This work explores both the difficulty and the utility of machine learning in biological sciences. While the efficacy of machine learning depends highly on the data available and the design of the model, it can be a powerful tool in identifying underlying contributors to pathology and predicting disease outcomes.
