Machine learning approaches for analyzing microbiome dynamics in ventilator-associated pneumonia

The National Institute of Allergy and Infectious Diseases awarded University of New Haven, Incorporated $126,351 on September 1, 2026, under the Allergy and Infectious Diseases Research program (CFDA 93.855) to develop machine learning approaches for analyzing microbiome dynamics in ventilator-associated pneumonia.

The project addresses VAP diagnostic delays and poor sensitivity in current clinical methods by developing interpretable machine learning models that integrate 16S rRNA sequencing data with clinical parameters to support VAP diagnosis. Aim 1 builds computational models using longitudinal microbiome data from 479 University of Pittsburgh Medical Center patients, with external validation on 83 University of Pennsylvania patients. Aim 2 employs genetic algorithms and state-space modeling to identify microbial taxa groups exhibiting collective dynamic trajectories associated with VAP development, potentially revealing early deviation points preceding clinical manifestation. The analytical framework incorporates Dirichlet mixture models, probabilistic graphical models with Markov blanket analysis, and novel genetic algorithm approaches for temporal pattern recognition. Expected outcomes include validated computational frameworks for temporal microbiome analysis and proof-of-concept evidence that coordinated microbial community shifts, integrated with clinical parameters, can identify VAP risk earlier and more accurately than existing methods.

Performance takes place in West Haven, Connecticut, with a period of performance from September 1, 2026, through August 31, 2030.

Abstract

Ventilator-associated pneumonia (VAP) affects up to 50% of mechanically ventilated patients, contributing to increased mortality, prolonged ICU stays, and healthcare costs exceeding $40,000 per case. Current diagnostic methods lack sensitivity and specificity, often requiring over 3 days for actionable results, leading to diagnostic uncertainty and delayed treatment decisions. This project leverages machine learning and longitudinal microbiome analysis to develop novel computational approaches for VAP assessment. We hypothesize that coordinated, collective patterns of microbial taxa dynamics, when integrated with clinical parameters, can identify VAP risk earlier and more accurately than existing methods. Rather than analyzing individual pathogens, we propose that microbial communities undergo distinct, coordinated shifts during mechanical ventilation that diverge between VAP and non-VAP patients. Aim 1 develops interpretable machine learning models integrating 16S rrna sequencing data with clinical parameters to support VAP diagnosis, using longitudinal data from 479 upmc patients with external validation on 83 upenn patients. Aim 2 employs genetic algorithms and state-space modeling to identify microbial taxa groups exhibiting collective dynamic trajectories associated with VAP development, potentially revealing early deviation points preceding clinical manifestation. The project utilizes established datasets with robust clinical validation across multiple VAP diagnostic criteria (CDC-nhsn, expert adjudication, treatment-based). Advanced computational methods include dirichlet mixture models, probabilistic graphical models with markov blanket analysis, and novel genetic algorithm approaches for temporal pattern recognition. Expected outcomes include validated computational frameworks for temporal microbiome analysis, proof-of-concept evidence for machine learning- supported VAP diagnostics, and identification of microbial biomarkers enabling predictive rather than reactive assessment. This work addresses critical clinical needs while providing authentic research experiences for undergraduate students in computational biology, bioinformatics, and translational medicine.