DEEP-GRAM: A Deep Learning Model for Gram Stain Species Prediction in Bloodstream Infections
CWSF · 2026 Disease & Illness Silver Medal
Overview
Bloodstream infections are life-threatening and frequently bacterial, requiring rapid initiation of species-specific antibiotic treatment. Gram staining provides initial bacterial characterization but cannot identify species. Definitive identification requires specialized testing that involves substantial resources, technical expertise, and can take two to three days. This delay can result in under-treatment or inappropriate broad-spectrum antibiotic use. To address this, I hypothesized that deep learning could automate Gram stain interpretation and extract species-specific morphological patterns imperceptible to the human eye. I generated an image dataset of Gram-stained blood cultures representing thirteen bacterial species. Then, I developed a multi-stage AI pipeline called DEEP-GRAM, comprising models for bacterial cell detection and isolation, Gram classification, morphology classification, species identification, and slide-level species prediction. The slide-level model achieved 92% accuracy for Gram type and morphology classification and 85% top-choice species accuracy. DEEP-GRAM demonstrates the potential for AI-assisted pathogen identification, reducing time to targeted treatment and improving patients outcomes.
Awards (2)
- Silver Medal
- Selected for CWSF 2026
Competition history
- CWSF 2026
Related projects
CWSF · 2026
Computer Vision For Bacterial Gram Stain Classification
CWSF · 2026
Pre-Symptomatic Sepsis Detection Using a Wearable Biosensing System and Deep Learning Framework
ISEF · 2022
A Novel Approach of Deep Learning on Detection and Classification of Leukemic Cells and BCR-ABL1 Gene
ISEF · 2020
Improving Hazard Characterization in Bacterial Pathogens: Predicting Efficiency of Antibiotics Using Machine Learning
Closest projects by meaning, across every fair and year in the corpus.