O2085: Rapid Identification of Pediatric Sepsis Using AI Detection Technology Based on Raman Spectroscopy

Background: Sepsis, a life-threatening systemic inflammatory response to pathogenic infections, demonstrates time-dependent mortality escalation in pediatric populations. Rapid and accurate diagnosis of pediatric sepsis is critical for reducing its associated high mortality. The Pediatric Sequential Organ Failure Assessment (pSOFA) score is widely employed for sepsis diagnosis in pediatric populations. However, its failure to incorporate infection-specific biomarkers necessitates supplementary time-consuming analyses, contributing to diagnostic delays and underrecognized cases. This study proposes an artificial intelligence (AI)-based Raman spectroscopy technique for the rapid diagnosis of pediatric sepsis.

Methods: Serum samples (2 μL) from 130 healthy pediatric controls and 180 sepsis patients underwent high-resolution Raman spectroscopic analysis, prospectively establishing a clinical spectral database encompassing both physiological and pathological signatures. The spectra underwent preprocessing through self-normalization, and datasets were split into training, validation, and test sets at a 7:1:2 ratio. Utilizing the ResNet18 residual neural network architecture as a foundation, the 2D convolution in the network was adapted to 1D convolution to accommodate Raman spectrum. Optimal parameters were ascertained via simulation experiments. The model’s weights were trained on the training set, followed by the selection of optimal hyperparameters grounded in the model’s performance on the validation set. The trained convolutional neural network model underwent testing with the test set to appraise its classification efficacy.

Results: Raman spectroscopy serves as a molecular fingerprinting modality, detecting sepsis-induced biochemical perturbations in pediatric serum including: Elevated levels of inflammatory molecules and cytokines (e.g., IL-6, TNF-α);Abnormal concentrations of acute-phase proteins (e.g., C-reactive protein, procalcitonin);Markers of metabolic disorders, such as lactic acid accumulation (indicating tissue hypoxia) and lipid peroxidation products (markers of oxidative stress);Pathogen-derived components, such as bacterial endotoxins (e.g., LPS) or fungal molecules (e.g., β-glucan).Raman spectroscopy can capture these sepsis-associated alterations with high sensitivity by detection of children's serum.In this work, one AI model was successfully constructed using serum Raman spectrum database of children with sepsis. The models achieved the following classification accuracies: 95.93% for distinguishing children with sepsis from healthy pediatric controls (Figure1-2).

Conclusion: This novel AI-based Raman spectroscopy technology provides a label-free, rapid, accurate, and cost-effective method for identifying children with sepsis. It holds significant promise to reduce the high mortality rate caused by sepsis.