A transformer-based deep learning model for identifying the occurrence of acute hematogenous osteomyelitis and predicting blood culture results

Yunhong Xia, Qiang Kang, 義典 橘高, Jiuhui SuFrontiers in Microbiology 2024 · DOI

TL;DR

Researchers have developed a deep learning architecture that combines Artificial Neural Networks (ANNs) with Transformers to diagnose acute hematogenous osteomyelitis (AHO) in children using routine laboratory data. Beyond just diagnosis, the model achieves high accuracy in predicting whether a blood culture will successfully grow pathogens—a critical insight that helps clinicians decide whether to pause antibiotics for testing or proceed with empirical therapy.

What Problem Does This Paper Address?

Acute hematogenous osteomyelitis (AHO)—a severe bone infection—is a significant challenge in pediatric orthopedics. Because children often struggle to articulate their symptoms, doctors must rely on a mix of physical exams, complex imaging (like MRI), and blood cultures.

The "gold standard" for treatment is identifying the specific bacteria via blood culture to tailor antibiotic therapy. However, there is a major clinical "catch-22": to get an accurate culture, samples should ideally be taken before antibiotics start. In reality, many children have already received antibiotics by the time they reach a specialist, which causes the success rate of blood cultures to plummet (often below 60%). This leaves clinicians in a bind: should they stop antibiotics to try and get a clean culture, risking disease progression, or continue blindly with broad-spectrum drugs?

Methods

The authors proposed a hybrid deep learning model designed to extract maximum signal from standard laboratory indicators (like white blood cell counts, hemoglobin, and liver function tests).

  • The Architecture: The model uses a two-stage approach. First, an ANN encoder (three linear layers with ReLU activation) processes the tabular medical data into feature vectors. Second, a Transformer decoder uses self-attention mechanisms to capture the "contextual" relationships between different laboratory markers. While Transformers are famous for Natural Language Processing, the authors argue they are uniquely suited here to identify how subtle shifts in one blood value (e.g., Hemoglobin) relate to another (e.g., Platelet count) in the context of infection.
  • The Tasks: The model was trained for two distinct purposes:
  • Multiclass Classification: Differentiating AHO from "look-alike" conditions, including fractures, benign bone tumors, and malignant bone tumors.
  • Culture Prediction: Predicting a "positive" or "negative" blood culture result specifically for AHO patients.

Datasets

The study performed a retrospective analysis of 634 pediatric patients (under age 18) hospitalized between 2016 and 2024.

  • Diagnostic Cohort: 100 AHO cases, 238 benign tumors, 70 malignant tumors, and 226 fractures.
  • Culture Cohort: A subset of 100 AHO patients (58 positive cultures, 42 negative).
  • Features: The researchers used 21 to 34 common variables, including complete blood counts (WBC, RBC, HGB, PLT), biochemical indicators (ALT, AST, GGT), and patient demographics (age, gender).

Key Insights and Ideas

  • The RBC-Recovery Link: One of the most interesting findings was that red blood cell (RBC) indicators—specifically Hemoglobin (HGB), Hematocrit (HCT), and Red Cell Distribution Width (RDW)—were more strongly correlated with the length of hospital stay than traditional inflammatory markers. This suggests that the "physiological state" of a patient's oxygen transport system is a primary driver of bone tissue repair and overall recovery speed.
  • Bypassing the Antibiotic Barrier: The model could predict culture success regardless of whether the patient had already started antibiotics. This implies that the systemic "fingerprint" left by the bacteria in the blood chemistry remains detectable by AI even when the live bacteria are too suppressed to grow in a petri dish.
  • Attention to Tabular Data: By using a Transformer, the model doesn't just look at values in isolation; it looks at the pattern of the entire blood panel. This mimics how an experienced clinician might "feel" a diagnosis based on a holistic view of a patient's labs, but with mathematical precision.

Key Results

  • Diagnostic Accuracy: In the multiclass task (AHO vs. tumors vs. fractures), the model achieved an accuracy of 1.0 on the test set after 32 epochs.
  • Culture Prediction: For the binary task of predicting blood culture results, the model also reached an accuracy of 1.0 on the test set.
  • Statistical Correlations: The study confirmed that RDW is positively correlated with hospital stay (higher RDW = longer stay), while HCT, HGB, and RBC counts are negatively correlated (lower values = longer stay/slower recovery).

Limitations and Future Work

While the results are statistically impressive, the authors acknowledge several hurdles:

  • Single-Center Bias: The data comes from one institution, which may limit how well the model performs at other hospitals with different patient demographics or lab equipment.
  • Data Modality: The current model relies solely on laboratory numbers. It does not yet incorporate medical imaging (X-rays/MRIs), which are staples of orthopedic diagnosis.
  • Pathogen Specificity: While the model predicts if a culture will be positive, it cannot yet predict which specific bacteria (e.g., S. aureus vs. Streptococcus) is present.

The researchers aim to expand this into a multi-center study and integrate imaging data to create a more robust, "multimodal" diagnostic tool for pediatricians.

Summary updated

Abstract

Laboratory indicators, as easily accessible medical information, can identify osteomyelitis in pediatric patients. They can also predict whether pathogenic bacteria can be successfully cultured, regardless of whether the patient has received antibiotics beforehand. This not only simplifies the diagnostic process for pediatricians but also provides a basis for deciding whether to use empirical antibiotic therapy or discontinue treatment for blood cultures.
BibTeX
@inproceedings{xia2024transformer,
  author = {Yunhong Xia and Qiang Kang and 義典 橘高 and Jiuhui Su},
  title = {{A transformer-based deep learning model for identifying the occurrence of acute hematogenous osteomyelitis and predicting blood culture results}},
  booktitle = {Frontiers in Microbiology},
  year = {2024},
  doi = {10.3389/fmicb.2024.1495709},
}
Figure