Large Language Model–Assisted Estimation of Left Double-Lumen Tube Insertion Depth Using Preoperative Chest Radiography

Technical Summary: Large Language Model–Assisted Estimation of Left Double-Lumen Tube Insertion Depth

Problem Statement
Accurate positioning of left-sided double-lumen endobronchial tubes (DLTs) is critical for successful one-lung ventilation (OLV) in thoracic surgery to prevent complications such as hypoxemia, inadequate lung isolation, and airway trauma. While fiberoptic bronchoscopy (FOB) is the gold standard for confirming placement, it requires specific equipment and expertise that may not be immediately available in all clinical settings. Consequently, anesthesiologists rely on estimated insertion depths prior to bronchoscopic confirmation. Existing estimation methods, primarily based on patient height, demonstrate statistical correlations but often lack precision for individual patients, particularly those with atypical body habitus. While imaging-based strategies like CT-guided measurements have shown improved accuracy, they are not always practical for real-time decision support due to data complexity and accessibility issues. There is a need for a clinically feasible, accurate method to guide initial DLT positioning using routinely available data.

Methodology
This retrospective, single-center observational validation study evaluated the accuracy of a pretrained large language model (LLM) in predicting optimal left-sided DLT insertion depths. The study included 154 adult patients undergoing thoracic surgery at a tertiary academic hospital.

• Reference Standard: The ground truth for all analyses was the intraoperative insertion depth (measured from the lip margin in centimeters) confirmed by flexible fiberoptic bronchoscopy (FOB).
• LLM Approach: Predictions were generated using a publicly available multimodal LLM (ChatGPT, OpenAI) without fine-tuning or retraining. Two distinct input strategies were tested:
  1. Demographic-Only: Standardized prompts containing patient age, sex, height, weight, and DLT size.
  2. Radiograph-Assisted: The same demographic data combined with descriptive anatomical information derived from preoperative posteroanterior (PA) chest radiographs.
• Protocol: To ensure consistency, a standardized prompt structure was used for all cases. Predictions were generated sequentially in a single session using default system settings, with no feedback provided to the model regarding actual outcomes. The first output for each case was recorded.
• Statistical Analysis: Agreement between LLM predictions and physician-confirmed FOB measurements was assessed using Bland–Altman analysis (calculating mean bias, standard deviation, and 95% limits of agreement). Differences among the three groups (Physician, LLM without PA, LLM with PA) were analyzed using the Friedman test, followed by Dunn's post-hoc pairwise comparisons with adjusted p-values.

Key Results

• Central Tendency: The median insertion depth was 29.00 cm across all three groups (Physician, LLM without PA, and LLM with PA).
• Statistical Agreement:
  • The Friedman test revealed a statistically significant overall difference among the three methods ($\chi^2(2) = 12.14, p = 0.0023$).
  • Post-hoc analysis showed a significant difference between physician measurements and LLM predictions using demographic data alone (adjusted $p = 0.0039$).
  • Crucially, no statistically significant difference was observed between physician measurements and LLM predictions when PA chest radiographs were incorporated (adjusted $p = 0.7634$).
• Bias and Limits of Agreement (LoA):
  • Demographic-Only: Mean bias of 0.20 cm with 95% LoA of approximately $\pm 2.8$ cm.
  • Radiograph-Assisted: Mean bias of 0.15 cm with 95% LoA of approximately $\pm 2.8$ cm.
  • Both approaches demonstrated minimal systematic bias.
• Precision: More than 90% of predictions in both LLM groups fell within $\pm 2$ cm of the physician-confirmed depth. The mean absolute error was approximately 1 cm for both approaches.

Significance and Claims
The paper claims that LLM-based prediction of DLT insertion depth demonstrates reasonable agreement with physician-confirmed measurements, characterized by minimal systematic bias and low absolute error. The study highlights that while demographic-only predictions showed measurable deviation from clinical standards, the incorporation of preoperative PA chest radiographs eliminated this statistical difference, suggesting improved concordance with clinician-determined positioning.

The authors position this technology not as a replacement for fiberoptic bronchoscopy, which remains indispensable, but as a supportive adjunct for initial DLT positioning. This is particularly relevant in clinical settings where bronchoscopy resources are limited or delayed. The use of PA radiographs is emphasized for its clinical feasibility compared to CT-based strategies, as PA films are universally available and standardized in thoracic surgery workflows.

Limitations and Scope
The authors explicitly note several limitations:

• The study is retrospective and single-center, requiring external validation in independent cohorts before broader implementation.
• Findings are specific to left-sided DLTs and may not generalize to right-sided devices.
• The sequential nature of model interaction and the potential for contextual influence between cases, alongside the evolving nature of LLM inference behavior, may affect reproducibility across different environments.
• The study did not involve real-time clinical guidance; predictions were generated retrospectively and did not influence actual tube placement decisions.

In conclusion, the paper suggests that LLM-assisted estimation, particularly when augmented with routine radiographic data, offers a promising, low-cost tool to support airway management, pending prospective validation under controlled inference conditions.