A study published in *Scientific Reports* presents a hybrid LLM and machine learning framework for early fire detection in subway tunnels. Sensitivity analysis on the Total dataset evaluated F1 Score, Detection Delay, and Pre-Alarm Rate (PAR) across decision thresholds (τ) and temporal persistence windows (k). The default operating setting was τ=0.50 and k=1.0s. Results showed a trade-off between detection sensitivity and premature-alarm suppression. At τ=0.25 and k=1.0s, SVM + LLM and GBM + LLM achieved F1 scores of 87.45% and 90.85%, but PAR values rose to 99.25% and 51.32%. At τ=0.75 and k=1.0s, PAR fell to 8.68% for SVM + LLM, 6.04% for RF + LLM, and 12.08% for GBM + LLM, but F1 scores dropped to 64.78%, 69.21%, and 69.66%, respectively. Increasing the persistence window from 1.0 to 2.0s or 3.0s generally reduced PAR, but excessive smoothing reduced F1. At τ=0.50 and k=2.0s, GBM + LLM maintained an F1 of 88.33% while reducing PAR from 47.55% to 28.68%.