Construction and Diagnostic Efficacy of Sepsis Early Warning Model Combined with Acoustic Biomarkers

ZHANG Xiaomeng; WANG Lili; PAN Zijie; YU Hongyang; LI Zhaofan; CHEN Li

doi:10.13730/j.issn.2097-2148.2025.08.005

Abstract：Objective To evaluate the predictive value of sepsis by integrating speech recognition parameters with vital signs.Methods The data of 343 patients who were admitted to the emergency room of the department of emergency medicine of the author's hospital from April 2022 to June 2024 were collected.According to the diagnostic criteria of sep-sis,the patients were divided into sepsis group(ni=67)and non-sepsis group(n=276).All patients' audio was collected according to the voice acquisition scheme,demographic information,medical history,laboratory and examination indica-tors,clinical information during treatment and hospitalization,acute physiology and chronic health evaluation Ⅱ(A-PACHE Ⅱ)and sequential organ failure assessment(SOFA)were also collected.Least absolute shrinkage and selection operator(LASSO)regression and Logistic regression analysis were used to screen the speech recognition parameters and vital signs.All patients were divided into training set and validation set according to the ratio of 8∶2 by stratified randomization method,the training set was used to con-struct the Logistic regression predictive model,the valida-tion set was used for model evaluation,and the area under the curve(AUC),sensitivity and specificity were calcu-lated by the receiver operating characteristic(ROC)curve,finally,the performance of the Logistic regression model was evaluated by calibration curve and decision curve analysis.Results The age,SOFA,APACHE Ⅱ score,the proportion of vasopressor use,maximum sound intensity and heart rate in the sepsis group were significantly higher than those in the non-sepsis group(all P＜0.05);fundamental frequency perturbation rate,differential perturbation,local am-plitude perturbation ①,local amplitude perturbation ②,three point period of amplitude perturbation,five point period of amplitude perturbation,eleven point period of amplitude perturbation,differential amplitude perturbation,mean sound intensity and oxygen saturation were lower in the sepsis group than in the non-sepsis group(all P＜0.05).According to Logistic regression multivariate analysis,local amplitude perturbation ② and harmonic-to-noise ratio in speech recognition parameters were protective factors against sepsis(all P＜0.05),heart rate was a risk factor for sepsis(P=0.001).The AUC of ROC of the predictive model constructed by the heart rate,local amplitude perturbation ②,harmonic noise ratio and maximum sound intensity was 0.805[95％confidence interval(CI):0.668-0.942],the sensitivity was 0.765 and the specificity was 0.824,the model had good predictive ability;the calibration curve showed that the model prediction was in good agreement with the actual prediction.Clinical decision curve analysis showed that when the threshold was 0.05-0.78,the predictive model showed that patients could gain clinical benefits.Conclusion In the speech recognition parameters,local amplitude perturbation ② and harmonic-to-noise ratio can be used as potential biomarkers for the pre-diction of sepsis.The prediction model combined with heart rate has moderate prediction efficiency and good calibration,which provides a new idea for the non-invasive screening of sepsis.