Machine learning predicts non-thyroidal illness syndrome (NTIS) occurrence and mortality in sepsis patients.

BACKGROUND

Non-thyroid disease syndrome (NTIS) occurs in various serious illnesses, with sepsis being a common cause. Early identification of NTIS risk in sepsis patients is crucial. This study aimed to develop machine learning-based predictive models for the NTIS occurrence and mortality in sepsis patients.

METHODS

Clinical data were collected from 963 recruited sepsis patients, and 890 of them were retained for NTIS occurrence prediction after data selection and processing, while 797 NTIS patients were selected for mortality prediction. LASSO regression and correlation analysis identified key clinical features. We evaluated the predictive capacity on both occurrence and mortality of eight machine learning algorithms using Receiver Operating Characteristic (ROC) and Sensitivity and Specificity. For mortality prediction, univariate and multivariate Cox regression analyses were performed to identify risk factors and construct a nomogram model, which was further validated by decision curve and survival analyses.

RESULTS

XGBoost performed best in NTIS occurrence prediction, while LASSO showed the highest efficiency for mortality prediction. Lower levels of T3, FT3, T4, and KAP/LAM were associated with higher NTIS risk. Mechanical ventilation, norepinephrine, abdominal infection, cardiopulmonary resuscitation with sudden cardiac arrest, multiple organ failure, novel coronavirus infection, and gender were associated with NTIS mortality. Additionally, univariate and multivariate Cox regression underscored age, mechanical ventilation, pulmonary infection, multiple organ failure, norepinephrine, UA, and FIB as independent risk factors for mortality.

CONCLUSIONS

XGBoost and LASSO models showed promising performance in predicting NTIS occurrence and mortality in sepsis patients, respectively, and integrating them with clinical risk factors may improve risk stratification and support clinical decision-making. However, the findings are derived from a single-center dataset, and the use of data imputation to address missing values may introduce model instability, warranting cautious interpretation and external validation.