摘要
Abstract
Objective To identify and validate the clinical phenotypes of patients with sepsis in the in-tensive care unit(ICU).Methods We applied unsupervised machine learning algorithms(K-means clustering and hierarchical clustering)to identify the phenotypes of sepsis patients in the Medical Information Mart for In-tensive Care Ⅳ(MIMIC-Ⅳ)2.2 database,based on 89 clinical features including demographic characteris-tics,laboratory indicators and treatment measures on the first day in ICU.Then,supervised machine learning algorithms(lightweight gradient boosting machine)were used for the prediction of the patient's phenotypes,and were further combined with SHAP(Shapely Additive eXplanations)for the identification of important features.Finally,traditional statistical methods were used to validate the differences in clinical characteristics and clinical outcomes among the phenotypes.Results We identified three phenotypes in 22 517 sepsis patients.The phenotype 1 patients had the highest risk of death(28-day mortality of 46.4％),dominated by abnormal renal function and elevated disease severity scores,while the phenotype 3 patients had the lowest risk of death(28-day mortality of 11.2％),and the best neurological function score.Using interpretable machine learning,we identified six features(all the worst value on the first day)that showed good performance in phenotypic identification(AUC≥0.89)and phenotypic prognostic prediction(AUC≥0.74):anion gap,blood urea ni-trogen,creatinine,Glasgow Coma Scale score,prothrombin time,and Sequential Organ Failure Assessment score.The mortality risk of phenotype 3 patients was the lowest at 28 days,60 days,90 days,and 1 year after ICU discharge(HR＜1).Conclusion Using machine learning methods,we successfully identified three clinical phenotypes of sepsis patients with different clinical characteristics and prognosis and screened out six key clinical features,which are expected to play an important role in the phenotype classification and prognostic assessment of sepsis and are conducive to individualized treatment.关键词
脓毒症/表型识别/机器学习/精准治疗Key words
sepsis/phenotypic recognition/machine learning/precision medicine分类
医药卫生
