关键词: 急性脑出血, 脑水肿, 预后, 机器学习法, 风险预测模型, 支持向量机

Abstract: BACKGROUND: Recent studies on the pathological mechanisms of acute intracerebral hemorrhage have shown that the occurrence of poor postoperative prognosis in patients with acute intracerebral hemorrhage is directly related to brain tissue edema caused by intracerebral hemorrhage. The severity of brain tissue edema is closely associated with a cascade reaction of inflammatory factors.
OBJECTIVE: To investigate the correlation between serum inflammatory factor protein levels and postoperative brain edema volume in older adult patients with acute intracerebral hemorrhage using machine learning algorithms to analyze their impact on the occurrence of poor postoperative prognosis.
METHODS: A total of 250 older adult patients with acute cerebral hemorrhage who underwent surgery at Yichun People's Hospital between June 2022 and June 2024 were included in this study. They were divided into a poor prognosis group and a good prognosis group based on their postoperative outcomes. Relevant patient data were collected to analyze the correlation between serum levels of matrix metalloproteinase-9, NOD-like receptor protein 3, and angiopoietin-like protein 2 and postoperative brain edema volume. A risk factor analysis was conducted using the occurrence of poor postoperative outcomes as the dependent variable. Risk prediction models for poor postoperative outcomes in older adult patients with acute cerebral hemorrhage were constructed using machine learning algorithms, including Logistic regression, Classification and Regression Tree, Back Propagation Neural Network, and Support Vector Machine. The receiver operating characteristic curve was used to assess the predictive efficacy of the different algorithms.
RESULTS AND CONCLUSION: (1) Among the 250 patients included in this study, 113 patients (45.20%) were assigned to the poor prognosis group, while 137 patients (54.80%) to the good prognosis group. (2) Multivariate analysis revealed that matrix metalloproteinase-9 (OR = 1.037, 95% CI = 1.010-1.064, P = 0.007), NOD-like receptor protein 3 (OR = 64.050, 95% CI = 5.139-798.325, P = 0.001), angiopoietin-like protein 2 (OR = 82.519, 95% CI = 6.961-978.225, P < 0.001), and brain edema volume (OR = 6.859, 95% CI = 2.109-22.309, P = 0.001) were independent factors associated with poor postoperative outcomes in older adult patients with acute cerebral hemorrhage. (3) The Classification and regression Tree algorithm indicated that NOD-like receptor protein 3, brain edema volume, and matrix metalloproteinase-9 were risk factors associated with poor postoperative outcomes. (4) The Back Propagation Neural Network algorithm ranked the influential factors as follows: Angiopoietin-like protein 2 > NOD-like receptor protein 3 > matrix metalloproteinase-9 > brain edema volume > tumor necrosis factor-alpha > National Institutes of Health Stroke Scale (NIHSS) score > history of alcohol consumption > history of hypertension > amount of blood loss > duration of illness. (5) The Support Vector Machine algorithm identified the top five influential factors as NOD-like receptor protein 3 (predictor importance = 0.25), angiopoietin-like protein 2 (predictor importance = 0.22), amount of blood loss (predictor importance = 0.14), tumor necrosis factor-alpha (predictor importance = 0.12), and brain edema volume (predictor importance = 0.10). (6) Among the four machine learning algorithms evaluated, the Support Vector Machine algorithm demonstrated the best predictive performance. (7) Results from this study suggest that serum levels of matrix metalloproteinase-9, NOD-like receptor protein 3, and angiopoietin-like protein 2 in older adult patients with acute cerebral hemorrhage are correlated with postoperative brain edema volume. These factors can be used to construct a risk prediction model for postoperative outcomes using machine learning algorithms, with the Support Vector Machine algorithm showing the best diagnostic efficacy.

Key words: acute cerebral hemorrhage, brain edema, prognosis, machine learning, risk prediction model, Support Vector Machine