Accuracy of machine learning-based interpretation of preterm brain maturity using electroencephalographic features
10.3760/cma.j.cn113903-20240731-00541
- VernacularTitle:基于脑电特征的机器学习方法判读早产儿脑成熟度的准确性
- Author:
Xiaoming LYU
1
;
Shuaiwen DING
;
Zhenyu LI
;
Ling LI
;
Jiahui LI
;
Hui WU
Author Information
1. 吉林大学第一医院儿童医院新生儿科，长春 130000
- Publication Type:Journal Article
- Keywords:
Electroencephalography;
Premature infant;
Machine learning;
Brain development;
Predictive value
- From:
Chinese Journal of Perinatal Medicine
2025;28(9):746-754
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop machine learning models for interpreting brain maturity in preterm infants based on electroencephalographic (EEG) features and analyze factors affecting interpretation accuracy.Methods:This prospective study enrolled preterm infants requiring bedside EEG monitoring in the Department of Neonatology at the First Hospital of Jilin University from January 2023 to March 2024. Data from each integer-corrected gestational age (GA) group were randomly split into training and testing sets (7∶3 ratio) using Python's sklearn.model_selection.train_test_split function. Three machine learning models, including support vector regression (SVR), random forest, and decision tree, were employed to analyze EEG signals. Model performance was evaluated against manually interpreted GA as the gold standard using prediction deviation, mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient ( r). Accuracy was defined based on the difference between predicted and manually interpreted GA (categorized into accurate and inaccurate groups), with a difference less than one week considered accurate. Statistical analyses included Chi-square test (or Fisher's exact test), t-test, Mann-Whitney U test, and multivariate logistic regression. Results:Among 241 preterm infants (training set: n=168; testing set: n=73), the random forest model demonstrated optimal performance: concordance rate 90.4% (66/73) with MAE 0.378 weeks, RMSE 0.577 weeks, and r=0.932 ( P<0.001). The decision tree model achieved 87.7% concordance (64/73) with MAE 0.316 weeks, while SVR showed 64.2% concordance (47/73) and MAE 0.840 weeks. Stratified by GA, random forest performed best in the 34 weeks group [concordance 100.0% (52/52), MAE 0.269 weeks], followed by the 32-34 weeks group [89.0% (81/91), MAE 0.448 weeks] and <32 weeks group [88.8% (87/98), MAE 0.561 weeks]. Compared to the accurate group ( n=205), the inaccurate group ( n=36) had higher rates of vaginal delivery [41.7% (15/36) vs. 19.5% (40/205), χ2=8.53], grade ≥Ⅱ intracranial hemorrhage [11.1% (4/36) vs. 2.4% (5/205), χ2=4.22], and periventricular echogenicity [33.3% (12/36) vs. 7.8% (16/205), χ2=17.03] (all P<0.05). Multivariate analysis identified vaginal delivery ( OR=0.190, 95% CI: 0.068-0.527), lower corrected GA ( OR=0.678, 95% CI: 0.488-0.941), and periventricular echogenicity ( OR=11.339, 95% CI: 3.250-39.559) as independent factors affecting accuracy (all P<0.05). Conclusion:The random forest-based model shows optimal accuracy for predicting brain maturity in preterm infants. Vaginal delivery, lower corrected GA, and periventricular echogenicity reduce its predictive accuracy.
