Combined model of radiomics and clinical features for predicting prognosis of term neonatal hypoxic-ischemic encephalopathy after one year: an exploratory study Authors: Jing Tang, Si-Ping He, Yu-Qing Liu, Yong-Hua Xiang, Ting Yi, Ke Jin Published: 2025 ABSTRACT AbstractPurposeTo establish a combined model integrating imaging-based radiomics features and clinical parameters to predict the prognosis of hypoxic-ischemic encephalopathy (HIE) in full-term newborns one year after birth.MethodsA total of 180 full-term neonates diagnosed with HIE were retrospectively analyzed. Based on cognitive and motor function scores at 12 months post-birth, patients were categorized into two groups: Group B, representing those with a good prognosis (n = 84), and Group W, representing those with a poor prognosis (n = 96). The dataset was randomly divided into a training dateset (n = 126) and a testing dateset (n = 54). Clinical characteristics were first compared between the two groups. Subsequently, three predictive models were developed: a clinical model, a radiomics model, and a combined model integrating both clinical and radiomics features. The predictive performances of these models were evaluated using receiver operating characteristic (ROC) curve analysis, and their discriminative abilities were quantified by calculating the area under the curve (AUC).ResultsThe Apgar scores at 1, 5, and 10 min after birth were significantly higher in Group B compared to Group W (P < 0.05). In the clinical model, the Apgar score at 10 min was identified as the strongest prognostic factor, yielding an AUC of 0.857 in the training datest and 0.737 in the testing datest. In the radiomics model, nine radiomics features were significantly associated with prognosis, achieving AUCs of 0.916 and 0.770 in the training and testing datests, respectively. In the combined model, seven radiomics features together with the 5-minute and 10-minute Apgar scores were identified as independent predictors of prognosis. This integrated model demonstrated superior predictive performance, with AUCs of 0.952 in the training datest and 0.823 in the testing datest.ConclusionsThe combined model incorporating MR-based radiomics signatures and clinical parameters demonstrates high predictive accuracy for assessing the one-year prognosis of full-term neonates with HIE, suggesting a promising framework for early risk stratification and individualized management of affected infants.Keywords: Hypoxic-ischemic encephalopathy, Neonatal, Radiomics, Magnetic resonance imaging, Prognosis Introduction Hypoxic-ischemic encephalopathy (HIE) refers to hypoxic-ischemic brain injury resulting from perinatal asphyxia [1]. It represents one of the leading causes of acquired neonatal brain injury and is the most common etiology of neonatal encephalopathy [2]. The incidence of HIE is estimated to range from approximately 1 to 8 cases per 1,000 live births [3, 4]. The underlying pathophysiology primarily involves hypoxia-induced cerebral cell necrosis and apoptosis. When the ischemic injury is severe and prolonged, it can lead to diffuse and extensive neuronal necrosis, resulting in irreversible brain tissue damage [5]. The prognosis of HIE is largely determined by the severity of the insult. Infants with mild HIE generally have a favorable outcome and can often recover fully without long-term neurological deficits, and those with moderate HIE may also achieve substantial recovery if timely and appropriate interventions are administered; however, some may still experience developmental sequelae such as cognitive delay or impaired motor function. In contrast, severe HIE is associated with high mortality, and surviving infants frequently suffer from permanent neurological complications, including cerebral palsy, intellectual disability, and epilepsy [6]. Accurate short- and long-term prognostic prediction is therefore essential for the management and follow-up of neonates with HIE. Reliable prognostic indicators enable clinicians to devise more precise and individualized treatment strategies and provide families with critical information regarding expected outcomes, thereby facilitating informed clinical and parental decision-making [7]. Therefore, identifying effective and objective prognostic predictors has become a major focus of contemporary research on neonatal HIE. Although conventional magnetic resonance imaging (MRI) is widely used for evaluating HIE, its diagnostic and prognostic utility remains limited when based solely on qualitative visual interpretation. Radiomics, an emerging and promising imaging analysis approach, allows for the high-throughput extraction and quantification of large numbers of imaging features from digital medical images, transforming them into analyzable data that can be used to develop predictive and diagnostic models [8]. While radiomics has been increasingly applied in various clinical settings, studies assessing its value in detecting ischemic changes among full-term neonates remain scarce. To date, no research has applied conventional MRI-based radiomics for evaluating the long-term prognosis of full-term neonatal HIE. Therefore, the present study was designed to develop and validate both clinical and radiomic models for predicting the one-year prognosis of full-term neonates with HIE and explore the potential of an integrated model that combines radiomic features with clinical indicators for improved prognostic accuracy. Materials and methods Patients and data collection A retrospective study was conducted to identify and include neonates diagnosed with HIE who were admitted to Hunan Children’s Hospital between January 2013 and May 2023. A total of 834 infants with HIE were initially screened for eligibility. The inclusion criteria were as follows: (a) a documented history of perinatal asphyxia; (b) a clinical diagnosis of HIE or an indication for MRI to exclude or evaluate brain injury; (c) full-term neonates with a gestational age between 37 and less than 42 weeks who underwent MRI within 10 days after birth; (d) complete imaging data including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and fluid-attenuated inversion recovery (FLAIR) sequences, with MRI-confirmed evidence of HIE; and (e) availability of clinical follow-up data including neurodevelopmental and motor function assessment at 12 months after birth. The exclusion criteria were: (a) suspected brain injury caused by factors other than perinatal asphyxia, including neonatal hyperbilirubinemia, intrauterine or intracranial infection, hypoglycemia, intracranial hemorrhage, congenital malformations, or metabolic disorders; and (b) poor-quality MRI images unsuitable for radiomic feature extraction or analysis. After applying these criteria, 193 infants met the initial eligibility requirements. Following image preprocessing and radiomic data quality control, 180 cases were found suitable for feature extraction and subsequent modeling analyses. The overall patient selection process is illustrated in Fig. 1. This retrospective study was approved by the Medical Ethics Committee of Hunan Children’s Hospital, Central South University (Approval No. HCHLL-2025-159). Flowchart illustrating the enrollment and selection process of the study population Comprehensive clinical data were collected for all enrolled neonates, including gestational age at birth, mode of delivery, sex, birth weight, head circumference, Apgar scores at 1, 5, and 10 min, and the postnatal age at which MRI was performed. At 12 months of corrected age, neurodevelopmental assessment was performed using the Bayley Scales of Infant and Toddler Development, Third Edition (Bayley-III) to evaluate cognitive performance, and the Gross Motor Function Classification System (GMFCS) to assess motor function. Based on the outcomes of these evaluations, infants were classified into two prognostic groups: Group B, representing those with a good prognosis, and Group W, representing those with a poor prognosis. In accordance with the Bayley-III and the GMFCS, and consistent with the criteria used in previous studies [9, 10], infants with a Bayley-III standard score ≥ 85 and GMFCS Level I were defined as having a good prognosis, whereas those with a Bayley-III standard score < 85 and GMFCS Levels II–V were defined as having a poor prognosis. Comparative analysis was subsequently performed between Group B and Group W for clinical characteristics, including gestational age at birth, mode of delivery, age, sex, birth weight, head circumference, and Apgar scores. MR image acquisition MRI examinations were performed once the infants’ general conditions had stabilized to a clinically acceptable level. All scans were conducted using a Siemens Skyra or Prisma 3.0 T MRI scanner equipped with an eight-channel head coil, and identical imaging parameters were applied across both systems to ensure uniformity. The imaging protocol included axial T1WI, axial T2WI, axial FLAIR, and sagittal T1WI sequences. For the sagittal T1WI, the repetition time (TR) and echo time (TE) were 400 ms and 8.1 ms, respectively. The axial T1WI sequence was acquired with a TR of 468 ms and a TE of 11 ms, while the axial T2WI sequence used a TR of 4000 ms and a TE of 101 ms. For the axial FLAIR sequence, the TR and TE were 9000 ms and 98 ms, respectively. The slice thickness was maintained at 4 mm with an inter-slice gap of 0.32 mm. Before segmentation and radiomic feature extraction, all MR images underwent standardized preprocessing to minimize potential discrepancies arising from variations between the two scanner systems. This normalization ensured the comparability of image intensity distributions and reduced systematic bias during subsequent quantitative analysis. Establishment of clinical model The clinical model was constructed using the clinical parameters that showed statistically significant differences between Group B and Group W. Feature selection was first performed using the Kruskal–Wallis test to identify variables most strongly associated with prognostic outcomes. Subsequently, a Support Vector Machine (SVM) algorithm was employed as the classification tool to establish the predictive model. Optimization of SVM hyperparameters was achieved through Leave-One-Out Cross-Validation (LOOCV) applied to the training dataset. Establishment of radiomics model ROI drawing and image analysis T2-weighted images were segmented by a radiologist with eleven years of experience using ITK-SNAP software (version 3.8.0, available at http://www.itksnap.org). Regions of interest (ROIs) were manually delineated slice by slice, and all delineated ROIs were subsequently merged to generate a three-dimensional volume of interest (VOI). The VOIs obtained from each patient’s T2-weighted images were then transferred to the corresponding T1-weighted and FLAIR sequences to ensure consistency in radiomic feature extraction. The anatomical regions included in the VOI comprised the bilateral basal ganglia, as well as the thalamus. Please refer to Fig. 2. Flowchart illustrating the construction process of the clinical, radiomics, and combined models To ensure the reproducibility and stability of the extracted radiomic features, inter-observer agreement was assessed using the intraclass correlation coefficient (ICC). A total of 30 randomly selected cases were independently re-segmented and re-analyzed by another senior radiologist. Features demonstrating good reproducibility, defined as ICC > 0.75, were retained for subsequent feature selection and modeling. Radiomics feature-extraction and model evaluation A total of 126 cases were selected as the training dataset (67 in Group W and 59 in Group B), and 54 cases (29 in Group W and 25 in Group B) were used as the independent testing dataset. To address class imbalance in the training dataset, the Synthetic Minority Oversampling Technique (SMOTE) was applied to balance the sample distribution between the two groups. The feature matrix was normalized by subtracting the mean value from each feature vector and dividing it by its length. To reduce redundancy due to high feature dimensionality, pairwise correlations between features were analyzed. When the Pearson Correlation Coefficient (PCC) between any two features exceeded 0.990, one was removed to ensure feature independence. Feature selection was then performed using Recursive Feature Elimination (RFE), which iteratively selected features based on classifier performance. SVM was used as the classifier because of its robustness in model construction. The kernel function allows mapping of features into higher dimensions to identify the optimal separating hyperplane. A linear kernel was adopted to facilitate the interpretation of the model coefficients. Model hyperparameters, including the number of selected features, were optimized using LOOCV within the training dataset. In each iteration, the model was trained on 125 samples and tested on 1 sample, and this process was repeated 126 times to obtain the cross-validation training and validation sets. Hyperparameters were determined according to model performance on the validation dataset. After model training and hyperparameter optimization, the final classification threshold was set on the entire training dataset by maximizing the Youden Index (J = Sensitivity + Specificity – 1). This threshold was then fixed and applied to the independent testing dataset to calculate the performance metrics, including accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), ensuring unbiased model evaluation. All procedures were implemented using Feature Explorer Pro (FAE, version 0.5.8) on Python (version 3.7.6). Establishment of combined model A combined model for predicting good and poor prognoses was developed using the same procedure applied in the radiomics model. This model integrated MR-based radiomic features with clinical factors to enhance predictive accuracy and provide a more comprehensive assessment of prognosis in neonates with HIE. Statistical analysis All statistical analyses were performed using SPSS software (version 25.0) and R statistical software (version 3.6.1). For continuous variables, the Kolmogorov–Smirnov and Levene tests were first conducted to assess the normality and homogeneity of variance. Variables conforming to a normal distribution were expressed as mean ± standard deviation (SD), and comparisons between groups were performed using the t-test. Categorical variables were presented as frequencies and percentages, and comparisons between groups were conducted using the χ² test. ROC curves were plotted to evaluate and compare the predictive performance of the clinical, radiomics, and combined models. The models’ discriminative abilities were assessed by calculating the AUC, along with accuracy, sensitivity, specificity, PPV, and NPV. The bootstrap method was used to estimate 95% confidence intervals for the performance metrics. A two-sided P < 0.05 was considered statistically significant. DeLong test was performed to compare the AUCs of three models in the training and testing dateset. A two-sided P-value of < 0.05 was considered statistically significant. Results Patient characteristics A total of 180 full-term neonates diagnosed with HIE were included in the study, comprising 84 infants in Group B (good prognosis) and 96 in Group W (poor prognosis). The detailed clinical characteristics of both groups are summarized in Table 1, which shows no statistically significant differences between the two groups in terms of age, sex, birth weight, head circumference, mode of delivery, or gestational age (P > 0.05). However, the Apgar scores at 1, 5, and 10 min after birth were all significantly higher in Group B than in Group W (P < 0.05), indicating a correlation between higher Apgar scores and favorable neurological outcomes. Table 1Comparison of clinical characteristics between group B and group WVariablesGroup B(N = 84) No. of patient (%)Group W(N = 96) No. of patient (%)P valueAge (days), mean ± SD6.75 ± 2.86.59 ± 3.140.71Gender0.26 Male64(76.19)66(68.75) Female20(23.81)30(31.25)Delivery method0.32 Vaginal delivery50(59.52)64(66.7) Cesarean section34(40.48)32(33.3)Gestational week39.06 ± 1.2839.54 ± 1.290.06Weight(kg)3.23 ± 0.543.33 ± 0.480.17Head circumference33.48 ± 1.2533.82 ± 1.310.08Apgar scores at 1 min5.66 ± 2.723.25 ± 2.01<0.01Apgar scores at 5 min7.88 ± 2.025.91 ± 1.88<0.01Apgar scores at 10 min9.06 ± 1.297.20 ± 1.56<0.01 Comparison of clinical characteristics between group B and group W Performance of clinical model The clinical model identified the Apgar score at 10 min as the most effective factor in differentiating Group B from Group W (Table 2). In the training dataset, the model achieved an AUC of 0.857 and an accuracy of 0.833. In the testing dataset, the AUC, accuracy, sensitivity, and specificity were 0.737, 0.778, 0.965, and 0.560, respectively (Table 3; Fig. 4A). These findings indicate that the Apgar score at 10 min had substantial discriminative ability for predicting prognosis in neonates with HIE. Table 2The coefficients of features in the modelFeaturesCoef in modelClinical modelApgar scores at 10 min-5.667Radiomics modelFLAIR_wavelet-LLL_glcm_Imc2T1WI_log-sigma-1-mm-3D_gldm_LargeDependenceLowGrayLevelEmphasisT1WI_logarithm_glcm_InverseVarianceT1WI_wavelet-HLL_firstorder_MeanT1WI_wavelet-LLH_firstorder_SkewnessT2WI_gradient_glszm_LargeAreaHighGrayLevelEmphasisT2WI_logarithm_gldm_LargeDependenceLowGrayLevelEmphasisT2WI_logarithm_glrlm_LowGrayLevelRunEmphasisT2WI_squareroot_gldm_LargeDependenceLowGrayLevelEmphasis-1.111-0.0141.129-2.1722.743-2.164-0.825-1.914-0.461Combined modelApgar scores at 10 minApgar scores at 5 minFLAIR_wavelet-LLL_glcm_Imc2T1WI_logarithm_glcm_InverseVarianceT1WI_wavelet-HLL_firstorder_MeanT1WI_wavelet-LLH_firstorder_SkewnessT2WI_gradient_glszm_LargeAreaHighGrayLevelEmphasisT2WI_logarithm_gldm_LargeDependenceLowGrayLevelEmphasisT2WI_logarithm_glrlm_LowGrayLevelRunEmphasis-2.566-0.632-0.8750.935–1.7872.205-1.362-0.768-1.969 The coefficients of features in the model FLAIR_wavelet-LLL_glcm_Imc2 T1WI_log-sigma-1-mm-3D_gldm_LargeDependenceLowGrayLevelEmphasis T1WI_logarithm_glcm_InverseVariance T1WI_wavelet-HLL_firstorder_Mean T1WI_wavelet-LLH_firstorder_Skewness T2WI_gradient_glszm_LargeAreaHighGrayLevelEmphasis T2WI_logarithm_gldm_LargeDependenceLowGrayLevelEmphasis T2WI_logarithm_glrlm_LowGrayLevelRunEmphasis T2WI_squareroot_gldm_LargeDependenceLowGrayLevelEmphasis -1.111 -0.014 1.129 -2.172 2.743 -2.164 -0.825 -1.914 -0.461 Apgar scores at 10 min Apgar scores at 5 min FLAIR_wavelet-LLL_glcm_Imc2 T1WI_logarithm_glcm_InverseVariance T1WI_wavelet-HLL_firstorder_Mean T1WI_wavelet-LLH_firstorder_Skewness T2WI_gradient_glszm_LargeAreaHighGrayLevelEmphasis T2WI_logarithm_gldm_LargeDependenceLowGrayLevelEmphasis T2WI_logarithm_glrlm_LowGrayLevelRunEmphasis -2.566 -0.632 -0.875 0.935 –1.787 2.205 -1.362 -0.768 -1.969 Table 3Performance of the clinical, radiomics, and combined models in the training and testing datasetsGroupAUC (95% CI)Cut-offAccuracySensitivitySpecificityPPVNPV Training dataset Clinical model0.857(0.789–0.925)0.6020.8330.8810.7800.8200.852Radiomics model0.916(0.870–0.962)0.3490.8410.9250.7460.8050.898Combined model0.952(0.919–0.985)0.6500.9000.8660.9320.9360.860 CV training dataset Clinical model0.856(0.850–0.862)0.5730.8280.8810.7760.8000.867Radiomics model0.912(0.908–0.917)0.5170.8330.8220.8450.8410.826Combined model0.951(0.948–0.953)0.6420.8910.8510.9310.9250.862 validation dataset Clinical model0.843(0.765–0.921)0.6100.8650.8810.8480.8680.862Radiomics model0.867(0.804–0.930)0.3890.8250.8810.7630.8080.850Combined model0.931(0.890–0.973)0.6600.8810.8510.9150.9190.844 testing dataset Clinical model 0.737(0.595–0.879) 0.6020.7780.9660.5600.7180.933Radiomics model 0.770(0.641–0.898) 0.2970.7780.8620.6800.7580.810Combined model 0.823(0.705–0.942) 0.3800.7960.8510.9310.9250.862 Performance of the clinical, radiomics, and combined models in the training and testing datasets Receiver operating characteristic (ROC) curve analyses of the three predictive models. (A) Clinical model; (B) Radiomics model; (C) Combined model integrating MR-based radiomic features and clinical parameters for prognosis prediction in full-term neonates with hypoxic-ischemic encephalopathy (HIE) Performance of radiomics model A total of 4,227 radiomic features were extracted from T1WI, T2WI, and FLAIR sequences. These included shape-based, first-order statistical, texture, and higher-order statistical features. After ranking all extracted features, the nine most significant and best-performing radiomic features were selected for model construction (Fig. 3A; Table 2). In the training dataset, the radiomics model achieved an AUC of 0.916 and an accuracy of 0.841. In the testing dataset, the AUC, accuracy, sensitivity, and specificity were 0.770, 0.778, 0.862, and 0.680, respectively (Table 3; Fig. 4B). The final selected radiomic features and their corresponding model coefficients are summarized in Table 2. Histogram of the importance of the selected features. The y-axis represents the nine selected features in the radiomics model (A) or the combined model (B), and the x-axis represents the corresponding radiomic feature coefficients Performance of combined model The combined model incorporated both radiomic and clinical predictors, identifying Apgar scores at 10 min, Apgar scores at 5 min, and seven significant radiomic features as independent predictors for distinguishing Group W from Group B (Fig. 3B; Table 2). The seven radiomic features included: wavelet LLL glcm imc2, logarithm glcm inverse variance, wavelet HLL first-order mean, wavelet HLL first-order skewness, gradient glszm large area high gray level emphasis, logarithm gldm large dependence low gray level emphasis, and logarithm glrlm low gray level run emphasis. In the training dataset, the combined model achieved an AUC of 0.952 and an accuracy of 0.900. In the testing dataset, the corresponding AUC, accuracy, sensitivity, and specificity were 0.823, 0.796, 0.862, and 0.720, respectively (Table 3; Fig. 4C). These results demonstrate that the combined model provided the best overall performance among the three models. Comparation of the AUCs of three models In the training and testing dataset, the combined model exhibited a larger AUC than both the clinical and radiomics model. The DeLong test indicated that the combined model outperformed the clinical model in the training dataset, but did not surpass the radiomics model.And in the testing dataset, no statistically significant differences in AUC were observed among any of the three models (p > 0.05). The results of the DeLong test are presented in Table 4. Table 4ROC comparison between clinical model, radiomics model, and combined model in the training and testing datesets using DeLong testDatasetsROC1ROC2 p TrainingRadiomics modelClinical model0.1532Radiomics modelCombined model0.0554Clinical modelCombined model0.0006TestingRadiomics modalClinical model0.7159Radiomics modelCombined model0.2116Clinical modelCombined model0.1378 ROC comparison between clinical model, radiomics model, and combined model in the training and testing datesets using DeLong test Discussion The development of accurate prognostic indicators has been a central focus of research in the field of neonatal HIE in recent years. At present, the known prognostic indicators primarily include serum proteins [11–13], neurophysiological examinations [14, 15], and neuroimaging studies. Among these, brain MRI, as a non-invasive technique, plays an essential role in the precise localization and assessment of brain lesions. However, one of the major challenges in prognostication using traditional anatomical MRI sequences, such as T1- and T2-weighted imaging, lies in their relatively low sensitivity. In many instances, neonates with HIE who later develop severe neurological impairments may show no apparent abnormalities on MRI during the neonatal period (the first 28 days after birth) [16–18]. Radiomics, which enables the extraction of sub-visual yet quantitative imaging features from radiological images [19, 20], can effectively compensate for this limitation by providing deeper quantitative information beyond what is perceptible to the human eye. In this study, there were no statistically significant differences in age, gender, weight, head circumference, mode of delivery, or maternal gestational age between Group B and Group W. However, the Apgar scores at 1, 5, and 10 min were consistently higher in Group B than in Group W, and these differences were statistically significant. The clinical model demonstrated that the Apgar score at 10 min was the most effective factor for distinguishing Group B from Group W, achieving an AUC of 0.843 and an accuracy of 0.865 in the validation dataset, and an AUC of 0.737 and an accuracy of 0.778 in the testing dataset, respectively. Previous studies have indicated that the risk of adverse outcomes increases when the Apgar score remains low at or beyond 10 min [21–23]. Vivek V. Shukla et al. further suggested that a standalone 10-minute Apgar score of 0 does not independently predict the risk of death or moderate-to-severe disability, although predictive accuracy improves when the 10-minute Apgar score is combined with additional risk variables available during resuscitation using classification and regression tree analysis [24]. In the present study, we also observed that the clinical model had the lowest AUC and accuracy among the three models, whereas the combined model that integrated the Apgar score with radiomic features achieved marked improvement in both AUC and accuracy, highlighting the additive value of radiomics for prognostic prediction. For establishing the radiomics model, we selected the bilateral basal ganglia and thalamus as the ROI, primarily because these structures are highly susceptible to hypoxic-ischemic injury in neonates with HIE [25]. HIE induces a cascade of biochemical alterations, and affected patients undergo a complex pathophysiological process involving hypoxia, ischemia, and subsequent reperfusion injury. Some researchers have suggested that during this process, when ischemic tissues regain blood supply, the resultant reperfusion injury may be even more severe than that caused by ischemia alone. This is because reperfusion can exacerbate the damage to cells and tissues that were already compromised by ischemia, thereby aggravating the overall condition. Several studies employing arterial spin labeling (ASL) to evaluate cerebral hemodynamic changes in neonates with HIE have consistently demonstrated early hyperperfusion of the basal ganglia and thalamus [26–28]. Moreover, the thalamus, being one of the most metabolically active regions of the brain, exhibits frequent cellular activity and high energy demands. Consequently, even mild hypoxic-ischemic insults can induce varying degrees of cellular apoptosis within the thalamus. In cases of mild HIE, conventional MRI often fails to directly detect white matter injury or damage to the deep brain nuclei by visual inspection. However, radiomics provides a quantitative approach capable of revealing subtle alterations in tissue characteristics that are not discernible to the naked eye, thereby offering additional diagnostic and prognostic information. In this study, we established a radiomics model for predicting the prognosis of HIE using T1WI, T2WI, and FLAIR sequences. This model was developed based on nine non-zero coefficient features extracted from the original MR images and from images processed with Gaussian filters and wavelet transformations. These features included two first-order statistical features and seven texture features, among which T1WI_wavelet-LLH_firstorder_Skewness was identified as the most predictive variable. T1WI_wavelet-LLH_firstorder_Skewness is a first-order statistical feature, and first-order statistics are derived directly from the gray-level intensity distribution of the original image. Skewness reflects the degree of asymmetry in the data distribution relative to the mean, thereby providing information about the balance of pixel intensity within the lesion region. The results of this study demonstrated that the radiomics model achieved good predictive performance for HIE prognosis, with AUC values of 0.916 in the training dataset, 0.867 in the validation dataset, and 0.770 in the testing dataset, confirming the capability of radiomics to predict HIE outcomes. Similarly, Youwon Shin et al. [29] investigated 46 preterm neonates who underwent brain MRI near or at term-equivalent age and assessed neurodevelopmental outcomes at a corrected age of 12 months. A binary classification model was constructed to predict the psychomotor developmental index, yielding AUCs of 0.925, 0.834, and 0.902 for models based on T1WI, T2WI, and combined T1WI and T2WI sequences, respectively. Despite 45 of the 46 neonates showing no evident white matter injury on visual inspection, the radiomics analysis of early brain MRI achieved good diagnostic performance (sensitivity 84.6%; specificity 78.8%) in predicting poor psychomotor outcomes. This observation aligns with our findings, suggesting that it is feasible to establish a prognostic prediction model for neonatal brain injury using conventional MRI sequences and that the radiomics-based prediction model provides superior prognostic performance compared with conventional MRI assessment. Furthermore, our results indicate that T1WI contributes more significantly to the radiomics model in predicting neonatal brain injury prognosis. Through feature selection, this study identified 2 clinical features and 7 radiomic features to establish a combined model. The clinical features included the Apgar scores at 5 min and 10 min. Among the seven radiomic features, one was extracted from FLAIR images, three from T1WI images, and three from T2WI images. The two most influential parameters in the combined model were the Apgar score at 10 min and T1WI_wavelet-LLH_firstorder_Skewness. The combined model showed excellent predictive performance in predicting the prognosis of HIE, with the highest AUC values of 0.952 and 0.823 in the training and testing datasets, respectively, which were higher than those of the radiomics model (training dataset: AUC = 0.916; testing dataset: AUC = 0.770) and the clinical model (training dataset: AUC = 0.857; testing dataset: AUC = 0.737). However, the difference was statistically significant with the clinical model only in the training dataset, while no significant difference was observed in the testing dataset. This underscores the importance of rigorous external validation and suggests that the added value of combining radiomics with clinical features in this specific dataset requires further investigation with larger sample sizes. The combined model did not significantly outperform the radiomics model, but it still had high value in terms of its accuracy (combined model vs. radiomics model: 0.892 vs. 0.833) in the testing dataset. The clinical model demonstrated the highest sensitivity (0.966) and NPV (0.933), whereas the combined model achieved the highest specificity (0.931) and PPV (0.925) in the testing dataset. These findings indicate that the clinical model was more effective in identifying neonates with a poor prognosis, while the combined model provided superior accuracy in evaluating those with a favorable prognosis. In a study conducted by Wagner M. W. et al. [30], a total of 166 preterm neonates born between 24 and 32 weeks’ gestation underwent early-life brain MRI, and radiomic features were extracted from T1-weighted and T2-weighted sequences. Motor, cognitive, and language outcomes were evaluated at corrected ages of 18 months, 33 months, and 4.5 years. The authors reported that radiomic features outperformed clinical variables in predicting adverse motor outcomes, and that the inclusion of clinical variables further enhanced the predictive performance of the radiomics model. Based on these findings and our present results, future research may explore the potential of radiomics combined with clinical factors for long-term prognostic prediction in neonates with HIE. Although this study is a single-center retrospective analysis, the single-center design ensures uniformity in data acquisition and processing procedures, thereby minimizing heterogeneity during the preliminary phase of model development. As an essential first step toward establishing prognostic prediction models for HIE, this design contributes to the methodological rigor required for future, more refined studies aimed at predicting the long-term prognosis of full-term newborns. A more precise understanding of each infant’s potential outcome allows clinicians to formulate individualized treatment strategies tailored to the specific needs and risks of each patient. For instance, infants predicted to develop severe neurological impairments may benefit from more aggressive therapeutic interventions, such as hypothermia therapy, neuroprotective pharmacological agents, or early rehabilitation programs, whereas those with milder prognostic profiles may follow less intensive management pathways. By accurately predicting specific neurological deficits or developmental delays, the model can facilitate the early implementation of targeted interventions, including physical therapy, occupational therapy, or speech therapy, thereby optimizing functional recovery and improving overall quality of life. Conversely, for infants predicted to have favorable outcomes, the model can assist clinicians in avoiding unnecessary or excessively aggressive treatments, thus reducing the risk of complications and enhancing patient comfort and well-being. Therefore, this work provides foundational insights and hypotheses that warrant further investigation through larger, prospective, and multicenter studies. There were still some limitations in this study. Firstly, similar to many previous investigations, this was a retrospective single-center study, which may introduce selection bias and limit the generalizability of the results [31]. To address this, we have initiated multicenter collaborations to externally validate the model. This work is currently in progress and will involve a larger and more diverse patient cohort to ensure a more robust evaluation. Secondly, in this study, image segmentation was performed through manual delineation, which is both time-consuming and inevitably subject to human variability. To enhance the objectivity and reproducibility of our methodology, we plan to explore semi-automatic and deep learning-based automatic segmentation techniques in future research. Thirdly, the current model was constructed exclusively using conventional MRI sequences, without incorporating the potentially richer biological and functional information available from advanced imaging sequences. Although this design improves the clinical feasibility of the model in routine settings, the integration of multiparametric MRI represents a meaningful direction for future development. We intend to include more sensitive MRI modalities, such as diffusion weighted imaging (DWI) and susceptibility weighted imaging (SWI), to further enhance the prognostic accuracy of the model. In addition, a comprehensive cost-effectiveness analysis should be conducted in the future to assess the true health-economic value of implementing the model in clinical practice. Despite these limitations, this study provides a promising radiomics-based tool for predicting outcomes in HIE and establishes a solid foundation for subsequent validation and translational research. In conclusion, both the clinical and radiomics models developed in this study demonstrated promising predictive performance for evaluating the prognosis of HIE. Building upon these findings, the combined model integrating clinical variables with MR-based radiomic features may help further improve prognostic accuracy. Since this model is objective, non-invasive and reproducible, it holds potential for clinical application in the early prediction and management of neonatal HIE outcomes. Acknowledgements Not applicable. Abbreviations Hypoxic-ischemic encephalopathy Receiver operating characteristic Area under the curve Magnetic resonance imaging T1-weighted imaging T2-weighted imaging Fluid-Attenuated Inversion Recovery Bayley scales of infant and toddler development-III Gross motor function classification system Repetition time Echo time Support vector machine Leave-one-out cross-validation Regions of interest Volume of interest Intraclass correlation coefficient Pearson Correlation Coefficient Recursive Feature Elimination Standard deviation Positive Predictive Value Negative Predictive Value Arterial spin labeling Diffusion weighted imaging Susceptibility weighted imaging Author contributions Guarantor of integrity of the entire study: Ke Jin and Ting Yi.Study concepts and design: Jing Tang, Ting Yi and Ke Jin.Literature research: Jing Tang and Ting Yi.Clinical studies: Jing Tang and Siping He; Experimental studies and data analysis: Ting Yi, Jing Tang and Yonghua Xiang.Statistical analysis: Ting Yi and Yuqing Liu.Manuscript preparation: Jing Tang.Manuscript editing: Jing Tang and Ting Yi. Funding This study was supported by the Hunan Provincial Science and Technology Innovation Plan (Natural Science Foundation of Hunan Province, Clinical Medical Technology Innovation Guidance Project, Grant No. 2021SK50511). Data availability The datasets generated and analyzed during the current study can be made available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate This retrospective study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Hunan Children’s Hospital, Central South University (Approval No. HCHLL-2025-159). The requirement for written informed consent was waived by the committee due to the use of anonymized data and the retrospective nature of this study, which is consistent with national regulations and does not adversely affect the rights or welfare of the participants. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Ting Yi, Email: yiting127@126.com. Ke Jin, Email: jinke001@sina.com. References Associated Data Data Availability Statement The datasets generated and analyzed during the current study can be made available from the corresponding author upon reasonable request.