Radiomics for predicting revised hematoma expansion with the inclusion of intraventricular hemorrhage growth in patients with supratentorial spontaneous intraparenchymal hematomas.

BACKGROUND: Previous radiomics analyses of hematoma expansion have been based on the traditional definition, which only focused on changes in intraparenchymal volume. However, the ability of radiomics-related models to predict revised hematoma expansion (RHE) with the inclusion of intraventricular hemorrhage expansion remains unclear. To develop and validate a noncontrast computed tomography (NCCT)-based clinical- semantic-radiomics nomogram to identify supratentorial spontaneous intracerebral hemorrhage (sICH) patients with RHE on admission.
METHODS: In this double-center retrospective study, data from 376 patients with sICH (training set: n=299; test set: n=77; external validation cohort: n=91) were reviewed. A radiomics model, a clinical-semantic model, and a combined model were then constructed based on the logistic regression machine learning approach. Radiomics features were extracted and selected by least absolute shrinkage and selection operator (LASSO) with 5-fold cross validation. Furthermore, the classical BRAIN scoring system was also constructed to predict RHE. Discriminative performance of the models was evaluated on the training and test set with area under the curve (AUC) and decision curve analysis (DCA).
RESULTS: The addition of radiomics to clinical-semantic factors significantly improved the prediction performance of RHE compared with the clinical-semantic model alone in the training (AUC, 0.94 vs. 0.81, P<0.05) and test (AUC, 0.84 vs. 0.71, P<0.05) sets, with similar results in the validation set (AUC, 0.83 vs. 0.69, P<0.05). Moreover, the discrimination efficacy of the BRAIN score was significantly lower than the other 3 models (AUC of 0.71 in the training set, P<0.05).
CONCLUSIONS: The clinical-semantic-radiomics combined model had the greatest potential for discriminating RHE, and significantly outperformed the classical BRAIN scoring system. 2022 Annals of Translational Medicine. All rights reserved.

Introduction

Spontaneous intracranial hemorrhage (sICH), accounting for 10–15% of all strokes, is the most devastating subtype of stroke, associated with high mortality, unfavorable outcome, and high treatment cost (1-3). Patients with sICH may continue to bleed after the initial hemorrhage, which is known as hematoma expansion (HE). HE, which occurs in about one-fifth of patients with sICH, is a well-established independent predictor of worse outcomes and early clinical deterioration (4,5). HE represents a plausible therapeutic target for anti-expansion treatment in many clinical trials due to the characteristic of being potentially preventable (6-8). So far, there is no consensus definition of HE in different trials (9-12). More recently, many research results have revealed that intraventricular hemorrhage (IVH) occurs in up to 50% of patients with sICH, which growth independently predicts poor outcome, and have recommended that it be incorporated into definitions of HE (13-16). Yogendrakumar et al. concluded that the revised definitions including new IVH development or expansion improve overall prediction of 90-day outcomes compared to a standard definition of HE (≥6 mL or ≥33%) (17). Recent trials failed to improve outcomes for ICH therapies, despite a reduction in HE (18-20). Therefore, extending the current HE definition to the revised hematoma expansion (RHE) definition with the inclusion of IVH growth may be conducive to anti-expansion treatment. Previous studies have validated that several noncontrast computed tomography (NCCT) markers such as black hole sign, bend sign, swirl sign, satellite sign, and island sign are effective predictors for HE (21-26). However, these qualitative predictors have relatively low accuracy and sensitivity (27), which may lead to missed diagnoses and misdiagnoses, affecting the achievement of precision medicine. As a noninvasive method for the objective assessment of lesion heterogeneity, radiomics is an emerging tool that extracts high-throughput quantitative features from medical images (28,29). By using machine learning (ML) methods such as logistic regression, random forest, support vector machines, and the Bayesian algorithm, radiomics analysis integrates quantitative imaging features and clinical findings. In conjunction with other clinical information, these quantitative features can support clinical decision making (29). Although radiomics analyses have been recently applied to predict HE (12,30-33), few studies have tried to predict RHE. In this study, we hypothesized that the nomogram model, which incorporates clinical-semantic and radiomics features, can be used to identify supratentorial sICH patients at high risk for RHE and can outperform the other 2 models alone as well as the BRAIN score model (9), serving as a promising clinical prediction model for HE.

We present the following article in accordance with the TRIPOD reporting checklist (available at https://dx.doi.org/10.21037/atm-21-6158).

Methods

All procedures performed in this study involving human participants were in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by Ethics Committee of Qilu Hospital of Shandong University (No. KYLL-2019014). Individual consent for this retrospective analysis was waived.

Patient selection and data collection

illustrates patient enrollment. The data of 1,790 patients with a primary diagnosis of sICH admitted to center 1 (Qilu Hospital of Shandong University, Qingdao) and center 2 (the Affiliated Hospital of Qingdao University) from January 2015 to December 2020 were retrospectively reviewed. Patients aged ≥18 years were eligible for our study if the following criteria were met: (I) Initial CT scan was completed within 24 h of ictus, and (II) follow-up CT scan was performed within 48 h after the baseline CT scan. The exclusion criteria were as follows: (I) secondary hemorrhage due to an external cause such as trauma, hemorrhagic transformation after acute ischemic stroke, tumor, aneurysm, cerebral aneurysm, or vascular malformation; (II) inadequate data or severe image artifacts; (III) surgical interventions before follow-up CT scan; (IV) subtentorial ICH; (V) multifocal ICH. Finally, 376 patients with sICH from center 1 and 91 patients with sICH from center 2 were included in the current study. Demographic information (age, gender), clinical data (baseline Glasgow Coma Scale (GCS), onset-to-baseline CT time), laboratory results (international normalized ratio (INR), glucose, D-dimer), medical history (intracerebral hemorrhage, ischemic stroke, antiplatelet use), and radiological data (ICH location, IVH, baseline hematoma volume, swirl sign, satellite sign) were recorded in a standardized data collection form.

Figure 1

Flowchart of patient enrollment and exclusion criteria.

Imaging acquisition

All NCCT scans were performed by scanners using a standard protocol, with details described in the Appendix 1.

Imaging analysis

RHE was defined as an increase in hematoma volume of >33% or >6 mL or any IVH at follow-up CT scan (17). Two neuroradiologists independently reviewed all images in a blinded manner. They assessed the NCCT markers on baseline CT scans according the guideline (34). Hematoma location was classified as deep (basal ganglia and thalamus) or lobar (34). Discrepancies were settled by joint discussion of the 2 readers and a third reader. Baseline and follow-up hematoma were semi-automatically detected and segmented by a computer-assisted software named “Dr. Wise Hemorrhagic Stroke Analyzer” with deep convolutional neural networks.

BRAIN score

Based on 5 independent predictors including baseline ICH volume, recurrent ICH, anticoagulation with warfarin at onset, intraventricular extension, and number of hours to baseline CT from symptom onset, BRAIN scores were obtained for every patient according to the definition of Wang et al. (9). ICH volumes were obtained with the aforementioned software automatically. Being the most widely utilized prognostic model for HE based on NCCT images in recent clinical trials (9,35), the BRAIN score was chosen as the comparison since there is currently no recognized prediction score for RHE.

Feature extraction

The workflow of the radiomics analysis of hematoma is shown in , which includes the steps of region of interest (ROI) segmentation, feature extraction, feature selection, model construction, model comparison, and evaluation.

Figure 2

The workflow of the radiomics analysis of hematoma.

We used the Deepwise multimodal research platform (https://keyan.deepwise.com, V1.6.2) to analyze the original DICOM format CT images of all cases.

Firstly, different image transformation methods were used to preprocess the original image, including high-pass filters, low-pass filters, wavelet filters, and Gaussian Laplacian filters with different λ parameters. Secondly, 3 types of radiomics features (first-order features, shape features, and texture features) were extracted from the original image and the preprocessed image, along with texture features consisting of Gray Level Co-occurrence Matrix (GLCM) features, Gray Level Run Length Matrix (GLRLM) features, Gray Level Size Zone Matrix (GLSZM) features, Gray Level Dependence Matrix (GLDM) features, and Neighboring Gray Tone Difference Matrix (NGTDM) features.

After feature extraction, each ROI extracted 2106 high-dimensional radiomics features, and the Z-score was used to standardize these features [Z = (X − mean)/SD].

Feature analysis

Firstly, we eliminated features with an intraclass correlation coefficient (ICC) value less than 0.8. We retained 1,492 features from the 2,106 radiomic features extracted from the segmented hematoma region of the initial CT image. Secondly, ANOVA was used to test the significance of a single feature between the hematoma-free and hematoma-enlarged group, and the statistically significant features were initially selected (P value <0.05). Then, the least absolute shrinkage and selection operator (LASSO) and logistic regression algorithms were used to select features and build ML models. The logistic regression machine learning was performed based on logistic regression analysis, searching the most fitting formula. Finally, 20 radiomics features with non-zero covariance became image predictors of HE. Details are presented in Table S1. In clinical and semantic feature analysis, univariate analysis was used to screen out features which had significant differences between groups (P<0.05).

Construction of multiple models

Firstly, we randomly divided the data set into a training set and a test set at a ratio of 8:2. In the training phase, in order to reduce the impact of the imbalance of positive and negative samples on the ML model, we first used the smote algorithm to resample the data and make the number of positive and negative samples consistent. In the training set, the model was trained after the data was divided by a 5-fold cross validation method. The performance of the model was tested on the test set.

In order to compare the prediction effects of clinical features, semantic features, radiomics features, and BRAIN scores on the classification of RHE and analyze their clinical value, we constructed 4 models, including: (I) BRAIN scores (BS); (II) clinical semantic model (CSM); (III) Radiomics model (RM); (IV) clinical semantic-radiomics model (CSRM), which contains significant clinical features, CT semantic features, and radiomics features.

We used logistic regression algorithms to train classification models. Five different ML algorithms, including the naïve Bayes (NB), random forest (RF), logistic regression (LR), support vector machine (SVM), and gradient boosting decision tree (GBDT) algorithms were used to construct models in the training cohort. These models were independently verified in the validation cohort. We used the output probability of the radiomics model as the Rad score, which can quantify the prediction probability of the radiomics model composed of high-dimensional features.

xk is one of the values of the radiomics features, ak is the coefficient of feature k.

In the CSRM, we combined clinical features, semantic features, and the Rad score.

Model evaluation

We used the area under the curve (AUC), accuracy, specificity, and sensitivity to evaluate the predictive performance of each model. Decision curves were drawn to display the relationship between the benefits and risks brought by different cut-off points in the different models. Cut off value was chosen as the biggest AUC value. Analysis of the calibration curves was performed for the calibration of the nomogram.

Statistical analysis

The statistical description of data was performed with R software (3.6.0) and Deepwise DxAI platform (https://dxonline.deepwise.com). ML algorithms and model evaluation were written in Python. Numerical variables were tested for normality. The independent sample t test was used for variables with a normal distribution, while the Wilcoxon rank sum test was used for variables with a non-normal distribution. Chi-square analysis or adjusted chi-square analysis were used for categorical variables. The DeLong test was used to compare the significance of the AUC difference of different models, and Spearman correlation was used to analyze the relationship between clinical features. In the bilateral significance test, P<0.05 was considered to be statistically significant.

Results

Patient characteristics

A total of 376 patients were enrolled in the study, including 236 males (62.8%) and 140 females (37.2%), with an average age of 59.9±13.3 years. Patients were assigned into 2 groups: a revised expander group (n=108) and a non-revised expander group (n=268). The patient characteristics in the training and test sets are summarized in .

Table 1

Clinical characteristics of the training set and test set

Variable	Training set (n=299)				Test set (n=77)
	No RHE (n=215)	RHE (n=84)	P value		No RHE (n=53)	RHE (n=24)	P value
Age (years)	58.479±12.931	62.048±13.812	0.036a		63.679±12.643	56.750±14.411	0.036a
Male (%)	139 (64.7)	50 (59.5)	0.409c		31 (58.5)	16 (66.7)	0.496c
GCS	13 (11, 15)	10 (6.75, 13)	0.001b		13 (10, 14)	12 (9, 12.25)	0.041b
Onset to follow-up CT, h	4.000 (2.000–13.000)	2.500 (1.000–5.000)	0.001b		4.000 (2.500–12.000)	1.650 (1.375–4.875)	0.001b
Serum glucose, mmol/L	6.670 (5.510–8.140)	6.750 (6.115–9.000)	0.088b		6.480 (5.720–7.160)	7.460 (5.820–9.310)	0.192b
Platelet count, 109/L	216.000 (181.250–258.500)	217.500 (176.500–255.000)	0.787b		201.000 (163.750–240.250)	229.500 (197.000–270.250)	0.031b
INR	1.060 (1.000–1.110)	1.030 (0.965–1.105)	0.179b		1.065 (1.000–1.130)	1.045 (1.018–1.062)	0.218b
APTT, s	29.400 (27.300–32.300)	30.400 (27.500–33.700)	0.186b		30.450 (28.275–32.325)	29.350 (28.025–30.825)	0.380b
D-dimer, ug/mL	0.330 (0.190–0.655)	0.510 (0.310–0.965)	0.005b		0.470 (0.247–0.797)	0.375 (0.180–1.163)	0.667b
Baseline ICH volume, mL, n (%)	12.577 (4.511–23.691)	32.544 (9.764–49.477)	0.000b		18.369 (8.181–26.096)	33.123 (24.125–55.783)	0.002b
Prior cerebral hemorrhage, n (%), present	24 (11.2)	7 (8.3)	0.471c		7 (13.2)	6 (25.0)	0.342c
Prior ischemic stroke, n (%), present	25 (11.6)	9 (10.7)	0.823c		8 (15.1)	4 (16.7)	1.000c
Diabetes, n (%), present	21 (9.8)	7 (8.3)	0.702c		7 (13.2)	5 (20.8)	0.606c
Hypertension, n (%), present	156 (72.6)	57 (67.9)	0.420c		40 (75.5)	17 (70.8)	0.667c
Coronary artery disease, n (%), present	20 (9.3)	18 (21.4)	0.005c		9 (17.0)	4 (16.7)	1.000c
Alcohol consumption, n (%), present	55 (25.6)	12 (14.3)	0.101c		14 (26.4)	3 (12.5)	0.297c
Smoking, n (%), present	47 (21.9)	16 (19.0)	0.757c		14 (26.4)	6 (25.0)	0.783c
Previous anticoagulant therapy, n (%), present	13 (6.0)	5 (6.0)	0.278c		2 (3.8)	2 (8.3)	0.779c
Previous antiplatelet therapy, n (%), present	24 (11.2)	25 (29.8)	0.000c		11 (20.8)	4 (16.7)	0.913c
Hematoma location, n (%), deep	169 (78.6)	60 (71.4)	0.188c		38 (71.7)	16 (66.7)	0.655c
Black hole sign, n (%), present	27 (12.6)	23 (27.4)	0.002c		4 (7.5)	5 (20.8)	0.194c
Blend sign, n (%), present	47 (21.9)	29 (34.5)	0.024c		21 (39.6)	11 (45.8)	0.609c
Swirl sign, n (%), present	69 (32.1)	50 (59.5)	0.000c		16 (30.2)	9 (39.1)	0.446c
Island sign, n (%), present	80 (37.2)	48 (57.1)	0.002c		20 (37.7)	18 (75.0)	0.002c
Satellite sign, n (%), present	51 (23.7)	34 (40.5)	0.004c		17 (32.1)	16 (66.7)	0.004c
Midline shift, n (%), present	160 (74.4)	36 (42.9)	0.000c		34 (64.2)	17 (70.8)	0.566c
IVH, n (%), present	45 (21.0)	35 (41.7)	0.000c		37 (69.8)	11 (45.8)	0.044c

a, two sample t test; b, Wilcoxon test; c, Chi-square test. RHE, revised hematoma expansion; GCS, Glasgow Coma Scale; INR, international normalized ratio; APTT, activated partial thromboplastin time; IVH, intraventricular hemorrhage.

BRAIN scores

The AUC of the training set was 0.7147, and the AUC of the test set was 0.6563 (). Comparing the model with the BRAIN score helps to better analyze the diagnostic significance of different models.

Table 2

Predictive performances of 4 models for hematoma expansion

Machine learning	Training					Test					Validation
	ACC	SEN	SPE	AUC		ACC	SEN	SPE	AUC		ACC	SEN	SPE	AUC
BS	0.6622	0.7738	0.5708	0.7147		0.6486	0.7917	0.5370	0.6563		0.6374	0.7714	0.4921	0.6437
CSM	0.7391	0.8095	0.6652	0.8065		0.6892	0.8333	0.5849	0.7054		0.6703	0.8000	0.5323	0.6864
RM	0.8696	0.8810	0.8304	0.9158		0.7568	0.8750	0.6731	0.8124		0.7582	0.8286	0.6557	0.7843
CSRM	0.9130	0.8929	0.8879	0.9429		0.8243	0.8148	0.9167	0.8447		0.8132	0.8571	0.7333	0.8287

BS, BRAIN score; CSM, clinical-semantic model; RM, radiomics model; CSRM, clinical-semantic radiomics model.

Radiomics models based on different ML algorithms

We tested 5 classic ML algorithms to screen out the best ML classifiers and found that the LR algorithm had the best performance on the training set (AUC =0.9158) and test set (AUC =0.8124). The results are shown in .

Table 3

Predictive performance of 5 models for RHE

Model	Training					Test
	ACC	SEN	SPE	AUC		ACC	SEN	SPE	AUC
SVM	0.8361	0.8452	0.7885	0.8545		0.7403	0.7500	0.6724	0.8012
RF	0.7559	0.7381	0.6949	0.8084		0.7013	0.7083	0.6271	0.7300
GBDT	0.7826	0.6905	0.7333	0.8204		0.6883	0.6667	0.6167	0.7240
LR	0.8696	0.8810	0.8304	0.9158		0.7568	0.8750	0.6731	0.8124
NB	0.7190	0.6429	0.6598	0.7328		0.6494	0.3333	0.6176	0.6349

ACC, accuracy; SEN, sensitivity; SPE, specificity; AUC, area under the curve; SVM, support vector machine; RF, random forest; GBDT, gradient boosting decision tree; LR, logistic regression; NB, naïve Bayes.

Analysis of clinical and semantic features

After univariate analysis of the clinical features and semantic features with significant differences between the groups, 7 clinical factors were finally selected. The results are shown in .

Table 4

Independent risk factors in the univariate analysis

Variable	Univariate analysis
	Odds ratio (95% CI)	P value
Rad score	12.492 (5.741–28.685)	<0.001
CT time from ICH onset, h	0.346 (0.203–0.581)	<0.001
Satellite sign, present	2.614 (1.551–4.422)	<0.001
Swirl sign, present	2.388 (1.432–4.009)	<0.001
D-dimer, ug/mL	0.983 (0.899–1.042)	0.0062
Baseline midline shift, present	3.398 (2.016–5.777)	<0.001
Serum glucose, mmol/L	1.049 (0.95–1.158)	0.00328
Baseline ICH volume, mL	2.204 (1.626–3.025)	<0.001

ICH, intracerebral hemorrhage; CI, confidence interval.

We constructed a classification model (CSM) based on clinical and semantic features. Its AUC on the training set was 0.81 and the AUC on the test set was 0.71 ().

Comparison between models

We compared the performances of all models on the training set and test set (), and tested their differences in AUC. On the training, test, and validation sets, the AUC of the CSRM was significantly higher than the RM, CSM, and BS (P<0.05). The AUC of the RM was higher than the CSM and BS (P<0.05). Among them, the CSRM had the best performance while the BS had the worst performance.

In order to intuitively quantify the probability of each factor in predicting the expansion of cerebral hematoma with the CSRM, we converted the value of each factor into a score, and calculated the total score of all risk factors to obtain the corresponding prediction probability of HE, as shown in . At the same time, we carried out model correction, and the results showed that the sample classification correctness of the CSRM was excellent.

Figure 3

Nomogram and calibration curves of the clinical semantic-radiomics model (CSRM). (A) Nomogram of the CSRM. (B) Calibration curves of the CSRM in the test set. (C) Calibration curves of the CSRM in the training set. (D) Calibration curves of the CSRM in the validation set.

shows the receiver operating characteristic curves (ROC curves) of the 4 models in the training set, test set, and validation set. The results showed that the AUC of the CSRM and RM were larger.

Figure 4

receiver operating characteristic (ROC) curves of different models. (A) ROC curves of different models in the training set. (B) ROC curves of different models in the test set. (C) ROC curves of different models in the validation set.

shows the decision curve analysis (DCA) curves of the 4 models on the training set and the test set. The clinical net benefits of the CSRM were always significantly higher than those of the RM, CSM, and BS on the training (), test (), and validation () sets.

Figure 5

Decision curve analysis (DCA) of different models. (A) DCA of different models in the training set. (B) DCA of different models in the test set. (C) DCA of different models in the validation set.

Discussion

In this retrospective study, we constructed a CSRM to discriminate early RHE after supratentorial sICH. This combined model can be completed in the ultra early phase of sICH and is only based on NCCT without contrast injection, which provides a fast and auxiliary method to identify patients with a high risk of RHE. Besides, with relatively lower examination cost and technical requirements, NCCT is widely used in almost every emergency room.

In the current study, several factors including midline shift, NCCT time from sICH onset, baseline ICH volume, D-dimer, serum glucose, swirl sign, and satellite sign were independently associated with RHE, which were pathologically plausible and constituted the CSM. Previous studies have indicated that midline shift was a predictor for poor outcome and HE in ICH patients (36,37), which is consistent with the findings of our study. The risk of RHE was associated with a short NCCT time from sICH onset, which may be due to further hematoma enlargement occurring within the first few hours after onset (38,39). The baseline hematoma volume was a potential predictor of HE, though with a contradictory effect. Our study result was consistent with most studies, in that the larger the volume, the higher the risk of HE (40). However, this is contrary to a small number of studies with a diametrically opposite conclusion (12). A larger baseline ICH may reflect multiple bleeding points or perilesional hemorrhage (9). D-dimer, as a marker of fibrin turnover reflecting an impairment in the coagulation and fibrinolysis pathways, may lead to early HE and may be related to intraventricular extension (41,42). Elevated serum glucose levels impair brain microvascular endothelial function and accelerate the destruction of the blood-brain barrier, which is likely to cause HE (43). Swirl sign, as a manifestation of heterogeneous density, constitutes more immature bleed areas leading to subsequent HE. This sign has been confirmed as an independent predictor of HE in previous studies (44), which is consistent with our results. Satellite signs have been explained as multifocal active bleeding from peripheral arterioles or reperfusion injury resulting from perihematomal edema, and were proven to be reliable predictors for HE and RHE (44). Our CSM demonstrated suboptimal discrimination performance, with AUCs of 0.81 in the training set and 0.71 in the test set.

This result indicates that it is not good enough to predict RHE based on extra-hematoma and visual assessment information alone. Therefore, quantitative predictors hidden in hematoma beyond vision have been sought.

In our study, the RM demonstrated better discrimination performance and higher sensitivity than the CSM in all 3 sets. By only requiring the delineation of the relatively well-defined boundaries of the hematoma rather than complex NCCT signs, the requirement for clinician experience is significantly reduced. An important aspect of constructing an NCCT radiomics-based model for prediction is eliminating unstable features caused by observer difference (45). After screening by the ICC value, ANOVA test, LASSO, and LR algorithms, only 20 of the 2,106 radiomics features associated with RHE of sICH were identified in this study. These 20 features contained the 3 types of aforementioned features as follows: first-order features, shape features, and texture features, which described the heterogeneity and shape of the hematoma associated with active bleeding and multifocal hemorrhage (46). The ability of only histogram-based texture features for quantifying hematoma heterogeneity in the prediction of HE has been previously verified (31). Subsequently, several studies based on R-score evaluated the risk of HE with AUCs of 0.69–0.89 (12,47), lower than our result of 0.92. Another study showed that the optimal accuracy (ACC) of RM in predicting HE was 0.88, which was slightly higher than our result of 0.87, but the sample size of 251 was significantly smaller than our study (32).

Furthermore, we compared the ability of 5 classic ML algorithms, LR, SVM, RF, GBDT, and NB, in constructing the RM to discriminate RHE after supratentorial sICH. As the tree-based ensemble model appeared to be overfitted on the training set, the regularized LR algorithm achieves a more robust performance in this study and the result is consistent with a previous study (48).

By adding the Rad score to the CSM, we developed the CSRM, which showed significantly better discriminating efficacy for RHE in sICH patients compared with the CSM in the training and test set (all P<0.05), indicating the statistical contribution of the R-score to the construction of the merged model. Additionally, the CSRM obtained more clinical net benefits in the prediction of RHE with almost all of the range of threshold probabilities compared with the CSM and RM. Compared with other research, more intra-hematoma and extra-hematoma characteristics are considered in our model, for more accurately predicting RHE. These results indicate that the combined application of intra-hematoma and extra-hematoma characteristics is expected to guide clinical practice and improve precision medicine. Compared with the test set, the results did not show a big change in the validation set, indicating that our model has good generalization performance.

The predictive ability of the BRAIN score for HE, which was derived from a prospective large sample study based on simple routinely available variables, has been proven (9), but has not been validated in RHE. In our study, the discrimination efficacy of the BS for RHE was only acceptable with an optimal AUC of 0.71 and an ACC of 0.66 in the training set and an AUC of 0.66 and ACC of 0.65 in the test set, which was significantly lower than the RM and CSRM (P<0.05). Furthermore, DCA showed its limitation in clinical practice. Although at this stage it is speculative and part of future studies, the proposed ML classifier may provide promising complementary results. The biggest drawback of the BS was the lack of NCCT markers describing the hematoma density, which might contribute to the relatively lower discernment. In this study, the top 5 predictors with first-order feature and second-order texture features in the RM are all linking to hematoma heterogeneity. Hence, the CSRM can be used as a supportive tool to augment conventional multidimensional scoring systems and to improve prognostic decisions for clinicians to achieve precision medicine.

One strength of our study is that the AUC and ACC of the RM and CSRM in the external validation set were similar to the test set, which means our models have stable predictive performance for RHE and can be further extended to other hospitals. Another strength of this study is the incorporation of the classical scoring system (BS) and the comparison of its prediction ability for RHE in supratentorial sICH patients with radiomics-related models (RM and CSRM). Since our quantitative feature analysis outperformed classical scoring systems, the application of the proposed ML approach may be of value for future randomized clinical trials. Challenges and opportunities to optimize clinical research and randomized trials in ICH are ongoing (49). The ML approach may simplify trial procedures by performing an imaging-based prediction of RHE. Furthermore, this approach may be applied in telemedicine and the remote prediction of RHE in poor regions lacking medical resources. Taken together, the proposed method integrates the advantages from quantitative radiomic and clinical semantic features and ML algorithms and relates the employed predictors to well-known imaging and clinical characteristics.

Several limitations deserve comment in this study. First, it was a double-center retrospective study with a relatively small sample size. Its reliability in other external cohorts needs to be validated with prospective, multi-center, large sample studies. Additionally, infratentorial hematomas were excluded. Therefore, the CSRM cannot be applied to patients with brainstem or cerebellar hemorrhage. Furthermore, perihematomal data such as edema was not included, which may limit the use of the CSRM and RM.

Conclusions

The addition of the Rad score to clinical-semantic factors can significantly improve the discriminating efficacy for RHE in sICH patients, and outperforms the classical BRAIN scoring system. These results support the potential of the CSRM to augment conventional image analysis, aid selection of those at risk of RHE, and simplify trial procedures. In the very near future, the CSRM may be incorporated into therapeutic trials once the challenges around predictive accuracy and variability of radiomics features have been settled.

The article’s supplementary files as