Clinically Applicable Machine Learning Approach to Predict Intracerebral Hematoma Expansion

Abstract

Background

Hematoma expansion (HE) is a significant risk factor for poor prognosis in patients with intracerebral hemorrhage (ICH). Accurately predicting HE is crucial for determining optimal treatment strategies.

Methods

This study enrolled 452 patients with ICH from 10 hospitals. To predict HE, 28 clinical variables available on patient arrival (including medical history, ICH location, and ICH volume) and 1142 radiomics features extracted from noncontrast computed tomography images of the ICH regions were used. Clinical variables and radiomics features were selected using gradient boosting and the least absolute shrinkage and selection operator. Three HE prediction models were built on clinical variables alone, radiomics features alone, and a third combining both. The models were compared using 5‐fold cross‐validation, and the mean area under the receiver operating characteristic curve was calculated for each. Additionally, the important features of HE prediction in the combined model were explored.

Results

The combined model demonstrated the highest performance for predicting HE with a 5‐fold mean area under the receiver operating characteristic curve of 0.77±0.05, compared with 0.70±0.06 for the clinical variables alone and 0.73±0.04 for the radiomics features alone. Permutation feature importance analysis suggested that anticoagulant treatment was the most predictive of HE.

Conclusions

A predictive model for HE was developed using the medical history, clinical features available on the patient’s arrival, imaging, and radiomics features extracted from computed tomography images. This prediction model will assist non–stroke care specialists in making treatment decisions for ICH in emergency settings.

Nonstandard Abbreviations and Acronyms

ES

emergent surgical treatment

HE

hematoma expansion

ICH

intracerebral hemorrhage

IVH

intraventricular hemorrhage

J‐ASPECT

Nationwide Survey of Acute Stroke Care Capacity for Proper Designation of Comprehensive Stroke Center in Japan

NCCT

noncontrast computed tomography

NGTDM

neighboring gray‐tone difference matrix

Clinical Perspective.

What Is New?

• The combined model demonstrated the highest performance for predicting hematoma expansion, compared with the clinical variables alone and the radiomics features alone in multicenter cross‐validation.

• Our hematoma expansion model depends only on medical history, clinical variables available on the patient’s arrival, imaging, and radiomics features extracted from computed tomography images.

What Are the Clinical Implications?

• Our hematoma expansion prediction model will assist non–stroke care specialists in making treatment decisions for intracerebral hemorrhage in the emergency setting.

Intracerebral hemorrhage (ICH) affects >1 million people annually worldwide and is the deadliest and most disabling type of stroke. ¹ Hematoma expansion (HE) is an independent risk factor of poor outcomes. ² , ³ Predicting HE is a crucial focus for clinicians to determine effective medical treatments, such as rapid intensive blood pressure lowering, for patients with ICH. ⁴

HE prediction has been studied extensively. Clinical variables such as patient demographics, laboratory data, ICH volume, and imaging signs have been used for HE prediction. ⁵ , ⁶ , ⁷ The presence of active contrast extravasation within the hematoma, known as the spot sign, is a validated imaging biomarker of HE on computed tomography (CT) angiography. ⁸ Several noncontrast computed tomography (NCCT) imaging predictors, such as satellite, island, swirl, blend, and black hole signs, and hypodensities ⁷ , ⁹ have gained attention as promising predictors of HE. A key area of relevance is the standardization of definitions and rating methods of these markers and demonstrating an independent association between NCCT biomarkers and functional outcomes.

Radiomics ¹⁰ is a method used to quantify the features and textures of images, and subsequent data mining is used for improved decision support. Previous studies demonstrated the usefulness of radiomics features of ICH on NCCT images in predicting HE. ¹¹ , ¹² , ¹³ , ¹⁴ These studies extracted the region of interest of ICH on NCCT images and calculated radiomics features. Some studies ¹⁵ , ¹⁶ used radiomics features for the region of ICH, where the features of the intermediate layer from U‐Net, ¹⁷ which is pretrained for the semantic segmentation task for ICH, were used. A study compared the predictive abilities of HE between models on the basis of radiomics features and clinical variables. ¹⁸ They found that radiomics features showed higher performance in predicting HE compared with clinical variables. A recent study in a single institution reported higher predictive abilities of the deep learning approach for HE when using image‐based features and clinical variables, probably because some patient background information was not captured by NCCT images. ¹⁹ However, these studies using a deep learning approach for HE did not determine what variables or features are important for improving HE prediction.

Therefore, this multicenter study aimed to develop and validate a model for predicting HE by using combined clinical variables and radiomics features and compare the predictive ability of the combined model with clinical variables only and with radiomics features only, respectively. Additionally, we analyzed the best prediction model to identify which clinical variables are crucial for improving prediction performance. Furthermore, we compared our results with recent HE prediction studies to provide a comprehensive view.

Methods

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Study Population

Patients admitted for primary spontaneous ICH at the 12 hospitals participating in the J‐ASPECT (Nationwide Survey of Acute Stroke Care Capacity for Proper Designation of Comprehensive Stroke Center in Japan) study between April 1, 2010, and March 31, 2016, were retrospectively selected. Patients’ clinical backgrounds were collected from the Diagnosis Procedure Combination database, while preonset medications and time from onset to the first CT image were gathered from an additional questionnaire. First and second NCCT images were also collected from the participating hospitals. All 12 participating hospitals (Table S1) were training institutions certified by the Japan Neurosurgical Society, offering 24/7 surgical treatment for ICH following the guidelines for treating stroke by the Japan Stroke Society.

Figure 1 shows the inclusion and exclusion criteria for the HE prediction data set. We selected patients with the following characteristics: (1) diagnosis of primary spontaneous ICH with baseline NCCT acquired within 6 hours from onset ²⁰ ; (2) age ≥18 years; (3) availability of good‐quality (having no artifacts and no lack of slices) baseline and follow‐up NCCT images; and (4) follow‐up NCCT performed within 27 hours from the baseline NCCT. We excluded the cases on the basis of the following criteria: (1) traumatic brain injury; (2) incomplete data (missing NCCT time, no paired images of the first and second NCCT, or poor‐quality NCCT images); (3) secondary ICH (excluded by the International Classification of Diseases, Tenth Revision [ICD‐10]); (4) hemorrhagic transformation of ischemic stroke; (5) primary intraventricular hemorrhage; (6) emergent surgical treatment (evacuation and drainage) of ICH before follow‐up NCCT; (7) multiple ICHs; and (8) cases with ICH represented only on 1 slice of the first NCCT.

Figure 1. Intracerebral HE prediction data set exclusion and inclusion criteria.

CT indicates computed tomography; HE, hematoma expansion; and ICH, intracerebral hemorrhage.

The data set of patients who had emergent surgical treatment (ES) after the first NCCT imaging (ES data set) built in the training institutions of the Japan Neurosurgical Society may provide a unique opportunity to develop a clinical decision support system for surgical indication of ICH for non–stroke care specialists. Therefore, clinical variables and radiomics features of the first NCCT of this ES data set were compared with those of the HE analysis data set. Finally, the HE prediction data set (HE analysis data set) included 452 patients from 10 hospitals.

Cross‐validation was performed in this study. All feature selection and all hyperparameter tuning processes were carried out within the training data before the performance of the prediction model was evaluated on the test data. The overview of experiments is illustrated in Figure S1.

Clinical Variables

We used the following clinical variables for HE prediction: age, sex, smoking history, level of consciousness on admission (awake, arousable, and unarousable on the basis of the Japan Coma Scale), the first NCCT time from onset, medical history (intracerebral hemorrhage, transient ischemic attack, myocardial infarction, alcoholism, bleeding tendency, heart failure, liver disease, renal disease), medications (antiplatelet treatment, anticoagulant treatment, anticoagulant neutralizers, NSAIDs, antihypertensives, antidiabetes drugs, antihyperlipidemia drugs, statins), Systolic blood pressure (SBP) and coagulopathy were managed following local protocols and guidelines for acute ICH management at each participating institution. SBP and diastolic blood pressure on admission were excluded from clinical variables because blood pressure management, but not blood pressure at admission, is a risk factor for HE. ²¹

ICH Features From NCCT Images

The NCCT imaging protocol was followed by each participating hospital. The spatial resolution of the NCCT image was 512×512. The slice thickness of the collected NCCT images ranged from 2 to 10 mm. To calculate radiomics features, it is necessary to annotate the region of ICH. Ren, Mukae, and Arimura used 3‐Dimensional Slicer (https://www.slicer.org/) to annotate ICH and intraventricular hemorrhage (IVH). Concurrently, the volume of ICH was calculated as the sum of the number of voxels multiplied by the size. The presence or absence of IVH was summarized. ICH anatomic location and the presence of ICH or IVH were annotated and checked by 3 board‐certified neurosurgeons with >10 years of experience in stroke care, who were blinded to the patient’s clinical data.

Feature Extraction and Selection

The total number of clinical variables was 28. For feature selection of clinical variables, we used a gradient boosting decision tree; particularly, LightGBM ²² was used. LightGBM was chosen due to its ability to handle missing values and offer better interpretability than other machine learning models. Although LightGBM is a high‐performance model, it has many hyperparameters. For hyperparameter tuning, we used Optuna, ²³ a tree‐structured parzen estimator.

For NCCT images, we calculated a total of 108 basic radiomics features, which include first‐order statistics, shape‐based features, gray‐level co‐occurrence matrix, gray‐level run length matrix, gray‐level size zone matrix, neighboring gray‐tone difference matrix (NGTDM), and gray‐level dependence matrix. The number of each radiomics feature is 19, 14, 24, 16, 16, 5, and 14, respectively. These radiomics features were derived from the original image, Laplacian of Gaussian filtered image $\left(\sigma =1.0,2.0,3.0\right)$, and 8 kinds of wavelet filtered images, on the region of interest of ICH. The total number of radiomics features was $1142=\left(19+24+16+16+5+14\right)\times \left(1+3+8\right)+14.$ The NCCT images were set at window width 100 and window level 40 before calculating each filtered image. We set the bin width to 4 for calculating gray‐level co‐occurrence matrix, gray‐level run length matrix, gray‐level size zone matrix, NGTDM, and gray‐level dependence matrix. We calculated all radiomics features by PyRadiomics version 3.0.1 (Harvard Medical School, Boston, MA).

We selected radiomics features using least absolute shrinkage and selection operator, ²⁴ and features with nonzero coefficients were included. The least absolute shrinkage and selection operator regularization coefficient was set to 1.0 on the basis of grid search (0.0001, 0.001, 0.01, 0.1, 1.0, 10) results.

Experiments and Model Analysis

We used LightGBM for the HE prediction model and Optuna for hyperparameter tuning. The tuning hyperparameters included learning rate, number of leaves, rate of sampling features, frequency of bagging, rate of positive data bagging, rate of negative bagging, coefficient of L1 regularization, coefficient of L2 regularization, and minimum data size in leaf. We used only clinical variables that were selected by LightGBM and radiomics features that were selected by least absolute shrinkage and selection operator. For the test, we applied early stopping and ensemble learning of the LightGBM model, choosing bagging as the ensemble method. The number of ensemble estimators was 100. The final output at inference was the average probability predicted by 100 ensemble models.

We implemented stratified nested 5‐fold cross‐validation. In each fold, we performed feature selection and tuning hyperparameters and evaluated the performance for test data. At splitting folds, we adjusted each fold to include all hospitals to maintain fair evaluation. The predictive ability of the 3 different models, on the basis of clinical variables only, radiomics features only, and their combined data, was evaluated using receiver operator characteristic (ROC) curves to compare the area under the curves of the 3 models. We evaluated the average of 5‐fold cross‐validation of test area under the ROC curves.

After training and testing the models, we analyzed the HE prediction models. We explored the relationships between the number of trees and splits of gradient boosting decision trees and model performance. Moreover, we analyzed features selected in all folds because these features are considered important for HE prediction. Additionally, we used permutation feature importance, ²⁵ known as 1 of the effective means in the machine learning model analysis for the combination model to reveal the importance of clinical variables.

Finally, we compared our HE prediction model with other models reported in recent studies considering data set variety, number of patients, rate of HE, variables or features, model interpretability, and prediction performance.

Statistical Analysis

We analyzed 2 ES and HE data sets. Categorical variables were expressed as count (percentage) and compared using the ${\mathrm{\chi }}^2$ test. Continuous variables were expressed as median (interquartile range) or mean±SD on the basis of their distribution and compared using the Mann–Whitney $U$ test, and t tests, respectively. We defined HE as an increase in ICH volume of >6 mL or >33% ¹ , ²⁰ from the first NCCT within 27 hours as the main outcome of interest of this analysis. ¹⁵ We considered P<0.05 as statistical significance. Statistical analysis was performed on JMP software version 16.2.0 (JMP Statistical Discovery LLC, Cary, NC).

Ethical Approval

Ethical approval was obtained from the Kyushu University Institutional Review Board and the National Cerebral and Cardiovascular Center, which waived the requirement for individual informed consent.

Results

Among 629 patients with ICH who underwent the first NCCT within 6 hours of symptom onset, 149 patients (23.7%) received ES after the baseline NCCT (ES data set; Table S2). The remaining 452 patients met the inclusion criteria for HE prediction analysis, of whom 90 (19.9%) experienced HE. Compared with the HE analysis data set, patients in the ES data set were associated with lower levels of consciousness, larger hematoma volume, and IVH on the first NCCT and were more likely to have anticoagulation, antidiabetes agents before onset, and anticoagulant neutralizers after admission and cerebellar hemorrhage, and were less likely to have brainstem hemorrhage (Table 1). Compared with the stable group, patients in the expansion group were more likely to be comatose and have ischemic heart disease and renal disease, larger hematoma volume on the first NCCT, more likely to have anticoagulation and antidiabetes agents before onset, use of anticoagulant neutralizers after admission, and were associated with lower SBP and diastolic blood pressure at admission (Table 1).

Table 1.

Demographics of the Hematoma Expansion Prediction Data Set

Variables	Included	Stable	Expansion	P value	Missing, %
Number	452	362	90
Demographics
Age*	69.0 (59–78)	70.0 (59–78)	66.0 (57–78)	0.3287
Sex, male	262 (58.0)	209 (57.7)	53 (58.9)	0.9052
Weight, kg*	56.4 (47–65)	56.8 (48–65)	55.5 (45–66)	0.5527	2.4
Smoking history	120 (30.5)	95 (29.3)	25 (36.2)	0.2540	13.1
Admission Coma Scale	320 (70.8)	267 (73.5)	54 (60.0)	0.0177†
	78 (17.3)	57 (15.7)	21 (23.3)
	54 (11.9)	39 (10.8)	15 (16.7)
Admission SBP, mm Hg*	180.0 (160–201)	181.0 (162–204)	175.0 (150–195)	0.0119†	1.3
Admission diastolic blood pressure, mm Hg*	99.0 (84–113)	100.0 (87–113)	94.0 (78–110)	0.0131†	1.3
First CT time from onset, h*	1.7 (1–2)	1.7 (1–2)	1.4 (1–2)	0.0492†
Comorbidity
Hypertension	293 (64.8)	239 (66.0)	54 (60.0)	0.3239
Diabetes	65 (14.4)	51 (14.1)	14 (15.6)	0.7378
Hyperlipidemia	36 (8.0)	30 (8.3)	6 (6.7)	0.8277
Ischemic heart disease	21 (4.6)	10 (2.8)	11 (12.2)	0.0007†
Atrial fibrillation	23 (5.1)	15 (4.1)	8 (8.9)	0.1020
Peripheral artery disease	2 (0.4)	1 (0.3)	1 (1.1)	0.3589
Alcoholic hepatitis	4 (0.9)	2 (0.6)	2 (2.2)	0.1788
Medical history
Intracerebral hemorrhage	54 (13.6)	44 (13.9)	10 (12.5)	0.8562	12.2
Transient ischemic attack	67 (16.8)	50 (15.7)	17 (21.0)	0.2480	11.5
Myocardial infarction	29 (7.3)	20 (6.3)	9 (11.1)	0.1523	12.2
Alcoholism	69 (18.6)	55 (18.8)	14 (17.9)	1.0000	18.1
Bleeding tendency	6 (1.5)	5 (1.6)	1 (1.2)	1.0000	12.4
Heart failure	28 (6.2)	20 (5.5)	8 (8.9)	0.2285
Liver disease	39 (8.6)	31 (8.6)	8 (8.9)	1.0000
Renal disease	64 (14.2)	44 (12.2)	20 (22.2)	0.0181†
Medication
Antiplatelet treatment	73 (16.2)	54 (14.9)	19 (21.1)	0.3589
Anticoagulant treatment	79 (17.5)	43 (11.9)	36 (40.0)	<0.0001†
Anticoagulant neutralizers	40 (8.8)	23 (6.4)	17 (18.9)	0.0006†
NSAID	17 (4.3)	11 (3.5)	6 (7.6)	0.1248	13.5
Antihypertension	395 (87.4)	317 (87.6)	78 (86.7)	0.8592
Antidiabetes	66 (14.6)	50 (13.8)	16 (17.8)	0.3227
Antihyperlipidemia	3 (0.7)	1 (0.3)	2 (2.2)	0.1025
Statin	27 (6.0)	21 (5.8)	6 (6.7)	0.8035
Hemorrhage
First ICH, volume, mL*	7.3 (2–15)	6.5 (2–13)	13.0 (4–36)	0.0001†
IVH	208 (46.0)	168 (46.4)	42 (46.7)	0.8133
Location
Lobar hemorrhage	97 (21.5)	72 (19.9)	25 (27.8)	0.1148
Putaminal hemorrhage	248 (54.9)	192 (53.0)	56 (62.2)	0.1252
Thalamic hemorrhage	146 (32.3)	119 (32.9)	27 (30.0)	0.7057
Brainstem hemorrhage	97 (21.5)	77 (21.3)	20 (22.2)	0.8860
Cerebellar hemorrhage	52 (11.5)	46 (12.7)	6 (6.7)	0.1389

Values are expressed as number (percentage) unless otherwise indicated. CT indicates computed tomography; ICH, intracerebral hemorrhage; IVH, intraventricular hemorrhage; and SBP, systolic blood pressure.

^*Median (interquartile range). Differences in clinical characteristics were compared by t test, Mann–Whitney U test, or ${\mathrm{\chi }}^2$ test.

^† P values denote statistical significance.

Figure 2 illustrates the average of area under the ROC curve in 5‐fold cross‐validation of the models on the basis of clinical variables only (0.70±0.06, blue line), radiomics features only (0.73±0.04, red line), and clinical variables and radiomics features combined (0.77±0.05, green line). The combination model achieved the highest performance.

Figure 2. Comparison of the average ROC and AUC of 5‐fold cross validation.

The blue, red, and green lines represent clinical variables alone, radiomics features alone, and both clinical variables and radiomics features, respectively. The clinical variables and radiomics features model are the highest area under the receiver operator characteristic curves. AUC indicates area under the curve; and ROC, receiver operating characteristic.

Figure 3 illustrates the ratio of variables used as a split of decision trees. The vertical axis represents variables built for HE prediction in descending order. The blue and red bars are clinical variables and radiomics features, respectively. The horizontal axis represents the ratio of variables used as splits in decision trees in all models of all folds. A higher ratio means the variables are often selected as the split. Clinical variables and radiomics features are used with equal frequency. This result indicates that some clinical variables are important in predicting HE. The average number of trees in the models are based on clinical variables only or radiomics features only, and their combinations are 8.8, 8.5, and 11.6, respectively. The average number of splits of these 3 models is 51.6, 27.1, and 51.8, respectively (Table S3). The largest number of trees and splits is noted in the combination model. The radiomics features model had the smallest number of trees and splits.

Figure 3. The ratio of clinical variables and radiomics features used as a split of decision trees.

The vertical axis represents variables for built HE prediction by descending order. The blue and red bar are clinical variables and radiomics features, respectively. The horizontal axis represents the ratio of variables used as splits in the decision trees in all models of all folds. Higher ratios mean the variables are often selected as the split. CT indicates computed tomography; ICH, intracerebral hemorrhage; IVH, intraventricular hemorrhage; and TIA, transient ischemic attack.

Figure 4 illustrates the result of permutation feature importance for clinical variables of the combination model. The horizontal axis represents the variation of area under the ROC curve from the baseline, with the baseline representing the area under the ROC curve for the original test data set. The larger the perturbation, the more important the features are. This result shows that anticoagulant treatment is the most important clinical variable. Moreover, age, putaminal hemorrhage, IVH existence, the first CT time from onset, the first ICH volume, and smoking history are important variables.

Figure 4. Results of the permutation feature importance for clinical variables in the combination model.

The horizontal axis represents the difference between average of the 5‐fold ROC AUC of the test data set and permutated test data set. Negative values indicate declining model performance. The vertical axis represents clinical variables in feature selection. Clinical variables are ordered by the absolute mean of the ROC AUC. AUC indicates area under the curve; CT, computed tomography; ICH, intracerebral hemorrhage; IVH indicates intraventricular hemorrhage; ROC, receiver operating characteristic; and TIA, transient ischemic attack.

Table 2 compares our best model with the best models in recent previous studies.^* Our model validated multicenter data; used simple patient information, quantitative radiomics features, and an interpretable machine learning model; and showed stable performance by cross‐validation.

Table 2.

Comparison of Recent Hematoma Expansion Prediction Model Reports

	Number of institutions	Number of patients	HE ratio, %	Clinical variables		CT images			Model	Performance
				Patient information	Laboratory data	Imaging biomarkers	Radiomics features	Deep learning features
Liu et al 7 2019	1	1157	21	〇		〇			SVM	AUC 0.89
Ma et al 11 2019	1	254	31				〇		LASSO	Acc. 0.82
Xu et al 12 2020	2	129	45				〇		Nomogram	AUC 0.86
Ma et al 16 2022	1	253	23					〇	Deep learning	AUC 0.93
Li et al 18 2022	1	258	—	〇			〇		Logistic regression	AUC 0.66 AUC 0.83
Tingting et al 19 2022	1	140	41	〇				〇	Deep learning	AUC 0.91
Teng et al 15 2021	84	3134	20				〇	〇	Gradient boosting	AUC 0.78
Feng et al 26 2023	3	231	28				〇	〇	Nomogram	AUC 0.83
Haider et al 27 2023	Multicentric	897	27			〇	〇		Logistic regression	AUC 0.64
Chen et al 28 2024	2	309	19				〇		Logistic regression	AUC 0.78
Tanioka et al 29 2022	4	422	20	〇	〇	〇			k‐Nearest neighbor	AUC 0.79
Du et al 30 2024	2	604	20	〇		〇			Gradient boosting	AUC 0.82
Our study	10	452	20	〇		〇	〇		Gradient boosting	AUC 0.77 ± 0.05

Acc, indicates accuracy; AUC, area under the curve; LASSO, least absolute shrinkage and selection operator; and SVM, support vector machine.

Discussion

In this multicenter study, we successfully developed an HE prediction model that delivers accurate results and helps select appropriate treatments for patients with ICH in hospitals without a neurosurgical specialist using 5‐fold cross‐validation. Due to the limited number of patients, we did not perform a new/unique test. Several studies have used clinical and imaging biomarkers. A previous study included laboratory data; however, we did not use laboratory data, as not all facilities conduct blood tests in off hours. ²⁹ Another study used CT sign as an imaging biomarker, but it is not easy to input as a variable in the absence of specialists. ³⁰ Therefore, it is more quantitative to calculate image features semiautomatically. Although previous studies have achieved high performance using deep learning, it is difficult to interpret the models. ¹⁶ , ¹⁹ Based on these considerations, we constructed a more interpretable machine learning model based on multicenter patient data, using simple patient information and semiautomatic quantitative radiomics features. In extracting radiomics features, experts manually segmented ICH, and we established a method for automatic hematoma extraction and determination of the anatomic location of hemorrhage using the segmented data obtained during this study. The results of this research are reported elsewhere.

Our HE prediction model, which combined clinical variables and radiomics features, demonstrated a reliable performance. This model was built by the data set using NCCT images and clinical variables of the patients with spontaneous ICH from 10 participating training institutions in Japan. Our performance is reliable and robust because we adopted 5‐fold cross‐validation to evaluate this model and found that the performance of this model is consistent.

Feature selection is an important step in machine learning. We explored the selected features of clinical variables and radiomics features in all folds. The most frequently selected clinical variables are age, smoking history, admission coma scale, the first CT time from onset, anticoagulant treatment, and the first ICH volume. These variables are consistent with the factors related to the HE in the previous studies. ⁵ , ⁶ , ³¹ The most frequently selected radiomics features are Surface Area, maximum 2‐dimensional diameter row (the length of ICH in the sagittal axis), NGTDM‐Busyness, Total Energy‐LLL, and NGTDM‐Contrast‐LLL. Surface Area and maximum 2‐dimensional diameter are shape‐based features representing a range of hematomas. The rest of the radiomics features, NGTDM‐Busyness, Total Energy‐LLL, and NGTDM‐Contrast‐LLL, are characteristics of CT values. NGTDM‐Busyness and NGTDM‐Contrast‐LLL represent the spatial CT value change. NGTDM‐Contrast‐LLL is considered to be important in calculating radiomics scores. ¹⁸ These features suggest that the size of ICH and the impurity of CT value (blood component) are related to HE.

We explored the frequencies of variables used as splits in the decision tree. If the addition of clinical variables does not contribute to performance improvement, the usage ratio of clinical variables may be less than that of radiomics features. However, the combination model used clinical variables and radiomics features at the same frequency despite the difference in the number of variables. This suggests that clinical variables also affect the performance improvement in this HE model.

We also explored the number of trees and splits in this model. The combination model had the highest number of trees and splits. Generally, as the model complexity increases, there is a decline in generalization performance because of overfitting. However, the combination model showed higher performance than the model that was based on clinical variables only. From these effects, we suggest that the combination model achieved the highest performance while preventing overfitting.

The result of permutation feature importance indicates that anticoagulant treatment is the most important clinical variable. Additionally, age, putaminal hemorrhage, the first CT time from onset, IVH, smoking history, the first ICH volume, and history of renal disease also influenced the performance of the clinical variables. Consistent with this, a study suggested that hypodensities, a risk factor of HE, may be caused by anticoagulant treatment. ³² Current smoking is also reported to be a key factor in HE. ³¹ Some reports suggest that the location of hemorrhage is associated with HE, but conflicting results are reported. ³³ , ³⁴ Our analysis indicated that hemorrhage location is a predictive factor of HE.

In our HE prediction data set, we excluded patients who underwent surgical treatment after the first CT. Although this criterion is consistent with previous studies, ⁵ , ⁶ , ⁸ , ¹³ , ¹⁸ we considered whether excluding these patients resulted in the loss of crucial data. The results of statistical analysis between the surgical treatment group and the HE prediction group, as well as between the HE and the stable group, showed similar tendencies. Clinical variables considered particularly important in clinical settings, such as age, admission coma scale, anticoagulant treatment, anticoagulant neutralizer, and first ICH volume, showed significant differences. A worse admission coma scale was associated with a higher likelihood of surgical treatment or HE. In the HE and stable data sets, admission SBP and diastolic blood pressure showed significant differences, which aligns with a previous study. ²¹ Our data set could not collect the SBP and diastolic blood pressure during hospitalization. In feature selection based on machine learning, anticoagulant treatment was selected rather than anticoagulant neutralizer. The location of hemorrhage was notably significant in the brainstem and cerebellum. Infratentorial hemorrhages are associated with surgical difficulties. IVH showed a significant difference between the surgical treatment group and the HE prediction group because the surgical treatment group included patients who underwent ventricular hemorrhage drainage. The first ICH volume in the surgical treatment group is significantly larger than in the HE prediction group. The median of the first ICH volume of the hematoma expansion group is also double that of the stable group. However, it is assumed that the large ICH size does not necessarily lead to expansion because the ICH size cannot increase indefinitely. Factors such as the time elapsed from onset and other patient background characteristics play a role. Machine learning, specifically the gradient boosting decision tree used in this research, can comprehensively analyze ICH volume, clinical background, and radiomics features.

This study has certain limitations. First, NCCT slice thickness was substantial, leading to lower precision in measuring ICH volumes compared with a previous study with 1‐mm thickness. ¹⁸ However, the slice thickness may not significantly impact performance, as some studies ¹¹ , ¹² , ¹³ , ¹⁴ have demonstrated high performance despite thicker slices. Second, radiomics features may lack robustness due to the subjective nature of human annotation. To calculate radiomics features, it is necessary to detect and annotate the region of ICH. In this study, board‐certified neurosurgeons manually annotated and checked all regions of ICH and IVH. Nevertheless, these annotations were time consuming, subjective, and susceptible to software manipulation errors. If more human resources are available in the future for manual annotation of ICH and IVH, a time‐consuming process, results of interrater reliability may provide more robust evidence of the HE prediction models. Furthermore, standardized parameter settings for calculating radiomics have not been established. Future studies should focus on extracting image features using convolutional neural networks ¹⁵ , ¹⁹ or unsupervised learning to develop more robust features. Third, there are differences in each hospital’s background, staff, and policies. In the study, cross‐validation included hospital‐based stratification, but hospital‐based stratification, residual center‐specific practices (eg, CT acquisition protocols, treatment strategies) may still introduce bias into the model. Unfortunately, we did not have any detailed information on such residual center‐specific practices. Despite the lack of such information, we think that our HE prediction performance mitigates these differences between hospitals by adjusting the number of hospitals of each fold for the HE prediction data set. Finally, it is important to note that reliance on cross‐validation only may potentially lead to inadequate calibration.

Conclusions

This multicenter study developed HE models by combining clinical variables and radiomics features from NCCT images. The combination of clinical variables and radiomics features showed higher performance compared with models based solely on the clinical variables or radiomics features. Our study identified key clinical variables and mechanisms that improve HE prediction performance. The information gathered is easily accessible in clinical settings and may aid treatment decision making.

Sources of Funding

This study was funded by the Practical Research Project for lifestyle‐related diseases, including cardiovascular diseases and diabetes managed by the Japan Agency for Medical Research and Development (JP18hk0102038, JP19ek0210088, JP20ek0210129, JP20ek0210147, JP21ek0210147, and JP 22ek0210147); Grants‐in‐Aid from the Japanese Ministry of Health, Labour and Welfare (H28‐Shinkin‐Ippan‐011, 19AC1003, 21FA1010, 22FA1015, 23FA1014, 24FA1015, and 24FA1016); KAKENHI grants (25 293 314, 18H02914, 22H03191, and 23K24450; principal investigator: K.I.) from the Japan Society for the Promotion of Science; Intramural Research Fund (20‐4‐10) for Cardiovascular Diseases of National Cerebral and Cardiovascular Center; and Idorsia Pharmaceuticals Japan Ltd. (Nos. D602 and D712).

Disclosures

Dr Iihara reports grant support from Idorsia Pharmaceuticals Japan Limited and Nxera Pharma Co., Ltd. The other authors report no conflicts.

Supporting information

Data S1