A Nomogram Based on CT Radiomics and Clinical Risk Factors for Prediction of Prognosis of Hypertensive Intracerebral Hemorrhage

Purpose To develop and validate a clinical-radiomics nomogram based on clinical risk factors and CT radiomics feature to predict hypertensive intracerebral hemorrhage (HICH) prognosis. Methods A total of 195 patients with HICH treated in our hospital from January 2018 to January 2022 were retrospectively enrolled and randomly divided into two cohorts for training (n = 138) and validation (n = 57) according to the ratio of 7 : 3. All CT radiomics features were extracted from intrahematomal, perihematomal, and combined intra- and perihematomal regions by using free open-source software called 3D slicer. The least absolute shrinkage and selection operator method was used to select the optimal radiomics features, and the radiomics score (Rad-score) was calculated. The relationship between Rad-score, clinical risk factors, and the HICH prognosis was analyzed by univariate and multivariate logistic regression analyses, and the clinical-radiomics nomogram was built. The area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA) were used to evaluate the performance of the clinical-radiomics nomogram in predicting the prognosis of HICH. Results A total of 1702 radiomics features were extracted from the CT images of each patient for analysis. By univariate and stepwise multivariate logistic regression analyses, age, sex, RBC, serum glucose, D-dimer level, hematoma volume, and midline shift were clinical risk factors for the prognosis of HICH. Rad-score and clinical risk factors developed the clinical-radiomics nomogram. The nomogram showed the highest predictive efficiency in the training cohort (AUC = 0.95, 95% confidence interval (CI), 0.92 to 0.98) and the validation cohort (AUC = 0.90, 95% CI, 0.82 to 0.98). The calibration curve indicated that the clinical-radiomics nomogram had good calibration. DCA showed that the nomogram had high applicability in clinical practice. Conclusions The clinical-radiomics nomogram incorporated with the radiomics features and clinical risk factors has good potential in predicting the prognosis of HICH.


Introduction
Hypertensive intracerebral hemorrhage (HICH) is one of the most common types of intracerebral hemorrhage (ICH) [1]. Brain CT scanning is the standard imaging for diagnosing HICH, which is a practical method to determine the location and volume of HICH [2]. On the unenhanced CT image, the hemorrhage mainly showed a high-density mass shadow of hematoma and a rim of hypodensity around the hematoma. functional outcomes and make better treatment decisions. Te so-called CT signs, such as "blend sign," "island sign," "black hole sign," and "swirl sign," have been proved to help predict hematoma expansion [7][8][9][10], whereas these signs have limited in predicting the prognosis of HICH, and these signs are vulnerable to inter-or intraobserver variations [11]. Radiomics is a new research method, which refers to the high-throughput extraction and analysis of a large number of high-dimensional quantitative image features from different modes of medical images [12]. Its advantage is to convert visual image information into deep-seated features for quantitative research [13]. Moreover, radiomics allows multiple imaging features to be studied in parallel, which can provide a combination of features, and the analysis of combined features is more promising than that of singlefeature analysis [14]. Although some studies have reported positive results in the analysis of radiomics features [15][16][17], a new statistical model integrating the combined radiomics features and clinical risk factors has not been developed to predict the prognosis of HICH. Nomogram is a forecasting tool, which can transform the complex regression equation into a visual graph, so as to provide accurate and personalized medical services [18,19]. In the past, most studies mainly focused on the intrahematomal region, ignoring the perihematomal region [15,20]. Te predictive value of radiomics features in the perihematomal region is not clear.
In the present study, we hypothesized that combining clinical risk factors and CT radiomics features (including radiomics features of intra-and perihematomal regions) could identify HICH patients with poor prognoses. To verify the feasibility of our hypothesis, we extracted the radiomics features based on brain CT and established a clinicalradiomics nomogram integrating radiomics features and clinical risk factors through multivariate logistic regression analysis.

Patients.
Te local institutional review board approved the retrospective study, and the requirement for informed consent was waived. Te CT images and clinical data of HICH patients treated in our hospital from July 2018 to January 2022 were collected. Te inclusion criteria were as follows: (1) age >18 years old; (2) a history of hypertension; (3) the frst baseline brain CT scanned within 6 hours after the onset of symptoms; (4) brain parenchymal bleeding; (5) complete clinical data. Te exclusion criteria were as follows: (1) secondary ICH, such as cerebral aneurysm, trauma, arteriovenous malformation, tumor, or hemorrhagic infarction; (2) patients with anticoagulant-associated ICH; (3) motion artifacts on CT images; (4) patients who refused follow-up after discharge. Te fow chart is shown in Figure 1.
In addition, the following relevant clinical information was obtained through the patient's inpatient medical record system and the emergency medical record system: (1) clinical data included age, sex, systolic blood pressure, diastolic blood pressure, smoking history, drinking history, and diabetes history; (2) admission laboratory parameters included platelet count, white blood cell count, red blood cell (RBC) count, lymphocyte count, monocyte count, hemoglobin, serum glucose, and D-dimer level.

Functional
Outcome Assessment. 90 days after discharge, the patients were followed up by standardized telephone interviews, and the clinical functional outcome of HICH patients was assessed by Glasgow Outcome Scale (GOS). Referring to a previous study [21], we classifed the prognosis of patients into two categories: unfavorable outcome (GOS 1, death; GOS 2, persistent vegetative state; GOS 3, severe disability) and favorable outcome (GOS 4, moderate disability; GOS 5, return to normal life).

CT Image Acquisition and Data
Collection. All patients underwent brain CT with GE Optima CT660 64 row spiral scanner. Patients were placed in the supine position. Te head was placed in the head frame of the examination table, and the two external ear holes were equidistant from the table. At 120 kV tube voltage, 300 mA tube current, 512 × 512 matrix, and 5 mm slicer thickness, scanning was performed in a continuous cross section from the skull base to the skull top after taking a positioning image.
According to the CT images, the CT plain scan signs, such as hematoma volume, hematoma location, the degree of midline shift, whether it broke into the cerebral ventricle, herniation, and ventricular entrapment were obtained. Te imaging data were independently evaluated by a senior radiologist who was blinded to the clinical information. Te hematoma volume was calculated by using formula A × B × C/2 [22]. A is the longest diameter on the largest hematoma slice, B is the most signifcant diameter perpendicular to A, and C is the number of bleeding layers in CT multiplied by the slice thickness. Te degree of midline shift was determined according to previous literature [23].

Image Segmentation and Feature
Extraction. Te brain CT images of 30 patients were randomly selected to evaluate the interobserver agreement of feature extraction. Two experienced radiologists (readers 1 and 2) independently and manually completed the hematoma contour blinded to clinical data. Reproducibility of interobserver for drawing region of interest (ROI) was assessed by intraclass correlation coefcient (ICC). ICC value above 0.75 was considered to have good consistency, and all the remaining images were completed by Reader 1. In addition, we also captured the information around the hematoma from the surrounding area 6 mm away from the hematoma surface. According to the contour of the intrahematomal ROI (intra-ROI), we used the the "dilation" algorithm to automatically reconstruct the perihematomal ROI (peri-ROI) and obtained a ring of brain parenchyma around the hematoma. Figure 2 shows an example of drawing intra-ROI and peri-ROI.
All radiomics features were extracted by using free opensource software called 3D slicer (version 4.13, https://www. slicer.org). A total of 851 features were extracted from each ROI, which can be summarized into the following four groups: 14 volume and shape features (2D and 3D), 18 frstorder features, 75 texture features, and 744 wavelet transform features. Tree groups of features were obtained from the intra-ROI, peri-ROI, and their combined ROI (intra-ROI + peri-ROI).

Radiomics Feature Screening and Rad-Score Building.
Te minimum-redundancy maximum-relevance (mRMR) and the least absolute shrinkage and selection operator (LASSO) method were used for feature selection. Initially, mRMR was applied to eliminate redundant and irrelevant features. Ten, the LASSO algorithm was conducted to select the optimized feature subset using ten-fold cross-validation to build the fnal model. Te radiomics score (Rad-score) was calculated for each patient by a linear combination of selected features weighted by their respective coefcients. Based on the selected features of intra-ROI, peri-ROI, and their combined ROI, three radiomics models, which were intrahematomalbased model (intra-model), perihematomal ring-based model (peri-model), and combined model, were then established. Te workfow of radiomics analysis of hematoma is shown in Supplementary Figure 1.

Radiomics Feature Selection and Rad-Score Construction.
First, 1702 radiomics features were extracted from the depicted ROI, including manually segmented intra-ROI and automatically segmented peri-ROI. Te interobserver ICC ranged from 0.751 to 0.997, so these features had good repeatability. Next, 11 optimal radiomics features were selected from intra-ROI and peri-ROI by mRMR and LASSO, respectively, and then, incorporating intra-and peri-ROI features, ten radiomics features (seven from intra-ROI and three from peri-ROI) were selected ( Figure 3). Finally, the Rad-score was calculated for each patient using the formula provided in Supplementary Table 1 from the selected features.

Establishment of Radiomics, Clinical, and Clinical-Radiomics
Model. Tree radiomics models for predicting the prognosis of HICH were built based on radiomics features of intra-ROI, peri-ROI, and their combined ROI, respectively. We drew the ROC curves to compare the predictive accuracy of three radiomics models. Te specifc results are shown in Supplementary Figure 2. In the validation cohort, the AUC of the combined model was 0.90 (95% confdence interval (CI), 0.82 to 0.98), which was superior to the intra-model (AUC = 0.88, 95% CI, 0.79 to 0.97) and the peri-model (AUC = 0.82, 95% CI, 0.71 to 0.93). By univariate and stepwise multivariate logistic regression analyses, age, sex, RBC, serum glucose, D-dimer level, hematoma volume, and midline shift were independent predictors for the prognosis of HICH (Table 2). Ten, seven clinical risk factors were used to establish the clinical model. In addition, because the combined radiomics model has the best performance, the clinical-radiomics model was further established by integrating the combined Rad-score and clinical risk factors. Figure 4 shows the performance of the clinical model, combined radiomics model, and clinical-radiomics model in predicting the prognosis of HICH in the training and validation queues. Te clinicalradiomics model showed the highest discrimination in the training cohort in identifying patients with excellent and poor prognoses, with an AUC of 0.95 (95% CI, 0.92 to 0.98). Te AUC of the clinical-radiomics model was signifcantly higher than that of the clinical model (AUC � 0.87, 95% CI, 0.81 to 0.92) and the combined radiomics model (AUC � 0.90, 95% CI, 0.85 to 0.95). In the validation cohort, the AUC of the clinical-radiomics model of 0.90 (95% CI, 0.82 to 0.98) was superior to that of the clinical model (AUC � 0.84, 95% CI, 0.73 to 0.94). Te clinical-radiomics model showed the best performance in predicting the prognosis of HICH, surpassing all the other models, and had the highest prediction accuracy. Moreover, the ROC comparison verifed by Delong test showed statistical signifcance between nomogram model and clinical model (Z � 3.56, P < 0.01), suggesting that the clinical predicted net return of nomogram was higher than that of clinical model. Based on this best model, we generated a visualized clinical-radiomics nomogram ( Figure 5).

Clinical Application.
Te calibration curve showed that the clinical-radiomics nomogram had a good consistency and high calibration degree in predicting the prognosis of HICH and the actual results ( Figure 6). Moreover, the DCA curves of clinical-radiomics nomogram, clinical model, and combined radiomics model showed that clinical-radiomics nomogram had more excellent clinical utility, which indicated that the nomogram was a reliable clinical tool. In addition, DCA showed that these models were better than the "all treatment" and "no treatment" indexes in the training cohort in predicting the prognosis of HICH (Figure 7).

Discussion
HICH is a common disease in neurosurgery, accounting for 70%-80% of all cases of ICH. Te prognosis is poor, which will endanger the lives of patients and seriously afect people's health and quality of life [24]. Early and accurate prediction of the prognosis of HICH patients is the key to personalized treatment of HICH. In this study, to identify patients with poor prognoses, we developed and validated the clinical-radiomics nomogram based on the radiomics features and clinical risk factors. Te nomogram showed good performance in training and validation cohorts and was an easy-to-use personalized decision-making tool.
In terms of clinical characteristics, through multivariate logistic regression analysis, our study found that age, sex, RBC, serum glucose, D-dimer level, hematoma volume, and midline shift were the clinical risk factors for predicting the prognosis of HICH. Previous studies have shown that age, sex, hematoma volume, and midline shift can be used to predict functional outcomes in ICH patients [23,25,26], which is consistent with our results. In addition, we also found that the risk of poor prognosis at 90 days was related to RBC, serum glucose, and D-dimer level. Low RBC levels are associated with poor ICH prognoses, which may be partly due to impaired cerebral oxygenation [27]. Béjot et al. [28] found that admission hyperglycemia was associated with 1-month mortality and poor functional recovery at discharge. Moreover, basic studies have also confrmed the  Notes. WBC, white blood cell; RBC, red blood cell; Rad-score, radiomics score; OR, odds ratio; CI, confdence interval; NA, not available. efect of hyperglycemia on early hematoma expansion, mainly manifested in neuron death, angiogenic brain edema, and aggravation of blood-brain barrier damage [29,30]. Zhou et al. [31] indicated that elevated plasma D-dimer levels after ICH were associated with mortality and poor functional outcomes. Te increase in D-dimer level is related to progressive bleeding injury, which may refect the disturbance of cerebral microcirculation and systemic hypercoagulability [32]. Based on these clinical risk factors, we developed a clinical model to predict the prognosis of HICH patients. Te diagnostic efect of this model was good, with AUCs of 0.87 and 0.84 in the training and validation cohorts, respectively. Radiomics methods have great potential in promoting clinical decision-making by improving the accuracy of clinical diagnosis, prognosis prediction, and treatment response [33]. In the present study, we attempted to apply a novel combined intra-and perihematomal radiomics method to predict the prognosis of HICH. Compared with the research of Xu et al. [15], we not only analyzed the radiomics features of the intra-ROI but also explored the radiomics features of the peri-ROI. Previous study has shown that the perihematomal microenvironment is related to the pathophysiological process of hematoma expansion and may provide some potential predictive information [34]. In our study, three of the ten best features in the combined radiomics model were from peri-ROI, which indicated that the perihematomal region might provide incremental information. In addition, our fndings showed that in the validation cohort, the AUC of the combined radiomics model, incorporating intra-and peri-ROI features, was 0.90 (95% CI, 0.82-0.98), which was better than the single intra- and perimodel, and yielded the overall best prediction performance. Tis fnding indicated that the radiomics features of the peri-ROI might have potential value and deserved further exploration.
We used machine learning methods (radiomics features) to evaluate the characteristics of the hematoma itself and around the hematoma in order to better assess the heterogeneity of the hematoma. Rad-score can be used to quantitatively refect the characteristics of the hematoma itself after radiomics analysis, and it can be concluded through the logistic regression analysis that Rad-score is an independent variable of HICH prognosis. By adding Radscore to the clinical model, a clinical-radiomics nomogram was developed to promote further clinical application and accurately predict the prognosis of HICH. Compared with other models, this nomogram has further improved the performance of predicting the prognosis of HICH and achieved higher accuracy. Te AUCs of the training and validation cohorts were 0.95 and 0.90, respectively, which outperformed the single clinical characteristics and the radiomics features. Te doctor can add the scores of each prediction variable and get the total score according to the individual diferences of patients, so as to better help clinical decision-making and enable clinicians to develop personalized treatment plans for HICH patients. In addition, the calibration curve and DCA curve showed that the nomogram had good consistency and potential clinical applicability, and the maximum beneft was obtained under all thresholds.
However, the present research still has some limitations. Firstly, only the radiomics of the 6 mm perihematomal area was analyzed. Te predictive ability of other perihematomal region radiomics models with diferent distances to the prognosis of HICH patient needs to be further analyzed and studied. Secondly, this study is a single-center retrospective study, and the sample size is relatively small, which inevitably has some deviations; hence, large sample prospective and external validation studies are required.

Conclusions
In conclusion, this study established a clinical-radiomics nomogram, composed of radiomics features (including radiomics features of intra-ROI and peri-ROI) and clinical risk factors to identify HICH patients with poor prognoses. It can assist doctors in making clinical treatment decisions for patients with poor prognoses. Moreover, the clinicalradiomics nomogram shows potential value in precision medicine and designs personalized treatment strategies to better achieve personalized precision treatment.

Data Availability
Te data used to support the fndings of this study are available from the corresponding author upon request.

Ethical Approval
Tis study was approved by the Ethical Review Committee of the First Afliated Hospital of Shandong First Medical University.

Consent
Considering that this work was a retrospective study, the ethics committee waived the requirement for informed patient consent.

Conflicts of Interest
Te authors declare that they have no conficts of interest.

Authors' Contributions
Caiyun Fang wrote the original draft and designed the study. Xiao An, Kejian Li, Hui Shang, and Tianyu Jiao collected the data. Caiyun Fang and Juntao Zhang analyzed the data and built the prediction models. Qingshi Zeng revised the manuscript. All the authors read and approved the fnal version of the manuscript.

Supplementary Materials
Supplementary Figure 1: the workfow of the radiomics analysis of hematoma. Supplementary Table 1