Construction method of deep learning-based short-term and long-term prognosis model for cerebral hemorrhage

Abstract

The application discloses a brain hemorrhage short-term and long-term prognosis model construction method based on deep learning, relates to the technical field of medical models, and has the technical scheme as follows: the method constructs a spontaneous brain hemorrhage population, evaluates the correlation between comprehensive multidimensional indexes such as comprehensive genes, biochemical indexes, CT / CTA signs within 24 hours after onset and demographic baseline data and the prognosis of the population, and comprehensively pays attention to the influencing factors of the brain hemorrhage prognosis; a prediction model for short-term and long-term poor prognosis of the spontaneous brain hemorrhage population is constructed by using deep learning and other machine learning methods, the model covers the prediction of short-term and long-term poor prognosis of brain hemorrhage, and is more comprehensive than the brain hemorrhage prognosis model in the prior art; the method performs internal and external verification on the prediction model, optimizes and improves the model, and improves the prediction accuracy and stability of the model to predict the prognosis of patients.

Description

Technical Field

[0001] This invention relates to the field of medical model technology, and more specifically, to a method for constructing short-term and long-term prognostic models of cerebral hemorrhage based on deep learning. Background Technology

[0002] Stroke, also known as cerebrovascular accident (CVA), is an acute cerebrovascular disease caused by the sudden rupture or blockage of blood vessels in the brain, resulting in brain tissue damage. It includes ischemic and hemorrhagic strokes. Ischemic stroke has a higher incidence than hemorrhagic stroke, accounting for 60%–70% of all strokes.

[0003] Spontaneous intracerebral hemorrhage (CHH) is a disease with high mortality and disability rates. It is a subtype of stroke, accounting for 10%-20% of all stroke types, second only to ischemic stroke in incidence. Hypertension is the most common cause and risk factor for CHH, which can lead to hyalinization of blood vessels, fibrinoid necrosis of smooth muscle cells, and the formation of microaneurysms, ultimately causing rupture of the vessel wall and resulting in hemorrhage. Other common causes of CHH include coronary artery disease (CAA), aneurysms, arteriovenous malformations, blood disorders (leukemia, aplastic anemia, etc.), and anticoagulation or thrombolytic therapy.

[0004] The high mortality, disability, and poor neurological recovery rates of intracerebral hemorrhage (IHH) have made its treatment and prognosis a major concern. Many factors influence prognosis, including level of consciousness, hemorrhage volume, age, and imaging findings. Previous prognostic models for IHH often have limited practicality due to their complexity; currently, there is no method to quickly and accurately determine or predict the prognosis of IHH. Prognostic imaging factors for spontaneous IHH involve multiple aspects, and short-term and long-term influencing factors differ. Furthermore, previously established models predicting poor prognosis for IHH are often difficult to replicate. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing short-term and long-term prognostic models for intracerebral hemorrhage based on deep learning. This method involves collecting genetic, biochemical, imaging, and baseline demographic data from a cohort of individuals with spontaneous intracerebral hemorrhage and conducting short-term (3 months) and long-term (12 months) follow-ups to obtain functional outcome data. Then, the correlation between various data and poor short-term and long-term prognoses of intracerebral hemorrhage is analyzed. Risk factors are selected, and machine learning methods such as deep learning are used to construct a predictive model for poor prognosis. Finally, the model is validated internally and externally.

[0006] The above-mentioned technical objective of this invention is achieved through the following technical solution: a method for constructing short-term and long-term prognostic models for cerebral hemorrhage based on deep learning, comprising the following steps:

[0007] S1. Collect relevant data on the spontaneous intracerebral hemorrhage cohort and conduct short-term and long-term follow-ups to obtain functional outcome data, with the mRS score as the criterion for determining functional outcome.

[0008] S2. Analyze the correlation between various data and poor short-term and long-term prognosis of cerebral hemorrhage, select risk factors to establish a deep learning-based intelligent prediction model for poor short-term and long-term prognosis of spontaneous cerebral hemorrhage, and use an mRS score ≥3 to determine poor short-term and long-term prognosis of cerebral hemorrhage. Use deep neural networks to learn and classify the model features.

[0009] S3. Validate the model using existing data, measure the performance of the prediction model, and evaluate the model, including but not limited to the evaluation of accuracy, sensitivity, and specificity.

[0010] The present invention is further configured such that: the relevant data collected in S1 from the spontaneous intracerebral hemorrhage cohort mainly include genetic, biochemical indicators, CT / CTA signs within 24 hours of onset, and demographic baseline data.

[0011] The present invention is further configured such that the gene refers to apolipoprotein E-APOE.

[0012] The present invention is further configured such that: the short-term follow-up in S1 is 3 months and the long-term follow-up is 12 months.

[0013] The present invention is further configured such that the specific operation process for establishing the deep learning-based intelligent prediction model for poor short-term and long-term prognosis of spontaneous intracerebral hemorrhage in step S2 is as follows:

[0014] 1) Data preprocessing: First, the collected data is preprocessed, including data cleaning, standardization and normalization, to reduce data noise and improve data interpretability;

[0015] 2) Feature extraction: Using deep learning methods, the collected data is converted into a format that can be processed by neural networks;

[0016] 3) Model building: Designing and building deep learning models to predict prognoses;

[0017] 4) Model training: Using training data to train the model;

[0018] 5) Model Evaluation: Use test data to evaluate the model's performance.

[0019] 6) Model optimization: Based on the model evaluation results, optimize the model, including but not limited to adjusting hyperparameters, increasing or decreasing the number of layers;

[0020] 7) Predictive applications: Use the optimized model to predict new prognostic models.

[0021] The present invention is further configured such that: the model verification in S3 is divided into internal verification and external verification according to the existing data of the model builder; internal verification is performed using its own existing data, and external verification is performed using data borrowed from other regions.

[0022] In summary, the present invention has the following beneficial effects:

[0023] 1. This method constructs a spontaneous intracerebral hemorrhage cohort and assesses the correlation between comprehensive and multidimensional indicators such as genes (apolipoprotein E-APOE), biochemical indicators, CT / CTA signs within 24 hours of onset, and demographic data and population prognosis, thus comprehensively considering the factors influencing the prognosis of intracerebral hemorrhage.

[0024] 2. A predictive model for poor short-term and long-term prognosis of spontaneous intracerebral hemorrhage was constructed using machine learning methods such as deep learning. This model covers the prediction of poor short-term and long-term prognosis of intracerebral hemorrhage and is more comprehensive than existing intracerebral hemorrhage prognosis models.

[0025] 3. This method performs internal and external validation on the prediction model, and optimizes and improves the model in a targeted manner to enhance the prediction accuracy and stability of the model in predicting patient prognosis. Attached Figure Description

[0026] Figure 1 This is a flowchart of data collection in an embodiment of the present invention;

[0027] Figure 2 This is a research roadmap for model construction in this embodiment of the invention;

[0028] Figure 3 This is the core of model construction in the embodiments of the present invention;

[0029] Figure 4 These are the verification and optimization steps of the model in the embodiments of the present invention;

[0030] Figure 5 This describes the specific implementation of the model in the embodiments of the present invention. Detailed Implementation

[0031] The following is in conjunction with the appendix Figure 1-5 The present invention will be described in further detail below.

[0032] Example: A Deep Learning-Based Method for Constructing Short-Term and Long-Term Prognostic Models for Cerebral Hemorrhage

[0033] Patients hospitalized with cerebral hemorrhage admitted to the Department of Neurology, Affiliated Three Gorges Hospital of Chongqing University from January 1, 2020 to September 1, 2022 were selected. Patients were selected to be over 18 years old, not pregnant, with the time from onset to admission less than 10 days, and who underwent imaging examinations (CT, CTA, etc.) after admission and were able to obtain original imaging data. Corresponding biochemical examinations and medical history collection were also completed after admission.

[0034] Exclude patients with secondary intracerebral hemorrhage-related factors, such as tumors, trauma, vascular structural abnormalities (arteriovenous malformations, aneurysms, cavernous hemangiomas, moyamoya disease, etc.), transformation of cerebral infarction into hemorrhage, etc.; isolated intraventricular hemorrhage or subarachnoid hemorrhage; hematoma of unknown origin or multiple hematomas occurring simultaneously; and patients with a history of craniotomy.

[0035] The diagnostic criteria for cerebral hemorrhage are based on the "2019 Chinese Guidelines for the Diagnosis and Treatment of Cerebral Hemorrhage" and the "2022 AHA-ASA Guidelines for the Management of Patients with Spontaneous Cerebral Hemorrhage".

[0036] The patient or their representative has been informed of and given verbal or written consent to all matters involved in this embodiment.

[0037] 1. Data Collection

[0038] Data collection process as follows Figure 1 As shown.

[0039] (1) ApoE genotype data

[0040] ApoE genotypes are considered categorical variables, divided into six types: ɛ2ɛ2, ɛ2ɛ3, ɛ2ɛ4, ɛ3ɛ3, ɛ3ɛ4, and ɛ4ɛ4. For allele frequencies, three binary groups are defined: 1) ɛ3ɛ3 group, consisting of the ɛ3ɛ3 genotype and the other five genotypes; 2) ɛ2+ group, consisting of the ɛ2ɛ2, ɛ2ɛ3, and ɛ2ɛ4 genotypes and the ɛ3ɛ3 genotype; 3) ɛ4+ group, consisting of the ɛ4ɛ4, ɛ2ɛ4, and ɛ3ɛ4 genotypes and the ɛ3ɛ3 genotype. Considering that the ɛ2ɛ4 genotype contains both ɛ2 and ɛ4 alleles, only its frequency is statistically analyzed. Its exclusion for other analyses will be determined based on specific circumstances.

[0041] (2) Biochemical indicators

[0042] 1) Serum biochemistry: a total of 55 items, including 25 items of complete blood count (including white blood cells (WBC, unit *10^9 / L), neutrophil ratio (NEUT%, unit %), lymphocyte ratio (LYM%, unit %), C-reactive protein (CRP, unit mg / L), etc.).

[0043] 2) Nine serum biochemical tests (including serum potassium (mmol / L), serum sodium (mmol / L), glomerular filtration rate (GFR, mL / min), anion gap (AG), uric acid (UA, umol / L), etc.).

[0044] 3) 26 liver and metabolic functions (including total protein (TP, g / L), lactate dehydrogenase (LDH, U / L), total bile acid (TBA, umol / L), total cholesterol (TC, mg / dL), triglyceride (TG, mg / dL), cystatin C (Cys C, mg / L), glycated hemoglobin (GHB, %), etc.).

[0045] 4) 15 coagulation and fibrinolysis function tests (including prothrombin time (PT, in seconds), international normalized ratio (INR), D-dimer (mg / L), etc.).

[0046] 5) The other 6 items (including B-type natriuretic peptide (BNP, ng / L), myosensitive troponin T (MTT, ug / L), myoglobin (Mb, ng / mL), creatine kinase isoenzymes (CKMB, ng / ml), erythrocyte sedimentation rate (ESR, g / L), and procalcitonin (PCT, ng / ml)).

[0047] (3) CT / CTA findings within 24 hours of onset

[0048] 1) Imaging signs of the first CT assessment within 24 hours of onset (hematoma heterogeneous density, hematoma irregular shape, island sign, satellite sign, vortex sign, black hole sign, mixed sign, fluid level sign).

[0049] 2) A follow-up CT scan should be performed within 24 hours of onset to assess hematoma enlargement;

[0050] 3) Assessment of CTA imaging findings: including spot signs, leakage signs, etc.;

[0051] (4) Demographic baseline data

[0052] 1) Basic information: Patient's name, mobile phone number, gender, age, and ethnicity;

[0053] 2) Past medical history: Past medical history (including hypertension, diabetes, coronary heart disease, atrial fibrillation, cerebral hemorrhage, cerebral infarction, gout, immune system diseases, respiratory diseases, digestive system diseases, urinary system diseases, history of tumors, infectious diseases, and contagious diseases), medical history (including history of brain trauma, brain surgery, brain tumor, other surgeries, and mitral valve surgery), and medication history (including history of taking antiplatelet drugs and anticoagulants).

[0054] 3) Personal history: smoking history, drinking history;

[0055] 4) Admission physical examination data: basic physical examination (including height, weight, and abdominal circumference measured upon admission), vital signs (including body temperature, respiratory rate, heart rate, pulse, blood pressure (systolic and diastolic), and blood oxygen saturation measured upon admission), and blood glucose (including random blood glucose measured by a blood glucose meter upon admission and the first fasting blood glucose).

[0056] 5) Time data: Specific time (recorded in year-month-day-hour format, including onset time, admission time (time of arrival at the outpatient, emergency, or inpatient department of the hospital), time of the first cranial CT / CTA examination after admission, time of the second cranial CT / CTA examination after admission without any treatment, start and end time of minimally invasive surgery, and discharge time), and time intervals (recorded in hours and minutes, including admission-onset time, first CT / CTA examination-onset time, first CT / CTA examination-admission time, second CT / CTA examination-onset time, second CT / CTA examination-admission time, interval between the first and second CT / CTA examinations, surgery start-onset time, surgery start-admission time, and surgery duration);

[0057] 6) Clinical scores: mRS score (before onset, after onset (at admission), at discharge), GCS score (at admission, 24 hours after admission, preoperative (if necessary), postoperative (if necessary), at discharge), NIHSS score (at admission, 24 hours after admission, preoperative (if necessary), postoperative (if necessary), at discharge), Activities of Daily Living (ADL) scale, ICH-APS acute physiological assessment of intracerebral hemorrhage, NRS-2002 nutritional risk screening scale, frog drinking water test grading, Padua score for VTE risk assessment, and APACHE II score.

[0058] (5) Clinical specialty data

[0059] 1) Hemorrhage-related data: initial hemorrhage volume (calculated using the ABC / 2 method on the largest hematoma section based on imaging results), hemorrhage location (hemorrhage locations were divided into four categories, including lobar hemorrhage: mainly affecting the cortex and lower white matter of the cerebral hemispheres; deep hemorrhage: mainly affecting the basal ganglia, ventricular white matter, thalamus, or internal capsule; cerebellar hemorrhage: including the cerebellar cortex and deep cerebellum; brainstem hemorrhage: affecting the brainstem. A total of 7 types of hemorrhage locations were statistically analyzed, including lobar, deep, cerebellum, brainstem, lobar + deep, cerebellum + brainstem, lobar / deep + cerebellum / brainstem), whether the hemorrhage ruptured into the ventricles, and whether the hemorrhage ruptured into the subarachnoid space;

[0060] 2) Etiology: SMASH-U classification: structural vascular lesions (S): vascular structural damage; medication (M): medication; amyloid angiopathy (A): cerebral angiopathy; systemic disease (S): systemic disease; hypertension (H): hypertension; undetermined (U): unknown cause.

[0061] 3) Simple record of treatment methods: Surgical treatment (including minimally invasive surgery and craniotomy), drug treatment (including intracranial pressure lowering, blood pressure lowering, anticoagulant, and antiplatelet drugs), and supportive treatment (including hypothermia therapy, oxygen therapy, and other symptomatic supportive treatments). If relevant cases involve specific treatment measures or drugs, all relevant information regarding the specific treatment measures or drugs will be recorded in the corresponding derivative database later.

[0062] (6) Inpatient outcome data

[0063] 1) Death: Record whether the patient died within 24 hours of symptom onset; whether the patient died within 72 hours of symptom onset; and whether the patient died during hospitalization.

[0064] 2) Functional recovery status, including mRS score at discharge; mental status at discharge (including clear, confused, delirious, drowsy, stuporous, light coma, moderate coma, deep coma, etc.); muscle strength at discharge (record the muscle strength of the left upper limb, left lower limb, right upper limb, and right lower limb respectively).

[0065] 3) Other situations: including length of hospital stay, complications and comorbidities (including pulmonary infection, venous thrombosis, epilepsy, cognitive impairment, etc., which will be recorded and described in detail in the corresponding derivative database later);

[0066] 2. Follow-up outcomes

[0067] All eligible cases were followed up. Follow-up methods included telephone, questionnaires, and outpatient visits. The start point of follow-up was discharge from the hospital, and the end point was death or one year after the onset of cerebral hemorrhage. The statistical data included mortality data within one year (time of death, cause of death), cerebral hemorrhage recurrence data within one year (time of recurrence, location of recurrence), mRS score, and BI index at the end of follow-up.

[0068] Poor / good prognosis at 3 months and 1 year were used as the primary outcome variables. Survival was the observation endpoint, with an mRS score <3 defined as good prognosis, and all other outcomes defined as poor prognosis (mRS score included death (6 points)).

[0069] 3. Statistical Analysis

[0070] Sample size calculation: In the initial design, the primary outcome included an effect size (difference) of 20 percentage points. Assuming the expected rate of poor prognosis in patients with cerebral hemorrhage at 3 months was 0.636, the O'Brien-Fleming cutoff rule was used to estimate the response rate within + / - 10% of the true value. The Type I error (a) was 0.05, and the power was 100x(1-B) 90%. A two-tailed test was used, and the calculated sample size was approximately 550 cases. Considering the loss to follow-up factor was controlled to within 10%, the final sample size was determined to be 600 cases.

[0071] Overall Description: Categorical variables were expressed as frequencies (N) and percentages (%), and comparisons were performed using the χ² test. Continuous variables were expressed as mean ± standard deviation or median and interquartile range (IQR), and comparisons between groups were performed using the Mann-Whitney U test or Student's T test. All calculations were performed using SPSS 22.0 as the statistical tool. A p-value < 0.05 was considered statistically significant. Poor prognosis was defined as an mRS score ≥ 3, and good prognosis as an mRS score ≤ 2. Poor / good prognosis were used as binary outcomes at 3 months and 12 months after onset, respectively. Univariate analysis was first performed to explore the correlation between all included clinical data and poor / good prognosis. Factors with p-values ​​< 0.05 in the univariate analysis were then included in multivariate analysis, and logistic regression analysis was performed to explore independent risk factors for poor / good prognosis.

[0072] 4. Model Establishment

[0073] (1) Using deep learning methods to predict the prognostic model of collected data is one of the current research hotspots. The following are the general steps:

[0074] 1) Data preprocessing: First, the collected data is preprocessed, including data cleaning, standardization and normalization, to reduce data noise and improve data interpretability.

[0075] 2) Feature Extraction: Using deep learning methods, the collected data is converted into a format that the neural network can process. This typically involves feature extraction, such as using gene expression profiles to extract features.

[0076] 3) Model building: Designing and building deep learning models to predict prognoses. Common models include fully connected neural networks, convolutional neural networks, and recurrent neural networks.

[0077] 4) Model Training: Use training data to train the model. During training, it is necessary to select appropriate loss functions, optimizers, and hyperparameters to optimize the model's predictive performance.

[0078] 5) Model Evaluation: Use test data to evaluate the model's performance. Evaluation metrics typically include accuracy, recall, and F1 score.

[0079] 6) Model optimization: Based on the evaluation results of the model, optimize the model, including adjusting hyperparameters, increasing or decreasing the number of layers, etc.

[0080] 7) Predictive applications: Use the optimized model to predict new prognostic models.

[0081] (2) Model building and training

[0082] I. Data Queue Construction

[0083] This invention focuses on patients with spontaneous intracerebral hemorrhage. First, a single-center clinical cohort was constructed, and four categories of core characteristic variables (A, B, C, and D) were systematically collected to provide a high-quality data foundation for model training. Category A variables focused on apolipoprotein E (APOE) gene data, covering key genetic markers such as genotyping. Category B variables were biochemical indicators, including laboratory test results reflecting the body's metabolic and pathophysiological state, such as complete blood count, liver and kidney function, coagulation function, blood glucose, and blood lipids. Category C variables were CT / CTA imaging findings within 24 hours of onset, including radiomics characteristics such as hemorrhage volume, hemorrhage location, presence of ventricular rupture, and degree of edema. Category D variables were demographic baseline data, involving basic clinical information such as age, sex, history of underlying diseases (hypertension, diabetes, atrial fibrillation, etc.), and precipitating factors.

[0084] At the same time, a standardized follow-up system was established to conduct short-term (3 months) and long-term (12 months) prognostic follow-ups on enrolled patients. The modified Rankin Scale (mRS) score was used as the standard for determining functional outcomes. An mRS score ≥3 was defined as a poor prognosis and used as the target outcome variable of the model to ensure the integrity of the data and the objectivity of the outcome determination.

[0085] II. Data Preprocessing and Feature Engineering

[0086] A multi-dimensional data cleaning strategy was employed, addressing missing values ​​according to the characteristics of different variable types: continuous variables (such as age, bleeding volume, and biochemical indicators) were imputed using multiple imputation, while categorical variables (such as gender, genotype, and positive imaging findings) were imputed using maximum likelihood estimation. Outliers were identified and corrected, with extreme outliers beyond three standard deviations removed using box plots, and logically contradictory data (such as conflicts between onset time and examination time) removed based on clinical judgment.

[0087] Feature standardization and encoding were performed: Z-score standardization was applied to continuous features to convert the data into a distribution with a mean of 0 and a standard deviation of 1, eliminating dimensional differences; one-hot encoding or label encoding was used for categorical variables to ensure model recognizability. Correlation analysis (Pearson correlation coefficient, Spearman rank correlation coefficient) was used to screen features statistically associated with poor short-term and long-term prognoses, and redundant features with strong multicollinearity (variance inflation factor VIF > 10) were eliminated to construct a concise and effective feature set.

[0088] III. Model Building and Training Process

[0089] Based on a deep learning algorithm framework, a deep neural network model (ABCD prediction model) suitable for predicting the prognosis of spontaneous intracerebral hemorrhage was constructed. The model structure adopts a multilayer perceptron (MLP) architecture, with the number of nodes in the input layer consistent with the selected feature dimensions; 3-5 hidden layers are set, and the number of nodes in each layer is determined by grid search optimization (50-200 nodes). The ReLU function is selected as the activation function to solve the gradient vanishing problem, and a Dropout layer (dropout rate=0.2-0.5) is introduced to prevent the model from overfitting; the output layer uses the Sigmoid activation function to output the probability value of poor short-term or long-term prognosis of the patient.

[0090] The preprocessed dataset was randomly divided into a training set and an internal validation set in a 7:3 ratio. The training set was used for model parameter learning, while the internal validation set was used for real-time evaluation of model performance and parameter tuning. An adaptive moment estimation (Adam) optimizer was used for model training, with an initial learning rate of 0.001, dynamically adjusted based on changes in the validation set loss function using a learning rate decay strategy (ReduceLROnPlateau). The loss function chosen was binary cross-entropy to reflect the deviation between the predicted probability and the actual outcome. Early stopping was employed during training; training was stopped when the validation set loss function showed no decrease for 10 consecutive epochs to prevent overtraining and a decline in generalization ability.

[0091] IV. Model Validation and Optimization

[0092] (a) Internal Validation

[0093] An internal validation set was used to conduct a preliminary performance evaluation of the model. Key evaluation metrics included accuracy, sensitivity, specificity, area under the receiver operating characteristic (AUC), precision-recall area under the curve (PR-AUC), and Brier score. A confusion matrix analysis was used to assess the degree of agreement between the model's predictions and actual outcomes, and positive predictive value (PPV) and negative predictive value (NPV) were calculated to comprehensively reflect the model's predictive performance under different prognostic states.

[0094] (ii) External verification

[0095] To verify the model's generalization ability, we will subsequently obtain external independent cohort data through multi-center collaborations or public databases (such as MIMIC, SINOMACR, etc.), strictly screen research subjects according to the model's inclusion and exclusion criteria, collect unified ABCD four-category characteristic variables and prognostic outcome data, conduct external validation of the model, and use evaluation indicators consistent with internal validation to test the model's applicability in different populations and different medical environments.

[0096] (III) Model Optimization

[0097] Based on internal and external validation results, a combination of grid search and cross-validation (5-fold cross-validation) was used to iteratively optimize the model's hyperparameters, including key parameters such as the number of hidden layers, number of nodes, dropout rate, and learning rate. Feature importance analysis (e.g., permutation importance, SHAP value analysis) was used to identify the core features that contribute most to prognostic prediction, further optimizing the feature set structure. If the model exhibits performance deficiencies (e.g., low sensitivity or insufficient specificity), the model architecture was adjusted (e.g., adding hidden layers, changing activation functions) or ensemble learning strategies (e.g., stacking, bagging) were employed to fuse multiple deep learning models, improving the stability and accuracy of prediction performance. Ultimately, this resulted in an optimal ABCD prediction model that balances accuracy, sensitivity, specificity, and generalization ability.

[0098] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A method for constructing short-term and long-term prognostic models for cerebral hemorrhage based on deep learning, characterized by: Includes the following steps: S1. Collect relevant data on the spontaneous intracerebral hemorrhage cohort and conduct short-term and long-term follow-ups to obtain functional outcome data, with the mRS score as the criterion for determining functional outcome. S2. Analyze the correlation between various data and poor short-term and long-term prognosis of cerebral hemorrhage, select risk factors to establish a deep learning-based intelligent prediction model for poor short-term and long-term prognosis of spontaneous cerebral hemorrhage, and use an mRS score ≥3 to determine poor short-term and long-term prognosis of cerebral hemorrhage. Use deep neural networks to learn and classify the model features. S3. Validate the model using existing data, measure the performance of the prediction model, and evaluate the model, including but not limited to the evaluation of accuracy, sensitivity, and specificity.

2. The method for constructing short-term and long-term prognostic models for cerebral hemorrhage based on deep learning according to claim 1, characterized in that: The data collected in S1 for the spontaneous intracerebral hemorrhage cohort mainly includes genetic and biochemical indicators, CT / CTA findings within 24 hours of onset, and baseline demographic data.

3. The method for constructing short-term and long-term prognostic models for cerebral hemorrhage based on deep learning according to claim 2, characterized in that: The gene in question refers to apolipoprotein E-APOE.

4. The method for constructing short-term and long-term prognostic models for cerebral hemorrhage based on deep learning according to claim 1, characterized in that: The short-term follow-up in S1 is 3 months, and the long-term follow-up is 12 months.

5. The method for constructing short-term and long-term prognostic models for cerebral hemorrhage based on deep learning according to claim 1, characterized in that: The specific operational procedure for establishing a deep learning-based intelligent prediction model for poor short-term and long-term prognosis of spontaneous intracerebral hemorrhage in S2 is as follows: 1) Data preprocessing: First, the collected data is preprocessed, including data cleaning, standardization and normalization, to reduce data noise and improve data interpretability; 2) Feature extraction: Using deep learning methods, the collected data is converted into a format that can be processed by neural networks; 3) Model building: Designing and building deep learning models to predict prognoses; 4) Model training: Using training data to train the model; 5) Model Evaluation: Use test data to evaluate the model's performance; 6) Model optimization: Based on the model evaluation results, optimize the model, including but not limited to adjusting hyperparameters, increasing or decreasing the number of layers; 7) Predictive applications: Use the optimized model to predict new prognostic models.

6. The method for constructing short-term and long-term prognostic models for cerebral hemorrhage based on deep learning according to claim 1, characterized in that: In S3, model validation is divided into internal validation and external validation based on the existing data of the model builder. Validation using one's own existing data is internal validation, while validation using data borrowed from other regions is external validation.

Images