Acute ischemic stroke prognosis prediction model and construction method

Abstract

The invention discloses an acute ischemic stroke prognosis prediction model and a construction method, and relates to the technical field of disease prediction, and the key points of the technical scheme are that the method comprises the following steps: S1, obtaining a data set which comprises baseline information, prognosis information and meteorological data; S2, processing the data: firstly, randomly dividing the data set into a training set and a test set, filling missing data, and balancing patient distribution; s3, constructing a model: constructing the model by using six ML algorithms, and training the model by using the training set; and S4, evaluation and screening: performing performance evaluation on all the models by using the test set, and screening out an optimal model. According to the method, the ML model fusing weather and clinical characteristics is constructed, so that the accuracy of predicting AI S prognosis is improved, and decision support is provided for a patient diagnosis and treatment strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of disease prediction, and more particularly to an acute ischemic stroke prognosis prediction model and a construction method thereof. Background Art

[0002] Patients with ischemic stroke (AIS) experience varying degrees of functional impairment, placing a heavy burden on their families and society. Prognosis is influenced by a complex interplay of patient-specific clinical characteristics, treatment modalities, and environmental conditions. Accurately assessing patient prognosis will help clinicians promptly identify patients at high risk of poor prognosis and develop more personalized treatment plans, which is crucial for improving treatment modalities and promoting patient recovery.

[0003] In recent years, increasing research has focused on the relationship between meteorological factors and the incidence and prognosis of ischemic stroke, particularly the impact of extreme temperatures on ischemic stroke. Studies have shown that long-term exposure to large daily temperature ranges (DTRs) can have long-term effects on all-cause mortality and cardiovascular morbidity. Other meteorological parameters, such as humidity and precipitation, are also closely associated with functional outcomes in ischemic stroke. Therefore, considering the synergistic effects of meteorological factors may help improve the accuracy of stroke prognostic assessment.

[0004] With advances in artificial intelligence, the application of ML algorithms in medicine is becoming increasingly widespread, bringing new possibilities for disease prediction and personalized treatment. In the field of meteorological medicine, ML algorithms can more easily integrate various variables, including demographics, clinical characteristics, radiomics features, and meteorological parameters, thereby improving the accuracy of disease prediction. Studies have shown that ML models that integrate clinical and radiomics features can accurately predict the prognosis of patients with acute ischemic stroke (AIS) and guide their rehabilitation treatment. In recent years, with the development of interdisciplinary fields, some studies have used ML algorithms to integrate meteorological data with clinical information to predict disease incidence. For example, a machine learning prediction model that integrates meteorological and temporal data has achieved daily predictions of out-of-hospital cardiac arrest rates. Other studies have used weather information and clinical data from heatstroke patients to predict the number of heatstroke cases, hospitalizations, and deaths based on ML models. However, in the field of stroke, although ML models based on clinical information have been widely used to predict AIS prognosis, further research is needed to integrate meteorological and clinical data to predict AIS prognosis. In light of this, the present inventors propose a prognosis prediction model for acute ischemic stroke and a method for its construction. Summary of the Invention

[0005] The purpose of the present invention is to provide an acute ischemic stroke prognosis prediction model and a construction method to solve the above problems.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions:

[0007] The present invention provides a method for constructing a prognosis prediction model for acute ischemic stroke, comprising the following steps:

[0008] S1. Obtaining a dataset: The dataset includes baseline information, prognostic information, and meteorological data;

[0009] S2. First, randomly divide the dataset into training and test sets, and then fill in missing data to balance the patient distribution;

[0010] S3. Model Construction: Using the 3-month functional prognosis of AIS patients as a variable, six ML algorithms were used to construct the model, and the training set was used to train the model.

[0011] S4. Evaluation and screening: Use the test set to evaluate the performance of all models and screen out the best model.

[0012] The present invention is further configured as follows: in step S1,

[0013] The baseline data included baseline demographics, clinical data, and laboratory test data of patients after admission;

[0014] The prognostic information is the patient's modified Rankin Scale score 3 months after discharge;

[0015] The meteorological data are meteorological data from the day of onset of AIS patients to 14 days before the onset of the disease, including daily average temperature, daily maximum temperature, daily minimum temperature, daily temperature difference, daily average air pressure, relative humidity, daily average precipitation, and daily average wind speed.

[0016] The present invention is further configured as follows: in step S2, the filling data is to fill the missing values ​​using the average value of the column data corresponding to the missing indicator, and the balanced patient distribution is to use the synthetic minority oversampling technology to ensure the balanced distribution of patients with good prognosis and poor prognosis in the training set.

[0017] The present invention is further configured as follows: in step S2, the ratio of the training set to the test set is 7:3.

[0018] The present invention is further configured as follows: in step S3, the six ML algorithms are logistic regression, support vector machine, random forest, decision tree, extreme gradient boosting and light gradient boosting machine.

[0019] The present invention is further configured as follows: in step S4, the performance evaluation indicators include accuracy, precision, recall, F1 score, and AUC.

[0020] The present invention also provides an acute ischemic stroke prognosis prediction model, which is constructed using any of the above methods.

[0021] In summary, the present invention has the following beneficial effects:

[0022] 1. The LightGBM machine learning model that integrates clinical and meteorological features, which was finally selected in this paper, showed good performance in predicting the 3-month functional prognosis of AIS patients. This suggests that meteorological factors have potential value in predicting the functional prognosis of AIS patients.

[0023] 2. SHAP interpretability analysis found that the clinical characteristics that have a significant impact on the poor prognosis of AIS patients are older age, higher NIHSS score, and female gender; laboratory indicators are higher PDW and white blood cell count; and meteorological variables are lower daily average temperature 3 days before onset and larger diurnal temperature range 7 days before onset. This provides guidance for personalized assessment of the prognosis of AIS patients.

[0024] 3. The results of the exposure-response curve showed that low temperature 3 days before onset and large daily temperature difference 7 days before onset were significantly correlated with increased PDW levels. That is, the lower the temperature and the greater the temperature difference, the higher the platelet activity, indicating an increased risk of thrombosis, which has an adverse effect on the prognosis of AIS patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a detailed flow chart of the research population selection process in an embodiment of the present invention;

[0026] Figure 2 This is a flow chart of a method for constructing a prognosis prediction model for acute ischemic stroke in an embodiment of the present invention;

[0027] Figure 3 is the performance evaluation of each model in the embodiment of the present invention (wherein A is the clinical model; B is the meteorological model; B is the receiver operating characteristic curve of the clinical + meteorological model; D is the confusion matrix of the LightGBM algorithm of the clinical + meteorological model);

[0028] Figure 4 is a SHAP value graph in an embodiment of the present invention;

[0029] Figure 5 Figure 2 is the exposure-response curve of Lag3 mean daily temperature (A) and Lag7 diurnal temperature difference (B) and PDW in the embodiment of the present invention (after adjusting for the confounding effects of air pressure, humidity, wind speed, age, and gender, the analysis was performed using a GAMM model. Shading indicates 95% confidence intervals. Lag3 mean Temp is the daily mean temperature 3 days before onset; lag 7DTR is the diurnal temperature difference 7 days before onset; PDW is the platelet distribution width). DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] First, the technical terms involved in the embodiments of the present invention are explained:

[0032] ML algorithm: machine learning algorithm;

[0033] Prognosis: The prediction of various possible outcomes and their probabilities after the occurrence of a disease, including cure, remission, deterioration, recurrence, complications and death.

[0034] Example 1:

[0035] Research subjects

[0036] Patients diagnosed with AIS who were hospitalized in the Department of Neurology, Tianjin Medical University Second Hospital from January 1, 2018, to December 31, 2022. The detailed process of the study population selection process is shown in Figure 1 .

[0037] This study adhered to the ethical standards of the Declaration of Helsinki and has been approved by the Ethics Committee of the Second Hospital of Tianjin Medical University.

[0038] (1) Inclusion criteria:

[0039] ①Age ≥ 18 years old;

[0040] ② Diagnosis of acute ischemic stroke, defined as acute onset of neurological dysfunction with signs or symptoms, confirmed by non-contrast computed tomography (CT) or magnetic resonance imaging (MRI) of the head on admission;

[0041] ③ Hospitalized and received routine treatment within 7 days of onset;

[0042] ④ Permanent residents of Tianjin (and have not left the city in the past two weeks);

[0043] (2) Exclusion criteria:

[0044] ① The onset time is unclear;

[0045] ② Receive intravenous thrombolysis or mechanical thrombectomy;

[0046] ③ Stroke classified as other definite causes or unknown causes according to the etiology classification of the Trial of Org 10172 in Acute Stroke Treatment (TOAST);

[0047] ④ Incomplete admission information (including unregistered residence information and missing NIHSS score);

[0048] ⑤Incomplete follow-up information;

[0049] A method for constructing a prognostic prediction model for acute ischemic stroke, such as Figure 2 As shown, the following steps are included:

[0050] 1. Get the dataset:

[0051] (1) Baseline information collection

[0052] Baseline demographic and clinical data were obtained from the hospital electronic medical record system, including patients' residential address, age, gender, medical history (history of stroke, hypertension, hyperlipidemia, diabetes, coronary heart disease, and atrial fibrillation), onset time, admission blood pressure, NIHSS score on admission, and TOAST classification. At the same time, the laboratory test data of patients after admission were collected, including white blood cell count (WBC), hemoglobin concentration (HGB), platelet count (PLT), thrombocytocrit (PCT), platelet distribution width (PDW), mean platelet volume (MPV); fasting blood glucose (FBG), glycosylated hemoglobin A1c (hemoglobin A1c), triglyceride (TG), total cholesterol (TC), high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), fibrinogen (Fbg), alanine aminotransferase (ALT), aspartate aminotransferase (AST) and glutamic acid transaminase (AST). transaminase (AST), total bilirubin (TBIL), direct bilirubin (DBIL), indirect bilirubin (IBIL), blood urea nitrogen (BUN), creatinine (Cr), uric acid (UA), and homocysteine ​​(HCY).

[0053] (2) Prognostic information collection

[0054] The modified Rankin Scale (mRS) scores were collected and recorded three months after discharge using an electronic follow-up system. Experienced neurologists assessed the patients' functional outcomes and recorded their scores during telephone interviews three months after discharge. Based on the mRS scores, patients were divided into a favorable prognosis group (mRS score ≤ 2) and a poor prognosis group (mRS score > 2).

[0055] (3) Meteorological data collection

[0056] Meteorological data for the Tianjin area were collected from meteorological stations and provided by the Tianjin Meteorological Bureau. Observations were obtained from the Tianjin Climate Monitoring Station, with the nearest monitoring station matched based on the residence provided by the patient's homepage. Based on previous epidemiological evidence and our team's previous research, meteorological factors have a lag effect on the risk of ischemic stroke and mortality, with this lag effect persisting for 14 days. Therefore, this study selected an exposure window of 0 to 14 days, collecting meteorological data from the day of onset of AIS (lag 0) to 14 days before onset (lag 14). This data primarily included daily average temperature, daily maximum temperature, daily minimum temperature, diurnal temperature range (defined as the difference between the daily maximum and minimum temperatures), daily average pressure, relative humidity, daily average precipitation, and daily average wind speed. Seasons of onset were categorized by month as spring (March–May), summer (June–August), autumn (September–November), and winter (December–February).

[0057] 2. Data Processing: To address missing laboratory test data (missing percentage <0.1%), we imputed missing values ​​using the mean of the corresponding column data. To address imbalanced outcomes, we used the Synthetic Minority Over-sampling Technique (SMOTE) to ensure a balanced distribution of patients with good and poor prognosis in the training set. The remaining patients formed the test set to validate the accuracy of the model classification.

[0058] 3. Model Construction: The primary outcome variable of this study was the three-month functional outcome (favorable vs. poor prognosis) in patients with AIS. Six machine learning algorithms were used to construct the model, including logistic regression (LR), support vector machine (SVM), random forest (RF), decision tree (DT), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). These algorithms were selected based on their ability to identify diverse feature combinations, including both linear and nonlinear ones. For example, XGBoost and LightGBM can identify two types of combinations, while LR is limited to linear combinations. In contrast, RF, DT, and SVM excel at detecting nonlinear patterns. The entire dataset was randomly split into training and test sets in a 7:3 ratio, with 70% of the data used for training and 30% for testing. The model was trained using a five-fold cross-validation method, randomly dividing the training set into five groups: four for model training and one for algorithm performance evaluation. All possible fold combinations of training and test groups in the training set were repeated five times.

[0059] 4. Evaluation and Screening: Use the test set to evaluate the performance of all models. Model performance evaluation metrics include accuracy, precision, recall, F1 score, and AUC. Filter out the best model.

[0060] The best model screened out was the LightGBM machine learning model that integrated clinical and meteorological features.

[0061] Statistical analysis

[0062] To determine performance differences between models, the Delong test was used to compare the AUC values ​​between models. All statistical tests were two-sided, and P < 0.05 was considered statistically significant. Machine learning models were constructed and evaluated using Python 3.10.13 for visualization.

[0063] Model test results:

[0064] Performance comparison of various prediction models

[0065] The ROC curves of each model are as follows Figure 3 As shown in AC, when using LR (0.75 vs 0.72 / 0.69), XGBoost (0.86 vs 0.83 / 0.65), and LightGBM (0.87 vs 0.84 / 0.68) algorithms, the AUC of the prediction model integrating meteorological and clinical features was higher than that of the model using only clinical or meteorological features, and the AUC value of the LightGBM algorithm was larger than that of the LR and XGBoost algorithms (0.87>0.86>0.75, P<0.05). Figure 3 D shows the confusion matrix plot of the LightGBM prediction model that integrates meteorological and clinical features.

[0066] The detailed performance evaluation of each ML model is shown in Table 1. The AUC value of the LightGBM prediction model using clinical features alone was 0.84, which was significantly higher than that of the other models (P < 0.05). The AUC value of the LR prediction model using meteorological features alone was 0.69, which was significantly higher than that of the other prediction models (P < 0.05). When clinical and meteorological features were fused, the AUC values ​​of the LR, XGBoost, and LightGBM algorithms were significantly increased compared with the individual models (P < 0.05). Among the three, LightGBM showed the highest AUC (0.87, P < 0.05), with an accuracy of 0.874, a recall of 0.550, a precision of 0.592, and an F1 score of 0.570.

[0067] Table 1 Model performance indicators

[0068]

[0069] Interpretability analysis of predictive models

[0070] After predicting poor prognosis through modeling, the best performing lightGBM model was selected for interpretable analysis. Figure 4 As shown, A is the absolute average SHAP value of each feature's overall impact on the model output; B is the SHAP summary diagram (B) of the distribution of each feature and how its SHAP value affects the model results; gender (1: female; 2: male), TOAST classification (1: cardioembolic stroke; 2: large artery atherosclerosis; 3: small artery occlusion), and stroke history (0: no; 1: yes). By calculating the SHAP value of lightGBM, the importance of the features that affect the prediction of poor prognosis is ranked (such as Figure 4 A). The overall impact of each feature is expressed by the feature SHAP general value, such as Figure 4 As shown in Figure B, the important clinical indicators for predicting poor prognosis are older age, higher NIHSS score, female sex, higher PDW, TOAST type (cardioembolic stroke), and higher white blood cell count; the important meteorological variables are lower daily average temperature 3 days before onset and larger DTR 7 days before onset.

[0071] Effects of daily mean temperature at Lag3 and daily temperature difference at Lag7 on PDW

[0072] The previous SHAP analysis showed that the important meteorological variables that have a poor prognosis for AIS are lower daily average temperature 3 days before onset and larger DTR 7 days before onset, while higher PDW is a laboratory indicator that affects poor prognosis. To explore the effects of lag3 daily average temperature and lag7 daily temperature difference on PDW, we used a generalized additive mixed model (GAMM) to adjust the effects of confounding factors such as air pressure, humidity, wind speed, age, and gender, and analyzed the exposure-response relationship between lag3 daily average temperature, lag7 daily temperature difference and PDW. The results are as follows Figure 5 As shown, the relationship between daily mean temperature and PDW was generally nonlinear three days before onset. At lower temperatures (approximately below 0°C), PDW increased significantly with decreasing temperature (P < 0.01). Within a more suitable temperature range, PDW remained close to baseline levels. At higher temperatures (approximately above 25°C), PDW decreased significantly with increasing temperature (P < 0.01). Seven days before onset, a potential threshold effect existed between the diurnal temperature range and PDW. When the diurnal temperature range was small (approximately below 15°C), PDW changes remained close to baseline levels. When the diurnal temperature range was large (approximately above 15°C), PDW changes showed a significant upward trend (P < 0.01).

[0073] In summary:

[0074] 1. The LightGBM machine learning model integrating clinical and meteorological features showed good performance in predicting the 3-month functional prognosis of AIS patients. This suggests that meteorological factors have potential value in predicting the functional prognosis of AIS patients.

[0075] 2. SHAP interpretability analysis found that the clinical characteristics that have a significant impact on the poor prognosis of AIS patients are older age, higher NIHSS score, and female gender; laboratory indicators are higher PDW and white blood cell count; and meteorological variables are lower daily average temperature 3 days before onset and larger diurnal temperature range 7 days before onset. This provides guidance for personalized assessment of the prognosis of AIS patients.

[0076] 3. The results of the exposure-response curve showed that low temperature 3 days before onset and large daily temperature difference 7 days before onset were significantly correlated with increased PDW levels. That is, the lower the temperature and the greater the temperature difference, the higher the platelet activity, indicating an increased risk of thrombosis, which has an adverse effect on the prognosis of AIS patients.

[0077] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for constructing a prognosis prediction model for acute ischemic stroke, characterized by: The steps include: S1. Obtaining a dataset: The dataset includes baseline information, prognostic information, and meteorological data; S2. Data processing: First, randomly divide the dataset into training and test sets, then fill in missing data to balance the patient distribution; S3. Model Building: Build a model using six ML algorithms and train the model using the training set. S4. Evaluation and screening: Use the test set to evaluate the performance of all models and screen out the best model.

2. The method for constructing a prediction model according to claim 1, wherein: In step S1, The baseline data included baseline demographics, clinical data, and laboratory test data of patients after admission; The prognostic information is the patient's modified Rankin Scale score 3 months after discharge; The meteorological data are meteorological data from the day of onset of AIS patients to 14 days before the onset of the disease, including daily average temperature, daily maximum temperature, daily minimum temperature, daily temperature difference, daily average air pressure, relative humidity, daily average precipitation, and daily average wind speed.

3. The method for constructing a prediction model according to claim 1, wherein: In step S2, the filling data is to fill the missing values ​​using the average value of the column data corresponding to the missing indicator, and the balanced patient distribution is to use synthetic minority oversampling to ensure the balanced distribution of patients with good prognosis and poor prognosis in the training set.

4. The method for constructing a prediction model according to claim 1, wherein: In step S2, the ratio of the training set to the test set is 7:

3.

5. The method for constructing a prediction model according to claim 1, wherein: In step S3, the six ML algorithms are logistic regression, support vector machine, random forest, decision tree, extreme gradient boosting and light gradient boosting machine.

6. The method for constructing a prediction model according to claim 1, wherein: In step S4, the performance evaluation indicators include accuracy, precision, recall, F1 score, and AUC.

7. A prognosis prediction model for acute ischemic stroke, characterized by: Constructed using the method described in any one of claims 1 to 6.

Images