Decision tree model-based hospitalization day prediction method after combined trabectomy

Abstract

The invention relates to the field of clinical pathway optimization, in particular to a combined type trabectomy post-hospitalization day prediction method based on a decision tree model, which comprises the steps of data preprocessing, decision tree model construction, variation detection, pruning and standard hospitalization day prediction. And predicting the hospitalization day duration of the patient by using the decision tree model. According to the method, the real condition and the treatment effect of the patient can be reflected more objectively, so that the accuracy and the consistency of medical decisions are improved, and the decision of discharge time is optimized.

Description

Technical Field

[0001] The present invention relates to the field of clinical pathway optimization, and specifically to a method for predicting the length of hospital stay after combined trabeculectomy based on a decision tree model. Background Art

[0002] Glaucoma is a disease mainly characterized by optic nerve atrophy and visual field defect. At present, the treatment methods for this disease include drug, laser and surgical treatments. Usually, combined trabeculectomy is clinically used as the best treatment method for glaucoma. In combined trabeculectomy, the scleral flap is sutured airtight during the operation, and the sutures are removed according to the intraocular pressure after the operation to reopen the scleral flap, restore aqueous humor drainage, and reduce the occurrence of scarring, thereby improving the success rate of glaucoma surgery. However, due to the shortening of the average length of hospital stay, combined with the fast turnover rate of ophthalmic patients and the large number of admissions and discharges, the postoperative nursing work of glaucoma patients faces many problems. At present, clinically evaluating the length of hospital stay after combined trabeculectomy requires comprehensive consideration of multiple factors before, during and after the operation. Preoperative evaluation includes the patient's basic conditions, glaucoma type and severity; during the operation, attention is paid to the surgical complexity and operation time; after the operation, the occurrence of complications, intraocular pressure control and visual acuity recovery are observed. Due to the different experiences of doctors and the complexity of the clinical data considered, there is a deviation between the evaluated length of hospital stay and the actual length of hospital stay.

[0003] Relying on information technology, the length of hospital stay after combined trabeculectomy can be predicted. For example, using the method of regression analysis, taking the length of hospital stay after the operation as the dependent variable, and selecting multiple independent variables related to the length of hospital stay (such as patient age, surgical complexity, postoperative complications, etc.), and using a multiple linear regression model to analyze the influence degree of each factor on the length of hospital stay. However, regression analysis usually assumes that the relationship between variables is linear. However, in actual medical data, the relationship between variables is often more complex. When the regression model processes complex and variable medical data, it lacks sufficient generalization ability, resulting in poor prediction effects on new data. Therefore, there is an urgent need for a length of hospital stay prediction model that is more suitable for processing medical data. Summary of the Invention

[0004] Aiming at the problems existing in the prior art, the present invention proposes a method for predicting the length of hospital stay after combined trabeculectomy based on a decision tree model by data mining.

[0005] The method for predicting the length of hospital stay after combined trabeculectomy based on the decision tree model of the present invention is obtained through the following steps:

[0006] 1) Data preprocessing: Normalize the feature data in the intraoperative medical records collected by the Max-Min method, and retain the effective feature data as the input data of the model. The effective feature data includes intraocular pressure, filtering bleb morphology and anterior chamber depth data;

[0007] 2) Construct a decision tree model: Aggregate the effective feature data into a sample data set, divide the data according to the complications in the postoperative examination report, and divide the sample data set into several subsets; Aggregate all samples in the subset under the same complication level to form a recursively constructed sub-tree;

[0008] 3) Perform mutation detection on the trained decision tree, and mark the data generated by the mutation as abnormal; Use the test set to evaluate the constructed decision tree model, and evaluate the accuracy, precision, recall rate, and F1 score of the data marked as normal;

[0009] 4) Perform post-pruning on the constructed decision tree model, remove the branches that are not helpful for improving the generalization ability of the model, and adjust the maximum tree depth and minimum sample splitting number of the model to improve the performance of the model;

[0010] 5) Input the intraocular pressure, bleb morphology, and anterior chamber depth data of the new surgical patients into the trained decision tree model. The model will predict the postoperative complication level based on the intraoperative medical record data input, and combine the average length of hospital stay of the same-level complications in the historical data to predict the standard length of hospital stay of the patients.

[0011] Specifically, the specific method of the normalization process described in step 1) is as follows: Take the intraocular pressure value (in mmHg) and anterior chamber depth value (in mm) on the day before surgery; At the same time, use the Kronfeld classification method to classify the bleb morphology, which is divided into microcystic type, diffuse flat type, scar type, and encapsulated type according to the degree, and assign values to the bleb morphology of each type, corresponding to 0, 1, 2, and 3 respectively; Then calculate the maximum value max and minimum value min of the intraocular pressure, bleb morphology, and anterior chamber depth in the feature data set, and use the Max-Min normalization formula to transform each feature data, that is:

[0012]

[0013] Among them, x′ is the normalized data, x is the original data, and each feature value in the original data set is replaced with the normalized value to form a new sample data set.

[0014] Specifically, the complication attributes described in step 2) include pain, hyphema, and corneal edema. To achieve data set division, the above complications are quantitatively graded, where:

[0015] The pain level is quantitatively graded by the VAS visual analogue scale;

[0016] Anterior chamber hemorrhage grade: Grade I indicates that the amount of blood accumulation is less than 1 / 3 of the anterior chamber; Grade II means the blood accumulation accounts for 1 / 3 - 1 / 2 of the anterior chamber; Grade III indicates that the blood accumulation is > 1 / 2 and < 3 / 4; Grade IV means the blood accumulation is > 3 / 4 to the entire anterior chamber (the tissues behind the cornea cannot be seen).

[0017] Corneal edema grade: Grade I indicates that the cornea is transparent without edema; Grade II means local hazy edema of the cornea, with a smooth corneal endothelium and clear iris texture still visible; Grade III indicates light gray edema of the cornea, a rough corneal endothelium, and blurred iris texture; Grade IV means diffuse gray - white edema of the cornea, with a cracked corneal endothelium and unclear iris texture; Grade V indicates milky white edema of the cornea, and the intraocular structures cannot be seen clearly.

[0018] Specifically, the specific method of data set division in step 2) is as follows: The sample data set is divided into multiple subsets with the complication attributes determined by the information gain ratio as decision nodes. A new decision node is created for each subset, and the data of each subset will be passed to this new node until all complication attributes are used for data division.

[0019] Furthermore, the steps for determining the information gain ratio are as follows:

[0020] 1) Calculate entropy: First, calculate the entropy of the sample data set, which represents the uncertainty of the data set. The formula for entropy is:

[0021] IG(D, A) = H(D) - H(D|A)

[0022] Among them, H(D) is the information entropy of the data set D, and H(D│A) is the conditional entropy of the data set D under the condition of feature A;

[0023] 2) Calculate the intrinsic value IV(X) of the feature:

[0024]

[0025] Among them, D i is the data subset corresponding to the i - th value of the feature X, │D i │ is the number of samples in the subset D i , │D│ is the total number of samples in the original data set, and n is the number of values of the feature X;

[0026] 3) Calculate the information gain ratio GR(D,A):

[0027]

[0028] Among them, IG(D,A) is the information gain of the feature A; IV(A) is the intrinsic value of the feature A.

[0029] Specifically, the specific method for mutation detection in step 3) is as follows: By calculating the Gaussian distribution probability density function corresponding to the feature data, it can be used to find the abnormal points in the feature data. By modifying the probability density threshold contamination, the sensitivity of abnormal point detection can be adjusted, and the abnormal points are marked as an independent category.

[0030] Specifically, post-pruning in step 4) uses cost complexity pruning. The cost_complexity_pruning_path function from sklearn.tree import DecisionTreeClassifier is used to determine the pruning path and determine the optimal α value.

[0031] The beneficial effects of the present invention are as follows:

[0032] 1) Compared with the traditional expert evaluation method, this method uses data analysis and decision tree models. The process of clinical pathway optimization is based on a large amount of historical data and actual medical records, and can more objectively reflect the real situation and treatment effect of patients; the decision tree model can automatically analyze various clinical information of patients, reduce human bias, thereby improving the accuracy and consistency of medical decisions, and thus optimizing the decision on the discharge time; the data-driven method can be updated and adjusted according to new clinical data, and can adapt to the changing medical environment and the discharge strategy of a specific hospital.

[0033] 2) The present invention uses the Max-Min method to normalize the feature data. The formula is simple and easy to understand, and intuitively converts the data into a specified range, which is convenient for medical staff to understand. However, considering that the Max-Min method is sensitive to abnormal points, and the feature data based on medical records has abnormal point problems caused by wide sources and complex data, therefore, through Gaussian distribution-based anomaly detection, these abnormal points are identified and processed to improve the accuracy of the model.

[0034] 3) Mutation detection in decision trees is not a common practice. In the prior art, cross-validation or other independent test sets are usually used to evaluate the model performance. The tree structure of the decision tree model determined by the information gain ratio has the advantage of visualization, and the generated rules are easy for medical staff to understand and can handle missing values in medical record data. However, in data with uneven class distributions, the decision tree may be biased and prone to constructing an overly complex model, resulting in overfitting. Therefore, cost complexity pruning is performed through the cost_complexity_pruning_path function, which can quickly find a model with generalization ability, and select the best pruning degree by weighing the impurity of the tree and the number of nodes to improve the model's prediction ability for new data.

[0035] 4) The present invention selects intraocular pressure, bleb morphology, and anterior chamber depth as characteristic data through correlation analysis and the opinions of domain experts, determines the types of complications such as pain, hyphema, and corneal edema that have a significant impact on the discharge date, and scientifically grades the complications to facilitate the determination of limited decision nodes; eliminates unnecessary complex characteristic data including patient age, medical history, etc. and special complication types, reducing the amount of data processing on the one hand and avoiding the generation of too many outliers and the construction of an overly complex model on the other hand. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a logic block diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] Example 1: This example describes a method for predicting the length of hospital stay after compound trabeculectomy based on a decision tree model. This method optimizes the clinical management of postoperative patients through steps such as data preprocessing, decision tree model construction, anomaly detection, model pruning, and predicting the patient's length of hospital stay.

[0038] Step 1) Data preprocessing: Collect data on intraocular pressure, bleb morphology, and the degree of anterior chamber depth opening of patients from intraoperative medical records. These data are original and have different dimensions and scales. To make these data suitable for the model, we use the Max-Min normalization method to process the data. Specifically, for each feature, calculate its maximum value (max) and minimum value (min) in the dataset, and then perform normalization. Each feature value is converted to the range of 0 to 1, forming a new sample dataset to prepare for subsequent model training.

[0039] Step 2) Construct a decision tree model: Integrate the normalized feature data and postoperative complication attributes into a sample dataset. Each complication attribute is converted into a numerical grade through a quantification method so that the model can process it. For example, the pain grade is divided using the VAS visual analogue scale, the hyphema grade is divided according to the proportion of blood accumulation, and the corneal edema grade is divided according to the degree of edema. Use these quantified complication attributes as decision nodes to divide the dataset into multiple subsets and recursively construct subtrees until all attributes are used for data division, forming a complete decision tree model.

[0040] Step 3) Anomaly detection: After the decision tree model training is completed, perform anomaly detection. By calculating the Gaussian distribution probability density function corresponding to the feature data, outliers can be identified. Adjust the probability density threshold contamination to control the sensitivity of outlier detection, and mark the detected outliers as an independent category for special processing in subsequent analysis.

[0041] Step 4) Model Pruning: To improve the generalization ability of the model, post-pruning is performed on the constructed decision tree model. Using the cost_complexity_pruning_path function, the pruning path can be determined, and the optimal α value can be found. The α value is used to balance the complexity of the model and the fitting degree of the training data, thereby avoiding overfitting and improving the performance of the model on new data.

[0042] Step 5) Predicting the Standard Length of Hospital Stay: Input the intraocular pressure, bleb morphology, and anterior chamber depth data of the new surgical patients into the trained decision tree model. The model predicts the postoperative complication grade based on these input data. Then, by combining the average length of hospital stay of the same-grade complications in the historical data, the standard length of hospital stay of the new patient is predicted, providing support for clinical decision-making.

[0043] Apply the method obtained above to the ophthalmology department of a tertiary hospital in Yunnan for the discharge assessment of inpatients undergoing combined trabeculectomy. First, select the medical records of patients who underwent combined trabeculectomy from January to December 2022. To avoid data sample bias, select the medical records of patients with an age range of 20 - 60, no previous medical history, and no drug allergy history, and try to avoid the influence of other factors on the complexity of the surgery and the length of hospital stay. Normalize the intraocular pressure, bleb morphology, and anterior chamber depth data in the medical records as the sample data set. Select postoperative complication attributes such as patient pain, anterior chamber depth, and corneal edema for quantitative grading. Use these complication attributes as decision nodes to divide the data set into multiple subsets and recursively construct subtrees until all attributes are used for data division to form a complete decision tree model. After completing mutation detection and model pruning, use the patient medical record data from January to March 2023 as the test set to test the model.

[0044] Preoperative conjunctival congestion may indicate the presence of inflammation or infection in the eye, resulting in the need for closer observation of the regression of conjunctival congestion during the postoperative care of combined trabeculectomy; postoperative care measures for the bleb, such as massage and suture removal, can effectively prevent premature scarring of the filtration pathway and maintain the function of the bleb; to ensure the quality of the model, the preoperative conjunctival congestion grade, the bleb scarring grade, and the conjunctival congestion grade at the two-week postoperative review of the patient are respectively introduced into the existing model and input into the decision tree model together with the existing feature data such as intraocular pressure, bleb morphology, and anterior chamber depth to mark the predicted results of the length of hospital stay output by the model. The results show that the indicators such as the accuracy, precision, recall rate, and F1 score of the model do not change significantly, indicating that the preoperative conjunctival congestion and postoperative bleb scarring data have no positive impact on the model performance, and the prediction performance of the original model is excellent.

[0045] Patient A underwent a combined trabeculectomy. The intraocular pressure was 18.5 mmHg. The filtering bleb was flat and diffuse, and the anterior chamber depth was 3.0 mm. During the operation, 2 mm * 1 mm of trabecula was resected, the width of the resected iris was 2 mm, COD was 0.04, OD was 10.3 mmHg. The conjunctiva of the right eye was congested, the aqueous humor was clear, and a circumferential incision above the iris was visible. Mild pain occurred after the operation. The amount of hyphema was less than 1 / 3 of the anterior chamber. There was local corneal haze, and the corneal endothelium was smooth. The predicted length of hospital stay obtained through the decision tree model was 2 days, and the actual length of hospital stay was 2 days. The prediction result was consistent with the actual result.

Claims

1. A prediction method for the length of hospital stay after compound trabeculectomy based on a decision tree model, characterized in that Obtained through the following steps: 1) Data preprocessing: In the collected intraoperative medical records, the feature data is normalized using the Max-Min method, and the valid feature data is retained as the input data for the model. The valid feature data includes intraocular pressure, bleb morphology, and anterior chamber depth data. 2) Construct a decision tree model: Aggregate the valid feature data into a sample data set, divide the data according to the complications in the postoperative examination report, and divide the sample data set into several subsets; Aggregate all samples in the subset under the same complication level to form a recursively constructed sub-tree. 3) Mutation detection: Perform mutation detection on the trained decision tree, and mark the data generated by the mutation as abnormal. Use the test set to evaluate the constructed decision tree model, and evaluate the accuracy, precision, recall, and F1 score of the data marked as normal. 4) Pruning: Perform post-pruning on the constructed decision tree model, remove the branches that are not helpful for improving the generalization ability of the model, and adjust the maximum tree depth and minimum sample split number of the model to improve the performance of the model. 5) Predict the standard length of hospital stay: Input the intraocular pressure, bleb morphology, and anterior chamber depth data of the new surgical patient into the trained decision tree model. The model will predict the postoperative complication level based on the input intraoperative medical record data, and combine the average length of hospital stay of the same-level complications in the historical data to predict the standard length of hospital stay of the patient.

2. The prediction method of the length of hospital stay after compound trabeculectomy based on the decision tree model according to claim 1, wherein The specific method of the normalization process described in step 1) is: Calculate the maximum value max and minimum value min of the intraocular pressure, bleb morphology, and anterior chamber depth features in the data set respectively, and use the Max-Min normalization formula to transform each feature data, that is: Among them, x′ is the normalized data, x is the original data, and each feature value in the original data set is replaced with the normalized value to form a new sample data set.

3. The prediction method of the length of hospital stay after compound trabeculectomy based on the decision tree model according to claim 1, wherein The complication attributes described in step 2) include pain, anterior chamber hemorrhage, and corneal edema. To achieve data set division, the above complications are quantified by level, where: The pain level is quantified by the VAS visual analogue scale. Anterior chamber hemorrhage level: Grade 1 is when the amount of accumulated blood is less than 1 / 3 of the anterior chamber, Grade 2 is when the accumulated blood accounts for 1 / 3 - 1 / 2 of the anterior chamber, Grade 3 is when the accumulated blood > 1 / 2 and < 3 / 4, and Grade 4 is when the accumulated blood > 3 / 4 to the entire anterior chamber (the tissue behind the cornea cannot be seen). Corneal edema level: Grade 1 is when the cornea is transparent without edema; Grade 2 is when the cornea has local haze-like edema, the corneal endothelium is smooth, and the iris texture is still clearly visible; Grade 3 is when the cornea has light gray edema, the corneal endothelium is rough, and the iris texture is blurred; Grade 4 is when the cornea has diffuse grayish-white edema, the corneal endothelium is cracked, and the iris texture is unclear; Grade 5 is when the cornea has milky white edema and the intraocular structure is unclear.

4. The prediction method for the length of hospital stay after compound trabeculectomy based on the decision tree model according to claim 1, wherein The specific method of data set division in step 2) is: The sample data set is divided into multiple subsets with the complication attribute determined by the information gain ratio as the decision node. A new decision node is created for each subset, and the data of each subset will be passed to this new node until all complication attributes are used for data division.

5. The prediction method for the length of hospital stay after compound trabeculectomy based on the decision tree model according to claim 1, wherein The steps for determining the information gain ratio are as follows: 1) Calculate entropy: First, calculate the entropy of the sample dataset, which represents the uncertainty of the dataset. The formula for entropy is: IG(D, A) = H(D) - H(D|A) where H(D) is the information entropy of dataset D, and H(D│A) is the conditional entropy of dataset D under the condition of feature A; 2) Calculate the intrinsic value IV(X) of the feature: Among them, D i is the data subset corresponding to the i-th value of feature X, │D i │ is the number of samples in subset D i , │D│ is the total number of samples in the original data set, and n is the number of values of feature X; 3) Calculate the information gain ratio GR(D, A): where IG(D, A) is the information gain of feature A; IV(A) is the intrinsic value of feature A.

6. The prediction method for the length of hospital stay after compound trabeculectomy based on the decision tree model according to claim 1, wherein The specific method of anomaly detection in step 3) is: By calculating the Gaussian distribution probability density function corresponding to the feature data, it can be used to find the anomaly points in the feature data. By modifying the probability density threshold contamination, the sensitivity of anomaly point detection can be adjusted, and the anomaly points are marked as an independent category.

7. The prediction method for the length of hospital stay after compound trabeculectomy based on the decision tree model according to claim 1, wherein The post - pruning in step 4) uses cost complexity pruning and uses the cost_complexity_pruning_path function from sklearn.tree import DecisionTreeClassifier to determine the pruning path and determine the optimal α value.

Images