Postoperative risk prediction method and device for hypertrophic obstructive cardiomyopathy based on a multi-modal deep learning model

Abstract

The application discloses a hypertrophic obstructive cardiomyopathy postoperative risk prediction method and device based on a multi-modal deep learning model, relates to the field of cardiomyopathy postoperative risk prediction, and comprises the following steps: acquiring preoperative static electrocardiograms, postoperative static electrocardiograms, preoperative dynamic electrocardiographic data and postoperative dynamic electrocardiographic data of a target hypertrophic obstructive cardiomyopathy patient, applying a risk prediction model based on deep learning, and predicting the postoperative risk probability of the target hypertrophic obstructive cardiomyopathy patient; the risk prediction model comprises two image processing branches and two numerical processing branches, a fusion layer connected with the four branch output ends, and a full connection layer connected with the fusion layer output end. The application adopts the preoperative and postoperative multi-modal data of the patient, and in combination with the above-mentioned risk prediction model, can accurately predict the postoperative risk probability of the patient.

Description

Technical Field

[0001] This application relates to the field of postoperative risk prediction for cardiomyopathy, and in particular to a method and device for postoperative risk prediction of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model. Background Technology

[0002] Hypertrophic cardiomyopathy (HCM) is one of the most common inherited cardiomyopathy, with an estimated prevalence of approximately 1 in 500 in the general population. It is characterized by asymmetric left ventricular hypertrophy, with about 60% of patients exhibiting left ventricular outflow tract obstruction (LVOTO), a phenotype known as hypertrophic obstructive cardiomyopathy (HOCM). Transaortic septal myocardectomy is one of the core treatments for HOCM, significantly relieving obstruction-related symptoms and improving cardiac function. However, due to the effects of cardiopulmonary bypass and surgical trauma, septal myocardectomy may further increase the risk of arrhythmias in HOCM patients. Persistent atrial fibrillation (AF) or new-onset atrial fibrillation (NOAF), malignant ventricular arrhythmias (MVA), and major adverse cardiovascular events (MACE) constitute a significant disease burden in postoperative HOCM patients and significantly impact their long-term prognosis. Therefore, the focus of clinical management for postoperative HOCM patients should shift to the prevention and intervention of arrhythmias and MACE.

[0003] Atrial arrhythmias (AFs) are a particularly significant burden in patients undergoing ventricular septal myocardiectomy (VSM). The perioperative incidence of AF can be as high as 33.7%, with 22.5% being new-onset atrial fibrillation (NOAF). During follow-up, delayed AF remains common, with an incidence ranging from 9% to 18.9%. Early postoperative NOAF is also associated with subsequent NOAF and adverse long-term outcomes. Furthermore, although VSM significantly reduces the overall burden of microarrhythmias (MVAs) and sudden cardiac death (SCD) in patients with holocytic ventricular myocardial infarction (HOCM), postoperative MVA events, while relatively rare, still have a very high lethality. In a cohort of 1,915 HOCM patients with a mean follow-up of 4.6 years, the incidence of SCD due to MVA was approximately 1%, but it accounted for 35.7% of all deaths, suggesting that once MVA occurs, it may directly determine long-term survival. Other cardiac events (MACEs) also impose a significant burden during the perioperative period and long-term follow-up. One study reported an overall perioperative MACE incidence of approximately 39.04%. In addition, nearly half of the patients experienced cardiovascular-related readmissions within 2 to 3 years after surgery, which may be a key determinant of prognosis in real-world situations.

[0004] In summary, patients with hypertrophic obstructive cardiomyopathy (HOCM) continue to face the risk of adverse cardiovascular events after ventricular septal myocardectomy, and there is currently a lack of reliable, non-invasive, and clinically applicable postoperative risk stratification tools. Summary of the Invention

[0005] The purpose of this application is to provide a method and device for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model. It can use the patient's preoperative and postoperative multimodal data and combine them with the constructed risk prediction model to accurately predict the patient's postoperative risk probability.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model, including: Acquire preoperative static electrocardiogram, postoperative static electrocardiogram, preoperative dynamic electrocardiogram data, and postoperative dynamic electrocardiogram data of patients with target hypertrophic obstructive cardiomyopathy; Based on preoperative static electrocardiogram (ECG), postoperative static ECG, preoperative Holter ECG data, and postoperative Holter ECG data, a deep learning-based risk prediction model was applied to predict the postoperative risk probability of patients with target hypertrophic obstructive cardiomyopathy. The risk prediction model includes two image processing branches, two numerical processing branches, a fusion layer connected to the outputs of the two image processing branches and the two numerical processing branches, and a fully connected layer connected to the output of the fusion layer. The inputs to the two image processing branches are the preoperative static ECG and the postoperative static ECG, respectively; the inputs to the two numerical processing branches are the preoperative Holter ECG data and the postoperative Holter ECG data, respectively.

[0007] Secondly, this application provides a postoperative risk prediction device for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model, comprising: A static electrocardiograph is used to collect preoperative and postoperative static electrocardiograms of patients with target hypertrophic obstructive cardiomyopathy. A dynamic electrocardiogram (ECG) recorder is used to collect preoperative and postoperative dynamic ECGs of patients with target hypertrophic obstructive cardiomyopathy. The controller is used to acquire preoperative static electrocardiograms, postoperative static electrocardiograms, preoperative Holter electrocardiograms, and postoperative Holter electrocardiograms, and to execute the above-mentioned risk prediction method for postoperative hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model.

[0008] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model.

[0009] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model.

[0010] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model.

[0011] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method and apparatus for predicting postoperative risk in hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model. The method includes: acquiring preoperative static electrocardiograms (ECGs), postoperative static ECGs, preoperative Holter ECG data, and postoperative Holter ECG data of a target patient with hypertrophic obstructive cardiomyopathy; applying a deep learning-based risk prediction model to predict the postoperative risk probability of the target patient based on these data; the risk prediction model includes two image processing branches, two numerical processing branches, a fusion layer connected to the outputs of the two image processing branches and the two numerical processing branches, and a fully connected layer connected to the output of the fusion layer; the inputs to the two image processing branches are the preoperative static ECG and the postoperative static ECG, respectively; the inputs to the two numerical processing branches are the preoperative Holter ECG data and the postoperative Holter ECG data, respectively. This application utilizes the patient's preoperative and postoperative multimodal data (static ECG and Holter ECG data) and combines them with the constructed risk prediction model to accurately predict the patient's postoperative risk probability. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is an application environment diagram of a postoperative risk prediction method for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model, as described in one embodiment of this application. Figure 2 A flowchart illustrating a method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model, provided in an embodiment of this application; Figure 3 A schematic diagram of the structure of a deep learning-based risk prediction model provided in an embodiment of this application; Figure 4 This diagram illustrates the discrimination ability, precision-recall performance, and calibration status of a multimodal model provided in an embodiment of this application. Figure 5 A schematic diagram illustrating the time dependency prediction performance of a multimodal model on a validation set, provided in an embodiment of this application. Figures 6-7 A schematic diagram showing the performance comparison of the combined ECG, 12-lead ECG, and Holter ECG models provided in an embodiment of this application on an internal validation set; Figures 8-11 A schematic diagram illustrating the time-dependent performance of a 12-lead ECG model, a dynamic ECG model, and a combined ECG model provided in an embodiment of this application on a validation set. Figure 12 A schematic diagram illustrating the comparison between a multimodal model and a traditional clinical risk score in a validation set, provided as an embodiment of this application; Figure 13 A schematic diagram of the Kaplan-Meier risk stratification method for a multimodal model in a validation set, provided as an embodiment of this application; Figure 14 This diagram illustrates the performance and clinical relevance of a test set for a multimodal model provided in an embodiment of this application. Figure 15 A schematic diagram illustrating the time dependency prediction performance of a multimodal model on a test set, provided in an embodiment of this application. Figure 16 A schematic diagram showing the comparison between a multimodal model provided in one embodiment of this application and a traditional clinical risk score on a test set; Figures 17-18 A schematic diagram comparing the performance of combined ECG, 12-lead ECG, and dynamic ECG models in a test set provided for an embodiment of this application; Figures 19-22 A schematic diagram illustrating the time-dependent performance of a 12-lead ECG model, a dynamic ECG model, and a combined ECG model provided in an embodiment of this application on a test set. Figure 23 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] Cardiac arrhythmias and MACE are critical aspects of postoperative management in hypertrophic obstructive cardiomyopathy (HOCM). Accurate prediction of arrhythmias, non-arrhythmias, microvascular arrhythmias (MVA), and MACE is essential for achieving high-quality long-term postoperative survival. Resting electrocardiography (ECG) offers advantages such as being non-invasive, low-cost, highly standardized, and readily accessible, making it a routine examination method for HOCM patients during the perioperative period and long-term follow-up. It effectively reflects changes in cardiac surface electrophysiology and often indicates early abnormalities before obvious clinical manifestations appear. However, ECG interpretation is highly observer-dependent, and many ECG changes are subtle and easily overlooked in routine clinical practice. Therefore, this study explores using artificial intelligence (AI) to directly perform deep learning on ECG images to improve operational feasibility and expand clinical deployment potential. However, most current AI-ECG prediction models are derived from HCM cohorts receiving drug therapy and do not consider ECG changes introduced by ventricular septal resection, thus limiting their generalization in the surgical population. Overall, existing AI-ECG models related to HOCM are still not refined enough and have limited predictive power for postoperative risks. Currently, there is a lack of reliable AI-ECG models for predicting long-term risks after HOCM surgery. To address this, this application proposes a method and device for predicting postoperative risks in hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model. This method introduces a risk prediction model based on deep learning, which can accurately analyze subtle changes in preoperative and postoperative static and dynamic ECG data, thereby accurately predicting postoperative risks.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] The postoperative risk prediction method for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the preoperative static electrocardiogram, postoperative static electrocardiogram, preoperative Holter monitoring data, and postoperative Holter monitoring data of the target hypertrophic obstructive cardiomyopathy patient to server 104. After receiving the preoperative static electrocardiogram, postoperative static electrocardiogram, preoperative Holter monitoring data, and postoperative Holter monitoring data of the target hypertrophic obstructive cardiomyopathy patient, server 104 applies a deep learning-based risk prediction model to predict the postoperative risk probability of the target hypertrophic obstructive cardiomyopathy patient based on the preoperative static electrocardiogram, postoperative static electrocardiogram, preoperative Holter monitoring data, and postoperative Holter monitoring data. The risk prediction model includes two image processing branches and two numerical processing branches, a fusion layer connected to the outputs of the two image processing branches and the two numerical processing branches, and a fully connected layer connected to the output of the fusion layer. The inputs of the two image processing branches are the preoperative static electrocardiogram and the postoperative static electrocardiogram, respectively; the inputs of the two numerical processing branches are the preoperative Holter monitoring data and the postoperative Holter monitoring data, respectively. Server 104 can feed back the obtained postoperative risk probability of the target hypertrophic obstructive cardiomyopathy patient to terminal 102. Furthermore, in some embodiments, the postoperative risk prediction method for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model can also be implemented independently by server 104 or terminal 102. For example, terminal 102 can directly perform postoperative risk prediction based on a multimodal deep learning model using the preoperative static electrocardiogram, postoperative static electrocardiogram, preoperative dynamic electrocardiogram data, and postoperative dynamic electrocardiogram data of the target hypertrophic obstructive cardiomyopathy patient. Alternatively, server 104 can obtain the preoperative static electrocardiogram, postoperative static electrocardiogram, preoperative dynamic electrocardiogram data, and postoperative dynamic electrocardiogram data of the target hypertrophic obstructive cardiomyopathy patient from the data storage system and perform postoperative risk prediction based on a multimodal deep learning model.

[0018] Among them, terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices and portable wearable devices, and server 104 can be implemented by independent servers or server clusters composed of multiple servers, or it can be a cloud server.

[0019] In one exemplary embodiment, such as Figure 2 As shown, a method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model is provided. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 202. Wherein: Step 201: Obtain the preoperative resting electrocardiogram (ECG) of the target hypertrophic obstructive cardiomyopathy patient (corresponding to...). Figure 3 "Pre-op ECG Image" and postoperative static ECG (corresponding to) Figure 3 "Post-op ECG Image" and preoperative dynamic electrocardiogram data (corresponding to) Figure 3 The “Pre-op Holter Features” and postoperative dynamic electrocardiogram data (corresponding to) Figure 3 (See "Post-op Holter Features"). Static ECG can use 12-lead ECG images; Holter ECG data can be extracted using a Holter recorder.

[0020] Step 202: Based on the preoperative static electrocardiogram, postoperative static electrocardiogram, preoperative Holter electrocardiogram data, and postoperative Holter electrocardiogram data, a deep learning-based risk prediction model is applied to predict the postoperative risk probability of the target hypertrophic obstructive cardiomyopathy patient. The risk prediction model includes two image processing branches and two numerical processing branches, a fusion layer connected to the outputs of the two image processing branches and two numerical processing branches, and a fully connected layer connected to the output of the fusion layer. The inputs to the two image processing branches are the preoperative static electrocardiogram and the postoperative static electrocardiogram, respectively; the inputs to the two numerical processing branches are the preoperative Holter electrocardiogram data and the postoperative Holter electrocardiogram data, respectively.

[0021] The postoperative risk probability output by the risk prediction model for patients with target hypertrophic obstructive cardiomyopathy includes at least one of the following: atrial fibrillation risk probability (risk probability of AF occurrence), new-onset atrial fibrillation risk probability (risk probability of NOAF occurrence), malignant ventricular arrhythmia risk probability (risk probability of MVA occurrence), and major adverse cardiovascular event risk probability (risk probability of MACE occurrence).

[0022] The specific definitions of AF, NOAF, MVA, and MACE are as follows: (1) AF is defined as a heart arrhythmia recorded on ECG after the patient is discharged, characterized by irregular RR intervals, no identifiable P waves, and a duration of more than 30 seconds. This definition includes persistent and permanent AF. NOAF is defined as: AF occurring after surgery in patients who have no AF record on ECG before surgery and have no history of anti-AF treatment.

[0023] (2) MVA is defined as the composite endpoint consisting of the first occurrence of any of the following events, with reference to previous criteria: ① Appropriate treatment with an implantable cardioverter defibrillator (ICD) or cardiac resynchronization therapy defibrillator (CRT-D), including ICD shock to terminate ventricular fibrillation or ventricular tachycardia, or antitachycardia pacing for sustained ventricular tachycardia; ② Successful cardiopulmonary resuscitation, defined as effective basic life support after cardiac arrest; ③ Sudden cardiac death, defined as witnessed sudden death (regardless of whether ventricular fibrillation is recorded), death within 1 hour of the onset of acute symptoms, or unexplained death at night without warning of clinical deterioration.

[0024] (3) MACE was defined as a composite endpoint consisting of the following events: stroke, new-onset arrhythmia (including AF and sustained ventricular tachycardia / fibrillation), NYHA functional deterioration (≥1 grade increase), heart transplantation or left ventricular assist device implantation, SCD, appropriate ICD or CRT-D therapy, and cardiovascular death. All events were determined by two independent cardiologists based on electronic medical record review and telephone follow-up.

[0025] In another exemplary embodiment of this application, such as Figure 3 As shown, a four-branch neural network architecture was designed to integrate heterogeneous ECG modalities. Two image processing branches process preoperative and postoperative 12-lead ECG images, respectively, while two numerical processing branches process preoperative and postoperative Holter-based feature extraction (dynamic ECG data), respectively. Each image processing branch uses a ResNet-50 convolutional neural network pre-trained on ImageNet to extract high-level visual representations. After global average pooling, each image processing branch outputs a 2048-dimensional feature vector. Holter-based feature extraction is processed by a multilayer perceptron, which consists of two fully connected layers (i.e., ...). Figure 3 The architecture consists of four linear layers (with batch normalization and ReLU activation functions), each outputting a 64-dimensional feature embedding. The feature embeddings from the four branches are concatenated in a fusion layer, then passed through a dropout layer (ratio = 0.3) and a fully connected layer, mapping to the logit outputs specific to four risk indicators (AF, NOAF, MVA, MACE). After sigmoid activation, the predicted probability for each risk indicator is obtained. This architecture supports both full multimodal inference and simplified configurations using only ECG images or only Holter extraction for features.

[0026] Based on the above, such as Figure 3As shown, the image processing branch includes: a ResNet-50 convolutional neural network, an adaptive average pooling layer, and a flattening layer connected in sequence. The numerical processing branch includes: multiple numerical processing units connected in sequence; each numerical processing unit includes a linear layer and a batch normalization and activation layer connected in sequence.

[0027] In another exemplary embodiment of this application, for training and validating the risk prediction model, electrocardiogram (ECG) data from 1,324 patients undergoing surgical ventricular septal myocardiotomy (HOCM) were used to construct a longitudinal dataset. ECG data were collected both preoperatively and postoperatively, allowing for comparison of electrophysiological characteristics before and after the surgical intervention at the patient level. For each patient, the ECG data included standard 12-lead ECG images and quantitative features extracted from Holter monitoring (ambulatory ECG data). ECG images acquired within 30 days preoperatively were included. If multiple qualifying ECG examinations were available at a given time point, the record closest to the surgery date was selected to maximize temporal relevance. Patients underwent standard 12-lead ECG and Holter monitoring before discharge. For postoperative ECG image analysis, the record closest to the discharge date was selected. To ensure the completeness and consistency of the input dimensions of the risk prediction model, missing Holter quantitative features were filled with zero values.

[0028] Data was randomly divided into training (70%), validation (15%), and test (15%) sets based on patient characteristics, ensuring that the same patient did not appear in multiple subsets. A fixed random seed was used to ensure reproducibility. ECG images were cropped to remove headers, footers, and borders, uniformly adjusted to a fixed resolution, and normalized using a fixed mean and standard deviation. Holter numerical features (dynamic ECG data) were cleaned by removing units and non-numeric symbols, and missing or unreasonable values ​​were set to 0. The mean and standard deviation of each feature were calculated based on the training set and used for z-score standardization across all subsets.

[0029] The process of training a deep learning-based risk prediction model using the training set defined above is as follows: (1) Obtain the training set; the training set includes preoperative and postoperative monitoring data of several patients with hypertrophic obstructive cardiomyopathy and the corresponding postoperative actual risk type (AF, NOAF, MVA, MACE); the preoperative and postoperative monitoring data include preoperative static electrocardiogram samples, postoperative static electrocardiogram samples, preoperative dynamic electrocardiogram data samples and postoperative dynamic electrocardiogram data samples.

[0030] (2) Using the preoperative and postoperative monitoring data in the training set as input and the corresponding postoperative actual risk type as label, train the initial risk prediction model until the loss error of the model converges, and obtain the risk prediction model based on deep learning.

[0031] To address the class imbalance problem, a weighted random sampling strategy is employed during training. Instead of directly reading samples sequentially during training, a sampling weight is first calculated for each training sample, and then a WeightedRandomSampler (a commonly used sampler in deep learning frameworks such as PyTorch) is used to randomly sample samples according to the weights. The specific steps are as follows: For a sample, its corresponding binary label is checked for positivity, and the number of positive labels is counted. Suppose that during a certain training session, a certain sample has a total of If there are 1 label, then the sampling weight of the sample is calculated according to... Calculate, where, A tag, This is an adjustable hyperparameter, defaulting to 10. When there are no positive labels in a sample, the weight is fixed at 1; once at least one positive label is present, the weight becomes greater than 1, and the more positive labels there are, the higher the weight. After calculating the weights of all samples, a weight vector is constructed and passed to WeightedRandomSampler, enabling weighted random sampling with replacement. In this way, samples are randomly drawn from each mini-batch according to their weight proportions. Positive samples are oversampled to increase their frequency of occurrence, while negative samples still participate in training with lower weights. This achieves a weighted random sampling strategy that dynamically adjusts the sampling weights based on the number of positive labels.

[0032] The risk prediction model is trained using the Adam optimizer with an initial learning rate of 1×10⁻⁶. -4 The batch size is 8. The learning rate is reduced to 0.5 times the original value every 5 epochs. Focal loss (α=0.25, γ=2) is used to further mitigate class imbalance. The model that performs best on the validation set is selected. For each risk metric, specificity thresholding is performed on the validation set by evaluating a threshold in the range of 0.05 to 0.95. Thresholds that meet the preset minimum sensitivity and specificity criteria (both ≥0.4) are included in the candidate list, and the one with the highest balancing accuracy is selected. If no threshold meets the above criteria, the threshold that maximizes the balancing accuracy is selected.

[0033] To evaluate the added value of multimodal integration, this application constructed baseline models (with consistent model structures, using only image processing or numerical processing branches) for comparison, i.e., performance comparisons with models using only two image processing branches or two numerical processing branches. All baseline models employed the same data partitioning, training process, and evaluation scheme as the multimodal model (i.e., the deep learning-based risk prediction model used in this application). After fixing optimized outcome specificity thresholds, model performance was evaluated on an independent test set, which was not used for any further parameter tuning. Predicted probabilities were converted to binary classification results based on these thresholds, and model classification performance was evaluated using sensitivity, specificity, positive predicted value, negative predicted value, accuracy, F1 score, area under the curve (AUC), mean precision, and Brillouin score. Model calibration was evaluated using calibration slope and intercept, estimated calibration index, etc. Statistical uncertainty was quantified by bootstrap resampling (1,000 times) to obtain 95% confidence intervals for AUC, mean precision, and estimated calibration index, etc. All experiments were conducted on a workstation equipped with a single NVIDIA RTX 4080 Super GPU.

[0034] In the validation set, model performance was summarized by classification ability, precision-recall analysis under class imbalance, and calibration. Classification ability was evaluated using receiver operating characteristic (ROC) curves. Complete discrimination, calibration, and threshold-related performance metrics for all endpoints (risk indicators), datasets, and model variants are shown in Tables 1 and 2.

[0035] Table 1 Overall Discrimination and Calibration Indicators for Different Risk Indicators, Datasets, and Model Variants

[0036] In Table 1, AP represents the average accuracy; ECI represents the estimated calibration index; Brier score is the Brier score; Calibration intercept represents the calibration intercept; and Calibration slope represents the calibration slope.

[0037] Table 2 Threshold-based classification metrics for different risk indicators, datasets, and model variants

[0038] In Table 2, PPV is the positive predictive value; NPV is the negative predictive value.

[0039] Figure 4 The results of the multimodal model on the validation set are shown, such as... Figure 4In Part (a), the AUC for AF was 0.827, with 95% confidence intervals (CI) ranging from 0.722 to 0.915. The AUC for NOAF was 0.814 (95% CI: 0.697–0.919). The AUC for MVA was 0.817 (95% CI: 0.694–0.936). The AUC for MACE was 0.829 (95% CI: 0.759–0.893). Under imbalanced outcome conditions, precision-recall curves were used to evaluate model performance. MACE had the highest mean precision, at 0.779 (95% CI: 0.673–0.861). The mean accuracy of AF was 0.620 (95% CI: 0.434–0.775), MVA was 0.532 (95% CI: 0.272–0.777), and NOAF was 0.499 (95% CI: 0.252–0.719). Figure 4 (Part (b)). Calibration was evaluated by logistic calibration fit and estimation of the calibration index. The estimated calibration index for AF was 0.032 (95% CI: 0.018–0.049) (e.g., Figure 4 (part (c)). NOAF was 0.012 (95% CI: 0.004–0.022), MVA was 0.007 (95% CI: 0.003–0.014), and MACE was 0.089 (95% CI: 0.064–0.117) (e.g. Figure 4 (part (c)).

[0040] Overall, these analyses show that the model has consistent discriminative power across all endpoints, maintains measurable accuracy under class imbalance conditions, and is generally consistent with the observed risk.

[0041] To evaluate the discriminative power of the multimodal model over time, this application examined its performance at different follow-up time windows in the validation set. At 1 year, the AUC of AF was 0.816 (95% CI: 0.646–0.947). Figure 5 (a) NOAF was 0.753 (95% CI: 0.571–0.914), MVA was 0.904 (95% CI: 0.703–1.000), and MACE was 0.777 (95% CI: 0.693–0.861) (e.g. Figure 5 (Part (a)). At 3 years, the AUC of AF was 0.785 (95% CI: 0.655–0.899) (e.g. Figure 5(part (b)); NOAF was 0.788 (95% CI: 0.642–0.907), MVA was 0.801 (95% CI: 0.595–0.985), and MACE was 0.815 (95% CI: 0.741–0.882) (e.g. Figure 5 (Part (b)). At 5 years, the AUC of AF was 0.839 (95% CI: 0.734–0.931) (e.g. Figure 5 (part (c)); NOAF was 0.788 (95% CI: 0.642–0.907), MVA was 0.857 (95% CI: 0.705–0.984), and MACE was 0.819 (95% CI: 0.747–0.887) (e.g. Figure 5 (Part (c)). The AUC follow-up trajectory for different endpoints was generally relatively stable (e.g., Figure 5 (part (d)). The trajectories of sensitivity and specificity over time are generally stable, and specificity is always higher than sensitivity (e.g., Figure 5 (e) and (f) of the present application). Compared with the single-modal model, the multimodal fusion of this application consistently improves the discriminative power and calibration performance for all endpoints in the validation set (e.g., Figure 6 and Figure 7 As shown). The time-dependent AUC, as well as threshold-based sensitivity and specificity, remained stable across different follow-up time windows (e.g.). Figures 8 to 11 (As shown). Detailed distinctions and calibration indices between multimodal and unimodal models in the validation set are shown in Table 1, and threshold-related classification indices are shown in Table 2.

[0042] This application compared the classification ability of multimodal models in the validation set with commonly used clinical risk scores using AUC, and employed the bootstrap test to perform pairwise comparisons of AUC between models. For AF, the multimodal models (corresponding to Figure 12 The AUC of AI-ECG was 0.827 (95% CI: 0.722–0.915), which was higher than several single clinical models, including HCM-AF score (AUC=0.627, 95% CI: 0.521–0.728, P=0.0060, as shown in the figure). Figure 12 (a) and (b) in the text), CHA2DS2-VASc score (AUC=0.460, 95% CI: 0.338–0.578, P=0.0020, as shown in ...b) in the text). Figure 12 (a) and (b) in the table), CHADS2 score (AUC=0.376, 95% CI: 0.263–0.490, P=0.0020, as shown in the table). Figure 12(a) and (b) in the text), CHARGE-AF score (AUC=0.591, 95% CI: 0.443–0.740, P=0.0320, as shown in (a) and (b) in the text), Figure 12 (a) and (b) in the table) and POAF score (AUC=0.629, 95% CI: 0.514–0.731, P=0.0100, as shown in the table). Figure 12 (a) and (b) in the table). For MVA, the AUC of the multimodal model was 0.817 (95% CI: 0.694–0.936), which was also higher than that of the HCM-SCD model (AUC = 0.560, 95% CI: 0.313–0.798, P = 0.0380, as shown in (a) and (b) in the table. Figure 12 (c) and (d) in the text.

[0043] In the validation set, Kaplan-Meier analysis was used to assess whether the risk tiers defined by the multimodal model corresponded to survival differences in the absence of events. For AF, NOAF, MVA, and MACE, the high-risk group showed earlier and more frequent event occurrences than the lower-risk group (e.g., Figure 13 (a) to (d) in the text. Figure 13 The differences between groups in the four figures (a) to (d) were all supported by the log-rank test, with P values ​​all < 0.0001. These results indicate that risk classification based on a multimodal model can correspond to separable follow-up event occurrence patterns.

[0044] In the test set, the performance of the multimodal models was summarized by classification ability, precision-recall analysis, calibration, comparison with clinical risk scores, and Kaplan-Meier risk stratification. The ROC curve showed that the AUC for AF was 0.807 (95% CI: 0.703–0.893, as shown in the figure). Figure 14 (a) In the above, NOAF was 0.820 (95% CI: 0.710–0.912), MVA was 0.802 (95% CI: 0.613–0.959), and MACE was 0.800 (95% CI: 0.729–0.868). Figure 14 (a)). Precision-recall analysis showed that MACE had the highest mean precision, at 0.775 (95% CI: 0.689–0.853, as shown in (a)). Figure 14 (b)). The mean accuracy of AF was 0.571 (95% CI: 0.379–0.729), MVA was 0.496 (95% CI: 0.183–0.773), and NOAF was 0.485 (95% CI: 0.220–0.723, as shown in (b). Figure 14(b)). The calibration curves support a good agreement between predicted and actual risks. The estimated calibration index for AF is 0.034 (95% CI: 0.019–0.052, as shown in (b)). Figure 14 (c) NOAF was 0.016 (95% CI: 0.007–0.027), MVA was 0.008 (95% CI: 0.003–0.014), and MACE was 0.090 (95% CI: 0.064–0.119). Figure 14 (c)). The test set's time-dependent discrimination ability and threshold-based sensitivity and specificity trajectories are as follows: Figure 15 As shown.

[0045] Comparative analysis of the test set shows that the multimodal model (corresponding to) Figure 14 The AI-ECG score in this study demonstrates higher discriminative power compared to several commonly used scoring methods. For AF, the AUC of the multimodal model is 0.807, while HCM-AF is 0.622 (P=0.0180), CHA2DS2-VASc is 0.504 (P=0.0020), CHADS2 is 0.391 (P=0.0020), CHARGE-AF is 0.556 (P=0.0040), and the POAF score is 0.646 (P=0.0440). Figure 14 (d)). For MVA, the AUC of the multimodal model is 0.802, while that of the HCM-SCD model is 0.400 (P=0.0100) (e.g. Figure 14 (e)). A detailed comparison of the C-statistics between the multimodal model in the test set and the AF and MVA clinical risk scores, such as... Figure 16 As shown. Kaplan-Meier analysis revealed a separation of event-free survival between the high-risk and low-risk groups defined by the model for all four endpoints. Log-rank tests supported inter-group differences in AF, MVA, and MACE (all P < 0.0001), while the P for NOAF was 0.0063 (as shown in the figure). Figure 14 (f)-(i)). In the independent test set, the multimodal model (corresponding to) Figure 17 The Combined ECG in the model is compared to the single-mode model (corresponding to...). Figure 17 The performance advantages of Holter ECG and 12-lead ECG are retained (e.g., Figure 17 and Figure 18 As shown), and the ability to distinguish time dependence shows a consistent trend across different follow-up time windows (e.g. Figures 19 to 22 (As shown). The corresponding performance and calibration parameters are shown in Tables 1 and 2.

[0046] This application proposes a multimodal AI-ECG model (the aforementioned deep learning-based risk prediction model) that integrates paired 12-lead ECG images from preoperative and postoperative periods with quantitative features extracted by Holter studies to assess the long-term risk of HOCM patients after ventricular septal myocardectomy. The model demonstrated stability in discriminative power, calibration, and risk stratification analysis across four postoperative outcomes. Acceptable calibration, robust ROC and precision-recall performance, and consistent Kaplan-Meier separation between model-defined risk levels collectively support the view that the model's risk predictions capture clinically significant differences in long-term event risk, rather than merely classification results at a certain threshold. Unlike previous AI-ECG models that primarily focused on non-surgical HCM cohorts, the model developed in this application was specifically developed for surgical HOCM cohorts. Ventricular septal myocardectomy alters the natural course of the disease but can also introduce potential confounding due to the acute physiological effects of cardiac surgery. To minimize this impact, this application sets the risk follow-up starting point at discharge and uses strict criteria to define postoperative atrial fibrillation, thereby reducing the interference of transient perioperative events. Overall, the results of this application indicate that paired preoperative and postoperative ECG phenotypes combined with dynamic rhythm characteristics can provide informative and clinically relevant signals for risk assessment after ventricular septal myocardectomy.

[0047] The performance analysis results of the multimodal model in this application have a reasonable clinical interpretation. First, under class imbalance, performance differences between different endpoints (risk indicators) are expected: MACE has the highest mean precision, consistent with its higher event burden and broader composite phenotypes; while atrial and ventricular arrhythmia indicators, although having similar areas under the ROC curve, have lower mean precision. This difference between the ROC indicator and the precision-recall indicator is consistent with the characteristic that precision depends on the incidence of the outcome, and also supports the simultaneous reporting of these two types of indicators for rare postoperative events. Second, among several endpoints, the model's discriminative ability generally remained stable or even improved over longer prediction time windows. This pattern is compatible with the following interpretation: as long-term events accumulate, the occurrence of events increasingly reflects substrate-related electrophysiological characteristics captured by the burden of underlying disease and postoperative remodeling; even after left ventricular outflow tract obstruction has been relieved, atrial arrhythmias and other adverse events remain clinically significant during follow-up. Third, the very high discriminative power of MVA in short time windows should be interpreted with caution, as extreme AUC estimates may occur when the number of events is small in a specific time window; reporting the number of events and sensitivity analysis for different time windows will help clarify the stability of the estimation results.

[0048] The model constructed in this application is specifically designed for postoperative risk prediction and has achieved stable discriminative power across multiple long-term endpoints, with AUCs of AF 0.807, NOAF 0.820, MVA 0.802, and MACE 0.800 on the test set. Unlike other methods based on single preoperative ECG-derived surrogate indicators, the framework in this application integrates preoperative and postoperative paired ECG images and Holter features, enabling direct prediction of multiple postoperative risk indicators. Other ECG- and Holter-based studies have also shown that single ECG indicators are associated with specific postoperative myocardial resection complications, further supporting the value of ECG-based postoperative risk stratification.

[0049] Preoperative and postoperative paired designs and multimodal fusion likely contributed to improved performance of risk prediction models. Including postoperative ECGs may enable more longitudinal patient assessments, particularly capturing changes in ECG characteristics from preoperative to postoperative stages. Combining preoperative and postoperative ECGs could expand the information dimensions of risk prediction models. Furthermore, the persistence of high-risk ECG features in sequential recordings may indicate clinically significant longitudinal changes, potentially relevant to future risk stratification.

[0050] Furthermore, incorporating postoperative electrophysiological changes and high-dimensional ECG morphological features may further enhance the discriminative power and stability of the risk prediction model.

[0051] This application proposes a risk prediction model that uses multimodal information as input, integrating preoperative and postoperative paired 12-lead electrocardiogram (ECG) images, as well as perioperative 24-hour Holter monitoring features, to predict the risk probabilities of AF, NOAF, MVA, and MACE after ventricular septal myocardiotomy. Data from 1,324 patients were analyzed, including 2,648 12-lead ECGs and 1,529 Holter records. The data were divided into training, validation, and test sets (70 / 15 / 15) by patient level. The risk prediction model constructed in this application showed good discriminative ability for all risk indicators in the validation set (AUC: 0.814–0.829) with acceptable calibration; its performance remained consistent in the independent test set (AUC: 0.800–0.820). Compared with commonly used clinical risk scores, the risk prediction model constructed in this application has better discriminative ability in predicting AF and MVA, and successfully classifies patients into different risk groups to predict all risk indicators.

[0052] This application also provides an application scenario in which the above-mentioned method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy (HOCM) based on a multimodal deep learning model is applied. Specifically, the method for predicting postoperative risk of HOCM based on a multimodal deep learning model provided in this embodiment can be applied in the scenario of postoperative risk prediction for HOCM patients. This scenario includes a data acquisition stage and a prediction stage; the data acquisition stage is used to collect preoperative static electrocardiograms, postoperative static electrocardiograms, preoperative dynamic electrocardiogram data, and postoperative dynamic electrocardiogram data of HOCM patients; the prediction stage is used to perform postoperative risk prediction of HOCM based on the collected preoperative static electrocardiograms, postoperative static electrocardiograms, preoperative dynamic electrocardiogram data, and postoperative dynamic electrocardiogram data, and to obtain the postoperative risk probability. The method for predicting postoperative risk of HOCM based on a multimodal deep learning model provided in this embodiment belongs to the prediction stage.

[0053] In one exemplary embodiment, a postoperative risk prediction device for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model is provided, comprising: A static electrocardiograph is used to collect preoperative and postoperative static electrocardiograms of patients with hypertrophic obstructive cardiomyopathy.

[0054] A dynamic electrocardiogram (ECG) recorder is used to collect preoperative and postoperative dynamic ECGs in patients with target hypertrophic obstructive cardiomyopathy.

[0055] The controller is used to acquire preoperative static electrocardiograms, postoperative static electrocardiograms, preoperative Holter electrocardiograms, and postoperative Holter electrocardiograms, and to execute the above-mentioned risk prediction method for postoperative hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model.

[0056] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 23As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores postoperative risk prediction data for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a postoperative risk prediction method for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model.

[0057] Figure 23 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0058] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0059] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0060] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.

[0061] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0062] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0063] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0064] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting postoperative risk in hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model, characterized in that, include: Acquire preoperative static electrocardiogram, postoperative static electrocardiogram, preoperative dynamic electrocardiogram data, and postoperative dynamic electrocardiogram data of patients with target hypertrophic obstructive cardiomyopathy; Based on the preoperative static electrocardiogram, postoperative static electrocardiogram, preoperative dynamic electrocardiogram data, and postoperative dynamic electrocardiogram data, a risk prediction model based on deep learning was applied to predict the postoperative risk probability of patients with target hypertrophic obstructive cardiomyopathy. The risk prediction model includes two image processing branches and two numerical processing branches, a fusion layer connected to the outputs of the two image processing branches and two numerical processing branches, and a fully connected layer connected to the output of the fusion layer. The inputs to the two image processing branches are the preoperative static electrocardiogram and the postoperative static electrocardiogram, respectively. The inputs to the two numerical processing branches are preoperative dynamic electrocardiogram data and postoperative dynamic electrocardiogram data, respectively.

2. The method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model according to claim 1, characterized in that, The image processing branch consists of a ResNet-50 convolutional neural network, an adaptive average pooling layer, and a flattening layer connected in sequence.

3. The method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model according to claim 1, characterized in that, The numerical processing branch includes: multiple numerical processing units connected in sequence; each numerical processing unit includes a linear layer and a batch normalization and activation layer connected in sequence.

4. The method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model according to claim 1, characterized in that, Postoperative risk probabilities for patients with target hypertrophic obstructive cardiomyopathy include at least one of the following: risk of atrial fibrillation, risk of new-onset atrial fibrillation, risk of malignant ventricular arrhythmia, and risk of major adverse cardiovascular events.

5. The method for predicting postoperative risk of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model according to claim 4, characterized in that, The training process of a deep learning-based risk prediction model includes: Obtain the training set; the training set includes preoperative and postoperative monitoring data of several patients with hypertrophic obstructive cardiomyopathy and the corresponding actual postoperative risk types; the preoperative and postoperative monitoring data include preoperative static electrocardiogram samples, postoperative static electrocardiogram samples, preoperative dynamic electrocardiogram data samples, and postoperative dynamic electrocardiogram data samples. Using preoperative and postoperative monitoring data from the training set as input and the corresponding actual postoperative risk type as label, an initial risk prediction model is trained until the model's loss error converges, resulting in a risk prediction model based on deep learning.

6. A postoperative risk prediction device for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model, characterized in that, include: A static electrocardiograph is used to collect preoperative and postoperative static electrocardiograms of patients with target hypertrophic obstructive cardiomyopathy. A dynamic electrocardiogram (ECG) recorder is used to collect preoperative and postoperative dynamic ECGs of patients with target hypertrophic obstructive cardiomyopathy. A controller is used to acquire preoperative static electrocardiograms, postoperative static electrocardiograms, preoperative Holter electrocardiograms, and postoperative Holter electrocardiograms, and to execute the postoperative risk prediction method for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model as described in any one of claims 1 to 5.

7. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the method for postoperative risk prediction of hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the postoperative risk prediction method for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model as described in any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the postoperative risk prediction method for hypertrophic obstructive cardiomyopathy based on a multimodal deep learning model as described in any one of claims 1-5.

Images