Survival prediction method for hemodialysis patients based on feature-filled autoencoder

Abstract

The patent of this invention discloses a method for predicting survival of hemodialysis patients based on feature-filled autoencoder, which solves the problems of imbalance and missing data features in short-term hemodialysis data. First, the method uses the manifold structure of non-missing features and the intrinsic relationship between features in high-dimensional data space to infer the value of missing features. Then, the distribution of missing data in the deep learning framework is simulated by generating feature discarding masks, and a feature-filled autoencoder is designed. This module uses easily available short-term data for unsupervised learning, enabling the encoder to reconstruct missing features and derive latent representations. Finally, a classifier based on latent representations is used to achieve mortality prediction. The present invention aims to use short-term hemodialysis data to provide more accurate mortality risk assessment for hemodialysis patients.

Description

Technical Field

[0001] The present invention relates to the field of medical information prediction technology, and more specifically to a hemodialysis patient survival prediction method based on a feature-filled autoencoder. Background Art

[0002] Patients with end-stage renal disease (ESRD) face a significantly higher mortality risk. The U.S. Renal Data System Annual Report shows that from 2019 to 2021, under the influence of COVID-19, the all-cause mortality rate of adult ESRD patients increased from 15.9% to 18.8%, of which 11.6% of patients died within 6 months. Although the survival rate of ESRD patients has improved, the mortality rate of chronic kidney disease (CKD) patients within 18 months of progression to ESRD is 24.7%.

[0003] Hemodialysis (HD) is a common and key treatment for ESRD, but it is associated with many complications and adverse reactions that can seriously affect health; for example, physical stress, nutritional problems (such as malnutrition and protein energy wasting), environmental factors (such as temperature fluctuations), and common problems such as blood pressure changes, arrhythmias, and cardiovascular problems, all of which can lead to increased morbidity and mortality. Analyzing blood tests and recent responses during hemodialysis is crucial to evaluating the effectiveness of hemodialysis treatment.

[0004] With the deep integration of information technology and healthcare, the volume of electronic health records (EHRs) continues to grow. Currently, using machine learning techniques for pattern recognition in EHRs has become a research hotspot. Studies have shown that by applying multiple machine learning models, data such as blood pressure and weight change in hemodialysis patients can be used to predict post-dialysis blood pressure. Furthermore, some studies have used machine learning to estimate the probability of hypotension using parameters collected during treatment. Other studies have used the WOA-CoxPH model to predict mortality based on the consistency index of eight key parameters. However, these methods primarily focus on clinical parameters during dialysis, and the variability of these parameters over time can lead to instability. To address this issue, some studies have used blood pressure as sequential data and applied a multi-layer long short-term memory (LSTM) network to significantly improve prediction accuracy. Similarly, some researchers have used convolutional neural networks (CNNs) combined with a two-stage training strategy to more effectively predict mortality risk in hemodialysis patients.

[0005] Existing techniques such as LSTM and CNN are known for their effectiveness on time series data, but they rely on having a long-term, complete dataset of hemodialysis patients, which is often not feasible due to incomplete and irregular data input. Due to different patient conditions, not all biochemical markers can be consistently measured, resulting in significant data gaps that affect the selection of predictive models and their accuracy. Studies have shown that the risk of death in hemodialysis patients is highest early in the treatment process, with survival rates decreasing from 91.9% in the first year to 28.5% in the fifth year. Changes in the patient's condition require long observation windows and possible hospital transfers, which increases the challenge and makes it impossible to perform rapid risk assessment using time series-based models.

[0006] Therefore, the present invention proposes a hemodialysis patient survival prediction method based on feature-filled autoencoder to solve the above problems. Summary of the Invention

[0007] To address the above problems, this paper proposes a method for predicting survival of hemodialysis patients based on feature-filled autoencoders, which uses readily available short-term data to quickly assess the mortality risk of hemodialysis patients. This method takes advantage of the unsupervised learning of autoencoders, making it suitable for scenarios where limited samples are available. By training the autoencoder with unlabeled data, we capture the latent representation of the incomplete data from an adaptive feature extraction module. The classifier then uses the latent representation to estimate the mortality risk of hemodialysis patients, providing valuable insights, especially in the early, high-risk stages of treatment.

[0008] The present invention provides the following technical solutions:

[0009] A method for predicting survival of hemodialysis patients based on a feature-filled autoencoder comprises the following steps:

[0010] Step 1. Patients are divided into three groups: a survival group, including patients with no death record during the prediction period and subsequent transfer or continued dialysis records; a death group, consisting of patients with non-accidental death records during the prediction period; and an unknown group, including patients with no death record during the prediction period and no subsequent transfer or dialysis records; an autoencoder for unsupervised representation learning is created, wherein the main components of the autoencoder for unsupervised representation learning include: an encoder and a decoder.

[0011] In step 1, the main components of the autoencoder for unsupervised representation learning: encoder and decoder are expressed as:

[0012] g=g(x)

[0013]

[0014]

[0015] In the formula, the input vector x is mapped to a hidden representation g by the encoder g, and the decoder f maps the hidden representation h back to the output vector Represents the loss function, which helps improve the encoder's ability to effectively compress high-dimensional data into a low-dimensional feature space, thereby promoting unsupervised learning of data features.

[0016] Step 2: Death prediction is performed on hemodialysis patient data with different degrees of feature missing by using the adaptive mortality prediction model based on feature filling autoencoder (AMPAE); first, missing data is resolved by exploiting the underlying manifold structure and relationship between available features in high-dimensional space; missing features are filled with the values ​​of complete nearest neighbor features, and then unsupervised feature extraction is performed through the autoencoder. A loss function is customized by combining the enhanced reconstruction loss. and a regularization term To improve the performance of the autoencoder, we promote semantic feature sparsity. Finally, we train the model. Training consists of two main stages. The adaptive feature extraction module is trained through unsupervised learning of the autoencoder, which utilizes unlabeled data. In the subsequent stage, the adaptive feature extraction module performs feature extraction on the labeled data, which is then fed into the classifier for supervised training. In addition, we implement an early stopping mechanism to prevent overfitting of the autoencoder. This method stops training based on monitoring the performance of feature extraction using a simple classifier on known training set samples. The model is trained for several rounds with randomly initialized weights, and the best-performing autoencoder is selected as the feature extractor.

[0017] In step 2, a loss function is designed by combining the enhanced reconstruction loss and regularization loss To improve the performance of the autoencoder to promote the sparsity of semantic features, the expression is:

[0018]

[0019] Where α is a non-negative hyperparameter, and each input feature x in the autoencoder i，k are associated with a data credibility mask, expressed as:

[0020]

[0021] Random feature dropout mask, the expression is:

[0022]

[0023] Among them, the output Processed by the autoencoder as g(f(x i ⊙m i ), the enhanced reconstruction loss is expressed as:

[0024]

[0025] Where N is the number of samples, ⊙ is the element-wise product, and β is a non-negative hyperparameter that amplifies the focus on reconstructing the real data during training. When β = 0, this simplifies to the mean squared error loss, which is expressed as:

[0026]

[0027] Regularization loss Helping to control model complexity and reduce overfitting, this strategic mixture of loss components enables our adaptive feature extraction module to effectively capture and utilize high-quality features in datasets with missing data.

[0028] Step 3: Test the quality of the predicted samples, use five classification methods as benchmarks, and use F1 score and average precision (AP) as the main indicators to evaluate the performance of the model.

[0029] In step three, five classification methods were used as benchmarks, and F1 score and average precision (AP) were used as the main indicators to evaluate the performance of the model, including traditional algorithms such as logistic regression, Gaussian Naive Bayes (Gaussian NB), support vector machine (SVM), and deep learning models such as XGBoost and LightGBM classifiers. In order to offset the imbalance of class data, all models were trained using oversampling technology. Given the significant differences between the survival group and the death group, relying solely on the accuracy of model evaluation proved to be insufficient. Therefore, F1 score and average precision (AP) were used as the main indicators to evaluate the performance of the model. The F1 score balances accuracy and recall, and is particularly effective in the case of class imbalance. The expression is as follows:

[0030]

[0031] Where TP is true positive, FP is false positive, and FN is false negative. F1 balances precision and recall, being the harmonic mean of the two, which can avoid misjudging the model under extreme conditions. On the other hand, AP provides a weighted average of precision at each threshold, using the increase in recall compared to the previous threshold as the weight, expressed as:

[0032]

[0033] Where R n and R n-1are the precision and recall at the nth threshold, respectively. This metric provides a nuanced view of precision and recall at different thresholds, offering a more comprehensive assessment of the performance and result quality of the retrieval model. A precision-recall (PR) curve is used for detailed performance comparison. While both the receiver operating characteristic (ROC) curve and the PR curve can effectively assess model sensitivity and specificity at different thresholds, the ROC curve can be misleading in the case of class imbalance, especially when there are few positive cases. However, the PR curve remains unaffected by class imbalance. For missing data, we performed data imputation on all data across different prediction windows to ensure consistency in data processing methods between AMPAE and other models. Grid search was used to optimize the hyperparameters of all models across different prediction windows. The final performance evaluation was based on 5-fold cross-validation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention but do not constitute a limitation of the present invention.

[0035] Figure 1 This is a flow chart of a method for predicting survival of hemodialysis patients based on feature-filled autoencoders according to the present invention;

[0036] Figure 2 This is a flowchart of the AMPAE model used for mortality prediction in an adaptive mortality prediction method based on feature-filled autoencoder of the present invention;

[0037] Figure 3 This is a performance evaluation diagram of the model of the present invention; DETAILED DESCRIPTION

[0038] The present invention relates to a method for predicting survival in hemodialysis patients based on a feature-filled autoencoder. This method can effectively address the problem of missing data and accurately predict the mortality risk of hemodialysis patients. The specific implementation steps of the present invention are as follows:

[0039] Step 1. Data collection and preprocessing: Medical data of hemodialysis patients were collected, including but not limited to age, gender, blood pressure during dialysis, hemoglobin level, serum ferritin level, etc.; the collected data were preprocessed, and cases with more than 50% missing data or no blood test or dialysis records during the observation period were excluded. The blood test results during the observation period were averaged, and the incidence of thrombolytic hypotension was calculated using 6 common definitions. The maximum, minimum and average values ​​of systolic and diastolic blood pressure during dialysis were recorded. Missing data were resolved by exploiting the underlying manifold structure and relationship between available features in high-dimensional space. The values ​​of complete nearest neighbor features were used to fill in missing features. The introduction of a large amount of unlabeled data helped to approximate the true distribution of the data, enabling the model to discover more meaningful insights from the available data.

[0040] Step 2: Build an autoencoder model: Design and implement an autoencoder neural network, which consists of two main components: an encoder and a decoder. The encoder maps input data to hidden representations, and the decoder maps hidden representations back to output data. The autoencoder is trained through unsupervised learning and uses unlabeled data to extract data features.

[0041] Step 3. Feature dropout mask generation: To simulate the distribution of missing data in deep learning frameworks, feature dropout masks are generated. These masks randomly disable certain features during the input of data to the encoder based on the historical probability of feature loss.

[0042] Step 4: Adaptive feature extraction: Leverage the unsupervised learning capabilities of the autoencoder to extract latent representations from short-term data. By using the feature loss mask, the autoencoder can accurately reconstruct the missing features.

[0043] Step 5: Classifier training: Using the latent representation generated by the encoder, train a classifier, such as LightGBM, to predict the mortality rate of hemodialysis patients. During the training process, an early stopping mechanism is used to prevent overfitting.

[0044] Step 6: Model Validation and Evaluation: Model performance was evaluated through 5-fold cross-validation, using F1 score and average precision (AP) as the main indicators. PR curves were also used to compare the performance of the models in detail.

[0045] Step 7. Feature Importance Analysis: Use the SHAP method to analyze the impact of different features on model prediction and identify features that are closely related to the risk of death, such as age, serum total calcium, international normalized ratio, and C-reactive protein.

[0046] The effect of the present invention is further verified by the following experiments:

[0047] The method used in the present invention utilizes short-term hemodialysis data to assess the mortality risk of hemodialysis patients. A dataset with a 30-day prediction window was selected as the baseline dataset, and the F1 and AP metrics were used to evaluate model performance. Because the dataset used was extremely imbalanced, the mortality rate for hemodialysis patients within a 30-day observation window was only 6.0%, making it unreliable to evaluate the model solely using accuracy. F1, however, comprehensively considers both precision and recall, providing a more comprehensive assessment of model performance in cases of class imbalance and exhibiting good stability. AP, by comprehensively considering both precision and recall, provides a more comprehensive assessment of the performance of the retrieval model and the quality of the results. The model performed well in assessing the mortality risk of hemodialysis patients across various time periods. When the evaluation window was 30 days, AMPAE achieved a significant F1 of 0.53 and an AP of 0.5, demonstrating significant advantages. AMPAE achieved the highest F1 and AP across all evaluation time periods, demonstrating good and stable overall results. This demonstrates that the proposed model is more effective in assessing the mortality risk of hemodialysis patients, and this advantage is even more pronounced when the prediction window is smaller.

Claims

1. A method for predicting survival of hemodialysis patients based on feature-filled autoencoders, comprising the following steps: Step 1: Collect medical data of hemodialysis patients, including age, gender, blood pressure during dialysis, hemoglobin level, serum ferritin level, systolic blood pressure, and diastolic blood pressure; preprocess the collected data, exclude cases with more than 50% missing data or no blood test or dialysis records during the observation period, average the blood test results during the observation period, and calculate the incidence of thrombolytic hypotension using 6 common definitions, and record the maximum, minimum, and average values ​​of systolic and diastolic blood pressure during dialysis; create an autoencoder for unsupervised representation learning, the main components of the autoencoder for unsupervised representation learning include an encoder and a decoder, which can be expressed as: h=g(x) In the formula, the input vector x is mapped to a hidden representation h by the encoder g, and the decoder f maps the hidden representation h back to the output vector Represents the loss function, this loss function It helps improve the encoder's ability to effectively compress high-dimensional data into a low-dimensional feature space, thereby promoting unsupervised learning of data features; Step 2: Death prediction is performed on hemodialysis patient data with different degrees of feature missing by using an adaptive mortality prediction model based on feature filling autoencoder. The adaptive mortality prediction model based on feature filling autoencoder designs a loss function Loss AE The expression of the loss function By combining the enhanced reconstruction loss and regularization loss To improve the performance of the autoencoder to promote the sparsity of semantic features, the expression is: Among them, Loss AE Compared with the loss function in step 1 For the same loss function, α is a non-negative hyperparameter, and each input feature x in the autoencoder i，k are associated with a data credibility mask, expressed as: The random feature dropout mask expression is: Output Processed by the autoencoder as g(f(x i ⊙m i ), the enhanced reconstruction loss is expressed as: where N is the number of samples, ⊙ is the element-wise product, and β is a non-negative hyperparameter that amplifies the focus on reconstructing the true data during training; when β = 0, this simplifies to the mean squared error loss: Regularization loss This strategic mixture of loss components helps control model complexity and reduce overfitting, enabling the adaptive feature extraction module to effectively capture and utilize high-quality features in datasets with missing data; Step 3: Detect the quality of the predicted samples. The detection and prediction sample quality uses five classification methods as benchmarks, and adopts F1 score and average precision (AP) as the main indicators to evaluate the performance of the model.

Images