Explainable transfer learning method and system for prognosis prediction of patients with stroke caused by rare diseases

Abstract

This invention discloses an interpretable transfer learning method and system for predicting the prognosis of stroke patients with rare causes, comprising the following steps: Step S1, collecting patient data, using stroke patient data as the source domain data and data from patients with stroke-related diseases of rare causes as the target domain data, randomly dividing the target domain data in a 7:3 ratio to obtain a training set and a test set; Step S2, aligning the feature spaces from the source domain to the target domain using a covariance alignment method based on the source domain data and the target domain training set, and performing cross-domain transfer learning; Step S3, embedding interpretability methods into the key steps of transfer learning, optimizing the feature transfer from the source domain to the target domain, and transforming post-hoc interpretability into interpretability of the transfer learning process; Step S4, comprehensively evaluating the final model based on the target domain test set, outputting and displaying performance indicators. This invention will help improve the accurate and rapid assessment of the prognosis of stroke patients with rare causes, assist clinical decision-making, and improve patient outcomes.

Description

Technical Field

[0001] This invention relates to the field of medical information, and in particular to an interpretable transfer learning method and system for predicting the prognosis of stroke patients with rare causes. Background Technology

[0002] Stroke imposes a heavy disease burden on Chinese residents, and prognostic research can help improve patient outcomes and reduce this burden. Prognostic studies on common causes of stroke (such as cerebral arteriosclerosis and atrial fibrillation) have been extensively conducted. However, for rare causes of stroke (such as moyamoya disease, autosomal dominant cerebral arteriosclerosis with subcortical infarction, and leukoencephalopathy), prognostic research is limited by factors such as the small number of patients and the complexity of the conditions. Traditional statistical methods and conventional machine learning algorithms cannot obtain reliable and efficient models. Rare causes of stroke have a relatively higher prevalence in younger patients, resulting in a heavier disability-adjusted life year burden. Therefore, accurate prognostic prediction for patients with rare causes of stroke has significant public health implications. Transfer learning (TL), as an emerging machine learning algorithm, utilizes existing knowledge to enhance the model's ability to acquire unknown knowledge, reducing the sample size required for model training and improving training efficiency. It is a viable method for solving small-sample prognostic prediction problems. However, the black-box nature of transfer learning leads to the lack of interpretability of the process and results, limiting its application in prognostic prediction. Optimizing transfer learning algorithms using interpretability methods to address the black-box problem and developing systems to implement these algorithms can improve their efficiency and effectiveness, facilitating practical applications. Summary of the Invention

[0003] The purpose of this invention is to provide an interpretable transfer learning method and system for predicting the prognosis of stroke patients with rare causes, in order to solve the problems mentioned in the background art.

[0004] To achieve the above-mentioned objectives, this invention provides an interpretable transfer learning (ITL) method for predicting the prognosis of stroke patients with rare causes, comprising the following steps:

[0005] Step S1: Collect patient data, using stroke patient data as source domain data and rare etiology stroke-related disease patient data as target domain data. Randomly divide the target domain data in a 7:3 ratio to obtain the target domain training set and the target domain test set.

[0006] Step S2: Based on the source domain data and the target domain training set, the feature space of the source domain to the target domain is aligned using the covariance alignment method, and cross-domain transfer learning is performed.

[0007] Step S3: Embed interpretability methods into key steps of transfer learning, optimize feature transfer from the source domain to the target domain, and transform post-hoc interpretability into interpretability of the transfer learning process.

[0008] Step S4: Perform a comprehensive evaluation of the final model based on the target domain test set, output and display the performance metrics.

[0009] Furthermore, in step S2, if the positive rate of the source domain outcome is 10% or 20%, the source domain data is first processed to remove class imbalance before transfer learning is performed.

[0010] Furthermore, step S3 includes the following steps:

[0011] Step S301: Perform Shapley Additive Explanations (SHAP) on the target domain training set to quantitatively evaluate the absolute contribution of each feature variable to the SHAP value. Optimize the SHAP threshold through grid search. The interval can be set to [0.01, 0.10] with a step size of 0.01 to screen the core feature subset that has a key impact on the model decision.

[0012] In step S302, the selected feature subset is simultaneously applied to the source and target domain training sets, driving the iterative optimization of the transfer learning process in step S2, forming an automated feedback mechanism of "feature interpretation → feature selection → transfer optimization → reinterpretation".

[0013] Furthermore, the research variables for patients with rare etiologies of stroke in step S1 include age, sex, race, atrial fibrillation, history of alcohol consumption, history of smoking, antiplatelet therapy, coronary artery disease, chronic obstructive pulmonary disease, diabetes, hyperlipidemia, hypertension, obesity, use of statins, mechanical ventilation, heart rate, systolic blood pressure, diastolic blood pressure, respiratory rate, body temperature, blood oxygen saturation, blood glucose, hemoglobin, platelets, white blood cells, anion gap, bicarbonate, blood urea nitrogen, calcium, chloride, sodium, potassium, creatinine, acute physiological score, Glasgow Coma Scale score, multiple organ failure score, Oxford Acute Severity of Illness Score (OASIS), and sequential organ failure score.

[0014] Furthermore, the performance metrics in step S4 include the area under the receiver operating characteristic curve (AUC), accuracy, F1 score, and Brill score.

[0015] Furthermore, using the interpretable transfer learning method of this invention, along with traditional transfer learning methods and logistic regression methods, a total of 5 feature variables were selected, namely hemoglobin, white blood cells, anion gap, sodium, and Oxford acute disease severity score.

[0016] Another aspect of the present invention provides an interpretable transfer learning system for predicting the prognosis of stroke patients with rare causes, comprising a data uploading module, a variable setting module, a transfer learning module, an analysis results module, and an interpretability analysis module, wherein:

[0017] The data upload module is used to upload data from the source and target domains.

[0018] The variable setting module is used to select the outcome variable and check the continuous variables in the dataset for standardization.

[0019] The transfer learning module is used for model training using ITL methods;

[0020] The analysis results module is used to compare and display the performance of different methods on various evaluation indicators in the form of radar charts.

[0021] The interpretability analysis module is used to perform interpretive analysis based on the optimal model and displays the top 10 features by importance using a swarm graph and bar chart.

[0022] Compared with existing technologies, this system and method have the following advantages:

[0023] 1. This invention uses stroke patient data as the source domain and rare cause stroke patient data as the target domain, employs transfer learning to achieve cross-domain feature space alignment, and embeds interpretability methods into the key steps of transfer learning. This enables real-time interpretability and optimization of the feature transfer process from the source domain to the target domain, thereby significantly reducing the sample size required for model training, improving training efficiency, and enhancing model interpretability. It also enables accurate and rapid assessment and risk warning of the prognosis of rare cause stroke patients, facilitating clinical decision support and improving patient prognosis.

[0024] 2. This invention provides an interpretable transfer learning system platform for predicting the prognosis of stroke patients with rare causes. With the help of this platform, users can upload source domain and target domain data as needed to complete transfer learning and intuitively view the model's predictive performance and its interpretability analysis results. Attached Figure Description

[0025] Figure 1 Flowchart of an interpretable transfer learning method for predicting prognosis in patients with rare etiologies of stroke.

[0026] Figure 2 Venn diagram for filtering variables for different methods.

[0027] Figure 3 The graph shows the performance of different methods.

[0028] Figure 4 This is a ranking graph of the feature importance of the optimal model.

[0029] Figure 5 SHAP beehive diagram of the optimal model.

[0030] Figure 6 This is an interface diagram of an online transfer learning system for predicting the prognosis of stroke caused by rare diseases. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1 The diagram shown is a flowchart of Embodiment 1 of the present invention. This embodiment provides an interpretable transfer learning method for predicting the prognosis of stroke due to rare causes. The present invention will help improve the accurate and rapid assessment of the prognosis of patients with stroke due to rare causes, assist clinical decision-making, and improve patient outcomes.

[0033] The specific steps are as follows:

[0034] Step S1: Collect patient data, using stroke patient data as source domain data and Moyamoya disease patient data as target domain data. Randomly divide the target domain data in a 7:3 ratio to obtain the target domain training set and the target domain test set.

[0035] This study applies interpretable transfer learning to predict mortality rates in patients with Moyamoya disease admitted to the intensive care unit (ICU). Three methods—interpretable transfer learning, transfer learning, and logistic regression—were used to predict the prognosis of Moyamoya disease patients.

[0036] Data were obtained from the MIMIC-IV database (V3.0) and included all patients admitted to the ICU aged ≥18 years whose primary diagnosis was stroke (source domain) and moyamoya disease (target domain). Stroke was identified using ICD-9 diagnostic codes 430, 431, 432, 432.0, 432.1, 432.9, 433.01, 433.11, 433.21, 433.31, 433.81, 433.91, 434.01, 434.11, 434.91, and 436, and ICD-10 diagnostic codes I60, I61, I62, and I63; moyamoya disease was identified using ICD-9 diagnostic code 437.5 and ICD-10 diagnostic code I67.5. The study outcome was 1-year mortality. Study variables included age (years), sex, race, atrial fibrillation, history of alcohol consumption, smoking history, antiplatelet therapy, coronary artery disease, chronic obstructive pulmonary disease, diabetes, hyperlipidemia, hypertension, obesity, use of statins, mechanical ventilation, heart rate (beats / min), systolic blood pressure (mmHg), diastolic blood pressure (mmHg), respiratory rate (breaths / min), body temperature (°C), oxygen saturation (%), blood glucose (mEq / L), hemoglobin (g / dL), platelets (×10⁹ / L), white blood cells (×10⁹ / L), anion gap (mmol / L), bicarbonate (mEq / L), blood urea nitrogen (mg / dL), calcium (mg / dL), chloride (mEq / L), sodium (mEq / L), potassium (mEq / L), creatinine (mg / dL), Acute Physiology Score III (APSIII), and Glasgow Coma Scale (GCS). The following scores were used: GCS (Gross Scale), Multiple Organ Dysfunction Score (MODS), Oxford Acute Illness Severity Scale, and Sequential Organ Failure Assessment (SOFA). All vital signs and laboratory tests were measured on the first day of ICU admission, and for multiple measurements of vital signs, the average was used.

[0037] Step S2: Based on the source domain data and the target domain training set, feature space alignment from the source domain to the target domain is achieved using the Covariance Alignment (CORAL) method, and cross-domain transfer learning is performed. If the positive rate of the source domain outcome is 10% or 20%, the source domain data is first imbalanced using the Synthetic Minority Oversampling Technique combined with Edited Nearest Neighbours (SMOTE-ENN) before transfer learning is performed.

[0038] Step S3: Dynamically optimize the transfer features based on interpretability.

[0039] Unlike traditional post-hoc interpretability analysis of models, this approach embeds interpretability methods into key steps of transfer learning, optimizing feature transfer from the source domain to the target domain and transforming post-hoc interpretability into interpretability within the transfer learning process itself. First, Shapley additive interpretability analysis (SHAP) is performed on the target domain training set to quantitatively evaluate the absolute contribution of each feature variable. The SHAP threshold is optimized using a grid search, with an interval set to [0.01, 0.10] and a step size of 0.01, to filter out a subset of core features that have a critical impact on model decisions. This selected feature subset is then simultaneously applied to both the source and target domain training sets, driving iterative optimization of the transfer learning process in step 2. This forms an automated feedback mechanism of "feature interpretation → feature selection → transfer optimization → re-interpretation." Through transfer optimization guided by interpretability analysis, the interpretability of the entire transfer learning process is achieved, enhancing the transparency and credibility of model decisions.

[0040] like Figure 2 As shown, the variable selection results of three methods (ITL, TL, and Logistic) are presented when stroke is the source domain and moyamoya disease is the target domain. ITL, TL, and Logistic selected 17, 19, and 14 feature variables, respectively. A total of 5 feature variables were selected by the three methods: hemoglobin, white blood cells, anion gap, sodium, and OASIS score.

[0041] Step S4: Based on the target domain test set, perform a comprehensive evaluation of the final model and output performance metrics (such as AUC, accuracy, F1 score, Brill score, etc.).

[0042] Figure 3This paper presents the comparative results of three methods (ITL, TL, and Logistic) on multiple evaluation metrics when using stroke as the source domain and Moyamoya disease as the target domain. Compared with Logistic Regression trained only on the target domain, ITL and TL show better overall discriminative ability. Logistic Regression has a sensitivity of 0 and a specificity of 1, indicating that the model classifies almost all samples as negative, resulting in insufficient positive identification ability and thus limiting the quality of positive predictions and overall metrics. In contrast, ITL and TL have significantly higher AUC and mean precision, and approach perfect scores in sensitivity and negative prediction, indicating that transfer learning can more effectively identify positive samples and reduce the risk of missed diagnoses. Further comparison of the two transfer methods shows that ITL is superior in most metrics, and its radar chart coverage is also larger, suggesting that its overall performance is better than TL; at the same time, ITL includes fewer variables, resulting in a simpler model structure. Considering AUC, F1 score, and Brill score, the ITL model is selected as the optimal model.

[0043] Figure 4 This study demonstrates the feature importance of the optimal model when stroke is the source domain and moyamoya disease is the target domain. Ranked by mean absolute SAP value, the top 10 features are: age, antiplatelet therapy, statins, chloride, MODS score, white blood cell count, hemoglobin, APSIII score, OASIS score, and mechanical ventilation.

[0044] Figure 5 This is the SHAP beehive diagram of the optimal model, which simultaneously presents feature importance and feature effect, with features arranged from highest to lowest importance. Each point in the diagram represents the SHAP value of a certain feature on a specific sample; its position on the Y-axis is determined by the feature, and its position on the X-axis is determined by the SHAP value. Colors indicate the variation of feature values ​​from low to high (blue for low values, red for high values). A SHAP value greater than 0 indicates a positive impact on the outcome, suggesting a higher predicted risk of death in this example. Taking statins as an example, compared to using statins (red dots), not using statins (blue dots) corresponds to a higher predicted risk of death, suggesting that statin use is a protective factor against patient mortality.

[0045] Example 2 is an online transfer learning system for predicting the prognosis of stroke due to rare causes, comprising five modules: data upload, variable setting, transfer learning, analysis results, and interpretability analysis. In the data upload module, users can customize the source and target domain data to upload; in the variable setting module, users need to select the outcome variable and check the continuous variables in the dataset for standardization; in the transfer learning module, an ITL method is embedded, and the platform will automatically complete model training after clicking "Start Training"; in the analysis results module, the platform compares and displays the performance of different methods on various evaluation indicators in the form of radar charts; in the interpretability analysis module, the platform conducts interpretive analysis based on the optimal model and displays the top 10 features by importance using bee colony plots and bar charts. Figure 6 The diagram shown is an operation interface of the online system for predicting the prognosis of stroke patients with rare causes in Example 2. With the help of this platform, users can upload source domain and target domain data as needed to complete transfer learning and intuitively view the model prediction performance and its interpretability analysis results.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An interpretable transfer learning method for predicting the prognosis of stroke patients with rare causes, characterized in that, Includes the following steps: Step S1: Collect patient data, using stroke patient data as source domain data and rare etiology stroke-related disease patient data as target domain data. Randomly divide the target domain data in a 7:3 ratio to obtain the target domain training set and the target domain test set. Step S2: Based on the source domain data and the target domain training set, the feature space of the source domain to the target domain is aligned using the covariance alignment method, and cross-domain transfer learning is performed. Step S3: Embed interpretability methods into key steps of transfer learning, optimize feature transfer from source domain to target domain, and transform post-hoc interpretability into interpretability of the transfer learning process. Step S4: Perform a comprehensive evaluation of the final model based on the target domain test set, output and display the performance metrics.

2. The interpretable transfer learning method for predicting the prognosis of stroke patients with rare causes as described in claim 1, characterized in that, If the positive rate of the source domain outcome is 10% or 20% in step S2, the source domain data is first imbalanced using the class imbalance processing method, and then transfer learning is performed.

3. The interpretable transfer learning method for predicting the prognosis of stroke patients with rare causes as described in claim 1, characterized in that, Step S3 includes the following steps: Step S301: Perform Shapley additive interpretation analysis on the target domain training set to quantitatively evaluate the contribution of the absolute value of each feature variable to the SHAP value. Optimize the SHAP threshold through grid search. The interval can be set to [0.01, 0.10] with a step size of 0.01 to screen the core feature subset that has a key impact on the model decision. In step S302, the selected feature subset is simultaneously applied to the source and target domain training sets, driving the iterative optimization of the transfer learning process in step S2, forming an automated feedback mechanism of "feature interpretation → feature selection → transfer optimization → reinterpretation".

4. The interpretable transfer learning method for predicting the prognosis of stroke patients with rare causes as described in claim 1, characterized in that, The study variables for patients with rare etiologies of stroke in step S1 include age, sex, race, atrial fibrillation, history of alcohol consumption, history of smoking, antiplatelet therapy, coronary artery disease, chronic obstructive pulmonary disease, diabetes, hyperlipidemia, hypertension, obesity, use of statins, mechanical ventilation, heart rate, systolic blood pressure, diastolic blood pressure, respiratory rate, body temperature, blood oxygen saturation, blood glucose, hemoglobin, platelets, white blood cells, anion gap, bicarbonate, blood urea nitrogen, calcium, chloride, sodium, potassium, creatinine, acute physiological score, Glasgow Coma Scale score, multiple organ failure score, Oxford acute illness severity score, and sequential organ failure score.

5. The interpretable transfer learning method for predicting the prognosis of stroke patients with rare causes as described in claim 1, characterized in that, The performance metrics in step S4 include the area under the receiver operating characteristic curve, accuracy, F1 score, and Brill score.

6. The interpretable transfer learning method for predicting the prognosis of stroke patients with rare causes as described in claim 1, characterized in that, The interpretable transfer learning method of this invention, together with traditional transfer learning methods and logistic regression methods, selects a total of 5 feature variables, namely hemoglobin, white blood cells, anion gap, sodium and OASIS score.

7. An interpretable transfer learning system for predicting the prognosis of stroke patients with rare causes, characterized in that, It includes a data upload module, a variable setting module, a transfer learning module, an analysis results module, and an interpretability analysis module, among which: The data upload module is used to upload data from the source and target domains. The variable setting module is used to select the outcome variable and check the continuous variables in the dataset for standardization. The transfer learning module is used for model training using interpretable transfer learning methods. The analysis results module is used to compare and display the performance of different methods on various evaluation indicators in the form of radar charts. The interpretability analysis module is used to perform interpretability analysis based on the optimal model and displays the top 10 features by importance using a beehive diagram and bar chart.

Images