Diabetes risk prediction method based on high-dimensional disease embedding and health data fusion

Abstract

The invention discloses a diabetes risk prediction method based on high-dimensional disease embedding and health data fusion. Comprising the steps of data acquisition and preprocessing, construction of a high-dimensional disease embedding space, calculation of an individual health condition vector, deduction of a health score vector based on physical examination data, a hybrid prediction model for risk assessment, and model training and deployment. According to the method, a hybrid machine learning architecture is adopted, firstly, a health condition vector, a health score vector and original physical examination data are effectively integrated through a special fusion module, and then fused features are input into an integrated prediction model to generate a highly accurate diabetes risk score; according to the method, the prediction performance is remarkably improved, and the area (AUC) under the working characteristic curve of a subject reaches 0.84.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bioinformatics technology, and in particular to a diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data. BACKGROUND

[0002] Currently, diabetes risk prediction mainly relies on applying statistical models such as logistic regression to a limited set of known risk factors, such as age, body mass index (BMI), family history, blood pressure, and fasting blood glucose. These models are effective to some extent, but they often fail to capture the complex, non-linear, and systemic nature of disease progression.

[0003] Current methods ignore a large amount of potential information contained in the complete diagnosis history of patients. An individual's medical history, including the sequence and combination of diseases, is often treated as isolated events rather than a holistic representation of an individual's health trajectory, resulting in the omission of valuable predictive signals.

[0004] Traditional methods are difficult to quantify the complex relationship network between diseases. Diabetes is not an isolated disease, and its development is often accompanied by a series of other diseases. Existing technologies lack a systematic method to represent and utilize these complex comorbidity patterns for risk prediction. A patient's medical history is essentially a list of sparse, high-dimensional, and categorical International Classification of Diseases (ICD) codes. Traditional models have difficulty directly utilizing this unstructured data. How to convert this disease sequence into a fixed-length, dense numerical vector that captures its inherent "semantics" is a key bottleneck to improving prediction performance. This is similar to the challenge faced in the natural language processing (NLP) field of converting a sentence (word sequence) into a vector.

[0005] Therefore, there is a need to develop a diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data to solve the above problems. SUMMARY

[0006] The purpose of the present application is to design a diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data to solve the above problems.

[0007] The present application achieves the above-mentioned purpose through the following technical solutions:

[0008] The diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data comprises the following steps:

[0009] S1, data acquisition and preprocessing;

[0010] S2, construction of high-dimensional disease embedding space;

[0011] S3. Calculation of the individual health status vector;

[0012] S4. Derivation of the health score vector based on physical examination data;

[0013] S5. Hybrid prediction model for risk assessment;

[0014] S6. Model Training and Deployment.

[0015] Specifically, step S1 includes:

[0016] S11. Collect electronic health record data from clinical databases, including patient identifiers, diagnosis codes, and diagnosis dates;

[0017] S12. For the same patient cohort, collect relevant physical examination data, including but not limited to: systolic blood pressure, diastolic blood pressure, urinalysis specific gravity, urinalysis pH, high-density lipoprotein, fasting blood glucose, red blood cell distribution width, platelet count, alanine aminotransferase, low-density lipoprotein, creatinine, uric acid, aspartate aminotransferase, total cholesterol, total white blood cell count, monocyte percentage, absolute monocyte count, total red blood cell count, lymphocyte percentage, absolute lymphocyte count, mean corpuscular volume, mean corpuscular hemoglobin content, mean corpuscular hemoglobin concentration, basophil percentage, absolute basophil count, eosinophil percentage, absolute eosinophil count, hemoglobin, neutrophil percentage, absolute neutrophil count, and BMI;

[0018] S13. Standardize and clean the collected data; specific operations include: mapping all diagnostic information to a unified coding system, deleting duplicate diagnostic entries; and integrating multiple medical records of each individual patient into a diagnostic sequence arranged in chronological order.

[0019] Specifically, step S2 includes:

[0020] S21. Define the corpus: Each patient's chronologically ordered sequence of diagnosed diseases is considered a "sentence", and the set of all patients' "sentences" constitutes the "corpus" used for model training.

[0021] S22. Apply word embedding algorithm: Use a continuous bag-of-words model or skip-word model to train the corpus defined in step S21. By processing disease "sentences", learn the vector representation of each unique disease "word".

[0022] S23, generating disease vector: the output of this step is a mapping table, in which each unique disease code corresponds to an N-dimensional dense vector; in the constructed N-dimensional space, diseases with similar comorbidity patterns and potential causes will show higher proximity in cosine similarity.

[0023] In particular, step S3 comprises:

[0024] S31, for a given individual, identify all diagnosed diseases before the prediction time point, for patients with diagnosed diabetes, the time point is before the diagnosis of diabetes; for individuals without diabetes, all their diseases are their entire medical history;

[0025] S32, retrieve the N-dimensional disease embedding vector corresponding to all diagnosed diseases of the individual from the disease embedding space constructed in step S2;

[0026] S33, calculate the N-dimensional health status vector of the individual ; the vector is obtained by weighted average of all disease vectors retrieved, and the weight is based on the frequency of each disease d in the diagnosis record D of the individual , the calculation formula of the i-th dimension of the vector is as follows: ;

[0027] wherein, represents the value of the vector of disease d in the i-th dimension, through this step, the complete medical history of the individual is converted into an accurate coordinate point in the disease coordinate system, i.e. the health status vector.

[0028] In particular, step S4 comprises:

[0029] S41, feature selection of physical examination indicators: to reduce the collinearity between indicators, calculate the Pearson correlation coefficient matrix between all available physical examination indicators; apply a threshold to retain a set of M relatively independent physical examination indicators; S42, associate physical examination indicators with disease space dimensions: for each of the N dimensions of the disease embedding space, linear regression analysis is performed with the M independent physical examination indicators selected in step S41 respectively to determine the effect value β ij , which quantifies the association strength between the jth physical examination indicator and the i-th dimension of the disease space;

[0030] S43, calculate the health score vector of the individual: for a given individual, according to its specific physical examination results, calculate its N-dimensional health score vector , the calculation formula of the i-th dimension of the score vector is as follows: ​​

[0031] ;

[0032] wherein, is the normalized value of the jth physical examination indicator of the individual, β ij is the effect value calculated in advance in step S42.

[0033] Specifically, step S5 includes:

[0034] S51, feature fusion sub-model: a neural network is used as a feature fusion module; the feature fusion module is designed with three independent input branches:

[0035] Branch A: receiving the N-dimensional health status vector from step S3;

[0036] Branch B: receiving the N-dimensional health score vector from step S44;

[0037] Branch C: receiving M original physical examination indicators, where the physical examination indicators are normalized;

[0038] Each branch is composed of one or more fully connected layers, activation functions, and regularization layers, which are used to learn the latent representation of each input type;

[0039] S52, attention mechanism-based fusion: the latent representations learned by the three branches are fused through an attention mechanism; the attention mechanism calculates the weights for the output of each branch, allowing the model to dynamically prioritize features with the largest amount of information according to the individual's situation, and then concatenating the weighted representations into a unified feature vector;

[0040] S53, integrated prediction sub-model: the fusion feature vector generated in step S52 is input into a hybrid integrated prediction model; the hybrid integrated prediction model combines the outputs of a gradient boosting machine and an independent neural network classifier;

[0041] S54, generation of the final risk score: the prediction results of the gradient boosting machine and the neural network classifier are combined through a meta-learning strategy to produce a final probability score, which is between 0 and 1, representing the individual's risk of developing diabetes.

[0042] Specifically, step S6 includes:

[0043] S61, training the integrated prediction sub-model described in step S53 on a balanced dataset, with a 1:1 ratio of diabetic patients to non-diabetic individuals in the dataset;

[0044] S62, save the trained model and deploy it on a server or computer system; by receiving new patient data, perform the prediction process to provide risk assessment results in real-time or batch processing.

[0045] The beneficial effects of the present application are:

[0046] 1. Construct a high-dimensional "disease embedding space", in which each disease is represented as a unique numerical vector. This vector can capture the potential clinical and etiological relationships between diseases, quantifying the complex associations between diseases.

[0047] 2. Calculate a quantitative "health status vector" for each individual by aggregating the embedding vectors of their diagnosed diseases. This converts the individual's qualitative medical records into a precise multi-dimensional space coordinate, achieving a holistic quantification of the individual's health status.

[0048] 3. Derive a novel "health score vector" by systematically relating the dimensions of the disease space to standard medical examination indicators. This vector links easily accessible medical examination data to the deep structure of the disease space, creating powerful predictive features.

[0049] 4. Use a hybrid machine learning architecture that first effectively integrates the health status vector, health score vector, and raw medical examination data through a dedicated fusion module, and then inputs the fused features into an integrated prediction model to generate highly accurate diabetes risk scores; this method significantly improves prediction performance, with an area under the receiver operating characteristic curve (AUC) of 0.84. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The flowchart of the present application;

[0051] Figure 2 The model curve area diagram of the present application;

[0052] Figure 3 The performance comparison diagram of the integrated model constructed by the present application and other models. DETAILED DESCRIPTION

[0053] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0054] Therefore, the following detailed description of the embodiments of the application provided in the accompanying drawings is not intended to limit the scope of the application claimed, but merely represents selected embodiments of the application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the application without creative labor fall within the scope of protection of the application.

[0055] It should be noted that similar reference numbers and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0056] In the description of the application, it should be understood that the terms "upper", "lower", "inner", "outer", "left", "right", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship commonly placed when the product of the application is used, or the orientation or positional relationship commonly understood by those skilled in the art, and are only for the convenience of describing the application and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the application.

[0057] In addition, the terms "first", "second", etc. are only used for differentiation in description and cannot be understood as indicating or implying relative importance.

[0058] In the description of the application, it should also be noted that unless otherwise explicitly specified and limited, the terms "provided", "connected", etc. should be broadly understood, for example, "connected" can be fixedly connected, or detachably connected, or integrally connected; can be mechanically connected, or electrically connected; can be directly connected, or indirectly connected through an intermediate medium, or the internal communication of two elements. For those of ordinary skill in the art, the specific meaning of the above terms in the application can be understood according to the specific circumstances.

[0059] The specific embodiments of the application will be described in detail below with reference to the accompanying drawings.

[0060] As shown in the figure, the diabetes risk prediction method based on high-dimensional disease embedding and fused health data includes the following steps: Figure 1

[0061] S1, data acquisition and preprocessing;

[0062] S11, collecting electronic health record data from a clinical database, the data content including patient identifier, diagnosis code and diagnosis date;

[0063] ​S12. For the same patient cohort, collect relevant physical examination data, including but not limited to: systolic blood pressure, diastolic blood pressure, urinalysis specific gravity, urinalysis pH, high-density lipoprotein (HDL), fasting plasma glucose (FPG), red blood cell distribution width (RDW), platelet count, alanine aminotransferase, low-density lipoprotein, creatinine, uric acid, aspartate aminotransferase, total cholesterol, total white blood cell count, monocyte percentage, absolute monocyte count, total red blood cell count, lymphocyte percentage, absolute lymphocyte count, mean corpuscular volume, mean corpuscular hemoglobin content, mean corpuscular hemoglobin concentration, basophil percentage, absolute basophil count, eosinophil percentage, absolute eosinophil count, hemoglobin, neutrophil percentage, absolute neutrophil count, and BMI.

[0064] S13. Standardize and clean the collected data; specific operations include: mapping all diagnostic information to a unified coding system, such as the International Classification of Diseases, 10th Revision (ICD-10); deleting duplicate diagnostic entries; and integrating multiple medical records of each individual patient into a diagnostic sequence arranged in chronological order.

[0065] S2, Construction of a high-dimensional disease embedding space;

[0066] S21. Define the corpus: Each patient's chronologically ordered sequence of diagnosed diseases is considered a "sentence", and the set of all patients' "sentences" constitutes the "corpus" used for model training.

[0067] S22. Apply word embedding algorithm: Use a continuous bag-of-words (CBOW) model or a skip-gram model, such as the word2vec algorithm, to train the corpus defined in step S21. The algorithm learns the vector representation of each unique disease "word" (i.e., ICD-10 encoding) by processing disease "sentences".

[0068] S23. Generating Disease Vectors: The output of this step is a mapping table where each unique disease code corresponds to an N-dimensional (e.g., N=20) dense vector. In the constructed N-dimensional space, diseases with similar comorbidity patterns and potential causes will show a higher degree of closeness in their corresponding vectors based on cosine similarity. This process is not just simple encoding, but rather the discovery of a new data-driven coordinate system for describing human diseases. Each dimension in this N-dimensional space represents a potential "meta-feature" of the disease. For example, a certain dimension may represent a continuous spectrum from digestive system diseases to ophthalmological diseases, and the values ​​in this dimension are significantly correlated with metabolic indicators such as fasting blood glucose and blood lipids, thus transforming abstract disease associations into quantifiable physiological state axes.

[0069] S3, Calculation of individual health status vector;

[0070] S31, For a given individual, identify all diagnosed diseases before a prediction time point, for patients with diagnosed diabetes, the time point is before their diagnosis of diabetes; for individuals without diabetes, all their diseases are their entire medical history;

[0071] S32, retrieve the N-dimensional disease embedding vector corresponding to all diagnosed diseases of the individual from the disease embedding space constructed in step S2;

[0072] S33, calculate the N-dimensional health status vector of the individual ; the vector is obtained by weighted average of all retrieved disease vectors, the weight is based on the frequency of each disease d in the diagnosis record D of the individual , the calculation formula of the i-th dimension of the vector is as follows:

[0073] ;

[0074] wherein, represents the value of the vector of disease d in the i-th dimension, through this step, the complete medical history of the individual is converted into an accurate coordinate point in the disease coordinate system, i.e. the health status vector.

[0075] S4, Derivation of health score vector based on physical examination data;

[0076] S41, Feature selection of physical examination indicators: to reduce the collinearity between indicators, calculate the Pearson correlation coefficient matrix between all available physical examination indicators; apply a threshold (for example, 0.5) to retain a set of M relatively independent physical examination indicators (for example, retain 6 indicators such as urine routine proportion, high-density lipoprotein, fasting blood glucose, red blood cell volume distribution width, etc.);

[0077] S42, Association of physical examination indicators with disease space dimensions: for each of the N dimensions of the disease embedding space, linear regression analysis is performed with the M independent physical examination indicators selected in step S41, this step aims to determine the effect value β ij , which quantifies the association strength between the j-th physical examination indicator and the i-th dimension of the disease space;

[0078] S43, Calculation of individual health score vector: for a given individual, according to its specific physical examination results, calculate its N-dimensional health score vector , the calculation formula of the i-th dimension of the score vector is as follows:

[0079] ;

[0080] wherein, is the normalized value of the j-th physical examination indicator of the individual, β ij is the effect value pre-computed in step S42. This step is a key innovation that creates powerful new predictive features by systematically linking readily available laboratory data to the deep underlying structure of the disease space, unlocking predictive power that would not be apparent from analyzing laboratory data alone.

[0081] The method of deriving the "health score vector" (step S4) is a non-obvious, creative step. It creates powerful new predictive features by systematically linking routine laboratory data to the deep underlying structure of the disease space, unlocking predictive power that would not be available from laboratory data alone.

[0082] S5, a hybrid prediction model for risk assessment;

[0083] S51, feature fusion sub-model: a neural network is used as the feature fusion module; the feature fusion module is designed with three independent input branches:

[0084] Branch A: receives the N-dimensional health status vector from step S3;

[0085] Branch B: receives the N-dimensional health score vector from step S44;

[0086] Branch C: receives M original physical examination indicators, which are normalized here;

[0087] Each branch is composed of one or more fully connected layers, activation functions (such as ReLU), and regularization layers (such as batch normalization, Dropout), which are used to learn the latent representation of each input type;

[0088] S52, attention mechanism-based fusion: the latent representations learned by the three branches are fused through an attention mechanism; the attention mechanism calculates the weight for the output of each branch, enabling the model to dynamically prioritize features with the most information according to the individual's situation, and then concatenates the weighted representations into a unified feature vector;

[0089] S53, integrated prediction sub-model: the fusion feature vector generated in step S52 is input into a hybrid integrated prediction model; the hybrid integrated prediction model combines the outputs of a gradient boosting machine (specifically XGBoost) and an independent neural network classifier;

[0090] S54, final risk score generation: combine the prediction results of the gradient boosting machine and the neural network classifier through a meta-learning strategy (such as stacking or weighted average method) to generate a final probability score between 0 and 1, representing the risk of the individual suffering from diabetes.

[0091] S6, model training and deployment:

[0092] S61, train the integrated prediction sub-model described in step S53 on a balanced dataset with a 1:1 ratio of diabetic patients to non-diabetic individuals;

[0093] S62, save the trained model and deploy it on a server or computer system; by receiving new patient data (i.e. the input of steps S3 and S4), performing the prediction process (step S5) to provide risk assessment results in real time or batch processing, Figure 2 The ROC curve of the model of the present application is shown, which shows the performance of the model; Figure 3 The performance of the integrated model constructed in the present application is compared with that of other models, and the performance under the same dataset highlights the superiority of the model performance in the present application.

[0094] In specific use, the method program of the present application is put into a computer system comprising a processor and a memory, which is configured to execute the method of the present application. The system receives electronic health records (EHR) and physical examination data as input, and outputs a numerical probability representing the risk of diabetes.

[0095] The present application adopts a special neural network for feature fusion, and combines a powerful neural network-XGBoost integrated model for prediction (step S5) to construct a robust system. The system can effectively learn and integrate from heterogeneous data types (vector representation and raw numerical data), thereby improving the performance and generalization ability of the model.

[0096] The method of the present application achieves an AUC value of 0.84, which is a leading performance level in the industry. By capturing the complex nonlinear patterns in patient data, it significantly outperforms traditional risk models.

[0097] The above is only the preferred embodiment of the present application, it should be pointed out that for those skilled in the art, without departing from the technical principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application.

Claims

1. A method for predicting diabetes risk based on high-dimensional disease embedding and fusion of health data, characterized in that, Includes the following steps: S1. Data acquisition and preprocessing; S2, Construction of a high-dimensional disease embedding space; S3. Calculation of the individual health status vector; S4. Derivation of the health score vector based on physical examination data; S5. Hybrid prediction model for risk assessment; S6. Model Training and Deployment.

2. The diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data according to claim 1, characterized in that, Step S1 includes: S11. Collect electronic health record data from clinical databases, including patient identifiers, diagnosis codes, and diagnosis dates; S12. For the same patient cohort, collect relevant physical examination data, including but not limited to: systolic blood pressure, diastolic blood pressure, urinalysis specific gravity, urinalysis pH, high-density lipoprotein, fasting blood glucose, red blood cell distribution width, platelet count, alanine aminotransferase, low-density lipoprotein, creatinine, uric acid, aspartate aminotransferase, total cholesterol, total white blood cell count, monocyte percentage, absolute monocyte count, total red blood cell count, lymphocyte percentage, absolute lymphocyte count, mean corpuscular volume, mean corpuscular hemoglobin content, mean corpuscular hemoglobin concentration, basophil percentage, absolute basophil count, eosinophil percentage, absolute eosinophil count, hemoglobin, neutrophil percentage, absolute neutrophil count, and BMI; S13. Standardize and clean the collected data; specific operations include: mapping all diagnostic information to a unified coding system, deleting duplicate diagnostic entries; and integrating multiple medical records of each individual patient into a diagnostic sequence arranged in chronological order.

3. The diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data according to claim 2, characterized in that, Step S2 includes: S21. Define the corpus: Treat the diagnosis sequence of each patient arranged in chronological order as a "sentence", and the set of "sentences" of all patients constitutes the "corpus" for model training; S22. Apply word embedding algorithm: Use a continuous bag-of-words model or skip-word model to train the corpus defined in step S21. By processing disease "sentences", learn the vector representation of each unique disease "word". S23. Generate disease vectors: The output of this step is a mapping table, in which each unique disease code corresponds to an N-dimensional dense vector; in the constructed N-dimensional space, diseases with similar comorbidity patterns and potential causes will show a higher degree of closeness in terms of cosine similarity in their corresponding vectors.

4. The diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data according to claim 3, characterized in that, Step S3 includes: S31. For a given individual, identify all diagnosed diseases prior to the predicted time point. For a patient diagnosed with diabetes, the time point is prior to the diagnosis of diabetes; for an individual without diabetes, all diseases are their entire medical history. S32. From the disease embedding space constructed in step S2, retrieve the N-dimensional disease embedding vectors corresponding to all diagnosed diseases of the individual. S33. Calculate the individual's N-dimensional health status vector. This vector is obtained by weighting all retrieved disease vectors, with the weights based on the frequency of each disease d appearing in the individual's diagnosis record D. The i-th dimension of the vector The calculation formula is as follows: ； in, The value of the vector representing disease d in the i-th dimension is used to transform an individual's complete disease history into a precise coordinate point in the disease coordinate system, i.e., the health status vector.

5. The diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data according to claim 4, characterized in that, Step S4 includes: S41. Feature selection of physical examination indicators: To reduce collinearity among indicators, calculate the Pearson correlation coefficient matrix among all available physical examination indicators; apply a threshold to retain a set of M relatively independent physical examination indicators. S42. Correlation between physical examination indicators and disease spatial dimensions: For each of the N dimensions of the disease embedding space, perform linear regression analysis with the M independent physical examination indicators selected in step S41 to determine the effect size β. ij This value quantifies the correlation strength between the j-th physical examination index and the i-th dimension of the disease space; S43. Calculate an individual's health score vector: For a given individual, calculate their N-dimensional health score vector based on their specific physical examination results. The i-th dimension of the rating vector The calculation formula is as follows: ； in, β is the normalized value of the j-th individual's physical examination index. ij It is the effect value calculated in advance in step S42.

6. The diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data according to claim 5, characterized in that, Step S5 includes: S51, Feature Fusion Sub-model: A neural network is used as the feature fusion module; this feature fusion module is designed with three independent input branches: Branch A: Receives the N-dimensional health status vector from step S3; Branch B: Receives the N-dimensional health score vector from step S44; Branch C: Receives M raw physical examination indicators, which are normalized. Each branch consists of one or more fully connected layers, activation functions, and regularization layers, used to learn the latent representation for each input type; S52. Fusion based on attention mechanism: The latent representations learned by the three branches are fused through an attention mechanism. The attention mechanism calculates weights for the output of each branch, enabling the model to dynamically prioritize the features with the most information based on individual circumstances, and then concatenate the weighted representations into a unified feature vector. S53, Integrated Prediction Sub-model: The fused feature vector generated in step S52 is used as input and passed to a hybrid ensemble prediction model; this hybrid ensemble prediction model combines the outputs of a gradient booster and an independent neural network classifier. S54. Generation of the final risk score: The prediction results of the gradient booster and the neural network classifier are combined through a meta-learning strategy to generate a final probability score between 0 and 1, representing an individual's risk of developing diabetes.

7. The diabetes risk prediction method based on high-dimensional disease embedding and fusion of health data according to claim 6, characterized in that, Step S6 includes: S61. Train the integrated prediction sub-model described in step S53 on a balanced dataset in which the ratio of diabetic patients to non-diabetic individuals is 1:

1. S62. Save the trained model and deploy it on a server or computer system; by receiving new patient data, execute the prediction process to provide risk assessment results in real time or batch processing.

Images