EP-4738380-A1 - PREDICTION OF CARDIOVASCULAR EVENT

Abstract

A computer-implemented method of prediction of a cardiovascular, CV, event of a patient, based on a retinal image from the patient, is provided. The method comprises extracting a feature data set from the retinal image using a neural network, the neural network being pretrained to output feature data sets based on retinal images; inputting the extracted feature data set to a machine learning model, ML model, the ML model being pretrained to output a CV event indicator, the pretraining being based on the following data: training feature data sets extracted from retinal images; and training CV event indicators, one training CV event indicator per each one of the training feature data sets; and obtaining a CV event indicator of the patient from the ML model. A computer-implemented method of training a deep learning model based on retinal images and CV event indicators wherein the deep learning model comprises a neural network, NN, and a machine learning, ML, model, is provided. A data processing system, a training data processing system and a computer program product are also provided.

Inventors

• HERNÁNDEZ PASCUAL, Cristina
• SIMÓ CANONGE, Rafael
• SIMÓ-SERVAT, Olga
• MASIP, David
• GONZÁLEZ BARRIADA, Rubén

Assignees

• Fundació Hospital Universitari Vall d'Hebron - Institut de Recerca
• Fundacio per a la Universitat Oberta de Catalunya (UOC)

Dates

Publication Date

20260506

Application Date

20241030

Claims (15)

1. A computer-implemented method of prediction of a cardiovascular, CV event of a patient, based on a retinal image from the patient, the method comprising: - extracting a feature data set from the retinal image using a neural network, the neural network being pretrained to output feature data sets based on retinal images; - inputting the extracted feature data set to a machine learning model, ML model, the ML model being pretrained to output a CV event indicator, the pretraining being based on the following data: - training feature data sets extracted from retinal images; and - training CV event indicators, one training CV event indicator per each one of the training feature data sets; and - obtaining a CV event indicator of the patient from the ML model.
2. The computer-implemented method of prediction of a CV event of claim 1, wherein the training feature data sets are extracted using the neural network pretrained to output feature data sets based on retinal images.
3. The computer-implemented method of prediction of a CV event of any of claims 1 to 2, wherein the neural network comprises an autoencoder.
4. The computer-implemented method of prediction of a CV event of claim 3 wherein the autoencoder is an autoencoder based on the Transformer model.
5. The computer-implemented method of prediction of a CV event of any one of claims 1 to 4 wherein the machine learning model comprises one of: - Logistic regression; or - K-Nearest Neighbors, KNN; or - Support Vector Machine, SVM; or - Random Forest; or - Multi-Layer Perceptron , MLP; or - AdaBoost; or - Bayesian classifiers ; or - Decision Trees.
6. The computer-implemented method of prediction of a CV event of any one of claims 1 to 5 wherein the CV event indicator comprises a binary value indicating presence or absence of risk of developing a CV event of a patient.
7. A computer-implemented method of training a deep learning model based on retinal images and cardiovascular, CV, event indicators, wherein the deep learning model comprises a neural network, NN, and a machine learning, ML, model; the method comprising: - receiving, by the deep learning model, one or more sets of training retinal images and respective training CV events indicator per retinal image; - obtaining training feature data sets from the NN based on the received one or more sets of training retinal images, the NN being pretrained to obtain feature data sets based on retinal images; and - training the ML model to map each obtained training feature data set, of the obtained training feature data sets, to a training CV event indicator by providing, to the ML model, the obtained training feature data sets and the respective training CV event indicators where each CV event indicators is associated to a respective feature data set.
8. The computer-implemented method according to claim 7 of training further comprising training the NN to extract feature data sets based on at least part of the sets of training retinal images by providing, to the NN, the sets of training retinal images.
9. The computer-implemented method of training according to claim 7 or 8 wherein the neural network is an autoencoder and the training of the autoencoder comprises: - mapping each retinal image into a respective feature data set by using an encoding section of the autoencoder; and - reconstructing each mapped retinal image from the respective mapped feature data set by using a reconstruction section of the autoencoder; - wherein each retinal image has an initial dimension; - wherein each one of the feature data sets has a lower dimension than the initial dimension; and - wherein the mapping and the reconstructing are performed using a loss function.
10. The computer-implemented method of training according to any one of claims 7 to 9 further comprising: - transforming retinal images of an initial set of retinal images by adding occlusions at different points of the retinal images; - combining the initial set of retinal images with the transformed retinal images; and - providing thereby at least one set of the one or more sets of training retinal images, the at least one set of retinal images comprising an increased number of retinal images compared to the initial set of retinal images.
11. The computer-implemented method according to claim 9 of training further comprising: eliminating, once the autoencoder has been trained, the reconstruction section of the autoencoder.
12. A data processing system, the data processing system comprising a processor configured to carry out the steps of a computer-implemented method according to any one of claims 1 to 6 of prediction of a cardiovascular, CV, event of a patient, based on a retinal image from the patient.
13. The data processing system of claim 12 further comprising: - a first interface to obtain retinal images; and - a second interface to output a CV event indicator.
14. A training data processing system comprising a processor configured to carry out the steps of a computer-implemented method of training a deep learning model according to any one of claims 7 to 11.
15. A computer program product comprising program instructions for causing a computing system to perform a method according to any one of claims 1 to 6 or 7 to 11.

Description

The present disclosure relates to methods and systems used for predicting a risk for a patient to suffer from a cardiovascular event. BACKGROUND Some patients may present cardiovascular event risk compared to other patients. The identification of patients at a relative risk of suffering from a cardiovascular event is a challenge for healthcare professionals. SUMMARY In a first aspect, a computer-implemented method of prediction of a cardiovascular, CV, event of a patient is presented. The method is based on a retinal image from the patient and the method comprises: extracting a feature data set from the retinal image using a neural network, NN, the NN being pretrained to output feature data sets based on retinal images; inputting the extracted feature data set to a machine learning model, ML model; and obtaining a CV event indicator of the patient from the ML model. The ML model is pretrained to output a CV event indicator, the pretraining being based on the following data: training feature data sets extracted from retinal images; and training CV event indicators, one training CV event indicator per each one of the training feature data sets. Inference The above method defines therefore a method for the inference of a risk of a cardiovascular event based on retinal images. The inference is performed by a deep learning model consisting of a combination of the NN and the ML model, wherein the output of the NN is an input for the ML model. The NN is pretrained to output feature data sets based on retinal images. The NN is pretrained to output one feature data set per each input retinal image. In inference mode, the NN outputs or provides feature data sets given sets of retinal images. The ML model is pretrained to output a CV event indicator based on the following data: training feature data sets extracted from retinal images; and training CV event indicators, one training CV event indicator per each one of the training feature data sets. In the present disclosure the feature data sets may be referred to as "samples for the ML model" since such "feature data sets" may be used for training the ML model. In examples the pretraining is based on training feature data sets extracted from the same NN as the neural network used for the inference or prediction of a CV event. A cardiovascular event may be understood in the present disclosure as any one of heart attack, angina, or stroke. The patients, in examples, may comprise patients who have suffered from a CV event and patients who have not suffered from a CV event, or control patients. Extracting a feature data set from the retinal image by a neural network may be performed by a convolutional autoencoder (CAE) or by a variational autoencoder (VAE); such neural networks may be used to extract relevant features from the retinal images in an unsupervised mode. The convolutional autoencoder (CAE) may present an architecture comprising convolutional layers to process the retinal images. CAE may be advantageous for capturing spatial hierarchies and features within the data. The CAE may undergo an unsupervised training, where the neural network learns to encode retinal images into a compressed representation or a latent space and then decode that representation to reconstruct the original retinal image. Through this process, the neural network automatically learns relevant features without the need of labeled data. The variational autoencoder (VAE) may provide a latent space modeled as a probability distribution. The VAE learns to map input retinal images to a continuous latent space, enabling smooth transitions between features. The VAEs may learn to generate new retinal images by sampling from the latent space and decoding the samples. The neural network learns a rich set of features by maximizing the likelihood of reconstructing the input. VAEs may be used for generating new images, through generative modeling, and for extracting meaningful, disentangled features from data in an unsupervised manner. Generative Adversarial Networks (GANs), particularly unsupervised GAN variants, may also learn useful feature representations by generating images. Neural networks can extract meaningful representations from images without needing explicit labels, making them powerful tools for unsupervised learning. The neural network of the computer-implemented method may be trained or pretrained in an unsupervised mode to provide feature data sets based on the retinal images. The neural network may comprise an autoencoder, for example a masked autoencoder or for example an autoencoder based on a Transformer model. A Transformer model may be understood as a deep-learning model that may adopt the mechanism of attention, differentially weighing the significance of each part of the input data. Transformer models may be used in the field of computer vision. A Transformer model does not necessarily process the data in order. The term Transformer is to be understood in line with the common und