Selleski, Franco (2021) Desarrollo e implementación de técnicas avanzadas de aprendizaje automático y procesamiento de señales para el análisis de registros electrocardiográficos en pacientes de riesgo cardíaco / Development and implementation of advanced machine learning and signal processing techniques for the analysis of electrocardiographic recordings in patients with cardiac risk. Proyecto Integrador Ingeniería en Telecomunicaciones, Universidad Nacional de Cuyo, Instituto Balseiro.
![[img]](/style/images/fileicons/application_pdf.png)
| PDF (Tesis) Español 22Mb |
Resumen en español
Las enfermedades cardiovasculares son la principal causa de muerte en el mundo. Debido a esto, el diagnóstico rápido y eficiente es de vital importancia en los pacientes con riesgo de sufrir estas afecciones. El Síndrome Coronario Agudo (SCA) es una patología comprendida dentro de las enfermedades cardiovasculares, cuyo diagnóstico inicial se realiza en base al análisis de la actividad eléctrica del corazón por medio de un electrocardiograma (ECG). Se estima que la sensibilidad del análisis del ECG es de entre 49 y 61 %, mas aumenta al 100% si se mantiene al paciente en observación entre 6 y 12 horas y se llevan a cabo estudios de laboratorio. Como consecuencia, resulta importante contar con herramientas basadas solamente en el análisis de registros ECG que ayuden al personal de guardia a estimar de forma rápida y confiable el riesgo que presenta un paciente de sufrir un SCA. Con este objetivo se diseñaron, desarrollaron e implementaron dos métodos de Deep Learning para predecir la posibilidad de que un paciente sufra un SCA a partir de una señal de ECG. El primer de método se basó en un autoencoder convolucional y para los datos de testing se obtuvo una exactitud del 57 %, con una sensibilidad del 55% y una especificidad del 60 %. El segundo método se basó en la estrategia de transfer learning, utilizando una red preentrenada para clasificar arritmias, y para los datos de testing se obtuvo una exactitud del 58 %, con una sensibilidad del 56% y una especificidad del 61 %. Los resultados de sensibilidad obtenidos en pacientes infartados son comparables a los obtenidos por especialistas al analizar sólo el registro de ECG, mas son inferiores a los obtenidos en los métodos analizados del estado del arte. Se encontró que la marginalidad de estos índices de desempeñó se debe a que el número de casos de la base de datos utilizada resultó insuficiente para que los métodos propuestos aprendieran a resolver satisfactoriamente el problema abordado. Sin embargo, es importante destacar que los métodos desarrollados podrán ser aplicados a medida que se registren nuevos casos en la base de datos local, y de esta manera definir en forma más precisa la cota inferior asociada al número de casos necesarios para obtener una mejora significativa en la capacidad de clasificación. Otro resultado relevante asociado a la comprensión del funcionamiento de los métodos de Deep Learning implementados se obtuvo mediante la aplicación de la técnica de Grad-CAM, para hallar las regiones de la entrada de mayor importancia al momento de realizar una predicción, de donde se pudo comprender que la red se enfoca en las mismas regiones para pacientes infartados y no infartados, y que no se enfoca en una de las regiones características del latido (segmento ST) que utilizan los cardiólogos para realizar un diagnóstico.
Resumen en inglés
Cardiovascular diseases are the leading cause of death worldwide. Because of this, a rapid and efficient diagnosis is of vital importance in patients with suspected risk of suffering from these conditions. Acute Coronary Syndrome (ACS) is a pathology included within cardiovascular diseases, whose initial diagnosis is made based on the analysis of the heart’s electrical activity by means of an electrocardiogram (ECG). The sensitivity of the ECG analysis is estimated to be between 49 and 61%, but increases to 100% if the patient is kept under observation for 6 to 12 hours and specialized laboratory studies are performed. As a consequence, it is important to have tools based solely on the analysis of ECG recordings to help emergency personnel to quickly and reliably estimate a patient’s risk of suffering an ACS. With this objective in mind, two methods were designed, developed and implemented to predict the possibility of a patient suffering an ACS from an ECG signal. The first method was based on a convolutional autoencoder and for testing data had an accuracy of 57%, with a sensitivity of 55% and a specificity of 60%. The second method was based on a transfer learning strategy, using part of a network trained to classify arrhythmias, and for testing data had an accuracy of 58%, with a sensitivity of 56% and a specificity of 61%. The sensitivity results obtained for infarcted patients are comparable to those obtained by specialists when analyzing the ECG records alone, but are lower than those obtained in the state-of-the-art methods analyzed. It was found that the marginality of these performance indices was related to the reduced number of cases available in the local database. However, it is worth noting that the application of the proposed methods, as new cases are added to the local database, could pave the way to significantly improve the classification performance. Grad-CAM technique was applied to find the most important regions of the input at the time of making a prediction, from which it was possible to understand that the network focuses on the same regions for infarcted and non-infarcted patients, and that it does not focus on one of the characteristic regions of the heartbeat (ST segment) used by cardiologists to make a diagnosis.
| Tipo de objeto: | Tesis (Proyecto Integrador Ingeniería en Telecomunicaciones) |
|---|---|
| Palabras Clave: | Machine learning; Aprendizaje automático; Patients; Pacientes; Electrocardiograms; Electrocardiogramas; [Acute coronary syndrome; Síndrome coronario agudo; Deep learning; Aprendizaje profundo] |
| Referencias: | [1] WHO: The top 10 causes of death, 2020. URL https://www.who.int/news-room/fact-sheets/detail/the-top-10-causes-of-death. 1 [2] WHO: Cardiovascular diseases (CVDs), 2021. URL https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds). 1 [3] Jesse, R. L., Kontos, M. C. Evaluation of chest pain in the emergency department. Current Problems in Cardiology, 22 (4), 149–236, 1997. URL https://www.sciencedirect.com/science/article/pii/S0146280697800072. 1, 11,40 [4] Daubert, M. A., Jeremias, A. The utility of troponin measurement to detect myocardial infarction: review of the current findings. Vascular health and risk management, 6, 691, 2010. 1, 11 [5] Diagram of the human heart (cropped). URL https://commons.wikimedia.org/wiki/File:Diagram_of_the_human_heart_(cropped)_es.svg. 5 [6] Esquema de colocaci´on de electrodos en electrocardiograma de 12 derivaciones. URL https://commons.wikimedia.org/wiki/File:ECGcolor.svg. 6 [7] Esquema de colocaci´on de electrodos en derivaciones precordiales. URL https: //commons.wikimedia.org/wiki/File:Precordial_Leads_2.svg. 8 [8] Pardal, S. F. Entiendo electrocardiograma. 9 [9] Lee, T. H., Cook, E. F., Weisberg, M., Sargent, R. K., Wilson, C., Goldman, L. Acute Chest Pain in the Emergency Room: Identification and Examination of Low-Risk Patients. Archives of Internal Medicine, 145 (1), 65–69, 01 1985. URL https://doi.org/10.1001/archinte.1985.00360010085013. 11 [10] Ejemplo de perceptr´on multicapa. URL https://commons.wikimedia.org/wiki/File:RedNeuronalArtificial.png. 15 [11] LeCun, Y., Bengio, Y., et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361 (10), 1995, 1995. 15 [12] Goodfellow, I., Bengio, Y., Courville, A. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org. 16 [13] Chollet, F. Deep Learning with Python. Manning, 2017. 16 [14] Nair, V., Hinton, G. E. Rectified linear units improve restricted boltzmann machines. En: Icml. 2010. 17 [15] Maas, A. L., Hannun, A. Y., Ng, A. Y., et al. Rectifier nonlinearities improve neural network acoustic models. En: Proc. icml, tomo 30, p´ag. 3. Citeseer, 2013.17 [16] Representaci´on esquem´atica de la arquitectura de un autoencoder. URL https://medium.com/autoencoder-for-anomaly-detection/ autoencoder-for-anomaly-detection-db6178ad07b2. 17 [17] Hinton, G. E., Srivastava, N., Krizhevsky, A., Sutskever, I., Salakhutdinov, R. R. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012. 17 [18] Validaci´on cruzada de K iteraciones con K=4. URL https://commons. wikimedia.org/wiki/File:K-fold_cross_validation.jpg. 18 [19] Selvaraju, R. R., Das, A., Vedantam, R., Cogswell, M., Parikh, D., Batra, D. Grad-cam: Why did you say that? visual explanations from deep networks via gradient-based localization. CoRR, abs/1610.02391, 2016. URL http://arxiv.org/abs/1610.02391. 20 [20] Acharya, U. R., Fujita, H., Oh, S. L., Hagiwara, Y., Tan, J. H., Adam, M. Application of deep convolutional neural network for automated detection of myocardial infarction using ecg signals. Information Sciences, 415, 190–198, 2017. 21, 22 [21] PTB Diagnostic ECG Database. URL https://physionet.org/content/ptbdb/1.0.0/. 21 [22] Lodhi, A. M., Qureshi, A. N., Sharif, U., Ashiq, Z. A novel approach using voting from ecg leads to detect myocardial infarction. En: Proceedings of SAI Intelligent Systems Conference, p´ags. 337–352. Springer, 2018. 21 [23] Lui, H. W., Chow, K. L. Multiclass classification of myocardial infarction with convolutional and recurrent neural networks for portable ecg devices. Informatics in Medicine Unlocked, 13, 26–33, 2018. 21 [24] AF Classification from a Short Single Lead ECG Recording - The PhysioNet Computing in Cardiology Challenge 2017. URL https://physionet.org/content/challenge-2017/1.0.0/. 21 [25] Baloglu, U. B., Talo, M., Yildirim, O., San Tan, R., Acharya, U. R. Classification of myocardial infarction with multi-lead ecg signals and deep cnn. Pattern Recognition Letters, 122, 23–30, 2019. 22 [26] Wu, C.-C., Hsu, W.-D., Islam, M. M., Poly, T. N., Yang, H.-C., Nguyen, P.- A. A., et al. An artificial intelligence approach to early predict non-st-elevation myocardial infarction patients with chest pain. Computer methods and programs in biomedicine, 173, 109–117, 2019. 22 [27] Nurmaini, S., Umi Partan, R., Caesarendra, W., Dewi, T., Naufal Rahmatullah, M., Darmawahyuni, A., et al. An automated ecg beat classification system using deep neural networks with an unsupervised feature extraction technique. Applied Sciences, 9 (14), 2921, 2019. 22 [28] Moody, G. B., Mark, R. G. The impact of the mit-bih arrhythmia database. IEEE Engineering in Medicine and Biology Magazine, 20 (3), 45–50, 2001. 22 [29] Polero, L. D., GARMENDIA, C. M., Echegoyen, R. E., de Lima, A. A., Bert´on, F., Lambardi, F., et al. Predicci´on de riesgo de sufrir un síndrome coronario agudo mediante un algoritmo de machine learning (angina). Revista Argentina de Cardiolog´ıa, 88 (1), 9–13, 2020. 23 [30] Ribeiro, A. H., Ribeiro, M. H., Paix˜ao, G. M. M., Oliveira, D. M., Gomes, P. R., Canazart, J. A., et al. Automatic diagnosis of the 12-lead ecg using a deep neural network. Nature Communications, 11 (1), Apr 2020. URL http://dx.doi.org/10.1038/s41467-020-15432-4. 23, 24, 29, 30, 41 [31] He, K., Zhang, X., Ren, S., Sun, J. Deep residual learning for image recognition. En: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2016. 23 [32] Cho, Y., Kwon, J.-m., Kim, K.-H., Medina-Inojosa, J. R., Jeon, K.-H., Cho, S., et al. Artificial intelligence algorithm for detecting myocardial infarction using six-lead electrocardiography. Scientific reports, 10 (1), 1–10, 2020. 24, 25, 26, 42 [33] Ribeiro, A. L. P., Paix˜ao, G. M., Gomes, P. R., Ribeiro, M. H., Ribeiro, A. H., Canazart, J. A., et al. Tele-electrocardiography and bigdata: The code (clinical outcomes in digital electrocardiography) study. Journal of Electrocardiology, 57, S75–S78, 2019. URL https://www.sciencedirect.com/science/article/pii/S0022073619304984. 27, 28 [34] Ribeiro, A. H., Paixao, G. M., Lima, E. M., Horta Ribeiro, M., Pinto Filho, M. M., Gomes, P. R., et al. CODE-15%: a large scale annotated dataset of 12-lead ECGs, jun. 2021. URL https://doi.org/10.5281/zenodo.4916206. 28 [35] Rude, R. E., Poole, W. K., Muller, J. E., Turi, Z., Rutherford, J., Parker, C., et al. Electrocardiographic and clinical criteria for recognition of acute myocardial infarction based on analysis of 3,697 patients. The American journal of cardiology, 52 (8), 936–941, 1983. 40 [36] Woo, S., Park, J., Lee, J.-Y., Kweon, I. S. Cbam: Convolutional block attention module. En: Proceedings of the European conference on computer vision (ECCV), págs.. 3–19. 2018. 44 |
| Materias: | Ingeniería en telecomunicaciones > Aprendizaje automático Ingeniería en telecomunicaciones > Redes profundas |
| Divisiones: | Gcia. de área de Investigación y aplicaciones no nucleares > Gcia. de Física > Física médica |
| Código ID: | 1114 |
| Depositado Por: | Tamara Cárcamo |
| Depositado En: | 07 Sep 2022 14:15 |
| Última Modificación: | 07 Sep 2022 14:28 |
Personal del repositorio solamente: página de control del documento
