Predictions with Deep Learning and System Dynamics - AIH
0 likes197 views
The document discusses the automation of artificial intelligence models for predicting deaths in hospital admissions for patients over 50 using machine learning and deep learning. It highlights the use of AutoML to improve prediction accuracy and reduce errors through the analysis of substantial historical data from hospital records. The project invites participation in enhancing predictive systems, emphasizing the potential application of derived functions in deep learning models.
Predictions with Deep Learning and System Dynamics - AIH
- 1. Automation of Artificial Intelligence Models Guttenberg Ferreira Passos Deep Learning and System Dynamics for Time Series forecasts of deaths in hospital admissions - AIH
- 2. Propósito • Automação dos Modelos de Inteligência Artificial para reduzir os erros nos testes com algoritmos de Machine Learning e Deep Learning. Fonte: https://www.analyticsinsight.net/top-automl-platforms-to- look-out-for-in-2020/
- 3. Artificial Intelligence • Artificial Intelligence (AI) is the study of how to make computers perform tasks in which, at the moment, people are better.
- 4. Fonte: https://goo.gl/p2SjrV Por Stefan Kojouharov • Machine Learning is a branch of Artificial Intelligence based on the idea that systems can learn from data, identify patterns and make decisions with minimal human intervention.
- 5. Fonte: https://goo.gl/p2SjrV Por Stefan Kojouharov
- 6. Deep Learning • Deep Learning is a type of Machine Learning that trains computers to perform tasks like humans, which includes speech recognition, image identification and predictions.
- 7. AutoML • AutoML is the automation of Machine Learning and Deep Learning models through the automatic creation of model codes. AutoML is in its infancy and its future will determine the future of Artificial Intelligence itself. Fonte: http://mentoriadsalc.com/2020/08/09/automl-automacao-do-trabalho-na-ia/#Resumo
- 8. Deaths in hospital admissions – AIH • The present project aims to predict the number of deaths of patients over 50 years of age in hospital admissions - AIH, through the study of Artificial Intelligence models, using time series, for the comparison of results and the analysis of predictions suggesting the best option with the least error. • Data were downloaded from the Datasus website, SIHSUS module, file type RD- AIH Reduced, years 2010 to 2020, UF MG, every month : • http://www2.datasus.gov.br/DATASUS/index.php?area=0901&item =1&acao=25
- 9. Deaths in hospital admissions – AIH • 129 files were downloaded containing the number of hospitalizations of people over 50 years of age who died. • These files were grouped with 5,231,829 records used to prepare a csv file. • The hospitalization data were totaled, making a total of 3,896 daily records. • The daily records were regrouped in 556 weekly records with date, number and total value of hospitalizations, data since 2010.
- 10. Challenge • Use AutoML to perform the automation of Predictive Systems, using as a basis the data the prediction of the number of deaths of patients over 50 years in hospital admissions - AIH, allowing a better accuracy in the forecast and reducing errors in tests with Machine Learning algorithms and Deep Learning. Prototype : https://pt2.slideshare.net/guttenbergpasso s/inteligncia-artificial-em-sries-temporais- na-arrecadao
- 11. Purpose of AutoML: decrease errors RSME RSME RSME Nº Arquitetura Modelo Setup treino validação teste 1 Métodos Estatísticos BASE 11 Método Naive 86,1777 2 Métodos Estatísticos Forecasting 12 Exponential Smoothing v1 92,9691 3 Métodos Estatísticos Forecasting 12 Exponential Smoothing v2 76,6730 4 Métodos Estatísticos ARIMA 13 ARIMA LOG (1, 0, 1) 149,2452 5 Métodos Estatísticos ARIMA 14 ARIMA LOG (1, 1, 1) 79,8950 6 Métodos Estatísticos ARMA 15 ARMA (1, 0) 150,6672 7 Métodos Estatísticos ARMA 15 ARMA (5, 5) 144,1510 8 Métodos Estatísticos ARMA 15 ARMA (12, 9) 100,0858 9 Métodos Estatísticos ARIMA 16 ARIMA (3, 1, 3) Forecast 80,4460 10 Métodos Estatísticos SARIMAX 17 SARIMAX (2, 1, 2) 2, 1, 0, 12) 197,0291 11 Métodos Estatísticos SARIMAX 18 SARIMAX (0, 1, 1) (0, 1, 1, 12) 81,8482 12 IA - Deep Learning LSTM 22 LSTM (3 repetições) 130,1589 13 IA - Deep Learning LSTM 22 LSTM (5 repetições) 120,1346 14 IA - Deep Learning LSTM 23 LSTM Otimizado 98,5724 15 IA - Deep Learning LSTM 24 Stacked LSTM 100,2641 16 IA - Deep Learning LSTM 25 LSTM Bidirecional 86,4822 17 IA - Deep Learning DeepAR 26 DeepAR 934,1830 18 Inteligência Artificial - IA RNA - MLP 1 MLP Vanilla 51,5066 57,1090 19 Inteligência Artificial - IA RNA - MLP 2 MLP e Método Window 47,8556 51,3619 20 IA - Deep Learning RNN - LSTM 1 LSTM Vanilla - 130 epochs 45,6769 67,4191 63,9893 21 IA - Deep Learning RNN - LSTM 1 LSTM Vanilla - 200 epochs 45,2500 63,4594 61,6839 22 IA - Deep Learning RNN - LSTM 2 LSTM e Método Window 48,5169 62,8968 69,1200 23 IA - Deep Learning RNN - LSTM 3 LSTM e Time Steps 46,1317 70,3126 73,4725 24 IA - Deep Learning RNN - LSTM 4 LSTM e Stateful 42,0199 60,9425 57,2971 25 IA - Deep Learning RNN - LSTM 5 LSTM e Stacked 41,7801 73,1018 72,6560 Scores de Óbitos em Internações Hospitalares - AIH
- 12. Purpose of AutoML: decrease errors
- 13. AWS AutoPilot • With the AWS SageMaker Studio service we can use AutoPilot to automate Predictive Systems and publish the results on the Data Science Academy - DSA team line to share the results.
- 14. Invitation to challenge • We invite everyone to take part in this challenge using AutoML to automate Predictive Systems and improve accuracy in predicting models by decreasing test errors.
- 15. Next challenge Next year we have a new challenge : Depois de uma camada oculta no Deep Learning é aplicada uma função de ativação para promover a não linearidade, imagine se fosse possível aplicar uma função derivada de um modelo de System Dynamics sobre COVID-19 nessa etapa! Fonte: https://www.systemdynamics.org/covid-19 Fonte: http://deeplearningbook.com.br/