The efficacy of machine learning models in predicting the spread of COVID-19

Boluwaji Odufuwa

Research Mentor: Fang-Yi Yu

Epidemiology

&

COVID-19

• Epidemiologist lacked accurate regression tool for COVID-19

​

• The advent of such a tool could direct more effective policy and thus save thousands of lives while maintaining national economic stability

Paper 1: ANFIS (Adaptive neuro fuzzy inference system )

Ardabili, S.F.; Mosavi, A.; Ghamisi, P.; Ferdinand, F.; Varkonyi-Koczy, A.R.; Reuter, U.; Rabczuk, T.; Atkinson, P.M. COVID-19 Outbreak Prediction with Machine Learning. Algorithms 2020, 13, 249.

Paper Summary

Context

Researchers were aiming to find/develop a predictive model for COVID-19 that was greater in efficacy than standard epidemiological models

Paper Summary

Context

Researchers were aiming to find/develop a predictive model for COVID-19 that was greater in efficacy than standard epidemiological models

​

Pros

• ANFIS utilizes multiple neural networks
• Inherently more accurate than linear regression

Paper Summary

Context

Researchers were aiming to find/develop a predictive model for COVID-19 that was greater in efficacy than standard epidemiological models

Pros

• ANFIS utilizes multiple neural networks
• Inherently more accurate than linear regression

Cons

• Uses static, unilayer data
• One of the first ML applications to COVID-19

Model Architecture

A type of Neural Network based on the Takagi-Sugeno fuzzy system. Fuzzy logic utilizes magnitudes of truth (Any real number between 0 and 1) instead of Boolean True(1) and False(0)

Challenge 2

Challenge 3

Model Architecture

A type of Neural Network based on the Takagi-Sugeno fuzzy system. Fuzzy logic utilizes magnitudes of truth (Any real number between 0 and 1) instead of Boolean True(1) and False(0)

This model is a kind of division and conquest method. Instead of using one neural network for all the input and output data, several networks are created in this model

Model Architecture

A type of Neural Network based on the Takagi-Sugeno fuzzy system. Fuzzy logic utilizes magnitudes of truth (Any real number between 0 and 1) instead of Boolean True(1) and False(0)

This model is a kind of division and conquest method. Instead of using one neural network for all the input and output data, several networks are created in this model

Multiple hidden layers within each neural network

Paper 2: NARX (Nonlinear autoregressive exogenous model)

Akhtar, M., Kraemer, M.U.G. & Gardner, L.M. A dynamic neural network model for predicting risk of Zika in real time. BMC Med 17, 171 (2019). https://doi.org/10.1186/s12916-019-1389-3

Paper Summary

Context

During the height of the Zika Virus in 2018, this paper utilized NARX (Nonlinear autoregressive models with exogenous inputs) neural networks and multilayer data to predictively map the spread of Zika

Pros

Cons

Paper Summary

Context

During the height of the Zika Virus in 2018, this paper utilized NARX (Nonlinear autoregressive models with exogenous inputs) neural networks and multilayer data to predictively map the spread of Zika

Pros

Cons

Pros

Pros

• 85%< Model Accuracy
• Accounts for time-series data and human mobility
• Inclusion of travel data improves result accuracy

Paper Summary

Context

During the height of the Zika Virus in 2018, this paper utilized NARX (Nonlinear autoregressive models with exogenous inputs) neural networks and multilayer data to predictively map the spread of Zika

Pros

Cons

Pros

Pros

• 85%< Model Accuracy
• Accounts for time-series data and human mobility
• Inclusion of travel data improves result accuracy

Cons

• Designed for Zika Virus
• Inclusion of non-travel related data did not improve prediction accuracy and decreased performance

Model Architecture

Unlike conventional recurrent neural networks, NARX neural networks require input from output neurons rather than hidden states and converge much faster with better generalization

Model Architecture

Unlike conventional recurrent neural networks, NARX neural networks require input from output neurons rather than hidden states and converge much faster with better generalization

Desired output (yk(t + N)), is a binary classifier of risk

Challenge 3

Model Architecture

Unlike conventional recurrent neural networks, NARX neural networks require input from output neurons rather than hidden states and converge much faster with better generalization

Desired output (yk(t + N)), is a binary classifier of risk

Before input variables were applied to NARX model, they were normalized to the range [0,1]

Hypothesis

ANFIS network will produce most accurate predictive data

Model Modifications

NARX

• One input variable (Week)
• Hidden layer with 15 nodes
• 80% of data points used for learning and 20% used for testing/validation

ANFIS

• Computed Mean Square Error (MSE) instead of Root Mean Square Error (RMSE)
• Normalized MSE by a ratio of 1/300,000,000
• Focused on USA MSE

Results

SIR Model: Baseline Comparison

NARX Neural Network

1.09281282729e-05

SIR Model

9.18515168041e-06

NARX Model

3.57848408333e-08

Comparing Mean Square Errors

ANFIS Model

Discussion:

• Limitations
• Conclusion
• Future Steps