A Novel Matrix Profile-Guided Attention LSTM Model for Forecasting COVID-19 Cases in USA

Qian Liu, Daryl L. X. Fung, Leann Lac, Pingzhao HuFrontiers in Public Health 2021 · DOI

TL;DR

This paper introduces a novel hybrid model that combines matrix profile-based anomaly detection with an attention-based LSTM for more accurate COVID-19 case forecasting. The proposed model significantly outperforms traditional statistical and baseline RNN models, offering a more robust tool for public health decision-making.

What Problem Does This Paper Address?

Accurately forecasting infectious disease outbreaks, like COVID-19, is crucial for effective public health interventions, resource allocation, and policy development. Traditional time series models often struggle to capture the complex, non-linear patterns and anomalies inherent in epidemic data. While Recurrent Neural Networks (RNNs), particularly Long Short-Term Memory (LSTM) networks, have shown promise, they can be limited in identifying which specific historical data points are most influential for future predictions. This paper aims to address this by developing a more sophisticated forecasting model that can better identify and leverage important patterns, including anomalies, within historical COVID-19 data.

Methods

The core of this research is a novel hybrid forecasting algorithm that integrates two key components:

  • Matrix Profile: This is an unsupervised time series anomaly detection technique. It works by comparing subsequences (or "snippets") of a time series to find the most similar ones. If a subsequence is significantly different from its nearest neighbors, it's flagged as a potential anomaly or "discord." The paper uses the matrix profile to identify these unusual patterns in the COVID-19 data.
  • Attention-Based LSTM: Long Short-Term Memory (LSTM) networks are a type of RNN well-suited for sequential data. The "attention mechanism" is a recent advancement that allows the model to dynamically weigh the importance of different parts of the input sequence when making a prediction. This means the model can learn to focus on the most relevant historical data points, rather than treating all past data equally.

The proposed model concatenates the normalized observed data with features derived from the matrix profile (specifically, the distance profile and relative position) and feeds this combined input into an LSTM. This LSTM then utilizes an attention mechanism within an encoder-decoder architecture to generate forecasts. The authors also compare their model against several baseline models, including SARIMA, simple RNN, ESN, GRU, and an LSTM without attention or matrix profile features.

Datasets

The study utilizes publicly available COVID-19 data from the Carnegie Mellon University Delphi Research Group. The data spans from March 1, 2020, to August 5, 2021, and includes three key indicators:

  • Adjusted percentage of daily admitted hospitalized COVID-19 cases (hospital admission)
  • Number of daily confirmed COVID-19 cases (confirmed cases)
  • Number of daily death cases caused by COVID-19 (death cases)

The data was pre-processed using z-normalization. The evaluation employed a "rep-holdout" cross-validation strategy, where the last 10%, 20%, 30%, 40%, and 50% of the time series data were used as test sets, with the remaining data serving as the training set.

Key Insights and Ideas

  • Anomalies Matter: The matrix profile successfully identified significant anomalies and turning points in the COVID-19 time series data, such as the peaks in hospital admissions and deaths. This suggests that explicitly accounting for these unusual events can improve forecasting.
  • Attention is Key: The attention mechanism allows the LSTM to focus on the most relevant historical data, leading to more accurate predictions. This is crucial for complex time series where not all past data points are equally informative.
  • Hybridization Pays Off: Combining the anomaly detection capabilities of the matrix profile with the sequence modeling power of an attention-based LSTM creates a synergistic effect, leading to superior forecasting performance.
  • Training Strategy Nuances: The study found that a larger training set doesn't necessarily guarantee better performance, highlighting the importance of the training strategy itself.

Key Results

The proposed matrix profile-assisted attention-based LSTM model (LSTM-MatAtt) demonstrated superior performance across all three COVID-19 indicators compared to baseline models. Specifically, the paper reports the following average performance metrics:

  • Hospital Admission: RMSE = 1.23, MAE = 0.95, MAPE = 0.25
  • Confirmed Cases: RMSE = 31612.81, MAE = 26259.55, MAPE = 1.06
  • Death Cases: RMSE = 467.17, MAE = 364.02, MAPE = 0.55

These results indicate a significant improvement in forecasting accuracy, suggesting the model's potential to aid policymakers in making informed decisions regarding prevention and resource allocation.

Limitations and Future Work

The authors acknowledge a few limitations:

  • Validation Dataset: The model was validated using a five-rep-holdout strategy but was not tested on a completely independent dataset.
  • Forecasting Uncertainty: Predicting future cases is inherently uncertain, and mathematical models can sometimes overpredict, especially when dealing with complex societal factors and policy interventions.

Future work could involve validating the model on additional independent datasets and exploring ways to better incorporate potential uncertain factors into the forecasting process. The authors also suggest that the forecasted data can provide valuable insights for decision-makers to manage the consequences of COVID-19.

Summary updated

Abstract

<b>Background:</b> The outbreak of the novel coronavirus disease 2019 (COVID-19) has been raging around the world for more than 1 year. Analysis of previous COVID-19 data is useful to explore its epidemic patterns. Utilizing data mining and machine learning methods for COVID-19 forecasting might provide a better insight into the trends of COVID-19 cases. This study aims to model the COVID-19 cases and perform forecasting of three important indicators of COVID-19 in the United States of America (USA), which are the adjusted percentage of daily admitted hospitalized COVID-19 cases (<i>hospital admission</i>), the number of daily confirmed COVID-19 cases (<i>confirmed cases</i>), and the number of daily death cases caused by COVID-19 (<i>death cases</i>). <b>Materials and Methods:</b> The actual COVID-19 data from March 1, 2020 to August 5, 2021 were obtained from Carnegie Mellon University Delphi Research Group. A novel forecasting algorithm was proposed to model and predict the three indicators. This algorithm is a hybrid of an unsupervised time series anomaly detection technique called matrix profile and an attention-based long short-term memory (LSTM) model. Several classic statistical models and the baseline recurrent neural network (RNN) models were used as the baseline models. All models were evaluated using a repeated holdout training and test strategy. <b>Results:</b> The proposed matrix profile-assisted attention-based LSTM model performed the best among all the compared models, which has the root mean square error (RMSE) = 1.23, 31612.81, 467.17, mean absolute error (MAE) = 0.95, 26259.55, 364.02, and mean absolute percentage error (MAPE) = 0.25, 1.06, 0.55, for <i>hospital admission, confirmed cases</i>, and <i>death cases</i>, respectively. <b>Conclusion:</b> The proposed model is more powerful in forecasting COVID-19 cases. It can potentially aid policymakers in making prevention plans and guide health care managers to allocate health care resources reasonably.
BibTeX
@inproceedings{liu2021novel,
  author = {Qian Liu and Daryl L. X. Fung and Leann Lac and Pingzhao Hu},
  title = {{A Novel Matrix Profile-Guided Attention LSTM Model for Forecasting COVID-19 Cases in USA}},
  booktitle = {Frontiers in Public Health},
  year = {2021},
  doi = {10.3389/fpubh.2021.741030},
}
Figure