Analysing the impact of comorbid conditions and media coverage on online symptom search data: a novel AI-based approach for COVID-19 tracking

Figures & data

Figure 1. Long Short-Term Memory Network structure.

Table 1. Long Short-Term Memory Network (LSTM) formula.

LSTM Component^a	Formula	Function
Input Gate^b	$i_t=\sigma (X_t U^i+h_{t-1} W^i)$	The Input Gate, controlled by the Sigmoid ($\sigma$) function, determines the amount of information added to the cell state by transforming the input value into a range of 0 to 1.
Forget Gate^c	$f_t=\sigma (X_t U^f+h_{t-1} W^f)$	The Forget Gate, controlled by the Sigmoid ($\sigma$) function, determines the proportion of information to be forgotten from the cell state by transforming the input value into a range from 0 to 1.
State Candidate Gate (Cell update gate)^d	$\widetilde{C_t}=\mathrm{\text{tanh}}{(X}_t U^g+h_{t-1} W^g)$	The State Candidate Gate, using the Tanh function, calculates a new candidate cell state by integrating both current and previous inputs.
Output Gate^e	$O_t=\sigma {(X}_t U^o+h_{t-1} W^0)$	The Output Gate is controlled by the sigmoid function that controls the current state, which is used to generate an output vector that weights the current cell state based on how relevant it is to the current output.
Output^f	$C_t=\sigma {(f}_t* C_{t-1}+i_t* \widetilde{C_t})$ $h_t=\mathrm{\text{tanh}}\left(C_t\right)* O_t$	Create a new cell state and hidden state

^a In this table, $X_t$​ denotes the daily COVID-19 case numbers, and $h_{t-1}$​ represents the output of the previous time step of the LSTM cell, which influences the function of each gate.

^b In the input gate, $U^i$ is the weight matrix applied to the current input (daily case numbers), adjusting how this data influences the cell state. The $W^i$ is the weight matrix for the previous state $h_{t-1},$ which helps in incorporating historical data trends into the current state analysis. The $i_t$ presents the extent of new information to be added to the cell state.

^c In the forget gate, $U^f$ is the weight matrix for the current input in the forget gate, influencing which parts of the new data should be prioritised or downplayed. The $W^f$ is the weight matrix for the previous state in the forget gate, determining the extent to which past data influences the current cell state. Lastly, the $f_t$ used to decide the amount of information from the cell state to be discarded.

^d In the state candidate gate, $U^g$ is the weight matrix for the current input (daily case numbers) in the state candidate gate, shaping the new state proposal. The $W^{g }$ is the weight matrix for the previous state in the state candidate gate, blending historical trends into the new state formulation. The $\widetilde{C_t}$ is used to calculate a potential new value for the cell state, considering current and previous inputs.

^e In the output gate, the $U^o$ is the weight matrix for the current input in the output gate, crucial for deciding how the current data influences the final output. The $W^0$ is the weight matrix for the previous state in the output gate, playing a role in how past trends and states affect the current output. The $O_t$ is used to control what part of the cell state is transmitted to the output.

^f In the output, the $C_t$ represents the predicted case at time t while $h_t$ is the new state of the LSTM cell after processing the current input.

Figure 2. Minimised the impact of comorbid conditions with similar symptoms in (a) Australia and (b) New Zealand, respectively. The pink area represents the search range for comorbidities with similar symptoms. Specifically, the comorbid conditions with similar symptoms were estimated using data collected during the non-COVID-19 period, from 6th January 2019 to 10th October 2019. Meanwhile, the blue area in the graph represents the COVID-19 related search.

Figure 3. Estimated components of COVID-19 symptom search in Australia and New Zealand with (a) (c) the quantile 10–90 scenario, (b) (d) the quantile 25–75 scenario, respectively. The green area represents minimal media influence on symptom search, while the grey area represents minimal infected person-led symptom search. The pink area in the figure represents the uncertainty zone, bounded by the lower and upper bounds of media-influenced symptom searches, which visualises the dynamic impact of media coverage on online symptom searches and highlights the range of uncertainty in the analysis.

Figure 4. Estimation of COVID-19 cases using the proposed method with the RNN-LSTM embedding model for (a) Australia and (b) New Zealand for the 10–90 and 25–75 quantile scenarios.

Table 2. Results of model evaluation.

Model	Australia			New Zealand
	R^2	RMSE	MAE	R^2	RMSE	MAE
Proposed Method with RNN-LSTM Embedding (Quantile 10–90)	0.8948	0.0519	0.0392	0.7746	0.1013	0.0780
Proposed Method with RNN-LSTM Embedding (Quantile 25–75)	0.9130	0.0472	0.0331	0.7813	0.0998	0.0780
RNN-LSTM	0.7724	0.4770	0.3911	0.3068	0.8325	0.6464
Linear regression	0.6843	0.5618	0.4887	0.1938	0.8978	0.6642
ElasticNet	0.4812	0.7202	0.5489	0.1002	0.9485	0.7660