MEDINFO 2004 M. Fieschi et al. (Eds) Amsterdam: IOS Press 2004 IMIA.

All rights reserved

We Did the Right Thing: An Intervention Analysis Approach to Modeling Intervened SARS Propagation in Singapore
Bin Han, Tze-Yun Leong
Medical Computing Laboratory, School of Computing, National University of Singapore, Singapore 117543 {hanb, leongty}@comp.nus.edu.sg

Bin Han, Tze-Yun Leong

Abstract In this paper, we adopt the Intervention Analysis approach to model an intervened natural process, i.e., propagation of the severe acute respiratory syndrome (SARS) in Singapore, which is affected not only by its own evolutionary history but also by the control measures taken. Using this model, the propagation trend of the epidemic and the effects of different control measures on the outcomes of this epidemic can be simulated and quantitatively analyzed. Based on the model, we have performed an evaluation and sensitivity analysis of the Singapore governments responses to this epidemic. Preliminary results have shown that the control measures taken are effective in controlling the outbreak. Keywords: Intervention Analysis, severe acute respiratory syndrome (SARS), Modeling and Simulation. could build a model characterizing the mechanics of propagation of this epidemic. Some recent research efforts have attempted to address these issues using mathematical and statistical analytic techniques. [29]. The difficulty of accurately modeling the propagation process of SARS in Singapore using these techniques is partly due to the strict control measures are taken during the epidemic, which makes it different from a natural epidemic spread process. Obviously these measures will affect the evolution of the epidemic, but they are discrete events and cannot be directly incorporated into an epidemic propagation model. In this paper, we adopt the Intervention Analysis to model the trend of SARS propagation in Singapore. In an intervention model, the propagation process is viewed as a dynamical process with an input. The input series is an indicator variable containing discrete values that flags the occurrence of events, i.e. the control measures, which could affect the response series. So in the intervention model, the propagation of the epidemic is not only related to its history, but is also affected by the control measures. Based on this model, the trend of the propagation of this epidemic is analyzed, and the effects of different measures on the outcomes of this epidemic are simulated and evaluated.

Introduction
On 31 May 2003, Singapore was removed from the list of areas with local transmissions of the severe acute respiratory syndrome (SARS). As of 3 June 2003, using a modification of the WHO case definition, a total of 206 probable cases of SARS has been reported in Singapore [1].The relatively rapid containment of the SARS outbreak in Singapore has generally been attributed to the decisive control measures implemented by the Singapore government such as contact tracing, home quarantine policy, and close temperature monitoring; occasional skeptics, however, remain as to how much of the outcome could be attributed to pure luck or natural progression of the disease. During the fight against SARS, many important and interesting questions arised, for example: How to predict the spread and course of an epidemic with uncertain etiologies and modes of transmission, nonspecific diagnostic measures, and experimental treatment regimes? Will the current public health measures, such as isolation of SARS cases and quarantine of their asymptomatic contacts, be enough to bring SARS under control? What would be the outcome if these measures have not been taken? And how to quantitatively evaluate the effectiveness of these measures taken by the government during the SARS epidemic? Even with the epidemic over, many of these questions still puzzle the public health workers, the policy makers, and the general public. Some of these important questions can be answered if we

Methods
If we do not know the mechanics of a process, e.g. the etiological factors and propagation modes of SARS, we could view it as a black box with inputs and outputs then identifying the process in terms of control theory. But unfortunately, unlike many industrial processes, the propagation of an epidemic is unrepeatable and the data does not satisfy the identification constrains. In this case, we treat the SARS propagation process as a time series. In order to introduce the control measures into the model, instead of the pure ARIMA model, one special kind of ARIMA model with input series called intervention model [10] is used. Data pre-processing Based on the daily updates from the Ministry of Health, Singapore, the cumulative number and incidence of SARS cases starting from 16 March 2003, when the fist case of SARS was reported in Singapore, are plotted in figure 11 and 2.
1. Data and plots are available at http://www.medcomp.comp.nus.edu.sg/sars/

1246

B. Han et al.

Intervention analysis
Series 220. 200. 1 80. 1 60. 1 40. 1 20. 1 00. 80. 60. 40. 20. 0.

The propagation process of SARS can be seen as a black box with input (Xt) and output (Yt). The output, cumulative number (incidence) of SARS case is not only related to its own history (a characteristic of the epidemic) but also affected by the control measures which are the input of this dynamic process. In Singapore, the control measures took effect as of 21 March 2003 (The sixth day from starting date). We refer to these measures as interventions. Since these input interventions are continuously, i.e. taking effect from 21, March to the end of the epidemic, we model the interventions Xt as a step function as shown in (1).
0 T Xt = St = 1 t<T tT (1)

1 0

20

30

40

50

60

70

Figure 1 - Cumulative number of the SARS patient

where T is the date 21 March 2003.

Series 1 4.

Now the Intervention analysis model can be written as (2) [7]

1 2.

1 0.

Yt =

j Xt j + Nj
j=0

(2)

8.

6.

4.

2.

0.

1 0

20

30

40

50

60

70

Nt denotes the noise which constitutes an ARMA process. As in the case of transfer function modeling [7], once {Xt} has been chosen, estimation of the linear filter {vt} is simplified by approximating the operator T(B)=jBj (B is the back-shift operator) with a rational operator of the form,
B ( B) T ( B ) = ---------------------(B)
s 2

Figure 2 - Incidence of the SARS patient

(3)

1 .00

Sample ACF

1 .00

Sample PACF

( B ) = 0 1 B s B ( B ) = 1 1 B r B
r

(4)

.80

.80

.60 .40

.60 .40

(5)

.20 .00

.20 .00

In the same way, the noise can be expressed as

(B) N ( B ) = ----------- a t (B)
0 5 1 0 1 5 20 25 30 35 40

-.20

-.20

-.40 -.60

-.40 -.60

(6)
q

-.80

-.80

-1 .00

1 0

1 5

20

25

30

35

40

-1 .00

( B ) = 1 1 B q B ( B ) = 1 1 B p B

(7)

Figure 3 - Sample ACF and Sample PACF All the modeling process is based on stationary time series, therefore, firstly we should convert a time series, if not stationary, to a stationary one. Obviously, there are trends in these series. To make them stationary we have to remove the trends first. After differencing twice, we draw the plots of the Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) of the first series (cumulative number of SARS patient) in figure 3. Since the ACF of the second difference of cumulative number of SARS patient has a spike at lag 1 and falls quickly within the 95% confidence limits, we assume the new series (after twice differences) are stationary. In the same way, we get the new series for the incidence of the SARS patients after one difference. All the latter analyses are based on these stationary series.

(8)

where at is the white noise. According to (3) to (8), we rewrite equation (2) as,
(B) (B) Y t = ------------ X t b + ----------- a i (B) (B) (9)

So our objective is to find (B), (B), (B), (B) and b, where b is the time lag, i.e. the input Xt influences output Yt+b, Yt+b+1,but has no effect on Yt, Yt+1, , Yt+b-1. Examining Figure 1, we can see that there is a turning point around the sixth day after the date of measures taking effect (the sixth day from starting). Before the turning point, the patient number increases approximately exponentially; after that increases approximately logarithmically. So b may be around 6.

1247

B. Han et al.

In many practical applications, the value of r and s do not exceed 2, and s< r [11]. Therefore, it is of particular interest to examine these models i.e. b around 6, s, r<2 and s < r , in more detail. Based on the preceding analysis, we estimate the parameters using a Kalman filter algorithm [12] in the SAS Proc ARIMA.

Results
Intervention model for the SARS propagation process The final model obtained for the incidence of SARS patient is
3, 4439 + 3, 12483B ( 1 B )Y t = --------------------------------------------------- X t 6 + ( 1 0, 67853B )a t 2 1 0, 4459AB (10)

The check for white noise residuals is shown in Table 1. The 2 tests show that we cannot reject the hypothesis that the residuals are uncorrelated. Thus, we conclude that the MA(1) model is adequate for the change in incidence of SARS patient series. In the same way, we get the model for the cumulative number of SARS patient shown as (11) and the corresponding residual check shown in Table 2.
3, 4559 + 3, 13553B ( 1 B ) ( 1 B )Y t = --------------------------------------------------- X t 6 + ( 1 0, 67859B )a t 2 1 0, 44546AB

Figure 4 - Predicted incidence of SARS patient based on the previous 30 days data (+++ Predicted *** Observed ____ 95% confidence limits) In Figures 4 to 8, the reference line perpendicular to the horizontal axis on the plots indicates the date we start to forecast. On the left side of the line, the result is one-step-ahead predictions, i.e. we predict the (n+1)th value of the variable based on n available observations, on the right side of the line, the predictions are hstep prediction, h is the number of period that we look ahead, e.g. Figures 4, 5 and 6 are 95, 85 and 40 step ahead predictions respectively . From these plots we can see that the predictions fit the observations very well. Examining the models and the plots again, we find that the propagation of the epidemic is a damping process. It implies that SARS is controllable and the control measures taken by the Singapore government have been effective. Simulation and sensitivity analysis of the control measures Once we have constructed the model, we can simulation the spread of this epidemic and evaluate the effects of different measures to be taken by the policy makers. In Singapore, the measures such as home quarantine and temperature monitoring took effect from 21 March 2003 (the effect date). Here we assume three scenarios, i.e. what would happen, if no measures were taken, if the measures were taken 10 days later than the effect date and 20 days later than the effect date. Using the models we have built, we simulate the three scenarios respectively, and the results are shown in Figures 7 and 8. From Figure 7, we can see that if no measures taken, the number of case increases exponentially. If the measures were taken 10 days later, the number of cases would be doubled before going into stable state (Actually, the real situation should be even worse than that shown in Figures 7 and 8. For example, in the first scenario, we assume that no measures were taken, but our predictions have to be based on the previous 30 days real observations, in which the outcomes have already been affected by these measures.) Similar conclusions can be drawn from Figure 8.

Table 1: Check for white noise residuals for MA (1) Model of incidence of SARS patient To lag 6 12 18 24 ChiSqare 6.27 11.16 14.14 14.60 Autocorrelation Check of Residuals DF Prob>ChSq. Autocorrelations 5 11 17 23 0.2811 0.4300 0.6568 0.8374 -0.032 0.144 -0.084 0.049 -0.090 0.040 0.041 -0.029 0.026 -0.033 0.161 -0.171 -0.137 -0.035 -0.166 -0.024 0.129 -0.026 0.040 0.060 -0.100 -0.073 -0.010 0.061

Using model (10), we predict the future incidence of SARS patient based on the previous 30 and 40 day data respectively. The results are shown in Figures 4 and 5. Using model (11), we get the forecast of the future cumulative number of SARS patient based on the previous 50 days data, the result is shown in Figure 6. Table 2: Check for white noise residuals for MA (1) Model of cumulative number of SARS patient To lag 6 12 18 24 ChiSqare 6.27 11.16 14.14 14.60 Autocorrelation Check of Residuals DF Prob>ChSq. Autocorrelations 5 11 17 23 0.2812 0.4315 0.6587 0.8386 -0.032 0.144 -0.083 0.049 -0.090 0.040 0.041 -0.029 0.025 0.162 -0.136 -0.166 0.129 0.040 -0.100 -0.010 -0.033 -0.171 -0.034 -0.024 -0.027 0.059 -0.073 0.061

1248

B. Han et al.

rough comparison. And the stochastic processes suggest by Dye and Gay [7] is not included in this model.

Figure 5 - Predicted incidence of SARS patient based on the previous 40 days data (+++ Predicted *** Observed ____ 95% confidence limits)

Figure 7 - Simulated cumulative number of SARS patient based previous 30 day data (*** Observations, +++ No measure, ooo measures taking effect 20 days later, ____ measures taking effect 10 days later)

Figure 6 - Predicted cumulative number of SARS patient based on the previous 50 days data (+++ Predicted *** Observed)

Discussion and Conclusion

Christl et al [2] conducted statistical analysis to conclude that the development of the epidemic in Hong Kong featured a period of exponential growth. Their results are a good summary of the development of the epidemic, but they did not take into account the dynamics that drives the development of the epidemic. Therefore, their results cannot be used to predict the spread of the disease. Unlike the intervention model, where the impact of the control measures is directly characterized by the transfer function, Lipstch et al.[3], Riley et al[4] and Lee et al [5] [6] developed the SEIR model, in which the effectiveness of the control measure is indirectly evaluated by the case reproduction number (R0). Compared with the intervention model, the SEIR models are more complex, and you have to assume some parameters before building models, e.g. the probability of infection and incubation period. But the prediction results are worse than ours in a

Figure 8 - Simulated incidence of SARS patient based on previous 30 day data (*** Observations, +++ No measure, ooo measures taking effect 20 days later, ___ measures taking effect 10 days later) Several on-going efforts that are based on traditional mathematical and statistical modeling have been reported in a recent Workshop on SARS Modeling held in Singapore, Bruce et al. and Kamalesh et. al. [8][9] separately proposed that the branching process model was formulated to describe the spread of SARS in Singapore. But building the branching process model should know the spread history of the epidemic, say the contact map; this is easier when the epidemic spreads in a small community, but when an epidemic spreads in a large region, drawing the contact map becomes very difficult. Also, this model did not take account of the control measures that actually are unnegligible. Kamalesh et al [9] also built the auto-regressive (AR) model and got good two and three days ahead prediction results. But just as pointed out above, those models lack the analysis of the effects of control measures, so those results attribute more to the

1249

B. Han et al.

inertia of the epidemic propagation than to the model. Actually, when only the AR model is used to describe the spread of the SARS, the results for ten or twenty days ahead predictions are very poor, because the development of the epidemic is not only related to its own history but greatly affected by the control measures. In this paper, we adopted the Intervention Analysis approach to model the SARS propagation in Singapore, which is affected not only by its own history but also by the control measures. The control measures are describe as an input step function, hence in this model the evolution of the epidemic is driven by both its own history and the input. Using this model, the propagation trend of the epidemic and the effects of different control measures on the outcomes of this epidemic can be simulated and quantitatively analyzed. Based on this model, we performed an evaluation and sensitivity analysis of Singapore governments responses to this epidemic. In this work, the effects of the individual control methods are modeled together as a step function; but actually these measures may cause separate or interactive effects gradually on of the yet unknown etiology of the SARS outbreak. Our future research agenda include building more complex models to more accurately characterize the control measures by including more epidemiological and clinical knowledge about SARS as they become available. Acknowledgments
We would like thank Wong Swee Seong, Xu Song Song, Rohit Joshi, and Sreeram Ramachandran for their insightful discussions during this research. This research is supported by Research Grant No. R-252-000111-112/303 from the Agency for Science, Technology, and Research (A*Star) and the Ministry of Education in Singapore.

[8] Brown B and Cheng LS. Some Branching Process Models and the Experience of SARS in Singapore. Workshop of the Mathematics and Statistics of SARS, May 29, 2003. http:// www.ims.nus.edu.sg/activities/wksars/. [9] Kamalesh V, Kuralmani V, Goh LP et al. Statistical modeling of SARS epidemic propagation via branching processes. Workshop of the Mathematics and Statistics of SARS, May 29, 2003.http://www.ims.nus.edu.sg/activities/wksars/. [10]Brockwell PJ and Davis RA. Introduction to Time Series and Forecasting. New York: Springer Press,1996. [11]Montgomery DC, Johnson LA. and Gardiner JS. Forecasting and Time Series Analysis. 2nd ed. New York: McGrawHill Book co., 1990. [12]Morf M, Sidhu GS. and Kailath T.Some new Algorithms for Recursive Estimation on Constant Linear Discrete Time Systems. IEEE Transactins on Automatic Control 1974: 19:315-323. Correspondence:
Han Bin School of Computing,, National University of Singapore 3 Science Drive 2, Singapore 117543 Email:hanb@comp.nu.edu.sg

References
[1] http://www.moh.gov.sg/sars/. [2] Donnelly, CA, Ghani, AC, Leung, GM, et. al., Epidemiological determinants of spread of causal agent of severe acute respiratory syndrome in Hong Kong. The Lancet, published online May 7, 2003, http://image.thelancet.com/ extras/03art4453web.pdf. [3] Lipsitch M, Cohen T, Cooper B et al. Transmission Dynamics and Control of Severe Acute Respiratory Syndrome. Science 2003(published online): 300: 1966-1970. [4] Riley S, Fraser C, Donnelly CA et al. Transmission Dynamics of the Etiological Agent of SARS in Hong Kong: Impact of Public Health Intervetentions. Science 2003(published online): 300: 1961-1966. [5] Dye C and Gay N. Modeling the SARS Epidemic. Science 2003(published online): 300: 1984-1985. [6] Liang KW, Lee SL, Pan XB et al. A multistage SIR Model. Workshop of the Mathematics and Statistics of SARS, May 29, 2003. http://www.ims.nus.edu.sg/activities/wksars/. [7] Liang KW, Lee SL, Pan XB et al. SEIR Model: SARS outside the Hospitals. Workshop of the Mathematics and Statistics of SARS, May 29, 2003. http://www.ims.nus.edu.sg/ activities/wksars/.

1250