Hidden Markov Model (HMM) method

This function implements on-line HMM detection of outbreaks based on the retrospective procedure described in Le Strat and Carret (1999). Using the function msm (from package msm) a specified HMM is estimated, the decoding problem, i.e. the most probable state configuration, is found by the Viterbi algorithm and the most probable state of the last observation is recorded. On-line detection is performed by sequentially repeating this procedure.

Warning: This function can be very slow - a more efficient implementation would be nice!

Usage

algo.hmm(disProgObj, control = list(range=range, Mtilde=-1, 
           noStates=2, trend=TRUE, noHarmonics=1,
           covEffectEqual=FALSE, saveHMMs = FALSE, extraMSMargs=list()))

Arguments

disProgObj

object of class disProg (including the observed and the state chain)

control

control object:

range: determines the desired time points which should be evaluated. Note that opposite to other surveillance methods an initial parameter estimation occurs in the HMM. Note that range should be high enough to allow for enough reference values for estimating the HMM
Mtilde: number of observations back in time to use for fitting the HMM (including the current observation). Reasonable values are a multiple of disProgObj$freq, the default is Mtilde=-1, which means to use all possible values - for long series this might take very long time!
noStates: number of hidden states in the HMM -- the typical choice is 2. The initial rates are set such that the noStatesth state is the one having the highest rate. In other words: this state is considered the outbreak state.
trend: Boolean stating whether a linear time trend exists, i.e. if TRUE (default) then $\beta_j \neq 0$
noHarmonics: number of harmonic waves to include in the linear predictor. Default is 1.
covEffectEqual: see details
saveHMMs: Boolean, if TRUE then the fitted HMMs are saved. With this option the function can also be used to analyse data retrospectively. Default option is FALSE
extraMSMArgs: A named list with additional arguments to send to the msm HMM fitting function. Note that the msm arguments formula, data, qmatrix, hmodel, hcovariates and hconstraint are automatically filled by algo.hmm, thus these should NOT be modified.

Value

algo.hmm gives a list of class survRes which includes the vector of alarm values for every timepoint in range. No

upperbound can be specified and is put equal to zero.

The resulting object contains a list control$hmms, which contains the "msm" objects with the fitted HMMs (if saveHMMs=TRUE).

Details

For each time point t the reference values values are extracted. If the number of requested values is larger than the number of possible values the latter is used. Now the following happens on these reference values:

A noStates-State Hidden Markov Model (HMM) is used based on the Poisson distribution with linear predictor on the log-link scale. I.e. $$Y_t | X_t = j \sim Po(\mu_t^j),$$ where $$\log(\mu_t^j) = \alpha_j + \beta_j\cdot t + \sum_{i=1}^{nH} \gamma_j^i \cos(2i\pi/freq\cdot (t-1)) + \delta_j^i \sin(2i\pi/freq\cdot (t-1))$$ and $nH=$noHarmonics and $freq=12,52$ depending on the sampling frequency of the surveillance data. In the above $t-1$ is used, because the first week is always saved as t=1, i.e. we want to ensure that the first observation corresponds to cos(0) and sin(0).

If covEffectEqual then all covariate effects parameters are equal for the states, i.e. $\beta_j=\beta, \gamma_j^i=\gamma^i, \delta_j^i=\delta^i$ for all $j=1,...,\code{noStates}$.

In case more complicated HMM models are to be fitted it is possible to modify the msm code used in this function. Using e.g. AIC one can select between different models (see the msm package for further details).

Using the Viterbi algorithms the most probable state configuration is obtained for the reference values and if the most probable configuration for the last reference value (i.e. time t) equals control$noOfStates then an alarm is given.

Note: The HMM is re-fitted from scratch every time, sequential updating schemes of the HMM would increase speed considerably! A major advantage of the approach is that outbreaks in the reference values are handled automatically.

Author

M. Höhle

References

Y. Le Strat and F. Carrat, Monitoring Epidemiologic Surveillance Data using Hidden Markov Models (1999), Statistics in Medicine, 18, 3463--3478

I.L. MacDonald and W. Zucchini, Hidden Markov and Other Models for Discrete-valued Time Series, (1997), Chapman & Hall, Monographs on Statistics and applied Probability 70

Examples

#Simulate outbreak data from HMM
set.seed(123)
counts <- sim.pointSource(p = 0.98, r = 0.8, length = 3*52,
                              A = 1, alpha = 1, beta = 0, phi = 0,
                              frequency = 1, state = NULL, K = 1.5)

if (FALSE) {
#Do surveillance using a two state HMM without trend component and
#the effect of the harmonics being the same in both states. A sliding
#window of two years is used to fit the HMM
surv <- algo.hmm(counts, control=list(range=(2*52):length(counts$observed),
                                   Mtilde=2*52,noStates=2,trend=FALSE,
                                   covEffectsEqual=TRUE,extraMSMargs=list()))
plot(surv,legend.opts=list(x="topright"))
}

if (require("msm")) {
#Retrospective use of the function, i.e. monitor only the last time point
#but use option saveHMMs to store the output of the HMM fitting
surv <- algo.hmm(counts,control=list(range=length(counts$observed),Mtilde=-1,noStates=2,
                          trend=FALSE,covEffectsEqual=TRUE, saveHMMs=TRUE))

#Compute most probable state using the viterbi algorithm - 1 is "normal", 2 is "outbreak".
viterbi.msm(surv$control$hmms[[1]])$fitted

#How often correct?
tab <- cbind(truth=counts$state + 1 ,
             hmm=viterbi.msm(surv$control$hmm[[1]])$fitted)
table(tab[,1],tab[,2])
}