--- title: "denim: deterministic model with memory" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{denim: deterministic model with memory} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r setup, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", fig.align="center" ) library(DiagrammeR) # for flowchart diagram library(denim) ``` ## 1. Simple SIR model with gamma distributed lengths of stay The SIR model uses 3 compartments: S (susceptible), I (infected), R (recovered) to describe clinical status of individuals. We use the simplest form of SIR model to demonstrate how to define the distribution of the lengths of stay distribution.
![](./images/denim_website-getting-started-basic.png){width="400"}
The model equations are: $$\frac{dS(t)}{dt} = -\lambda S(t) = -\frac{\beta I(t)}{N}S(t)$$ $$\frac{dI(t)}{dt} = \frac{\beta I(t)}{N}S - \gamma I(t)$$ $$\frac{dR(t)}{dt} = \gamma I(t)$$ - $N$: total population size, $N = S + I + R$ - $\beta$: the product of contact rates and transmission probability; usually we define $\lambda =\frac{\beta I(t)}{N}$ as the force of infection - $\gamma$: recovery rate For mathematical simplicity, modelers typically assume that the recovery rate $\gamma$ is constant, this will leads to an exponentially distributed length of stay i.e most individuals recover after 1 day being infected. ```{r, echo=FALSE} rates <- c(0.5, 1, 1.5) x <- seq(0, 5, 0.001) y <- dexp(x = x, rate = rates[1]) y2 <- dexp(x = x, rate = rates[2]) y3 <- dexp(x = x, rate = rates[3]) col_codes <- c("#374F77", "#EF9F32", "#6ECBD7") plot(x, y, type = "l", col = col_codes[1], lty = 1, lwd = 3, xlab = "Length of stay (days)", ylab = "", ylim = c(0, 1.5), yaxt = 'n') lines(x, y2, col = col_codes[2], lty = 1, lwd = 3) lines(x, y3, col = col_codes[3], lty = 1, lwd = 3) legend("right", legend = c(0.5, 1.0, 1.5), col = col_codes, lty = 1, lwd = 3, bty = "\n") ``` A more realistic length of stay distribution can look like this, of which most patients recovered after 4 days. We defined this using a gamma distribution with shape = 3 and rate = 1/2. ```{r, echo=FALSE} x <- seq(0, 20, 0.001) y <- dgamma(x = x, shape = 3, rate = 1/2) plot(x, y, type = "l", col = col_codes[1], lty = 1, lwd = 3, xlab = "Length of stay (days)", ylab = "", yaxt = 'n') ``` The following diagram demonstrates the model with a more realistic dwell time. The dashed arrow indicates transition described by dwell time distribution, while the solid arrow indicates transition rate.
![](./images/denim_website-getting-started-sir.png){width="400"}
**Model specification** *Model transition* We have two transitions `S -> I` and `I -> R` in this case. The transitions are specified in a list follow this format `"transition" = expression`, of which expression is defined with one of our functions for waiting time distribution [or]{.underline} a mathematical expression. ```{r} transitions <- list( "S -> I" = "beta * S * I / N", "I -> R" = d_gamma(rate = 1/2, 3) ) ``` Another option to define the transitions is by using denim's DSL. Refer to the [Model definition in denim](./denim_dsl.html) article for more information. ```{r} transitions <- denim_dsl({ S -> I = beta * (I/N) * S I -> R = d_gamma(rate = 1/2, shape = 3) }) ``` *Initial state* Use a vector to define the compartments with their assigned names and initial values in the format `compartment_name = initial_value`: ```{r} initialValues <- c( S = 999, I = 1, R = 0 ) ``` *Model parameters* If we use a math expression, any symbols except the compartment names are parameters, and would be defined by constant values. There are two constant parameters in our example: `beta` and `N`: ```{r} parameters <- c( beta = 1.2, N = 1000 ) ``` **Model application** *Time step specification* We run the model for 30 days and give output at 0.01 daily intervals. The default interval (time step) is 1 if not declared explicitly. ```{r} simulationDuration <- 30 timeStep <- 0.01 ``` ```{r, fig.width = 6} mod <- sim(transitions = transitions, initialValues = initialValues, parameters = parameters, simulationDuration = simulationDuration, timeStep = timeStep) head(mod) plot(mod) ``` ## 2. How the algorithm work? The central idea of our algorithm is to split the compartment into a chain of sub-compartments, allowing more control over the dwell time distribution. The number of sub-compartments depends on the [maximum duration]{.underline} an individual can stay in the compartment and the [duration of time step]{.underline}. At each time step, we calculate the out-going population from each sub-compartment based on the provided dwell-time distribution. For example, if we expect a disease with a maximum 4 days of infection, and the duration of time step is 1 day, we will end up with 4 sub-compartments $I_1, I_2, I_3, I_4$. At each time step, a proportion $q_i$ of individuals in $I_i$, where $i<4$ will move to $R$ while the remaining $1-q_i$ will move to $I_{i+1}$. Assuming we want the infectious period to follow a discrete distribution $P_I = \{p_1, p_2, p_3, p_4\}$. Our goal is to estimate $q_{1}$, $q_{2}$, $q_{3}$ such that the resulting dwell time distribution matches $P_I$. This can be obtained by setting: $$q_{i} = \frac{p_i}{1 - \sum_{j=1}^{i-1}p_j}$$ The following diagram illustrates the flow between sub-compartments. Note that all individuals in $I_4$ will transition to $R$ (i.e., $q_4 = 1$)
```{r, echo=FALSE} # DiagrammeR::grViz("digraph { # graph [layout = dot, rankdir = LR] # # node [shape = rectangle] # # S -> I1 [label = 'βS(I@_{1} + I@_{2} + I@_{3} + I@_{4})/N'] # I1 -> R1 [label = 'γ@_{1}I@_{1}'] # I2 -> R2 [label = 'γ@_{2}I@_{2}'] # I3 -> R3 [label = 'γ@_{3}I@_{3}'] # I4 -> R4 [label = 'I@_{4}'] # I1 -> I2 [label = '(1-γ@_{1})I@_{1}'] # I2 -> I3 [label = '(1-γ@_{2})I@_{2}'] # I3 -> I4 [label = '(1-γ@_{3})I@_{3}'] # }", height = "100%", width = "100%") ``` ![](./images/denim_website-getting-started-denim-sir.png){width="700"}
For a given dwell-time distribution, we identify the maximum length of stay using its cumulative distribution function. Because cumulative distribution function is asymptotic to 1, we need to set a value that is acceptable to be rounded to 1. If we want a cumulative probability of 0.999 to be rounded as 1, we set the error tolerance threshold as `1 - 0.999 = 0.001` (specified by the argument `errorTolerance = 0.001`). The time when cumulative probability = 0.999 will be set as the maximum length of stay of the compartment. Default `errorTolerance` of `denim` is set at `0.001`. *Initialize population in sub-compartments* By default, the initial population is always assigned to the first sub-compartment (for example, if the initial value for I compartment is `I = 1`, denim will initialize `I1 = 1` while `I2 = I3 = I4 = 0`). User can also choose to distribute initial population across sub-compartments based on the specified distribution, i.e. with initial population `I = n`, then denim will initialize `I1 = n*p1`, `I2 = n*p2`, `I3 = n*p3`, `I4 = n*p4`. To set up the initial population in this way, simply specify the parameter `dist_init = TRUE` (only applicable for distribution transitions). ```{r eval=FALSE} transitions <- denim_dsl({ S -> I = beta * S * I / N I -> R = d_gamma(rate = 1/2, 3, dist_init=TRUE) }) ``` ## 3. Waiting time distribution `denim` offers 2 ways to define a transition: either by a waiting time distribution in the origin compartment, or by a mathematical expression of the outgoing rate of the origin compartment. Current available distributions in this package including: - `d_exponential(rate)`: Discrete **exponential distribution** with parameter `rate` - `d_gamma(rate, shape)`: Discrete **gamma distribution** with parameters `rate` and `shape` - `d_weibull(scale, shape)`: Discrete **Weibull distribution** with parameters `scale` and `shape` - `d_lognormal(mu, sigma)`: Discrete **log-normal distribution** with parameters `mu` and `sigma` - `nonparametric(waitingTimes)`: A **vector of values**, could be numbers, percentages, density of the length of stay based on real data, `denim` will convert it into a distribution. These values should: (i) corresponding to 1 time step, (ii) be chronologically ordered according to the dwell-time. Mathematical expression: Transition can also be described using a math expression such as `beta * S * I / N`. You will need to define parameters that are not compartment names in the `parameters` argument ## 4. Multiple transitions from a compartment In `denim`, transitions between one compartment to multiple compartments are modeled as either (i) multinomial transition or (ii) competing risks. Consider this example:
![](./images/denim_website-getting-started-multiple-transitions.png){width="500"}
There are two scenarios in this example: - Susceptible individuals can be infected or vaccinated. The assumption here is they will be infected first (`S -> I`), and then the rest of them who were not infected will get vaccinated (`S -> V`). - Infected individuals can recover or die. If the mortality probability is known, we can implement it into the model. For example, here, we specify that 90% of the infected individuals will recover (sub-population $I_R$) and 10% will die (sub-population $I_D$). By doing so, we ensure that the mortality probability is 10%, while also define the length of stay of individuals at the infected state before recover or die follows gamma or log-normal distribution, respectively. We can define the model for this example as follows: ```{r, fig.width = 5} transitions <- denim_dsl({ S -> I = beta * S * I / N S -> V = 5 0.9 * I -> R = d_gamma(1/3, 2) 0.1 * I -> D = d_lognormal(2, 0.5) }) initialValues <- c( S = 999, I = 1, R = 0, V = 0, D = 0 ) parameters <- c( beta = 1.2, N = 1000 ) simulationDuration <- 20 timeStep <- 0.01 mod <- sim(transitions = transitions, initialValues = initialValues, parameters = parameters, simulationDuration = simulationDuration, timeStep = timeStep) head(mod) plot(mod, ylim = c(0, 1000)) ``` For a more detailed explanation for transitions to multiple states, refer to [this article](./multinomial.html) ## 5. Additional examples ### Example 1
![](./images/denim_website-getting-started-extra-1.png){width="700"}
```{r, fig.width = 5} initialValues <- c(S = 999, I = 1, R = 0, V = 0, IV = 0, D = 0) modelStructure <- denim_dsl({ S -> I = beta * S * (I + IV) / N S -> V = d_exponential(0.01) 0.1 * I -> D = d_lognormal(2, 0.5) 0.9 * I -> R = d_gamma(1/3, 2) V -> IV = 0.2 * beta * V * (I + IV) / N IV -> R = nonparametric(iv_r_dist) IV -> D = d_weibull(scale = 2, shape = 1.5) }) parameters <- list( beta = 0.9, N = 1000, iv_r_dist = c(0, 0.15, 0.15, 0.05, 0.2, 0.2, 0.25) ) simulation <- sim(transitions = modelStructure, initialValues = initialValues, parameters = parameters, simulationDuration = 30, timeStep = 0.5) plot(simulation) head(simulation) ``` ### Example 2 ![](./images/denim_website-getting-started-extra2.png){width="500"} ```{r echo=FALSE} # ------ Code to generate the distribution to demonstrate nonparametric --------- # Helper function compute_dist <- function(dist_func,..., timestep=0.05, error_tolerance=0.0001){ maxtime <- timestep prev_prob <- 0 prob_dist <- numeric() while(TRUE){ # get current cumulative prob and check whether it is sufficiently close to 1 temp_prob <- ifelse( dist_func(maxtime, ...) < (1 - error_tolerance), dist_func(maxtime, ...), 1); # get f(t) curr_prob <- temp_prob - prev_prob prob_dist <- c(prob_dist, curr_prob) prev_prob <- temp_prob maxtime <- maxtime + timestep if(temp_prob == 1){ break } } prob_dist } timestep <- 0.05 # create a multimodal first_dist <- compute_dist(pweibull, scale = 5, shape = 5, timestep = timestep) second_dist <- compute_dist(pweibull, scale = 2.5, shape = 4, timestep = timestep) second_dist <- c(second_dist, rep(0, length(first_dist) - length(second_dist))) ir_dist <- first_dist + second_dist ``` Assuming we have the distribution of time for the infected to recover ($P_{IR}$) as followed ```{r} timestep <- 0.05 plot(seq(0, by = 0.05, length.out = length(ir_dist)), ir_dist, type = "l", col = "#374F77", lty = 1, lwd = 3, xlab = "Length of stay (days)", ylab = "", yaxt = 'n') ``` This distribution can be used directly for modeling with the `nonparametric()` function. Note that when using `nonparametric()`, we must define the simulation `timeStep` to match the time resolution of the distribution data. For example, the provided distribution is given with the interval of 0.05 day, we must set `timeStep = 0.05`. ```{r} transitions <- denim_dsl({ S -> I = beta*I/N*S I -> R = nonparametric(ir_dist) I -> D = d_exponential(d_rate) }) parameters <- list( beta = 0.7, ir_dist = ir_dist, d_rate = 0.1, N = 1000 ) initialValues <- c( S = 999, I = 1, R = 0, D = 0 ) simulationDuration <- 30 timeStep <- 0.05 mod <- sim(transitions = transitions, initialValues = initialValues, parameters = parameters, simulationDuration = simulationDuration, timeStep = timeStep) plot(mod, ylim = c(0, 1000)) ```