"There are many diseases in our environment but generally, if you can't understand a computer model of a disease, then it's probably not worth understanding."
"Black box algorithms [ie. hidden mathematical formulae] are 'black box' because no one can tell how they work. Biomodels represent something better."
Epidemiologists know that there are problems with mathematical models and the above quotes reflect some of the current issues [1] involved. One of the issues that appears to be most damaging in terms of public confidence in mathematical modelling (and for computer models of disease) is that the models fail in their predictions [2]. There are however, new advances that provide significant strides in improving the accuracy of computer models, and these involve looking more closely at the biology of disease [3]. To appreciate the latest approach, there's merit in looking briefly at the older models for comparison.
Originally when computer models were constructed, a random factor was added to make the models more realistic and the randomness created a range of possible outcomes for disease spread, simulating where the disease would spread to. However a problem centred around the level of guessing that was required to convert the biology into maths ie. incubation periods (measured in days) were guessed or averaged because incubation periods were not a single, stable or set value. Other biological parameters were also guessed or averaged, resulting in models that were frequently inaccurate and not useful [4]. Examples of the additional and variable biology that was guessed, included the infectious periods, the effects of age and herd management on disease transmission rates, the effects of pregnancy and lactation on immunity, etc., [14]. These were first generation maths models.
It's important to note that getting the first generation models to mirror past epidemics wasn't a difficulty [examples:
] because the internal workings of a model could be changed, and this would in turn adjust the model output to what was required. However, first generation models have seldom been on track in terms of predicting what would happen in the future. If first generation models are used to investigate the likely success of different disease control programmes, their problem with the predictive accuracy resurfaces ie. if the model isn't accurate in predicting the course of a disease, then its simulation of disease control programmes is unlikely to be accurate. Hence, without predictive accuracy, first generation models aren't useful in simulating the likely success of disease contingency plans. First generation models do remain useful for teaching purposes.
Second generation computer models focus upon the biology of a disease to predict the future course of a disease epidemic, and also the duration of an epidemic. This becomes possible because many diseases have a subclinical form that doesn't show clinical signs, and is therefore undetected as it spreads. A large number of individuals can become infected long before the disease is diagnosed, so effectively the disease is seeded and spread throughout a group or region before it is recognised. This seeding of subclinical disease (SCD) determines the prevalence of an outbreak (at the local level) or an epidemic (at the regional level) ie. the seeding determines the total number of individuals that will become infected. The seeding also determines the duration of an epidemic [12]. Knowing how long an epidemic will last allows administrators to decide whether control measures such as vaccination will be effective (for humans) or cost effective (for animals).
Moreover, a useful aspect of second generation models is that they can provide insights into how and when various control measures should be used against a disease. Not all individuals and groups will become infected by subclinical disease, but conversely they will show acute clinical signs that are easily recognized. Individuals and groups with acute signs of a disease should be treated quickly (ie. for humans and animals) or culled quickly (ie. for animals). Hence, the future spread of disease can be controlled [9].
The disease ratio (DR) between individuals showing mild clinical signs (which follow on from subclinical disease) and acute clinical signs, is optimally 1:1. When a disease ratio is 1:1, the predictive capabilities of a biomodel is optimally balanced alongside the disease control capabilities of the biomodel. Hence a biomodel holding a DR value of 1, permits optimal predictions to be made about the spread of a disease, whilst at the same time the disease demonstrates epidemiological characteristics that permit a rapid control of the spread.
Biomodels are reliant upon a close examination of historical data, but they are not reliant upon mathematical guesswork with respect to the biology or the epidemiology of a disease; they are equally applicable to both human and animal diseases. Biomodels are therefore representative of a second generation of mathematicallybased disease models [13].
