IV. Discussion
When modelling the course of disease it is important to consider as many disease-specific risk factors as possible in order to provide an informed view of outcomes that may occur. The word 'may' is important because no model can predict any outcome with 100% accuracy. Modelling techniques are evolving in response to criticism aimed at improving their predictive abilities. Discrete event simulation further contributes to the field. While Markov models have served – and continue to serve – the scientific and decision-making communities well, we are of the opinion that DES also offers additional possibilities for modelling patients with depression and their progression through the healthcare system. A frequent criticism of depression models is that they are often too short and, therefore, unable to accurately reflect the true progression of the disorder. In the present case, the DES timeframe was large enough to capture all events occurring during the disease span and even beyond (periods of recovery). As such, distinction between relapses and recurrences (according to whether the patient is experiencing a new episode of depression or not) were shown to be important issues to be taken into consideration in order to more precisely assess the capacity of a given strategy to delay further risk of developing depressive symptoms. The number of previous depressive events, their duration and severity together with patients' adherence to therapy were also proven to be key factors that needed be taken into account in the computational framework.
Despite their lack of memory, Markov models managed to handle the problem of patient history by specifying various health states defined according to risk levels (low, moderate, high risk of relapse). The issue of disease persistence was also addressed, but at the expense of defining multiple temporary health states. A major drawback, however, persists in the handling of multiple events. If analysts truly seek to portray reality as closely as possible, they should consider scenarios that are more complicated and that take into account, for example, suicidal behaviour and patients' attitudes towards treatment. In such situations they may be more likely to employ more elaborate modelling methods such as discrete event simulation.
Health service research in general, and economic evaluation in particular, is commonly associated with a lack of adequate data. In the present case, the intention was to validate the use of a DES model in a conceptual manner, i.e. in terms of its computational validity. To numerically quantify the benefits of DES over Markov models, empirical validation is required. Future research would necessitate comparing simulated results with those obtained from observational data (i.e. perform an external validation). Further research should focus on both the internal and external validity of the conceptual model [47], by collecting adequate data (after systematic review of the literature), then choosing appropriate statistical distributions on parameters, and finally by calibrating the model with reference to results obtained from naturalistic studies. However, when reliable data is not available, DES may be a highly suitable information system that could be used to run a series of different "what if?" scenarios, allowing the user to understand the interaction of the model parameters, and their effects on the output of interest. For example, in the context of exploratory analyses whose purpose would be to identify preferred health outcomes for inclusion alongside a clinical trial, DES may help in defining the requirements of a definitive economic analysis and determine a data collection strategy.
There are certain limits to DES that deserve to be pointed out. First, greater flexibility may only be reached at the expense of supplementary specialist analytic knowledge, which may reduce the evaluator's direct access to the model. Also, it may take time to develop, implement and verify the conceptual model. Moreover, individual-based models like DES models are highly time-consuming, as multiple replications are needed to get good estimates of mean effects. However, variance reduction methods are available that can reduce the number of replications and time needed [48]. Finally, DES may induce over-specification, whereby possible patient pathways become more complex than necessary, thus implying an increase in data requirements.
We deliberately chose here to focus on the methodological aspects of the modelling methods, regardless of the therapeutic strategies and without any costing purpose. However, costing would be equally feasible in both methods: DES models would use variables associated to each event experienced, while Markov models would associate a monetary value to health states.
In order to provide decision makers with a fully specified tool aimed at prioritizing actions for relapse prevention in depression, further work should incorporate, in the form of a DES model, both clinical and economic data in accordance with national and international clinical and pharmacoeconomic guidelines. A practical example of discrete event simulation model for depression, together with judicious distribution choice on parameters (among Weibull, Log-logistic and more generally Gamma distributions) will be the next step of this research, with an aim towards benchmarking results from a DES model with those from standard simulation models.
Answers to all your Biology Questions
