The Thanzi La Onse model – a brief introduction
We developed a model of RTIs using an IBM approach. The road traffic injury model is part of a collection of models, forming the Thanzi La Onse (TLO) multi-disease population health and health system model. Much of the RTIs model presented in this paper makes use of the existing TLO modelling framework, which we provide a brief description of here. A more in-depth description of the TLO modelling framework can be found at https://www.tlomodel.org/.
In the model, the agents are fictional representatives of individuals of the population of Malawi. Each person has a number of attributes assigned to them which represents the demographic, lifestyle and health characteristics of the population of Malawi. For example, each person has an age and sex assigned to them, a smoking and excessive alcohol consumption status and disease infection statuses. A full list of the attribute assigned to each individual related to the road traffic injury model is given in the appendix, Table A 1 The model's population go through life over a predetermined length of simulation time, over the course of which some will develop or be afflicted by diseases/injuries. Some of those with health problems will seek care at one of the model’s representations of the Malawian health system. Here the health system is represented as a group of resources available for consumption, such as staff time, bed capacity and physical consumables. A detailed description of the model’s health system can be found (see link above). Whether they seek care or not is dependent on lifestyle and demographic characteristics each person possesses and the disease/injury they are affected by, with the health seeking behaviour model being based studies of health seeking behaviour in Malawian children and adults (Ng’ambi et al. 2020; Ng’ambi et al. 2020).
Those who seek healthcare interact with the model’s representation of Malawi’s health system. The health system model provides healthcare to the population at differing facility levels, representing the variety of levels of health care such as village level health facilities, district hospitals, regional hospitals and national hospitals. Each facility level has a finite number of resources which can be used to treat patients. For many diseases modelled the interaction with the health system will in turn affect the health outcomes for the patient, this health system-epidemiology interaction is captured with the TLO modelling framework.
Road traffic injuries model
The RTI model essentially asks a series of questions of the model population each month:
- Who is injured in a road traffic accident this month?
- Did the injured persons die on the scene of the crash?
- If they didn’t die on scene, what injuries did each injured person receive?
- Did they go on to seek health care for their injuries?
- If they sought health care for their injuries, what do they need from the health system for their treatment?
- Based on their choice to seek or not seek health care, what health outcomes did each injured person experience (mortality, morbidity or recovery)?
In an attempt to make simplifications to the model which would not affect the overall level of health burden and health system usage, we do not explicitly model individual vehicle collisions. This means that we do not for account for properties of the physical world in our model (e.g. speed, road surface, vehicle type, number of vehicles in a crash, the number of people in an individual crash or the relationship between people involved in crashes).
A schematic of the model is given in Figure 1 and we go through each question (Q1-Q6) asked by the model in turn below. A full list of parameters is provided in the Appendix. A detailed explanation of the RTI model including the code used in the model can be found on the TLO website.
Figure 1: Road traffic injury model diagram, Q1 through 6 related to the questions asked by the model to determine the health burden and corresponding health system usage caused by RTIs in Malawi. Owing to the large number parameters used to determine the answer to these questions, we have listed the parameters in the appendix in Table A 2, organised by the question asked.
Q1 Who is injured in a road traffic accident this month?
To determine who will be injured in a road traffic collision in the model, we assume that the members of the model’s population are involved in a road traffic collision at fixed rate ‘base_rate_injrti’. We then increase the likelihood of males, those in certain age groups and those who consume excessive alcohol being injured in a road accident. We calibrated the value of the parameter ‘base_rate_injrti’ to produce an incidence of RTIs matching the ten-year average of the GBD’s predicted incidence of RTI. We calibrated the parameter ‘rr_injrti_male’ to the GBD’s estimated gender ratio. The several age-related parameters were also calibrated to the GBD estimated age distribution of those with RTIs. Another risk factor for RTIs is alcohol consumption, for example, roughly 25% of the road traffic injury patients treated at Kamuzu Central Hospital either tested positive for alcohol or reported using alcohol before their injuries (Sundet et al. 2020). We used the relative risk of alcohol consumption in a Tanzanian study to parameterise ‘rr_injrti_excessalcohol’ (Staton et al. 2018).
Q2 Did the injured persons die on the scene of the crash?
Some RTIs will invariably be fatal. We assume that pre-hospital mortality occurs in a fixed proportion of those involved in road traffic accidents before the allocation of injuries. The parameter value of ‘imm_death_proportion_rti’ was calibrated to the incidence of on scene mortality reported by Malawi’s police (Schlottmann et al. 2017).
Q3 What injuries did each injured person receive?
In determining the health outcomes for RTIs, the number of injuries, anatomic location of the injury and type of injury have been shown to be important factors in determining mortality and morbidity (Gabbe et al. 2014). The injuries each person receives will also determine what they require from the health system for their treatment. To decide the exact injuries each injured person has, the model assigns injuries in a three-step process. The model determines how many injuries each person has, where the injuries are anatomically located on the body, and based on the anatomic location, what these injuries are. When designing the injury assigning section of the model, we limited the injuries that were assigned to those which would warrant some form of healthcare. This approach was chosen as some of the injuries received from road accidents will be minor and not produce a significant health burden, an assumption shared with the GBD study (Haagsma et al. 2016).
To determine the number of injuries assigned to each injured person, we developed a negative exponential distribution and calibrated our resulting average number of injuries reported in a paediatric study from Malawi’s Kamuzu Central Hospital (Sundet et al. 2018). Our process of developing the distribution is given in the appendix. The resulting percentages of single injuries in those with RTIs ranged from 71% to 76%. This falls within the ranges reported in several studies from Sub Saharan Africa (SSA), with the percentage of single injuries in RTI patients ranging from 66% to 81% (Akinpelu et al. 2007; Ganveer and Tiwari 2005; Madubueze et al. 2010; Sanyang et al. 2017; Thanni and Kehinde 2006).
The anatomic location of the injuries is determined by the average distribution of anatomic injury location found in (Otieno et al. 2004; Ranti et al. 2015), and finally the exact injury each person receives is informed by several studies. The numerous parameter values used to determine which of the 76 injuries accounted for by the model each person has are given in appendix.
Q4 Did they go on to seek health care for their injuries?
The model’s predicted health seeking behaviour (HSB) is determined in part by the results of (Ng’ambi et al. 2020; Ng’ambi et al. 2020). The HSB model uses demographic and lifestyle factors, along with the symptoms the person has as a result of their injury/illness to determine HSB. Ng’ambi et al. did not specifically focus on RTIs, instead focusing on injuries in general. We assume that at a certain level of injury, people will always seek healthcare for their injuries, we used the injury severity score (ISS) (Baker et al. 1974), to determine this threshold. We assume that there exists a level of injury severity which will cause a person to automatically seek healthcare, below this level of severity, HSB is determined by the results of Ng’ambi et al. (2020). This severity level is determined by the parameter ‘rt_emergency_care_ISS_score_cut_off’. We established a parameter space which produced an overall level of HSB for RTIs which fell in the bounds reported in other SSA countries (Zafar et al. 2018).
Q5 If they sought health care for their injuries, what do they need from the health system for their treatment?
To best represent the provision of treatment for RTIs in Malawi, we based the provided treatments on those that are described in the Malawian treatment guidelines (Ministry of Health 2015), Malawi’s essential health package (EHP) or treatments that have been reported in academic literature. To simplify the model, we assume that if a treatment is described in the treatment guidelines or Malawi-based academic literature then the treatment is available at local hospitals.
The treatment plan for some injuries must be determined on an individual basis. An example of this is the use of several treatment methods to treat lower extremity fractures in Malawi (Chagomerana et al. 2017). We account for intricate differences in potential treatment plans with fixed probabilities that certain treatment options will be used for each patient. A full list of treatments being provided by the model and references to show evidence they are used in Malawi, is given in the appendix (Table A 3).
Q6 Based on their choice to seek or not seek health care, what health outcomes did each injured person experience (mortality, morbidity or recovery)?
We used the severity of a person’s injuries to determine the mortality with and without seeking healthcare. To quantify the health burden of a person’s injuries, we use a number of commonly used injury severity metrics. To quantify the severity of singular injuries we used the abbreviated injury score (AIS) (Gennarelli and Wodzin 2006), using the R library ‘InjurySeverityScore’ to convert ICD-9 diagnosis codes to a corresponding AIS score (D. Tian 2019). To quantify the severity of multiple injuries, we used the injury severity score (ISS) (Baker et al. 1974), which makes use of the AIS score of the person’s injuries. We also used the military abbreviated injury score (MAIS) to quantify the severity of injuries and used this to predict the probability of mortality without medical intervention (Champion et al. 2010). The disability burden posed by RTIs was quantified using DALY weights. Each injury has a corresponding DALY weight, sourced from the GBD study (Salomon et al. 2015), see appendix. Where the GBD studies DALY weights was too broad or had missing injuries we used another source (Gabbe et al. 2016).
For those who seek healthcare and receive treatment (conditional on its availability), we use the ISS score to determine mortality. There was limited information of the probability of death based on the ISS score in Malawi or elsewhere in Africa, as such we used results from a non-African study to establish a relationship between ISS score and mortality. We used the same score boundaries reported in the study (Kuwabara et al. 2010), but scaled the reported probability of mortality in each ISS score boundary so that the overall in-hospital mortality predicted by the model matched the overall in-hospital mortality reported in a national-scale Tanzanian study (Sawe et al. 2021).
For those who did not seek healthcare and when healthcare was not available, we assume that mortality is only considered for those with an injury above a certain threshold. In these circumstances we use their MAIS score to determine mortality. This assumption that the MAIS can be used to predict mortality without healthcare is tenuous, however information is limited for the probability of mortality from injuries without medical intervention and this was the one of the few mortality-predicting scoring systems in a setting with limited health care provision. The person’s MAIS score corresponded with a probability of mortality, if the person hadn’t sought care after a week of model simulation time, the model used this probability to determine whether they had died from their injuries.
For morbidity, we used DALY weights to quantify health burden from each injury. This health burden was applied to the person when the person’s injuries were assigned. Once a person had sought healthcare, after a period of recovery time post treatment this DALY weight was removed (if applicable as some injuries have an associated long-term DALY weight). For those who didn’t seek healthcare and survived their injuries, we assumed that the DALY weight associated with their injuries would still be removed, but after a longer duration than if they had sought healthcare for their injuries. The duration of time which a person experiences a health burden associated with their injuries is dependent on the injuries sustained, a full list of the assumed heal time associated with each injury is given in the appendix.
A full list of injuries modelled, the health burden associated with the injury, the treatment used, properties of the population modelled and parameters used in the model and their source/calibration process is given in the appendix.
Calibration
To calibrate the sections of our model described above, we created multiple scenarios of road traffic injury epidemics. Within each scenario a particular parameter was incrementally changed, producing a change in the model’s behaviour in an area we desired to calibrate. For example, a change in the parameter ‘rr_injrti_male’ would either increase or decrease the relative risk of being injured in a road collision if male and would in turn change the overall percentage of those in RTIs who are male.
For each scenario and associated parameter value, we ran the model multiple times over ten years of simulation time. The average results of the runs associated with each scenario are an indication of the model’s behaviours for that parameter value. By incrementally changing the parameter values used in the model and taking the average model output per parameter value, we found parameter values which produced model output that matched our various calibration targets, stated in the previous section.
Some parameters were independent of other model parameters, and as such we could find a single parameter value which produced our calibration target. Other parts of the model were interdependent and required scenarios where multiple parameters were changed in combination with one another to find parameter value which produced calibration targets in each model area.
For some sections of the model, we had no particular estimate to calibrate the model to. For example, we had a range of values for the overall percentage of HSB, rather than a specific percentage of HSB to calibrate to. In this case, we established a parameter space for the parameter ‘rt_emergency_care_ISS_score_cut_off’ which produced an overall level of HSB which fell within our calibration targets.
Accounting for uncertainty
There is an inherent uncertainty in IBM that is caused by the stochasticity between runs. Each model run is unique, meaning that in order to produce clearer picture of the model’s behaviour, we have to run the model multiple times, taking the average of the results produced in each run. In each scenario we ran the model with a population of 20,000 for ten years of simulation time, running the simulation 4 times each. When running individual model runs over a 10-year period with a population of 20,000, all individual model runs produced an average incidence of RTIs which fell within the 95% confidence interval predicted by the GBD study. We performed 4 model runs per scenario as this managed computational time whilst producing results which consistently fell within our calibration targets (see appendix).
Comparison between single injury and multiple injury RTI models.
To investigate the effect of single and multiple injuries in other areas of population health, we compare two forms of our model. One where we only give those injured in road accidents single injuries and one where we give out multiple injuries. In both forms of the model, the incidence of RTI in the population is calibrated to the average incidence of RTI predicted by GBD for Malawi in 2010-2019. By comparing the results of the two models we see the influence of considering multiple injuries on the health outcomes for RTIs.
Estimating the reduction in health burden attributed to the health system
To demonstrate the usefulness of explicitly modelling the health system, we investigate the reduction in harm to population health caused by the health system. We ran the model as normal, with health care being provided by the health system; then ran the model without health care being provided by the health system (effectively all injured persons going through the ‘didn’t seek care’ route of the model, see Figure 1). We compared the results of each scenario to find the reduction in DALYs and deaths attributed to the health system.