A multi-disease simulation swaps its road-injury module for a generative model, cutting runtime 86.6% while DALY error stays at 13%

In a multi-disease agent-based simulation, the event chain inside one disease module can be replaced wholesale by a generative model that samples outcomes directly, leaving the rest of the simulation intact. Earlier emulators needed a large sample of existing simulation outputs for training, which is prohibitively expensive for costly models, and they did not reduce the simulation's own cost. In the Thanzi La Onse model, replacing the road traffic injury module with the generative model improved mean CPU runtime by 86.6% across 50 independent runs, with a 95% bootstrap confidence interval of [85.2, 88.1]; the discussion section of the same work instead states up to 83%, so the two figures disagree.

After the swap, the distribution of yearly disability-adjusted life years (DALYs, a health-burden measure combining early death and time lived with disability) is still recovered by age and sex. Fatal road injuries are assumed to occur at only 2%, a rare outcome for a generator to capture. Emulated means fall within or at the boundary of the central 68.2% of the original distribution in all age groups for both sexes, and the mean error averaged over age, sex and year is 13%.

Interactions between diseases survive the replacement, so the individual-level perspective is preserved. Adding burden estimates from different diseases after the fact double-counts the same people's morbidity and mortality. When mortality from other causes among 50-year-old males is artificially raised 250-fold, the simulation still reproduces the expected drop in road-injury DALYs for that group.

The mechanism samples all outcomes once at onset with a conditional tabular generative adversarial network (CTGAN, a generative model that learns the joint distribution of tabular data and generates samples directly). The original model sampled and scheduled a chain of stochastic events covering symptoms, care seeking, resource access, health changes and resolution. Training data came from 53 service-configuration scenarios with 10,000 individuals each, 530,000 outcome samples in total; 43 scenarios were used for training and in-sample validation and 10 were held out for out-of-sample testing.

The demonstration covers one non-communicable, event-intensive module, road traffic injuries, in the Thanzi La Onse model; output comparisons use a simulated population of 50,000 over a 10-year period across 50 independent runs, with service provision set to Malawi's current set of all 11 road-injury services. The method applies to detachable event chains whose intermediate states affect only that agent's later events, not other agents or the environment.

The runtime gain varies by application, and road injuries are a particularly favourable test case; for infectious diseases, intermediate events such as the end of an incubation period or the timing of treatment initiation may need to be kept, and the gain falls as more events must be retained. Conditional sampling currently relies on rejection sampling, and the authors report that in a benchmark with 11 conditioning variables generating 100,000 samples, unconditional generation needed only about 6% of the CPU time, so direct conditional sampling could add further gains. The 13% error appears as a systematic upward bias, which the authors attribute to a marginal difference in the probability of death and estimate at roughly a 0.26% case-fatality discrepancy given the 2% fatality rate.