Modeling Therapy Sequence for Advanced Cancer: A Microsimulation Approach Leveraging Electronic Health Record Data

Many patients with advanced cancers undergo multiple lines of treatment. Optimal therapy sequence (ie, what therapies are given and in what order) is often unknown. We developed methods for estimating quality-adjusted outcomes and cost-effectiveness of therapy sequences, informed by patient-level longitudinal data from electronic health records (EHRs).

We developed microsimulation models with a discrete-time health-state transition framework and propose 2 methods: one using multistate models to estimate transition probabilities and one using observed patient trajectories through the health states, with bootstrap resampling to estimate standard errors. We create synthetic EHR-like datasets to evaluate these methods where within-patient transition times depend on covariates and a copula generator and compare them with Markov cohort models. We provide a demonstration with treatment sequences for advanced bladder cancer (cisplatin- or carboplatin-based therapy followed by immunotherapy), incorporating external information on costs, utilities, and expected adverse events.

Both methods produced well-calibrated results, although the trajectory approach was often superior. The multistate model approach generated lower standard errors but was biased when compared with known results. The sum of squared errors for the base-case multistate model approach was 0.515 (P value for lack of fit = 0.18), whereas the trajectory approach was 0.056 (P = 0.51). The trajectory approach mostly produced confidence intervals that covered known values. In the example, both methods resulted in a net monetary benefit >0 for the cisplatin-based sequence with a willingness to pay of $100,000 per quality-adjusted life-year.

Both microsimulations produce well-calibrated results and offer superior performance to a homogeneous Markov cohort. Where available, patient data should be considered to inform cost-effectiveness models while considering the likelihood of unmeasured confounding in the observational EHR data.