gms | German Medical Science

Background: In the medical history of a cancer patient different intermediate events such as local recurrence (lr) or distant metastasis (dm) can occur which alter the prognosis. Standard survival analysis can analyse separately different endpoints but fails to give insight into what happens to a patient after a first event. A multi-state model is an extension of classical survival analysis in which all individuals start in one (or more) starting state(s) and eventually may end up in one (or more) absorbing state(s). In between, intermediate states can be visited, representing clinical events or interventions.

Materials and methods: We re-analyzed data from the RCT FOGT-2 (n=796) by using a multi-state model. The original study was set up to optimize adjuvant chemotherapy (CTx) for locally advanced rectal cancer (5-FU alone vs. 5-FU+FA vs. 5-FU+INF-a administred for 12 months). We included n=472 patients with complete set of baseline covariates (Table 1 [Tab. 1]). We designed a multi-state model with 6 states (begin of CTx, 6 months<CTx<12 months, CTx=12 months, LR, DM or both LR and DM and death). By dynamic prediction we updated the initial prediction by the incorporation of information about the arrival in other states at later timepoints.

Results: A patient with tumor stage UICC IIIc has a higher risk to develop dm or both dm and lr (state 5) than a patient in stage UICC II after the begin of CTx (state 1) and after 6 months<CTx<12 months (state 2), as well as after the completion of 12 months CTx (state 3) with HR 2.5 (95%-CI [1.2;5.3], p=0.01), HR 4.9 (95%-CI [2.1;11.5], p<0.001), HR 3.4 (95%-CI [1.8;6.2], p=0.001). This patient also has a higher risk to die (state 6) after the development of dm or both dm and lr (HR 1.7, 95%-CI [1.0;2.8], p=0.02). UICC IIIb (vs. UICC II) was a significant prognostic factor for the transition from state 3 to 5 (HR 2.4, 95%-CI [1.3;4.3], p=0.004). Patients who underwent an anterior resection have a 50% risk reduction to develop dm or both dm and lr in comparison to patients who underwent an abdominoperineal amputation (HR 0.5, 95%-CI [0.3;1.0], p=0.04). After development of lr (state 4) a woman has a 4.2 higher risk to die (state 6) in comparison to a man (HR 4.2, 95%-CI [1.5;13.9], p=0.006). Finally, bmi was also significant in the transition from begin of CTx to death with HR 1.2 (95%-CI [1.0;1.3], p=0.008). We then illustrated in a graph (Figure 1 [Fig. 1]) how the 5-year predicted survival probabilities change when a patient completed at least 6 months of CTx (black line), 12 months of CTx (red line) and experiences a dm 2 years later (green line). At this time (2 years) the patient has a 5-year survival probability of 15%. This probability would have been about 70% if at the same time point the patient who completed 12 months CTx would not have developed dm (dashed red line), but 40% if CTx would not have been fully completed (dashed black line).

Conclusion: Multi-state models are a useful tool in analyzing survival data where different intermediate events can occur. They help to gain additional insight in the complex events after start of treatment and thus to extract more information from trial data. Dynamic prediction showed how survival probabilities change by progression of the clinical history of the patient. CTx improves the survival probability by avoiding the development of DM, but when a patient develops DM survival probability is the same independently from the duration of CTx.