Application Details
The purpose of this application is to facilitate the interpretation of PSA kinetics in patients on active surveillance, in order to improve decision making with respect to the decision to maintain a patient on surveillance or to intervene with definitive local treatment. We have modeled the PSA kinetics in a large cohort of patients on active surveillance, using the General Linear Mixed Model (GLMM). This model uses the PSA variation of the entire cohort to reduce the error in calculating the PSA doubling time (DT) in an individual patient. It also compensates for the effect of baseline PSA on the PSA DT. The application allows the clinician to compare a patient's PSA kinetics with a stable cohort and a cohort who demonstrated biochemical or pathologic progression over time, and determine which phenotype his patient most closely resembles.
Modelling PSA kinetics with GLMM approach
Among 231 patients in the surveillance study (Zhang et al., 2006), 93 patients (40%) fulfilled the criteria for high risk for disease progression, and 138 patients (60%) were categorized as low risk at the time of last follow-up. Serial PSA measurements, before any clinical therapeutic intervention, were used to estimate the evolution of PSA over time in high risk and low risk patients. Using the GLMM approach, a quadratic model (fixed effect) with random intercept (PSA at baseline) and random linear slope was needed to adequately capture PSA evolution in the group of patients at high risk. Similarly, in the group of patients at low risk, both random intercept and random linear slope were required in the quadratic model. Moreover, two additional baseline covariates were significantly associated with the evolution of PSA levels in the group of patients at low risk, namely age and Gleason score.
High risk:
Low risk:
In the group of patients at high risk, the corresponding average of PSA DT was calculated as 2.97 years (95% confidence interval [CI]: 2.24-4.41 years). In the group of patients at low risk, the corresponding average of PSA DT was 6.54 years (95% CI: 4.81-12.3 years). Patients at high risk for progression have much shorter periods of PSA DT, on average, than patients at low risk.
From GLMM modeling, the 'high risk' and 'low risk' lines are generated for an individual patient given his baseline covariates. Observed PSA values are then plotted against two lines. Any "marked" trend above the high risk line would alert the physician to consider definitive intervention. For the 'typical' patient, definitive therapy should be recommended if more than 40% of the observed points lie above the predicted high risk line. By contrast, less stringent monitoring should be recommended if more than 35% of the observed points lie below the predicted low risk line.
By applying the dynamic prognostic rule to all patients, a decision to initiate the intervention was optimally made about 2.3 years (8 measurements) after surveillance was initiated. Importantly, this is a small proportion of the 10-12 year lead-time estimated to occur with PSA screening. A delay in treatment of 2.3 years is unlikely to alter the outcome in more than a very small proportion of patients. Our estimate is that, amongst a favorable risk population managed with active surveillance and selective delayed intervention, 1-1.5% of patients may suffer a preventable prostate cancer death. Since the disease is typically slow growing, those deaths will, in most cases, come near the end of the patient's natural life.
Using this strategy, in combination with serial biopsy, about 70% of patients will avoid radical therapy. The remainders are unlikely to have their opportunity for cure adversely affected. This approach is currently being validated in a prospective randomized trial, the 'START' trial.
