Developing a prognostic risk score for patients with newly diagnosed aggressive Non-Hodgkin Lymphoma : clinical and statistical issues
Optimisation of treatment decisions in aggressive Non-Hodgkin lymphomas (NHL) is necessarily needed as many patients still do not respond well to standard therapy — a chemotoxic regimen consisting of cyclophosphamide, doxorubicin, vincristine, prednisone, and, optionally, rituximab. Using data from the Positron Emission Tomography–Guided Therapy of Aggressive Non-Hodgkin Lymphomas (PETAL) trial, this thesis introduces three prognostic risk scores built from statistical modelling approaches and compares them with the well-established International Prognostic Index (IPI). The main clinical feature of these scores is that they in contrast to the IPI utilise positron emission tomography imaging to consider early response to therapy as assessed by the metabolic activity of the tumour after the first two cycles of chemotherapy. From statistical point of view, the three modelling approaches logistic regression, Cox regression, and the multivariable fractional polynomial time (MFPT) approach vary in their methodological assumptions. That is, in contrast to the other two, the latter technique can model non-linear relationships between prognostic variables and outcome while it also allows for the modelled effect to be time-dependent. The results of this thesis propose that early response to therapy measured as the relative reduction between pre- and post-chemotherapy standardised uptake value of a radioactive tracer does have an additional prognostic value beyond the IPI and that the modelling-based scores appear to be worth the effort as compared to simple one-point penalty scores like the IPI. With respect to the MFPT approach, the possibility to use non-linear transformations seems to improve discrimination and calibration performance while time-dependent effects appear to play a minor role in the aggressive NHL setting. The prognostic risk score obtained by the MFPT approach reveals good performance in identifying patients with an unfavourable prognosis under standard treatment and clearly outperforms the IPI. Therefore, it may be helpful in guiding decisions on treatment intensification in aggressive NHL. Nevertheless, any prognostic risk score developed in this thesis needs to be validated in an independent patient population before it shall be used in clinical routine.