Predicting outcomes and estimating treatment effects among patients with lymphoma are subjects of interest to a broad range of healthcare professionals. New real-world evidence can be obtained by combining the increasing amount of real-world data available from the Nordic countries with advanced statistical methods. The overall aims of this thesis were to develop new prognostic tools that can provide more accurate prognosis at the time of diagnosis and to estimate treatment effectiveness among patients in a real-world setting. Currently available prognostic models are often simplistic and were built based on individuals receiving outdated treatment; therefore, there is a need for continuous effort to update these models. Moreover, estimating treatment effectiveness is important, especially for patients who is unlikely to participate in a randomized trial.

In the first paper, two predictive machine learning models were developed to predict overall survival and progression-free survival in patients with advanced stage classical Hodgkin lymphoma. Notably, the machine learning models outperformed existing prognostic indices upon both internal and external validation but yielded only slightly higher performance measures compared to the newly developed Advanced Stage Hodgkin Lymphoma International Prognostic Index (A-HIPI) model.

In the second study, young patients with diffuse large B-cell lymphoma (DLBCL) who received R-CHOEP were matched 1:1 to patients receiving RCHOP who were similar in a large range of clinical characteristics. This study found no evidence of R-CHOEP having a superior effect over R-CHOP in terms of overall survival or progression-free survival. Analysis of the late effects among patients achieving complete remission revealed no evidence of an increased risk of cardiovascular disease or secondary primary malignancies in patients treated with R-CHOEP versus R-CHOP.

Protecting patients' right to privacy is important. But there is a risk of re-identifying individuals from predictive models. To address this subject, a new method was introduced for training survival models across sites using federated learning within the third study. The methods utilized gradients and Hessians to estimate covariate effects and included a kernel smoothing process for a Cox proportional hazard model. The novel proposed methods ensuring higher protecting of patient data and achieved similar performance measures compared to the case where data are pooled at a single site, using both simulated data and real-world data.