A Comparison of Different Models to Recommend Variables for Research Data Requests to German Cancer Registries

Cancer registry data is an important resource in cancer research, e.g. to formulate hypotheses for new interventions, to identify recruitment potentials for clinical trials or to evaluate side effects of new interventions on a large population, but requesting cancer registry data is an error-prone process. A system utilizing machine learning could assist cancer researchers with their data request, but it is unclear which machine learning method is best suited. In this paper, we will compare seven different machine learning models for variable recommendation for research data requests to German cancer registries. We collected data from five German cancer registries containing descriptions of and selected variables for analyses done by the cancer registry. Model training uses nested cross-validation to mitigate poor hyperparameter selection. Candidates for machine learning models are a neural network and variants suitable for multi-label classification of decision trees, random forests, and gradient boosting. The neural network performed best on the per-label F2 score with almost 0.3 while a gradient boosting variant performed best on the per-sample F2 score with 0.82. A random forest variant balances both per-sample and per-label F2 score with 0.78 and 0.26 respectively. The other four models were outperformed by those three models. Based on the comparison, the random forest seems to be best suited, if a single model is chosen from the candidates. An ensemble built from neural networks and gradient boosting models might outperform the random forest, though this adds complexity.