PanelomiX: A threshold-based algorithm to create panels of biomarkers

Abstract

In order to increase their predictive power, medical biomarkers can be combined into panels. However, the lack of ready-to-use tools generating interpretable results and implementing rigorous validation standards hampers the more widespread application of panels and their translation into clinical practice. The computational toolbox we present here – PanelomiX – uses the iterative combination of biomarkers and thresholds (ICBT) method. This method combines biomarkers and clinical scores by selecting thresholds that provide optimal classification performance. To speed up the calculation for a large number of biomarkers, PanelomiX selects a subset of thresholds and parameters based on the random forest method. The panels’ robustness and performance are analysed by cross-validation (CV) and receiver operating characteristic (ROC) analysis. Using 8 biomarkers, we compared this method against classic combination procedures in the determination of outcome for 113 patients with an aneurysmal subarachnoid haemorrhage. The panel classified the patients better than the best single biomarker (p < 0.005) and compared favourably with other off-the-shelf classification methods. In conclusion, the PanelomiX toolbox combines biomarkers and evaluates the performance of panels to classify patients better than single markers or other classifiers. The ICBT algorithm proved to be an efficient classifier, the results of which can easily be interpreted.

Ähnliche Arbeiten