Fig. 5: Benchmark of different methods for anomaly detection.
From: Investigation of measurable residual disease in acute myeloid leukemia by DNA methylation patterns
The samples of BA-seq were classified as normal or abnormal by four different methods and the percentages are demonstrated in the corresponding confusion matrix: a,b AML-score, c,d shallow learning by random forest, e,f deep learning by autoencoder without clustering, and g,h deep learning by autoencoder with clustering. All methods used the same training dataset and the same validation dataset of controls (blue line indicates the threshold of 99% percentile in the training set). All anomaly ratios range from 0 to 1.
