Figure 4. Classification of MDS Patients Using Autoantibody Reactivity Levels.

The standardized reactivity of each of the three proteins of interest was used as a three-dimensional coordinate to classify patient subgroups. (a) Classification into retrospective s-MDS and healthy subsets was the most successful. The example figure indicates the original data points (green: healthy samples, blue: s-MDS), superimposed on the corresponding classification results, kernel density estimates, as computed using Kernel Discriminant Analysis (KDA). Introducing all retrospective categories (b) did not affect s-MDS or healthy classification, though t-MDS and AML displayed lower classification based on the protein of interests. Subject total numbers are as indicated for Stage II in Table 1. The actual data points are superimposed over the densities computed by the KDA for all MDS subgroups, showing more overlap than in (a). (c) Equivalent classification into IPSS risk groups does not perform as well as classification into retrospective subsets. (d) All shown classifications, (a–c), were performed by KDA, using AKT3, FCGR3A and ARL8 standardized autoantibody reactivities to represent three coordinates, for assigning each sample to a point in a three-dimensional space. The classification involved 5-fold cross-validation and 1,000 repetitions for each data partitioning to assess variance and median classification (see also Supplementary Tables S4–S6). N.B. The above classification matrices display classification medians, which might lead to lower sums due to rounding, mismatching the total sum of samples used in each row.