Table 1:

Estimating batch effects by analyzing intra-cohort and inter-cohort performance in all sub-cohorts in the international cohort.

	train on TCGA n=426 15% MSI	train on QUASAR n=1770 14% dMMR	train on DACHS n=2013 14% MSI	train on NLCS n=2197 10% dMMR
test on TCGA (US)	0.74 [0.66, 0.80]	0.76 [0.70, 0.79]	0.77 [0.73, 0.79]	0.72 [0.71, 0.78]
test on QUASAR (UK)	0.67 [0.64, 0.68]	0.89 [0.86, 0.91]	0.71 [0.68, 0.75]	\ 0.76 [0.73, 0.78]
test on DACHS (DE)	0.81 [0.79, 0.83]	0.68 [0.65, 0.72]	0.92 [0.91, 0.94]	0.80 [0.78, 0.82]
test on NLCS (NL)	0.77 [0.74, 0.79]	0.80 [0.78, 0.81]	0.82 [0.79, 0.83]	0.90 [0.89, 0.91]

Main performance measure was area under the receiver operating curve, shown as mean with lower and upper bounds in a 10-fold bootstrapped experiment. Intra-cohort-performance was estimated by three-fold cross-validation. US = United States, UK = United Kingdom, DE = Germany, NL = Netherlands, MSI = microsatellite instability, dMMR = mismatch repair deficiency.