A deep neural network approach to predicting clinical outcomes of neuroblastoma patients

BMC Medical Genomics

Table 4 Parameter optimization for all classifiers.

Algorithm	Parameters	Balanced accuracy
Clinical outcome = ‘Death from disease’,
Data=Fischer-M, centralities
DNN [8,8,8,2]	o=Adam, lr=1e-3, d=0.3	87.3% (+0.0)
GEDFN^a	lr=1e-2, h=[64,16], b=8	79.5% (+8.6)
SVM	t=RBF, c=64, g=0.25	75.4% (+5.9)
RF	n=100	75.1% (+3.1)
Clinical outcome = ‘Disease progression’,
Data=Fischer, centralities
DNN [4,2,2,2]	o=Adam, lr=1e-3, d=0.3	84.7% (+0.0)
GEDFN^a	lr=1e-4, h=[16,4], b=32	81.2% (+0.4)
SVM	t=RBF, c=16, g=0.0625	81.8% (+2.0)
RF	n=100	78.1% (+3.2)

One row corresponds to the best model for a given clinical outcome and algorithm. The optimal parameter values are provided (o: optimizer, lr: learning rate, d: dropout, h: sizes of the second and third GEDFN hidden layers, b: batch size, t: SVM kernel type, c: cost, g: gamma, n: number of trees). The gain in balanced accuracy with respect to the models run with default parameters is indicated between parentheses (from Table 3 for DNN)
^afor GEDFN, the corresponding omics data is used as input instead of the topological features

ISSN: 1755-8794