Survival prediction models since liver transplantation - comparisons between Cox models and machine learning techniques

Table 3 The 12 most prognostic factors for the neural networks with 1 and 2 hidden layers (Rel-Imp: relative importance) and for the Random Survival Forest (VIMP: variable importance). Note that the NN utilises time intervals as one of the input variables (check the contribution of time intervals in Table 1 of Additional file 1). For RSF importance is measured for each variable without distinction for each level

Neural network 1h	Rel-Imp	Neural network 2h	Rel-Imp	RSF	VIMP
Re-transplantation	0.035	Re-transplantation	0.028	Donor age	0.010
Life-support	0.025	HCV ^(d) serology status	0.025	Re-transplantation	0.009
Pre-treatment status ICU ^(a)	0.023	Life-support	0.024	Life support	0.007
Donor type DCD ^(b)	0.023	Donor age	0.023	HCV ^(d) serology status	0.007
Race Black ^(c)	0.022	Diabetes	0.021	Pre-treatment status	0.006
HCV ^(d) serology status	0.022	Pre-treatment status ICU ^(a)	0.020	Recipient age	0.004
Diabetes	0.020	Working income	0.020	Aetiology	0.003
Donor age	0.020	Race Black ^(c)	0.019	log(Last serum creatinine)	0.003
Working income	0.018	Previous abdominal surgery	0.015	Functional status	0.002
Functional status Total assistance ^(e)	0.017	Donor pre-recovery diuretics	0.015	log(Total cold ischemic time)	0.002
Aetiology HCV	0.017	Aetiology Cholestatic	0.011	Race	0.002
Hypertensive bleeding	0.017	Functional status Total assistance ^(e)	0.015	Diabetes	0.002

(a): Intense Care Unit vs Non-hospitalised/Hospitalised (b): Donor type DCD (Donor Circulatory Dead) vs DBD (Donor after Brain-Dead), (c): Race Black vs White, (d): Chronic hepatitis C virus, (e): Total assistance vs No assistance

ISSN: 1471-2288