Artificial neural networks versus proportional hazards Cox models to predict 45-year all-cause mortality in the Italian Rural Areas of the Seven Countries Study

Table 2 Multilayer perceptron models predicting 45-year all-cause mortality (1447 deaths among 1591 men) by two methods

Rank	Gini	Variable	Keep
NEURAL NETWORK FORCED
1	0.15950	Age (AGE0: years)	1
2	0.60615	Cigarettes smoked per day (CIG0: N)	1
3	0.64475	Family history of CVD(famcv0: codes 0–1)	1
4	0.64790	Mean blood pressure (MBP: mmHg)	1
5	0.65872	Serum cholesterol (CHOL0: mmol/l)	1
6	0.67052	Corneal arcus (gero0: codes 0–1)	1
7	0.67484	Mother status (Mother0: codes 0–1)	1
8	0.67680	Prevalence DIAB (pdiab0: codes 0–1)	1
9	0.68081	Forced expiratory volume (fev0trans0: l/m²)	1
10	0.68321	Heart rate (hr0: beats/min)	1
11	0.68362	Father status (Father0: codes 0–1)	1
12	0.68414	Body mass index (BMI0: Kg/m²)	1
13	0.68516	Minor ECG abnormalities (MinorECG0: codes 0–1)	1
14	0.68524	Prevalence CVD (Pcvd0: codes 0–1)	1
15	0.68579	Arm circumference (midclean0: cm)	1
16	0.68664	Body mass index² (BMIsq0) [(Kg/m²)²]	1
17	0.68813	Prevalence cancer (pcan0: codes 0–1)	0
18	0.68872	Physical activity (PHYAC0: codes 1-2-3)	0
NEURAL NETWORK STEPWISE
1	0.17824	Age (years)	1
2	0.60661	Cigarettes smoked per day (N)	1
3	0.65370	Mean blood pressure (mmHg)	1
4	0.67844	Mother status (codes 0–1)	1
5	0.67918	Corneal arcus (codes 0–1)	1
6	0.68104	Heart rate (beats/min)	1
7	0.68347	Father status (codes 0–1)	1
8	0.68468	Minor ECG abnormalities (codes 0–1)	1
9	0.69231	Physical activity (codes 1-2-3)	1
10	0.69329	Family history (codes 0–1)	1
11	0.69480	Prevalence DIAB (codes 0–1)	1
12	0.69603	Serum cholesterol (mmol/l)	1
13	0.69854	Prevalence CVD (codes 0–1)	1

Gini = (ROC-0.5)*2 [see text for further details]; Keep = 1: the variable may stay in the model; Keep = 0: removing the variable improves the training set model.
CVD = cardiovascular disease; DIAB = diabetes. Variables’ codes are included in parentheses to interpret the Figure in Additional file 1: Appendix 2.

ISSN: 1471-2288