Artificial neural networks in the recognition of the presence of thyroid disease in patients with atrophic body gastritis

Table 1 Dataset of the 253 patients with atrophic body gastritis: original 29 input variables used by the standard artificial neural network (ANN) and input variables selected as the most informative by the T&T and IS System and the TWIST Protocol during the input data optimisation procedures

Input variables	Standard ANN	T&T and IS system	TWIST protocol
Positive clinical history for peptic ulcer	X	X	X
Active H pylori infection	X	X	X
Male	X	X
Female	X	X	X
Age at diagnosis of ABG1	X	X
ABG without associated neoplastic lesions	X
ABG with associated neoplastic gastric lesions	X	X	X
Referred for anemia	X	X
Referred for long-standing dispepsia	X
Referred for dermatological disorders	X	X	X
Referred for endocrinological disorders	X		X
Referred for neurological disorders	X	X
Presence of metaplastic atrophy	X	X	X
Presence of multifocal atrophic gastritis	X	X
Actual smoker	X	X
Previous smoker	X	X
Never smoked	X	X
Positive familiy history for gastric cancer	X
Positive family history for peptic ulcer	X	X
Presence of extrathyroidal autoimmune diseases	X	X	X
Presence of extragastric neoplasms	X	X	X
Presence of pernicious anemia	X	X	X
Presence of iron deficiency anemia	X	X	X
Absence of anemia	X	X	X
Hemoglobin1	X	X	X
Mean corpuscular volume1	X	X	X
Fasting gastrin1	X
Pepsinogen I1	X	X	X
Presence of parietal cell antibodies	X	X	X
Total number of input variables used	29	24	16

X = selected variables;

¹continous variables.

Table 2 Predictive performance of linear discriminant analyses (LDA) and artificial neural networks (ANN) in recognizing atrophic body gastritis patients with and without thyroid disease (%)

	Sensitivity	Specificity	Mean accuracy
LDA	60.8	58.1	59.4
Standard ANN (29 variables)	69.0	59.8	64.4
T&T and IS (24 variables)	78.8b	70.5	74.7
TWIST protocol (16 variables)	81.8d	69.9	75.8

^bP < 0.01 vs LDA;

^dP < 0.001 vs LDA and vs standard ANN P < 0.05.