Using Saliva Epigenetic Data to Develop and Validate a Multivariable Predictor of Esophageal Cancer Status

Figures & data

Figure 1. CONSORT diagram.

Figure 2. Plot of Kolmogorov–Smirnoff distance metric for biological duplicates showing distances between duplicates (x-axis) and distances from each array to a summarized average array (y-axis).

Table 1. Risk stratification of patients and controls in the discovery, testing and independent validation datasets.

		Discovery		Testing		Validation
	Weight factor	Cancer (n = 51)	Non-cancer (n = 107)	Cancer (n = 18)	Non-cancer (n = 34)	Cancer (n = 19)	Non-cancer (n = 27)
Diagnoses		IMC: 16 Invasive: 35	HV: 47 NPD: 27 NDBE: 33	IMC: 6 Invasive: 12	HV: 15 NPD: 8 NDBE: 11	IMC: 6 Invasive: 13	HV: 12 NPD: 7 NDBE: 8
Sex: female	5	14	34	4	8	2	9
Sex: male	5	37	73	14	26	17	18
Mean age (years)	3	70.9 (± 8.80)	63.2 (± 13.4)	69.0 (± 10.8)	60.5 (± 12.6)	70.9 (± 7.01)	63.0 (± 9.26)
BMI (kg/m^2)	3	28.2 (± 4.83)	26.8 (± 4.67)	27.0 (± 4.66)	26.7 (± 4.63)	26.9 (± 4.08)	28.0 (± 3.23)
PPI intake / pill-years	3	1.72 (± 2.73)	1.24 (± 1.83)	1.25 (± 1.72)	1.40 (± 2.20)	1.92 (± 3.35)	1.33 (± 1.44)
Cigarette smoking / pack-years	1	7.43 (± 14.40)	3.42 (± 7.98)	8.22 (± 10.50)	3.37 (± 7.33)	7.90 (± 10.80)	4.20 (± 8.63)
Total alcohol intake / drink-years	0.5	12.07 (± 10)	11.46 (± 9.5)	9.98 (± 10.1)	10.80 (± 9.5)	12.50 (± 11.9)	10.25 (± 7.7)
Heartburn incidence / incidence years	3	1.14 (± 2.3)	1.24 (± 2.3)	1.56 (± 2.6)	1.20 (± 2.21)	1.02 (± 2.7)	1.04 (± 2.1)

Having shown that the groups had similar epigenetic profiles, IMC and invasive cancer were clustered together, as were HVs, patients who had undergone normal endoscopy and had NPD, and patients with NDBE.

All covariates are shown together with the mean values and standard deviation in brackets for each group.

HV: Healthy volunteer; IMC: Intramucosal esophageal adenocarcinoma; NDBE: Nondysplastic Barrett’s esophagus; NPD: No positive diagnosis; PPI: Proton pump inhibitor; SD: Standard deviation.

Figure 3. Heat map plot of p-values of first ten principal component analysis components (A) before and (B) after plate-fitted residual correction for experiment batch, array row, age, sex, disease diagnosis and epithelial cell content.

The components were statistically tested against experiment, array plate row, age, sex, disease status and epithelial cell content. The plate-fitted residuals show p = 1.00. Blue = low p-value; red = high p-value.

Figure 4. Module membership after repeated classification with modification of five samples (3%) of the discovery cohort.

Table 2. Area under the curve for training set and the two holdout sets.

	Base model: IMC and invasive cancer combined			Base model: IMC only	Base model: invasive AdCa only	Cost-sensitive model: IMC and invasive AdCa combined			Cost-sensitive model: IMC only	Cost-sensitive model: invasive AdCa only
	AUC	Sensitivity	Specificity	Sensitivity	Sensitivity	AUC	Sensitivity	Specificity	Sensitivity	Sensitivity
Training	0.73 (0.69–0.75)	0.28 (0.26–0.30)	0.94 (0.91–0.98)	0.55 (0.51–0.59)	0.18 (0.15–0.21)	0.72 (0.69–0.76)	1.00 (0.98–1.00)	0.10 (0.06–0.16)	1.00 (0.96–1.00)	1.00 (0.97–1.00)
Test	0.72 (0.68–0.74)	0.11 (0.09–0.15)	0.97 (0.94–1.00)	0.25 (0.21–0.29)	0.00 (0.00–0.03)	0.72 (0.68–0.75)	0.88 (0.84–0.91)	0.31 (0.26–0.36)	0.78 (0.74–0.82)	1.00 (0.96–1.00)
Validation	0.73 (0.68–0.75)	0.19 (0.16–0.23)	0.97 (0.04–1.00)	0.29 (0.25–0.32)	0.14 (0.11–0.18)	0.63 (0.59–0.68)	0.86 (0.82–0.90)	0.23 (0.18–0.29)	0.86 (0.82–0.91)	0.86 (0.81–0.89)

All results are mean (95% CI).

Showing tenfold cross validation for the base model and after a cost-sensitive wrapper algorithm is applied which generates a high cost for missing cancer cases.

AdCa: Adenocarcinoma; AUC: Area under the curve; IMC: Intramucosal esophageal adenocarcinoma.

Figure 5. Receiver operating characteristic curve of the most successful classifiers for the discovery set, testing set and validation dataset.

Table 3. Area under the curve using only probes and only covariates of age and sex.

	Probes only	Covariates only
	AUC	AUC
Discovery	0.65 (0.60–0.69)	0.63 (0.59–0.68)
Testing	0.64 (0.60–0.70)	0.61 (0.56–0.66)
Validation	0.64 (0.59–0.70)	0.60 (0.54–0.65)

All results are mean (95% CI).

Table for training set (tenfold cross validation) and the two holdout sets.

AUC: Area under the curve.

Data sharing statement

Raw data, processed data and accompanying metadata have been deposited to the Gene Expression Omnibus (GEO) database under the accession code GSE232332.