Development of a prediction model for pancreatic cancer in patients with type 2 diabetes using logistic regression and artificial neural network models

Figures & data

Table 1 Distribution of train and test sets

All patients	Training set	Test set
1,358,634	1,324,669	33,965

Table 2 Baseline characteristics of T2DM patients with and without pancreatic cancer

Variable	Pancreatic cancer				P-value
	No		Yes
	N=1,355,542		N=3,092
	n	%	n	%
Age group (years)					<0.001
≤49	422,146	31.1	389	12.6
50–64	523,033	38.6	1,197	38.7
65+	410,363	30.3	1,506	48.7
Mean (SD) (years)*	57.3	14.2	63.8	11.4	<0.001
Gender					<0.001
Women	642,176	47.4	1,341	43.4
Men	713,366	52.6	1,751	56.6
Underlying disease
Acute pancreatitis	40,578	2.99	331	10.7	<0.001
Chronic pancreatitis	13,124	0.97	182	5.89	<0.001
Alcohol-related illness	143,856	10.6	307	9.93	0.22
Gallstone	147,231	10.9	596	19.3	<0.001
Cholecystectomy	53,533	3.95	179	5.79	<0.001
Cirrhosis	632,546	46.7	1,681	54.4	<0.001
COPD	383,509	28.3	894	28.9	0.25
Helicobacter pylori infection	21,838	1.61	57	1.84	0.31
Hepatitis B	129,275	9.54	290	9.38	0.77
Hepatitis C	74,671	5.51	162	5.24	0.51
Hypertension	1,001,683	73.9	2,279	73.7	0.81
Hyperlipidemia	912,371	67.3	1,784	57.7	<0.001
Nephropathy	26,796	1.98	40	1.29	0.006
Obesity	71,808	5.30	86	2.78	<0.001
CCI score*					<0.001
0	845,298	62.4	1,774	57.4
1	245,041	18.1	595	19.2
2	129,974	9.59	378	12.2
3 or more	135,231	9.98	345	11.2
Diabetes complication (components of the aDCSI)
Retinopathy	279,890	20.7	487	15.8	<0.001
Nephropathy	489,087	36.1	881	28.5	<0.001
Neuropathy	405,625	29.9	785	25.4	<0.001
Cerebrovascular	248,489	18.3	523	16.9	<0.001
Cardiovascular	686,634	50.7	1,622	52.5	0.045
Peripheral vascular disease	371,646	27.4	658	21.3	<0.001
Metabolic	61,492	4.54	125	4.04	0.19
Change in aDCSI score per year					<0.001
0–0.1	691,408	51.1	1,831	59.2
0.1–0.3	3,733,385	27.6	535	17.3
>0.3	290,749	21.5	726	23.5
Mean aDCSI score (SD)*
Onset	1.44	1.70	1.42	1.62	0.60
End of follow-up	2.62	2.18	2.23	1.99	<0.001
Medications
Statin	716,701	52.9	1,183	38.3	<0.001
Insulin	449,011	33.1	1,044	33.7	0.45
Sulfonylureas	782,389	57.7	1,970	63.7	<0.001
Metformin	868,824	64.1	1,985	64.2	0.90
Other antidiabetic drugs	371,333	27.4	786	25.4	<0.001
TZD	226,441	16.7	471	15.2	<0.001

Notes: Chi-squared test.

* t-Test comparing subjects with and without pancreatic cancer.

Abbreviations: aDCSI, adapted Diabetes Complication Severity Index; CCI, Charlson comorbidity index; T2DM, type 2 diabetes; TZD, thiazolidinediones.

Table 3 Accuracy analysis of LR and ANN models across all data set

Data set	Model	F1	Precision	Recall	AUROC	SE of AUC	95% CI of AUC
All data (n=1,358,634)	LR	0.997	0.995	0.998	0.727	0.004	0.718–0.735
	ANN	0.871	0.996	0.775	0.605	0.005	0.595–0.615
Training set (n=1,324,669)	LR	0.997	0.995	0.998	0.726	0.004	0.718–0.735
	ANN	0.932	0.996	0.876	0.606	0.005	0.596–0.617
Test set (n=33,965)	LR	0.996	0.995	0.998	0.707	0.029	0.650–0.765
	ANN	0.930	0.995	0.873	0.642	0.034	0.576–0.708

Abbreviations: ANN, artificial neural network; AUROC, area under the receiver operating characteristic curve; LR, logistic regression; SE, standard error; AUC, area under the curve.

Figure 1 The ROC curve of the LR model.

Note: The AUC across all data for the LR model is 0.727.

Abbreviations: AUC, area under the ROC curve; LR, logistic regression; ROC, receiver operating characteristic.

Figure 2 The ROC curve of the ANN model.

Note: The AUC curve across all data for the ANN model is 0.605.

Abbreviations: ANN, artificial neural network; AUC, area under the ROC curve; ROC, receiver operating characteristic.