Utility of a deep learning model and a clinical model for predicting bleeding after endoscopic submucosal dissection in patients with early gastric cancer

Table 1 Baseline characteristics of patients in entire cohort

Variable	Non-PEB, n = 5304	PEB, n = 325	P value1
Age2	64 ± 10	63 ± 11	0.065
Sex			0.001
Female	1245 (23.5)	49 (15.1)
Male	4059 (76.5)	276 (84.9)
Hypertension	1413 (26.6)	120 (36.9)	< 0.001
Diabetes mellitus	936 (17.6)	74 (22.8)	0.024
Liver cirrhosis	93 (1.8)	4 (1.2)	0.629
Chronic kidney disease	299 (5.6)	33 (10.2)	0.001
Aspirin	515 (9.7)	42 (12.9)	0.074
P2Y12RA	181 (3.4)	23 (7.1)	0.001
Warfarin	22 (0.4)	7 (2.2)	< 0.001
DOAC	31 (0.6)	6 (1.8)	0.017
Cilostazol	47 (0.9)	3 (0.9)	1.000
NSAIDs	28 (0.5)	3 (0.9)	0.583
Preprocedure management of AT			< 0.001
No indication	4605 (86.8)	264 (81.2)
Interruption	676 (12.7)	53 (16.3)
Replacement or heparin bridge	23 (0.4)	8 (2.5)
Tumor
Multiple	284 (5.4)	28 (8.6)	0.018
Location			< 0.001
Upper	433 (8.2)	22 (6.8)
Middle	1728 (32.6)	157 (48.3)
Lower	3143 (59.3)	146 (44.9)
Size2, mm	17 ± 10	21 ± 13	< 0.001
Undifferentiated type	125 (2.4)	7 (2.2)	0.963
Piecemeal resection	64 (1.2)	4 (1.2)	1.000
Laboratory data
Albumin2, g/dL	4.3 ± 0.3	4.4 ± 0.4	0.345
INR2	1.0 ± 0.1	1.0 ± 0.1	0.106

¹P value calculated using Student’s t-test for continuous variables or Pearson’s chi-square test for categorical variables for overall data.

²mean ± SD presented for continuous variables.

Values are expressed as n (%) unless otherwise specified. AT: Antithrombotic; DOAC: Direct oral anticoagulant; INR: International normalized ratio; NSAIDs: Non-steroidal anti-inflammatory drugs; PEB: Post-endoscopic submucosal dissection bleeding; P2Y12RA: P2Y12 receptor antagonist.

Table 2 Logistic regression analysis for predictors of bleeding after endoscopic submucosal dissection in development set

Variables		Multivariable
		OR	95%CI	P value	ß regression coefficient
Age		0.98	0.96–0.99	0.001	-0.024
Sex	Female/male	1.54	1.09–2.19	0.015	0.435
Hypertension	No/yes	1.35	1.00–1.82	0.049	0.299
Diabetes mellitus	No/yes	1.27	0.92–1.75	0.145	0.238
Liver cirrhosis	No/yes	0.59	0.18–1.95	0.385	-0.532
Chronic kidney disease	No/yes	1.78	1.12–2.84	0.015	0.578
Aspirin	No/yes	1.51	0.62–3.69	0.363	0.414
P2Y12RA	No/yes	2.26	1.05–4.88	0.037	0.818
Warfarin	No/yes	1.51	0.28–8.07	0.629	0.413
DOAC	No/yes	3.24	0.76–13.82	0.113	1.174
Cilostazol	No/yes	1.35	0.35–5.18	0.662	0.300
NSAIDs	No/yes	2.65	0.77–9.14	0.124	0.973
Preprocedure management of AT	No indication	1
	Interruption	0.63	0.24–1.67	0.353	-0.464
	Replacement orHeparin bridge	3.32	0.47–23.60	0.231	1.199
Multiple	No/yes	1.48	0.92–2.38	0.104	0.393
Location	Upper	1
	Middle	1.97	1.14–3.41	0.015	0.680
	Lower	1.11	0.64–1.91	0.711	0.103
Size		1.04	1.03–1.05	< 0.001	0.036
Undifferentiated type	No/yes	0.56	0.20–1.57	0.271	-0.579
Piecemeal	No/yes	0.98	0.30–3.22	0.976	-0.019
Albumin, g/dL		1.33	0.89–2.00	0.168	0.286
INR		2.04	0.37–11.08	0.410	0.711

Clinical model = 1/[1 + exp(-1 × [-0.024 × Age in years + 0.435 × Sex (0: female, 1: male) + 0.299 × Hypertension (0: no, 1: yes) + 0.578 × Chronic kidney disease (0: no, yes: 1) + 0.818 × P2Y12RA (0: no, 1: yes) + 0.680 × Middle location (0: no, 1: yes) + 0.036 × Size in mm])]. AT: Antithrombotic; DOAC: Direct oral anticoagulant; ESD: Endoscopic submucosal dissection; INR: International normalized ratio; NSAIDs, Non-steroidal anti-inflammatory drugs; P2Y12RA: P2Y12 receptor antagonist; OR: Odds ratio; CI: Confidence interval.

Table 3 Utility of deep learning model and clinical model

	Deep learning model	Clinical model	P value
Sensitivity (%)	64.3 (45.8–84.1)	69.6 (54.2–80.8)
Specificity (%)	74.0 (50.6–89.2)	71.0 (68.5–79.5)
PPV (%)	11.4 (7.4–18.1)	11.1 (8.0–15.4)
NPV (%)	97.5 (96.4–98.7)	97.8 (96.6–98.7)
AUC (95%CI)	0.71 (0.63–0.78)	0.70 (0.62–0.77)	0.730

One thousand times for bootstrapping were conducted to measure 95% confidence intervals. P value for statistical significance between area under the curves was derived from Delong’s test. AUC: Area under the curve; CI: Confidence interval; NPV: Negative predictive value; PPV: Positive predictive value.

Table 4 Decile of risk probability based on deep learning model and clinical model

Decile	Risk categories	Deep learning				Clinical model
		Score1	Patients	Bleeding	Rate (%)	Score1	Patients	Bleeding	Rate (%)
Development set
1	Low	25.7	451	11	2.4	8.4	451	12	2.6
2		29.1	451	15	3.3	12.2	451	15	3.3
3		32.5	451	6	1.3	12.4	451	12	2.6
4		35.9	450	17	3.8	12.7	450	20	4.4
5	Intermediate	40.2	450	27	6.0	14.8	450	34	7.6
6		45.3	450	29	6.4	16.6	450	15	3.3
7		50.8	450	36	8.0	23.3	450	24	5.3
8		57.5	450	24	5.3	24.6	450	32	7.1
9	High	67.2	450	41	9.1	31.0	450	54	12.0
10		197.0	450	63	14.0	122.0	450	51	11.3
Validation set
	Low	35.9	411	9	2.2	12.7	956	38	4.0
	Intermediate	57.5	466	18	3.9	24.5	137	12	8.8
	High	147.0	249	29	11.6	155.0	33	6	18.2

¹The score is calculated with probability multiplied by 1000 and presented as maximum cutoff in each decile.

Decile 1^st to 4^th: Low-risk category. Decile 5^th to 8^th: Intermediate-risk category. Decile 9^th to 10^th: High-risk category. The Cochran–Armitage test for trend was performed.