A Deep Learning Model Combining Multimodal Factors to Predict the Overall Survival of Transarterial Chemoembolization

Figures & data

Figure 1 Workflow of TACE patient survival prediction model based on deep learning (DLOPcombin).

Table 1 Baseline Demographic and Clinical Characteristics

	Training Cohort (n=241)	Validation Cohort (n=60)	P
Age	61.00(57.00–71.00)	60.00(58.00–68.00)	0.5609
Gender			0.8527
Male	182(75.5)	46(76.7)
Female	59(24.5)	14(23.3)
Portal Hypertension			0.8527
No	182(75.5)	46(76.7)
Yes	59(24.5)	14(23.3)
Diabetes			0.9519
No	196(81.3)	49(81.7)
Yes	45(18.7)	11(18.3)
Cirrhosis			0.2223
No	84(34.9)	26(43.3)
Yes	157(65.1)	34(56.7)
Hepatitis			0.0024
No-HBV	33(13.7)	0 (0)
HBV	208(86.3)	60(100)
AFP	142.84(27.72–415.59)	130.88(10.76–689.58)	0.6993
AST	65.00(27.00–169.00)	59.00(34.00–137.00)	0.5855
ALT	42.00(28.00–61.00)	35.50(24.50–56.50)	0.2044
ALB	37.80(34.00–41.50)	38.20(33.65–41.05)	0.8036
TBIL	20.40(13.20–29.50)	19.45(11.70–27.70)	0.4248
Tumor number			0.6227
> 2	100(41.5)	27(45.0)
≤ 2	141(58.5)	33(55.0)
Tumor size	48.00(31.00–78.00)	42.25(24.50–65.50)	0.1446
Tumor location			0.1549
Left	45(18.7)	7(11.7)
Right	153(63.5)	46(76.7)
Others	43(17.8)	7(11.7)
MVI			0.1124
No	109(45.2)	34(56.7)
Yes	132(54.8)	26(43.3)
ALBI grade			0.2147
1 grade	130(53.9)	27(45.0)
2~3 grade	111(46.1)	33(55.0)
Child-Pugh score			0.6436
A	172(71.4)	41(68.3)
B~C	69(28.6)	19(31.7)
BCLC B subclassification			0.0622
1	99(41.0)	20(33.3)
2	89(36.9)	19(31.7)
3	25(10.4)	14(23.3)
4	28(11.7)	7(11.7)

Abbreviations: AFP, alpha-fetoprotein; AST, aspartate aminotransferase; ALT, alanine aminotransferase; ALB, albumin; TBIL, total bilirubin; HBV, hepatitis B virus; MVI, microvascular infiltration; ALBI, albumin-bilirubin.

Table 2 Univariate and Multivariate Cox Regression Analysis of Clinical Factors in the Training Cohort

Variable	Univariate Analysis		Multivariate Analysis
	HR (95% CI)	P	HR (95% CI)	P
Age	0.97(0.96–0.99)	< 0.05	0.99 (0.97–1.02)	0.62
Gender	1.00 (0.78–1.40)	0.78
Portal Hypertension	0.79 (0.59–1.10)	0.11
Diabetes	0.89 (0.65–1.2)	0.47
Cirrhosis	0.87 (0.68–1.1)	0.28
Hepatitis	0.75 (0.51–1.1)	0.14
Child-Pugh score	1.447 (1.078–1.941)	0.01	1.36 (0.98–1.88)	0.67
AFP	1.59(0.78–3.26)	0.20
AST	1.93(0.71–5.26)	0.20
ALT	2.58(0.69–9.74)	0.16
ALB	0.74(0.36–1.54)	0.42
TBIL	1.16(0.48–2.75)	0.75
Tumor number	1.12 (0.88–1.43)	0.35
Tumor size	1.00 (1.00–1.00)	0.74
Tumor location
Left	Reference
Right	0.18 (0.86–1.7)	0.56
Others	0.18 (0.79–1.8)	0.56
MVI	1.742 (1.325–2.290)	<0.0001	1.42 (1.05–1.92)	0.02
ALBI grade	2.125 (1.618–2.791)	<0.0001	1.89 (1.41–2.54)	<0.0001
BCLC B subclassification
Reference
2	1.347 (1.078–1.941)	0.0543	1.53 (1.11–2.10)	0.01
3	1.866 (1.172–2.971)	0.0086	2.25 (1.39–3.65)	0.001
4	4.704 (2.926–7.562)	<0.0001	3.74 (2.29–6.10)	<0.0001
Treatment response	0.52 (0.37–0.72)	< 0.05	0.04 (0.293–0.969)	0.04

Note: P < 0.05 indicates statistically significant.

Abbreviations: HR, hazard ratio; CI, confidence interval.

Table 3 Performance of the Model in Predicting the Prognosis of HCC Patients in the Training Cohort and the Validation Cohort

Variable	Training Cohort						Validation Cohort
	TD AUC (95% CI)	ACC	PRE	REC	PPV	NPV	TD AUC (95% CI)	ACC	PRE	REC	PPV	NPV
DLOPCombin	0.98 (0.94, 1.00)	0.96	0.81	0.97	0.92	0.98	0.96 (0.88, 1.00)	0.93	0.84	0.92	0.90	0.92
DLOPCTR	0.92 (0.88, 0.95)	0.87	0.75	0.91	0.82	0.91	0.89 (0.80, 0.98)	0.85	0.72	1.00	0.8	0.95
DLOPCox	0.91 (0.86, 0.96)	0.86	0.72	0.91	0.85	0.93	0.89 (0.81, 0.97)	0.84	071	1.00	0.92	0.81
Radiomics	0.70 (0.61, 0.79)	0.74	0.47	0.69	0.69	0.68	0.63 (0.49, 0.76)	0.70	0.43	0.50	0.65	0.68
Clinical	0.66 (0.60, 0.72)	0.64	0.40	0.85	0.82	0.70	0.62 (0.59, 0.65)	0.59	0.42	0.78	0.76	0.65

Notes: Data are presented as time-dependent areas under the receiver operating characteristic curve (TD AUC) and Harrell C-Index, with 95% CI in parentheses. Deep Learning Model combining CT Images and Clinical Factors (DLSPcombin); Deep Learning Survival Prediction Model Based on CT Images (DLSP_CTR); DLSP_CTR as a separate variable combined with clinical factors in a multifactor COX model (DLSP_COX).

Table 4 Discrimination Performance and Calibration of the Survival Prediction Model

Parameter	Training Cohort			Validation Cohort
	Harrell C index	P	P^+	Harrell C index	P	P^+
DLOPCombin	0.90(0.85, 0.95)			0.88(0.80, 0.96)
DLOPCTR	0.78(0.72, 0.84)	0.04	0.05	0.75(0.75, 0.89)	0.03	0.04
DLOPCox	0.79(0.69, 0.89)	0.02	0.04	0.77(0.70, 0.84)	0.02	0.03
Radiomics	0.61(0.51, 0.71)	< 0.01	0.03	0.59(0.50, 0.68)	< 0.01	0.03
Clinical	0.60(0.45, 0.75)	< 0.01	0.03	0.56(0.48, 0.64)	< 0.01	0.03

Notes: Discrimination refers to Harrell C index, student test was used for comparison between C indices, and FDR was used for calibration, Data in parentheses are 95% confidence intervals. P values indicate that DLOP_Combin is used as the benchmark for comparison with other models. P ⁺values indicate FDR calibration of P value.

Figure 2 Performance of different survival prediction models. Time-dependent receiver operating curves under the area for 3-year survival of patients (A) training cohort (B) validation cohort. Calibration curves of the model (C) training cohort (D) validation cohort.

Figure 3 Kaplan-Meier estimates of overall survival according to the Youden index of DLSPcombin categorized high- and low-risk groups (A) training cohort (B) validation cohort.

Figure 4 Results of the DeLong test between the different models (A) training cohort (B) validation cohort. Results of the FDR after calibration of the p-value (C) training cohort (D) validation cohort.

Figure 5 Decision curve analysis of the (A) training cohort (B) validation cohort.

Figure 6 Examples of CT images (A), Heatmap (B), and their corresponding gradient-class activation maps (C) on the validation cohort. On both the heat map and the gradient class activation map, red areas indicate that the area may be associated with a high probability of survival for the patient, while blue lesions indicate a low probability of survival. (A) A 64-year-old woman with 4 years of survival after TACE and Ct-based deep learning survival prediction model activation focused on the tumor region. (B) A 52-year-old male with no internal tumor activation and a disease-free survival of 240 days. (C) The patient with the incorrect prediction, a 48-year-old male, survived 397 days after surgery. The model predicted low risk; however, death occurred after 1254 days.