Study Population
A total of 943 patients with ACLF enrolled in our analysis. The baseline characteristics of the enrolled patients can be seen in Table 1. The majority of patients (86.43%) were men. The median [IQR] age was 46 years (range, 38–55 years). We observed 32.87% of patients had hepatic encephalopathy, 4.56% had gastrointestinal bleeding, 43.80% had peritonitis, 19.41% had ascites, and 44.33% had pneumonia. Most (61.74%) non-survivors had liver cirrhosis. Non-survivors had higher TBil, BUN, Cr, INR levels, and prothrombin time, and lower serum alanine aminotransferase, aspartate transaminase, and fibrinogen levels. We found that 31.6% of the patients dead at the end of the follow-up.
Table 1 Baseline characteristics of patients.
Variable
|
Total
|
outcome=0
|
outcome=1
|
P
|
n
|
943
|
645(68.40%)
|
298(31.60%)
|
|
Gender
|
|
|
|
0.019
|
Female
|
128(13.57%)
|
99(15.35%)
|
29(9.73%)
|
|
Male
|
815(86.43%)
|
546(84.65%)
|
269(90.27%)
|
|
Hepatic encephalopathy
|
|
|
|
<0.001
|
0
|
633(67.13%)
|
493(76.43%)
|
140(46.98%)
|
|
1
|
310(32.87%)
|
152(23.57%)
|
158(53.02%)
|
|
Gastrointestinal bleeding
|
|
|
|
<0.001
|
0
|
900(95.44%)
|
630(97.67%)
|
270(90.60%)
|
|
1
|
43(4.56%)
|
15(2.33%)
|
28(9.40%)
|
|
peritonitis
|
|
|
|
<0.001
|
0
|
530(56.20%)
|
388(60.16%)
|
142(47.65%)
|
|
1
|
413(43.80%)
|
257(39.84%)
|
156(52.35%)
|
|
Ascites
|
|
|
|
<0.001
|
0
|
760(80.59%)
|
543(84.19%)
|
217(72.82%)
|
|
1
|
183(19.41%)
|
102(15.81%)
|
81(27.18%)
|
|
cirrhosis
|
|
|
|
<0.001
|
0
|
441(46.77%)
|
327(50.70%)
|
114(38.26%)
|
|
1
|
502(53.23%)
|
318(49.30%)
|
184(61.74%)
|
|
pneumonia
|
|
|
|
<0.001
|
0
|
525(55.67%)
|
389(60.31%)
|
136(45.64%)
|
|
1
|
418(44.33%)
|
256(39.69%)
|
162(54.36%)
|
|
Age
|
46.0(38.00-55.00)
|
45.0(36.00-54.00)
|
48.0(41.00-57.00)
|
<0.001
|
Anteroposterior diameter of Left liver
|
61.0(53.00-67.00)
|
61.0(53.00-67.00)
|
61.0(54.00-68.00)
|
0.173
|
Vertical diameter of Left liver
|
56.0(43.00-68.00)
|
56.0(43.00-68.00)
|
54.0(42.25-65.75)
|
0.121
|
Thickness of Right liver
|
101.0(94.00-108.00)
|
102.0(95.00-108.00)
|
100.0(90.25-107.00)
|
<0.001
|
Portal vein diameter
|
12.0(11.00-12.95)
|
12.0(11.00-12.50)
|
12.0(11.00-13.00)
|
0.086
|
Intercostal thick diameter of spleen
|
39.0(34.00-45.00)
|
38.0(34.00-45.00)
|
40.0(35.00-47.00)
|
0.013
|
length of spleen
|
117.0(104.00-132.00)
|
117.0(104.00-130.00)
|
119.0(105.00-136.00)
|
0.046
|
The inner diameter of the splenic portal vein
|
10.3(10.00-11.20)
|
10.6(10.00-11.60)
|
10.0(10.00-11.00)
|
0.198
|
WBC
|
7.14(5.26-9.68)
|
6.96(5.30-9.28)
|
7.48(5.23-10.31)
|
0.052
|
NEUT
|
0.7(0.63-0.77)
|
0.69(0.62-0.75)
|
0.73(0.65-0.79)
|
<0.001
|
Hb
|
119.0(101.00-134.00)
|
122.0(105.00-136.00)
|
112.0(92.25-129.00)
|
<0.001
|
PLT
|
104.0(70.00-148.00)
|
113.0(80.00-155.00)
|
87.5(57.00-131.00)
|
<0.001
|
AST
|
198.0(100.00-499.00)
|
218.0(106.00-564.00)
|
157.5(93.00-344.75)
|
<0.001
|
ALT
|
220.0(74.00-755.50)
|
280.0(87.00-853.00)
|
147.5(56.25-531.00)
|
<0.001
|
CHE
|
3811.0(2746.00-5069.50)
|
3849.0(2746.00-5052.00)
|
3716.5(2779.25-5078.50)
|
0.243
|
ALB
|
33.0(29.80-36.25)
|
32.9(29.60-36.10)
|
33.0(30.30-36.48)
|
0.196
|
TBil
|
362.8(250.00-487.90)
|
327.0(224.10-439.10)
|
448.92(332.47-575.98)
|
<0.001
|
DBil
|
240.0(158.45-321.85)
|
216.8(144.70-300.50)
|
285.0(201.93-373.40)
|
<0.001
|
Cr
|
71.0(60.00-87.00)
|
70.0(60.00-83.00)
|
73.0(60.00-102.75)
|
0.002
|
BUN
|
3.8(2.74-5.55)
|
3.71(2.73-5.25)
|
4.07(2.75-7.20)
|
0.003
|
Na
|
137.0(134.00-139.00)
|
137.0(134.00-139.00)
|
135.95(132.93-139.00)
|
<0.001
|
PT
|
26.4(21.90-33.10)
|
24.8(21.30-29.50)
|
31.3(24.60-40.20)
|
<0.001
|
INR
|
2.42(1.92-3.24)
|
2.24(1.84-2.80)
|
2.98(2.26-4.14)
|
<0.001
|
Fib
|
1.52(1.13-1.97)
|
1.6(1.27-2.04)
|
1.27(0.90-1.77)
|
<0.001
|
PTA
|
31.0(23.00-41.00)
|
34.0(27.00-43.00)
|
25.0(18.00-34.00)
|
<0.001
|
Characteristics of Prediction Models
A time-dependent Cox regression model was employed to analyze the dynamic changes and determine whether the evolution of clinical parameters correlated with the prognosis of ACLF. Sex, age, hepatic encephalopathy, inner diameter of the splenic portal vein, absolute neutrophil count, hemoglobin, cholinesterase, albumin, TBil, direct bilirubin, Cr, BUN, and PTA were significant risk factors (Table S1). Sex, age, hepatic encephalopathy, TBil, PTA, and BUN were selected stepwise for the JM (Table S2). Since TBil, BUN and PTA do not obey the normal distribution, they were transformed, using BoxCox conversion for TBil and BUN (lambda values were 0.52 and -0.32, respectively), and log conversion for PTA, followed by the mixed effect model (Table S3). Age, hepatic encephalopathy, BoxCox (TBil), BoxCox (BUN), and log (PTA) tended to be associated with death (P < 0.05) according to the JM, with death as the outcome (Table 2).
Table 2. Coefficients, standard error of coefficients and p-values for the joint model predicting death.
Variable
|
coef
|
std
|
p.value
|
HR(95% CI)
|
Gender
|
0.4642
|
0.3438
|
0.1640
|
1.5907(0.8329, 3.2088)
|
Age
|
0.0158
|
0.0076
|
0.0387
|
1.0159(1.0009, 1.03)
|
Hepatic encephalopathy
|
0.5896
|
0.1669
|
<0.001
|
1.8033(1.2987, 2.4461)
|
value(boxcox(TBil))
|
0.0230
|
0.0098
|
0.011
|
1.0233(1.0041, 1.0423)
|
value(boxcox(BUN))
|
1.2626
|
0.2451
|
<0.001
|
3.5346(2.2347, 5.7048)
|
value(log(PTA))
|
-2.0378
|
0.2156
|
<0.001
|
0.1303(0.0877, 0.2076)
|
Performance of Prediction Models
The probability of survival can be obtained at a certain time point during follow-up based on this JM. This method is based on calculation of the posterior predictive survival function. That is, discriminating between patients who will develop the event in the interval (t0, t0 + Dt), (i.e., in our case Tj∈ (7,28)) from patients who will survive at least 28 days (i.e., Tj > 28). In Figure 4B, we calculated the components of the ROC curve using up to seven pieces of information and were interested in events occurring within a 28-day window with an AUC of 0.83. Figure 2 and Table 3 show the prediction performance of the JM at 28 and 90 days, at baseline, and during follow-up. The ACLF-JM accurately predicted 28-day mortality, with AUCs ranging from 0.808 to 0.840 (Figures 2 and 3).
Table 3 The performance of joint model at different time point.
Tstart
|
Tstop
|
AUC
|
Brier
|
1
|
28
|
0.8078
|
0.1538
|
7
|
28
|
0.8124
|
0.1215
|
14
|
28
|
0.8409
|
0.0825
|
1
|
89
|
0.7466
|
0.2048
|
7
|
89
|
0.7621
|
0.2017
|
14
|
89
|
0.8074
|
0.1838
|
21
|
89
|
0.811
|
0.1788
|
28
|
89
|
0.7717
|
0.1834
|
35
|
89
|
0.7508
|
0.1806
|
Next, we calculated the MELD score to predict the 90-day mortality and compared it with the JM. The performance of JM was significantly superior to that of the MELD score at all time points and studied outcomes. The AUC of the MELD curve was 0.723. The AUCs for the JM model ranged from 0.747–0.811 for 90-day mortality prediction in patients with ACLF. The results indicated that the performance of the JM remained significantly superior to that of the MELD score at all time points. The linearity of the calibration curves was good, with Brier scores ranging from 0.0825–0.2048 (Figures 2 and 3), where the lower the Brier score, the better the model prediction.
Personal Dynamic Survival Prediction
Two patients with different outcomes were randomly selected. Figures 4 and 5 show the dynamic predictions of 28- and 90-day survival for Patients 4 and 147, respectively, employing the JM framework. Based on an individualized dynamic prediction map, the survival of patients from each follow-up period up to 28 days was derived. From Figure 4, we can see that the survival curve does not present significant changes as more clinical indices are measured over time for Patient 4. In Figure 4A–C, from follow-up to days 7 and 14 the survival probability of patients on day 28 was estimated to be 86.15% (95% CI: 0.5346–0.9666), 91.34% (95% CI: 0.7573–0.9763) and 96.61% (95% CI: 0.9348–0.9869), respectively. This can be explained by the fact that Patient 4 was relatively stable and had a higher survival probability. In contrast, Patient 147 had a steeper curve for expected survival. In Figure 5A–C, from follow-up to days 7, 14 and 21 the survival probability of the patient on day 89 was estimated to be 20.10% (95% CI: 0.0000–0.6816), 3.46% (95% CI: 0.0000–0.3129), 30.24% (95% CI: 0.0000–0.7295), and 36.90% (95% CI: 0.0011–0.7603), respectively. The patient’s survival probability was < 50% after his last visit, revealed that the patients need to be monitored constantly and be prioritized on the transplantation waiting list.