- Email: [email protected]
A Diagnostic Prediction Model of Acute Symptomatic Portal Vein Thrombosis
Clinical Research A Diagnostic Prediction Model of Acute Symptomatic Portal Vein Thrombosis Kun Liu,1,2 Jun Chen,1 Kaixin Zhang,1 Shuo Wang,1 and Xiaoqiang Li,3 Suqian and Suzhou, Jiangsu, China
Background: The aim of this study was to develop a diagnostic prediction model to improve identification of acute symptomatic portal vein thrombosis (PVT). Methods: We examined 47 patients with PVT and 94 controls without PVT in the Second Affiliated Hospital of Soochow University and Suqian People’s Hospital of Nanjing, Gulou Hospital Group. We constructed a prediction model by using a support vector machine (SVM) classifier coupled with a least absolute shrinkage and selection operator (LASSO). We applied a 10-fold cross-validation to estimate the error rate for each model. Results: The present study indicated that acute symptomatic PVT was associated with 11 indicators, including liver cirrhosis, D-Dimer, splenomegaly, splenectomy, inherited thrombophilia, ascetic fluid, history of abdominal surgery, bloating, C-reactive protein (CRP), albumin, and abdominal tenderness. The LASSO-SVM model achieved a sensitivity of 91.5% and a specificity of 100.0%. Conclusions: We developed a LASSO-SVM model to diagnose PVT. We demonstrated that the model achieved a sensitivity of 91.5% and a specificity of 100.0%.
INTRODUCTION According to epidemiological evidence, portal vein thrombosis (PVT) showed a worldwide prevalence of 1.0%. Of patients with PVT, 28% had cirrhosis, 23% primary and 44% secondary hepatobiliary malignancy, 10% major abdominal infectious or inflammatory disease, and 3% had a myeloproliferative disorder. The highest PVT risk was showed in
1 Department of Vascular Surgery, The Affiliated Suqian Hospital of Xuzhou Medical University, Suqian People’s Hospital affiliated to Nanjing Drama Tower Hospital Group, Suqian, Jiangsu, China. 2
Department of Vascular Surgery, The Second Affiliated Hospital of Soochow University, Suzhou, Jiangsu, China. 3 Department of Vascular Surgery, Affiliated Nanjing Drum Tower Hospital of Nanjing University Medical School, Nanjing, China.
Correspondence to: Xiaoqiang Li, Department of Vascular Surgery, Affiliated Nanjing Drum Tower Hospital of Nanjing University Medical School, 321 Zhongshan Road, Gulou District, Nanjing City, Jiangsu Province 210000, China; E-mail: [email protected] Ann Vasc Surg 2019; -: 1–6 https://doi.org/10.1016/j.avsg.2019.04.037 Ó 2019 Elsevier Inc. All rights reserved. Manuscript received: December 19, 2018; manuscript accepted: April 22, 2019; published online: - - -
patients with both cirrhosis and hepatic carcinoma.1 PVT has been regarded as an important marker of decompensated cirrhosis.2 In general, PVT was usually caused by a combination of factors. Among them, local factors accounted for 70% and systemic factors accounted for 30%. PVT is divided into acute and chronic PVT. But currently, there is no definitive definition. Generally, cases with acute PVT showed no obvious collateral vessel or spongy morphosis, few splenomegaly, and few compensatory esophageal varices according to images.3 PVT was divided into acute and chronic PVT according to the onset. Patients showing clinical symptoms within 60 days before admission were considered having acute PVT.4 The clinical manifestations of acute PVT could be abdominal pain, bloating, diarrhea, blood in the stool, nausea, vomiting, loss of appetite, fever, splenomegaly, and sepsis.5 However, some of these features are extremely hidden. In severe PVT, the entire portal vein system is widely affected, and there is no good compensation for venous blood vessel formation. The unresolved obstruction might contribute to intestinal perforation, peritonitis, infectivity, 1
2 Liu et al.
shock, and multiple organ failure. When the onset enters the chronic phase, extrahepatic portal hypertension (EPH) occurred and manifested as gastrointestinal bleeding. These sequelae will seriously affect the quality of life of patients and even the survival of patients. Studies6 have shown that early anticoagulant therapy could promote PVT recanalization. The sooner the treatment was obtained, the better the anticoagulant treatment would achieve a 69% complete recanalization rate within a week. In addition, the complete recanalization rate would decrease to 25% when anticoagulant treatment was started up at the beginning of the second week.7 Therefore, early diagnosis is very important. Many scholars at home and abroad8e10 have proposed some predictive indicators for the diagnosis of PVT. But because of the nonspecific clinical manifestations, the predictive effect of these indicators obtained from different studies was still controversial. A retrospective analysis was performed in the Second Affiliated Hospital of Soochow University and Suqian People’s Hospital of Nanjing, Gulou Hospital Group in the present study. The study aimed to find suitable indicators for early clinical diagnosis and to formulate relevant clinical diagnostic models for acute symptomatic PVT. However, the traditional logistic regression analysis and other research methods were often constrained because of the small sample size. With the rapid development of modern statistics and the increasing popularity of high-speed electronic computers, linear models are becoming more and more widely used in various fields. In 1996, Tibshirani proposed the least absolute shrinkage and selection operator (LASSO).11 It is a new variable selection technique that has the advantage of being able to perform continuous selection variables and model parameter estimation compared with conventional algorithms. In 1995, Cortes and Vapnik first proposed support vector machine (SVM), which is a new machine learning technology based on statistical learning theory and different from traditional neural network technology.12 SVM has become a new trend and research hotspot in the field of machine learning. Recently, some scholars have explored the research application of LASSO and SVM methods. For studies with small sample sizes, LASSO-SVM is appropriate for exploring the clinical diagnosis model for acute symptomatic PVT.
METHODS Participants The present study was performed in the Second Affiliated Hospital of Soochow University and
Annals of Vascular Surgery
Suqian People’s Hospital of Nanjing, Gulou Hospital Group between January 2005 and June 2016. Fourty-seven patients with PVT and 94 controls without PVT were included in the study. Our study was approved by Human Participants Ethics Committee of the two hospitals. Written informed consents were attained from these participants. Inclusion Criteria and Exclusion Criteria All patients with PVT met the diagnostic criteria, which included the following: (1) All the participants aged between 18-80 years; (2) Sufficient information (clinical date such as history, clinical manifestations, physical sign, and laboratory examination) was available from medical record; (3) All patients underwent a contrast-enhanced computed tomography (CT) of the abdomen; (4) Patients were diagnosed with PVT in the main trunk or major branches of the portal vein by two senior researchers (expert in imaging). In addition, we included 94 matched controls (inpatients with abdominal symptoms and providing sufficient results of a contrast-enhanced CT). We excluded subjects who had insufficient information, renal dysfunction, persistent hypotension despite fluid resuscitation, portal vein tumor thrombus, allergic reactions to contrast medium, and pregnant woman. Statistical Analysis We applied SPSS software 20.0 (SPSS, Inc., Chicago, IL) for statistical analyses. All statistical tests were two-sided, and a P value < 0.05 was considered statistically significant. For general information, data of continuous variables were showed as mean (standard deviation [SD]), and differences were evaluated using t tests. Categorical variables were expressed as number (percentage) and tested by Pearson’s chi-squared test. We applied the general logistic regression model to calculate the odds ratios (ORs). Covariates in the model included all distinctive features (43 clinical data showed in Appendix Table I). Recently, penalized likelihood-based methods have drawn much attention.13 Because of generally numerous variables, a suitable variable selection operator should be proposed to search the most important predictive factors associated with PVT and avoid using a complex model. Thus, we applied the LASSO14,15 to select the most important predictive factors for PVT in the present study. We used the glmnet package within R version 3.1.3 to establish LASSO logistic regression model.
Volume
-, -
2019
A diagnostic prediction model of acute symptomatic PVT 3
Suppose that we have an input n p matrix XT ¼ (X1, X2, ., Xp) and aim to predict an n 1 binary response vector Y, the logistic regression is defined as log
i
p ðyi ¼ 1 jX Þ i
1 p ðyi ¼ 1 jX Þ
¼ hb X i
P where hbðXi Þ ¼ b0 þ pj ¼ 1xijbj , b0 is the intercept and bj is the parameter corresponding to xj. The LASSO estimator is defined as Xn b b ¼ argmin b i¼1 yi hb Xi log 1 Xp
þ exp hb Xi þ l j¼1 bj
where l is the tuning parameter. Making l sufficiently large will cause some of the coefficients to be exactly zero. Thus the LASSO performs a continuous subset selection. We computed parameter estimations in the LASSO via a cyclical coordinate descent algorithm.16 We applied a 10-fold cross-validation to estimate the error rate for each model. A simulation study was conducted to confirm the robustness of the LASSO method. We randomly resampled 100 sets of bootstrapped and permutated data with various sample sizes from the survey data. After that, we specified various sample sizes (n ¼ 100, 250, and 500). After conducting the LASSO logistic regression model, the significant covariates and the corresponding frequency outputs of covariates were recorded.
RESULTS General Information The PVT group including 30 men and 17 women had an average age of 46.11 ± 14.27 years. The median time from onset to clinical diagnosis was 7.00 (4.00, 12.00) days. In addition, 60 men and 34 women non-PVT group were included in the nonPVT group. The average age of them was 49.83 ± 13.99 years. Other clinical data and laboratory test results of these participants were shown in Appendix Tables IeII. Identification of PVT-related Factors The cross-validation error rates and numbers of selected covariates at a grid of values of l for the LASSO models is shown in Figure 1. According to the results of 10-fold cross-validation, the most suitable tuning parameter l that gives minimum crossvalidated error was 3.81 (log scale), the number of
Fig. 1. Ten-fold cross-validation misclassification error rates with error bars of the LASSO logistic regression model across different values of the tuning parameter l (log scale).
selected covariates was eleven, and the coefficients of covariates is shown in Table I. The corresponding regularization path is presented in Figure 2. Logistic’s b and multivariable adjusted ORs for the eleven factors and PVT were presented in Table I. Compared with no liver cirrhosis, liver cirrhosis was independently associated with increased risk of PVT. The adjusted odds ratio (OR) was 3.68. In addition, the adjusted OR was shown as follows: 2.88 for D-Dimer, 2.82 for splenomegaly, 2.58 for splenectomy, 1.89 for inherited thrombophilia, 1.49 for ascetic fluid, 1.17 for recent history of abdominal surgery, 1.09 for bloating, 1.01 for Creactive protein (CRP), 1.00 for albumin, and 0.56 for abdominal tenderness. The LASSO-SVM model achieved a sensitivity of 91.5% (95% confidence interval [CI]: 78.7%, 97.2%), a specificity of 100.0% (95% CI: 95.1%, 100.0%), a positive predictive value of 100.0% (95% CI: 89.8%, 100.0%), and a negative predictive value of 95.9% (95% CI: 89.3%, 98.7%). Receiver operating characteristics (ROC) curve is shown in Figure 3.
DISCUSSION The present study indicated that acute symptomatic PVT was associated with 11 indicators, including liver cirrhosis, D-Dimer, splenomegaly, splenectomy, inherited thrombophilia, ascetic fluid, history of abdominal surgery, bloating, CRP, albumin, and abdominal tenderness. The LASSO-SVM model achieved a sensitivity of 91.5% and a specificity of 100.0%.
4 Liu et al.
Annals of Vascular Surgery
Table I. The estimates of factors from the logistic regression model and associations between these factors and PVT Number
Items
x1 a6 x15 x2 x3 x9 x4
Liver cirrhosis D-Dimer Splenomegaly Splenectomy Inherited thrombophilia Ascitic fluid Recent history of abdominal surgery Bloating CRP Albumin Abdominal tenderness
x10 a5 a16 x11
Logistic’s b
OR
1.3031 1.0577 1.0367 0.8075 0.6377 0.3975 0.1610
3.6810 2.8797 2.8200 2.5803 1.8920 1.4881 1.1746
0.0845 0.0090 0.0028 0.5832
1.0882 1.0090 0.9972 0.5581
CRP, C-reactive protein; OR, Odds ratio; PVT, portal vein thrombosis.
Fig. 2. The path of the parameter estimated over a grid of values for l.
According to previous studies, both the LASSO algorithm and the SVM algorithm were widely used in the studies of small sample data. The LASSO algorithm is superior to the traditional regression methods in the variable selection. However, the LASSO algorithm could not be used for nonlinear models. The combination of SVM algorithm and LASSO algorithm could be applied to deal with nonlinear problems. Thus, the present study combined the LASSO algorithm and the SVM algorithm to explore the clinical diagnosis model for acute symptomatic PVT. In the present study, almost half of patients with PVT had liver cirrhosis (a rate of 44.7%), whereas few of the control subjects had liver cirrhosis (a rate of 5.3%). Previous studies indicated that patients with liver cirrhosis are more prone to PVT,
Fig. 3. ROC curves are represented for LASSO-SVM model. LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristics; SVM, support vector machine.
with a PVT rate of approximately 1% in patients with compensated cirrhosis,17 and 8e25% in patients awaiting liver transplantation.18 The mechanisms included the following: (1) local factors: intrahepatic structural changes, increased portal venous resistance, slowed blood flow, lymphangitis around the portal vein; (2) systemic factors: anticoagulation, coagulation imbalance, promotes blood coagulation; (3) other congenital or acquired factors.19 Normally, no bleeding or thrombosis occurs when the blood coagulation function is still balanced.20 But when severe liver damage occurs, the production of thrombin is affected,21 causing a hypercoagulable state that eventually causes PVT. The present study showed that liver cirrhosis might be one of the main predictive factors of PVT. PVT has been thought to be a rare complication of splenectomy.22 Recent studies9,23 have found a higher incidence of PVT after splenectomy in patients with cirrhotic portal hypertension as imaging levels increase. The exact cause of PVT after splenectomy is unclear. The causes include blood flow in the portal vein after splenectomy changes, the hypercoagulable state, the blind end formation after splenic vein ligation,24 the pathological changes of local blood vessels, local inflammatory response, and unreasonable use of hemostatic drugs. In addition, PVT might also be caused by platelet elevation after splenectomy.25,26 PVT after other abdominal surgery except splenectomy rarely happens. Recently, reports showed PVT rates of 5%e16% and 1%e2% after liver transplantation and pancreaticoduodenectomy, respectively.27e29 Cases with thrombophilia have genetic or acquired defects in coagulation-related genes, leading
Volume
-, -
2019
to a higher tendency for thromboembolism.30 PVT and deep venous thrombosis of lower extremities, coronary heart disease, and cerebral infarction are all thrombotic diseases. We speculate that these thrombotic diseases may have the same pathological mechanism. Thus, the history of cardiovascular and cerebrovascular diseases was included as risk factors in the study. However, no significant differences in history of cardiovascular and cerebrovascular diseases were detected between PVT group and nonPVT group in the present study. CRP is an acute phase protein of hepatocyte synthesis, which is associated with inflammation. It is often used to diagnose and identify acute infectious diseases; monitor postoperative infections; observe the efficacy of antibiotics; and determine the course of disease and prognosis. In general, CRP begins to rise for several hours after the onset of inflammation, reaching a peak in about 48 hours. After the lesion subsides and the tissue structure and function recover, CRP can gradually drop to normal levels. Clinically, CRP is widely detected, and it is gratifying that radiotherapy and chemotherapy and corticosteroid treatment do not affect the results of CRP. CRP can be a good predictor when judging the extent of PVT thrombosis, the severity of symptoms, etc.31 PVT is one of the manifestations of venous thromboembolism events in visceral vessels (in rare or special areas). CRP could be elevated in PVT, or as an indicator of its severity. Recent studies suggest that the occurrence of venous thromboembolism is associated with infection, inflammation, and immunity.32 So we speculate that CRP may be included as a diagnostic predictor of PVT in this study. The increase in plasma concentration of D-dimer reflects the secondary fibrinolysis process in the body. Ddimer is a specific marker in the fibrinolysis process. An increase in D-dimer means the presence of secondary fibrinolysis in the body. It has been confirmed to be a valid predictor in the diagnosis of venous thromboembolism events.32 The level of D-dimer concentration can be used to assess the clinical severity of PVT and to predict the prognosis of PVT.33 In summary, we introduce machine learning algorithms into medical research and solve the problem. When the sample size is small, there are many independent variables. Eleven variables were selected by LASSO variable selection, which has been discussed in the previous researches. The results of the selected variables are basically consistent with the risk factors described in the previous literature. It indicates that the LASSO algorithm can be used to filter the variables for the PVT diagnostic model. The model obtained by the joint
A diagnostic prediction model of acute symptomatic PVT 5
method of LASSO screening variable output into SVM has very high sensitivity and specificity. Machine learning algorithms have the advantages of traditional statistical methods, such as handling a large number of high-dimensional data. The PVT diagnostic prediction model that we finally get is not static and can be regularly revised based on new data so that machine learning continues to evolve. There are some limitations in the present study. Firstly, the study is a retrospective study; data collected from the hospitalized medical records of patients might lack of some data, such as P-selectin, thrombus precursor protein, which were important indicators. Secondly, the sample size was limited.
CONCLUSIONS We developed a LASSO-SVM model to diagnose PVT. We demonstrated that the model achieved a sensitivity of 91.5% and a specificity of 100.0%.
SUPPLEMENTARY DATA Supplementary data related to this article can be found at https://doi.org/10.1016/j.avsg.2019.04. 037. REFERENCES 1. Ogren M, Bergqvist D, Bjorck M, et al. Portal vein thrombosis: prevalence, patient characteristics and lifetime risk: a population study based on 23,796 consecutive autopsies. World J Gastroenterol 2006;12:2115e9. 2. Qi X, Bai M, Yang Z, et al. Occlusive portal vein thrombosis as a new marker of decompensated cirrhosis. Med hypotheses 2011;76:522e6. 3. Chawla Y, Duseja A, Dhiman RK. Review article: the modern management of portal vein thrombosis. Aliment Pharmacol Ther 2009;30:881e94. 4. Chen H, Lv Y, Han G. Anticoagulation for portal vein thrombosis in liver cirrhosis: not only recanalize the portal vein. Gastroenterology 2018;154:758e9. 5. Sogaard KK, Astrup LB, Vilstrup H, et al. Portal vein thrombosis; risk factors, clinical presentation and treatment. BMC Gastroenterol 2007;7:34. 6. Hoekstra J, Janssen HL. Vascular liver disorders (II): portal vein thrombosis. Neth J Med 2009;67:46e53. 7. Ponziani FR, Zocco MA, Campanale C, et al. Portal vein thrombosis: insight into physiopathology, diagnosis, and treatment. World J Gastroenterol 2010;16:143e55. 8. Wang J, Chai JS, Zhang YM. Anticoagulation treatment of portal vein thrombosis in a patient with cirrhosis awaiting liver transplantation: a case report. Medicine 2018;97: e11183. 9. Wu S, Wu Z, Zhang X, et al. The incidence and risk factors of portal vein system thrombosis after splenectomy and pericardial devascularization. Turkish J Gastroenterol 2015;26: 423e8.
6 Liu et al.
10. Dai J, Qi X, Peng Y, et al. Association between D-dimer level and portal venous system thrombosis in liver cirrhosis: a retrospective observational study. Int J Clin Exp Med 2015;8:15296e301. 11. Tibshirani RJ. Regression shrinkage and selection via the LASSO. J R Stat Soc B J R Stat Soc 1996;58:267e88. 12. Cortes C, Vapnik V. Support-vector networks. Machine Learn 1995;20:273e97. 13. Chen Q, Wang S. Variable selection for multiply-imputed data with application to dioxin exposure study. Stat Med 2013;32:3646e59. 14. Tibshirani R. Regression shrinkage and selection via the lasso: a retrospective. J R Stat Soc 2011;73:273e82. 15. Guo P, Xu X, Huang B, et al. Blood lead levels and associated factors among children in Guiyu of China: a populationbased study. PloS one 2014;9:e105470. 16. Friedman J, Hastie T, Tibshirani R. Pathwise coordinate optimization. Ann Appl Stat 2007;1:302e32. 17. Okuda K, Ohnishi K, Kimura K, et al. Incidence of portal vein thrombosis in liver cirrhosis. An angiographic study in 708 patients. Gastroenterology 1985;89:279e86. 18. Francoz C, Belghiti J, Vilgrain V, et al. Splanchnic vein thrombosis in candidates for liver transplantation: usefulness of screening and anticoagulation. Gut 2005;54:691e7. 19. Danila M, Sporea I, Popescu A, et al. Portal vein thrombosis in liver cirrhosis - the added value of contrast enhanced ultrasonography. Med ultrasonography 2016;18:218e33. 20. Monroe DM, Hoffman M. The coagulation cascade in cirrhosis. Clin Liver Dis 2009;13:1e9. 21. Tripodi A, Primignani M, Chantarangkul V, et al. Thrombin generation in patients with cirrhosis: the role of platelets. Hepatology 2006;44:440e5. 22. Ellison EC, Fabri PJ. Complications of splenectomy. Etiology, prevention, and management. Surg Clin North America 1983;63:1313e30. 23. Kinjo N, Kawanaka H, Akahoshi T, et al. Risk factors for portal venous thrombosis after splenectomy in patients
Annals of Vascular Surgery
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
with cirrhosis and portal hypertension. Br J Surg 2010;97:910e6. Ikeda M, Sekimoto M, Takiguchi S, et al. High incidence of thrombosis of the portal venous system after laparoscopic splenectomy: a prospective study with contrast-enhanced CT scan. Ann Surg 2005;241:208e16. Francoz C, Valla D, Durand F. Portal vein thrombosis, cirrhosis, and liver transplantation. J Hepatol 2012;57: 203e12. Zocco MA, Di Stasio E, De Cristofaro R, et al. Thrombotic risk factors in patients with liver cirrhosis: correlation with MELD scoring system and portal vein thrombosis development. J Hepatol 2009;51:682e9. Smoot RL, Christein JD, Farnell MB. Durability of portal venous reconstruction following resection during pancreaticoduodenectomy. J Gastrointest Surg 2006;10:1371e5. Ansari D, Ansorge C, Andren-Sandberg A, et al. Portal venous system thrombosis after pancreatic resection. World J Surg 2013;37:179e84. Rodriguez-Castro KI, Porte RJ, Nadal E, et al. Management of nonneoplastic portal vein thrombosis in the setting of liver transplantation: a systematic review. Transplantation 2012;94:1145e53. Donadini MP, Dentali F, Ageno W. Splanchnic vein thrombosis: new risk factors and management. Thromb Res 2012;129(Suppl 1):S93e6. Lisman T, Kamphuisen PW, Northup PG, et al. Established and new-generation antithrombotic drugs in patients with cirrhosis - possibilities and caveats. J Hepatol 2013;59:358e66. Zhou L, Mao Y, Wang L, et al. Comparison of peripheral blood T lymphocyte immune function among venous thromboembolism patients with and without infection and patients with simple infection. Int J Clin Exp Med 2015;8: 6585e91. Wells PS, Anderson DR, Rodger M, et al. Evaluation of D-dimer in the diagnosis of suspected deep-vein thrombosis. New Engl J Med 2003;349:1227e35.
Background: The aim of this study was to develop a diagnostic prediction model to improve identification of acute symptomatic portal vein thrombosis (PVT). Methods: We examined 47 patients with PVT and 94 controls without PVT in the Second Affiliated Hospital of Soochow University and Suqian People’s Hospital of Nanjing, Gulou Hospital Group. We constructed a prediction model by using a support vector machine (SVM) classifier coupled with a least absolute shrinkage and selection operator (LASSO). We applied a 10-fold cross-validation to estimate the error rate for each model. Results: The present study indicated that acute symptomatic PVT was associated with 11 indicators, including liver cirrhosis, D-Dimer, splenomegaly, splenectomy, inherited thrombophilia, ascetic fluid, history of abdominal surgery, bloating, C-reactive protein (CRP), albumin, and abdominal tenderness. The LASSO-SVM model achieved a sensitivity of 91.5% and a specificity of 100.0%. Conclusions: We developed a LASSO-SVM model to diagnose PVT. We demonstrated that the model achieved a sensitivity of 91.5% and a specificity of 100.0%.
INTRODUCTION According to epidemiological evidence, portal vein thrombosis (PVT) showed a worldwide prevalence of 1.0%. Of patients with PVT, 28% had cirrhosis, 23% primary and 44% secondary hepatobiliary malignancy, 10% major abdominal infectious or inflammatory disease, and 3% had a myeloproliferative disorder. The highest PVT risk was showed in
1 Department of Vascular Surgery, The Affiliated Suqian Hospital of Xuzhou Medical University, Suqian People’s Hospital affiliated to Nanjing Drama Tower Hospital Group, Suqian, Jiangsu, China. 2
Department of Vascular Surgery, The Second Affiliated Hospital of Soochow University, Suzhou, Jiangsu, China. 3 Department of Vascular Surgery, Affiliated Nanjing Drum Tower Hospital of Nanjing University Medical School, Nanjing, China.
Correspondence to: Xiaoqiang Li, Department of Vascular Surgery, Affiliated Nanjing Drum Tower Hospital of Nanjing University Medical School, 321 Zhongshan Road, Gulou District, Nanjing City, Jiangsu Province 210000, China; E-mail: [email protected] Ann Vasc Surg 2019; -: 1–6 https://doi.org/10.1016/j.avsg.2019.04.037 Ó 2019 Elsevier Inc. All rights reserved. Manuscript received: December 19, 2018; manuscript accepted: April 22, 2019; published online: - - -
patients with both cirrhosis and hepatic carcinoma.1 PVT has been regarded as an important marker of decompensated cirrhosis.2 In general, PVT was usually caused by a combination of factors. Among them, local factors accounted for 70% and systemic factors accounted for 30%. PVT is divided into acute and chronic PVT. But currently, there is no definitive definition. Generally, cases with acute PVT showed no obvious collateral vessel or spongy morphosis, few splenomegaly, and few compensatory esophageal varices according to images.3 PVT was divided into acute and chronic PVT according to the onset. Patients showing clinical symptoms within 60 days before admission were considered having acute PVT.4 The clinical manifestations of acute PVT could be abdominal pain, bloating, diarrhea, blood in the stool, nausea, vomiting, loss of appetite, fever, splenomegaly, and sepsis.5 However, some of these features are extremely hidden. In severe PVT, the entire portal vein system is widely affected, and there is no good compensation for venous blood vessel formation. The unresolved obstruction might contribute to intestinal perforation, peritonitis, infectivity, 1
2 Liu et al.
shock, and multiple organ failure. When the onset enters the chronic phase, extrahepatic portal hypertension (EPH) occurred and manifested as gastrointestinal bleeding. These sequelae will seriously affect the quality of life of patients and even the survival of patients. Studies6 have shown that early anticoagulant therapy could promote PVT recanalization. The sooner the treatment was obtained, the better the anticoagulant treatment would achieve a 69% complete recanalization rate within a week. In addition, the complete recanalization rate would decrease to 25% when anticoagulant treatment was started up at the beginning of the second week.7 Therefore, early diagnosis is very important. Many scholars at home and abroad8e10 have proposed some predictive indicators for the diagnosis of PVT. But because of the nonspecific clinical manifestations, the predictive effect of these indicators obtained from different studies was still controversial. A retrospective analysis was performed in the Second Affiliated Hospital of Soochow University and Suqian People’s Hospital of Nanjing, Gulou Hospital Group in the present study. The study aimed to find suitable indicators for early clinical diagnosis and to formulate relevant clinical diagnostic models for acute symptomatic PVT. However, the traditional logistic regression analysis and other research methods were often constrained because of the small sample size. With the rapid development of modern statistics and the increasing popularity of high-speed electronic computers, linear models are becoming more and more widely used in various fields. In 1996, Tibshirani proposed the least absolute shrinkage and selection operator (LASSO).11 It is a new variable selection technique that has the advantage of being able to perform continuous selection variables and model parameter estimation compared with conventional algorithms. In 1995, Cortes and Vapnik first proposed support vector machine (SVM), which is a new machine learning technology based on statistical learning theory and different from traditional neural network technology.12 SVM has become a new trend and research hotspot in the field of machine learning. Recently, some scholars have explored the research application of LASSO and SVM methods. For studies with small sample sizes, LASSO-SVM is appropriate for exploring the clinical diagnosis model for acute symptomatic PVT.
METHODS Participants The present study was performed in the Second Affiliated Hospital of Soochow University and
Annals of Vascular Surgery
Suqian People’s Hospital of Nanjing, Gulou Hospital Group between January 2005 and June 2016. Fourty-seven patients with PVT and 94 controls without PVT were included in the study. Our study was approved by Human Participants Ethics Committee of the two hospitals. Written informed consents were attained from these participants. Inclusion Criteria and Exclusion Criteria All patients with PVT met the diagnostic criteria, which included the following: (1) All the participants aged between 18-80 years; (2) Sufficient information (clinical date such as history, clinical manifestations, physical sign, and laboratory examination) was available from medical record; (3) All patients underwent a contrast-enhanced computed tomography (CT) of the abdomen; (4) Patients were diagnosed with PVT in the main trunk or major branches of the portal vein by two senior researchers (expert in imaging). In addition, we included 94 matched controls (inpatients with abdominal symptoms and providing sufficient results of a contrast-enhanced CT). We excluded subjects who had insufficient information, renal dysfunction, persistent hypotension despite fluid resuscitation, portal vein tumor thrombus, allergic reactions to contrast medium, and pregnant woman. Statistical Analysis We applied SPSS software 20.0 (SPSS, Inc., Chicago, IL) for statistical analyses. All statistical tests were two-sided, and a P value < 0.05 was considered statistically significant. For general information, data of continuous variables were showed as mean (standard deviation [SD]), and differences were evaluated using t tests. Categorical variables were expressed as number (percentage) and tested by Pearson’s chi-squared test. We applied the general logistic regression model to calculate the odds ratios (ORs). Covariates in the model included all distinctive features (43 clinical data showed in Appendix Table I). Recently, penalized likelihood-based methods have drawn much attention.13 Because of generally numerous variables, a suitable variable selection operator should be proposed to search the most important predictive factors associated with PVT and avoid using a complex model. Thus, we applied the LASSO14,15 to select the most important predictive factors for PVT in the present study. We used the glmnet package within R version 3.1.3 to establish LASSO logistic regression model.
Volume
-, -
2019
A diagnostic prediction model of acute symptomatic PVT 3
Suppose that we have an input n p matrix XT ¼ (X1, X2, ., Xp) and aim to predict an n 1 binary response vector Y, the logistic regression is defined as log
i
p ðyi ¼ 1 jX Þ i
1 p ðyi ¼ 1 jX Þ
¼ hb X i
P where hbðXi Þ ¼ b0 þ pj ¼ 1xijbj , b0 is the intercept and bj is the parameter corresponding to xj. The LASSO estimator is defined as Xn b b ¼ argmin b i¼1 yi hb Xi log 1 Xp
þ exp hb Xi þ l j¼1 bj
where l is the tuning parameter. Making l sufficiently large will cause some of the coefficients to be exactly zero. Thus the LASSO performs a continuous subset selection. We computed parameter estimations in the LASSO via a cyclical coordinate descent algorithm.16 We applied a 10-fold cross-validation to estimate the error rate for each model. A simulation study was conducted to confirm the robustness of the LASSO method. We randomly resampled 100 sets of bootstrapped and permutated data with various sample sizes from the survey data. After that, we specified various sample sizes (n ¼ 100, 250, and 500). After conducting the LASSO logistic regression model, the significant covariates and the corresponding frequency outputs of covariates were recorded.
RESULTS General Information The PVT group including 30 men and 17 women had an average age of 46.11 ± 14.27 years. The median time from onset to clinical diagnosis was 7.00 (4.00, 12.00) days. In addition, 60 men and 34 women non-PVT group were included in the nonPVT group. The average age of them was 49.83 ± 13.99 years. Other clinical data and laboratory test results of these participants were shown in Appendix Tables IeII. Identification of PVT-related Factors The cross-validation error rates and numbers of selected covariates at a grid of values of l for the LASSO models is shown in Figure 1. According to the results of 10-fold cross-validation, the most suitable tuning parameter l that gives minimum crossvalidated error was 3.81 (log scale), the number of
Fig. 1. Ten-fold cross-validation misclassification error rates with error bars of the LASSO logistic regression model across different values of the tuning parameter l (log scale).
selected covariates was eleven, and the coefficients of covariates is shown in Table I. The corresponding regularization path is presented in Figure 2. Logistic’s b and multivariable adjusted ORs for the eleven factors and PVT were presented in Table I. Compared with no liver cirrhosis, liver cirrhosis was independently associated with increased risk of PVT. The adjusted odds ratio (OR) was 3.68. In addition, the adjusted OR was shown as follows: 2.88 for D-Dimer, 2.82 for splenomegaly, 2.58 for splenectomy, 1.89 for inherited thrombophilia, 1.49 for ascetic fluid, 1.17 for recent history of abdominal surgery, 1.09 for bloating, 1.01 for Creactive protein (CRP), 1.00 for albumin, and 0.56 for abdominal tenderness. The LASSO-SVM model achieved a sensitivity of 91.5% (95% confidence interval [CI]: 78.7%, 97.2%), a specificity of 100.0% (95% CI: 95.1%, 100.0%), a positive predictive value of 100.0% (95% CI: 89.8%, 100.0%), and a negative predictive value of 95.9% (95% CI: 89.3%, 98.7%). Receiver operating characteristics (ROC) curve is shown in Figure 3.
DISCUSSION The present study indicated that acute symptomatic PVT was associated with 11 indicators, including liver cirrhosis, D-Dimer, splenomegaly, splenectomy, inherited thrombophilia, ascetic fluid, history of abdominal surgery, bloating, CRP, albumin, and abdominal tenderness. The LASSO-SVM model achieved a sensitivity of 91.5% and a specificity of 100.0%.
4 Liu et al.
Annals of Vascular Surgery
Table I. The estimates of factors from the logistic regression model and associations between these factors and PVT Number
Items
x1 a6 x15 x2 x3 x9 x4
Liver cirrhosis D-Dimer Splenomegaly Splenectomy Inherited thrombophilia Ascitic fluid Recent history of abdominal surgery Bloating CRP Albumin Abdominal tenderness
x10 a5 a16 x11
Logistic’s b
OR
1.3031 1.0577 1.0367 0.8075 0.6377 0.3975 0.1610
3.6810 2.8797 2.8200 2.5803 1.8920 1.4881 1.1746
0.0845 0.0090 0.0028 0.5832
1.0882 1.0090 0.9972 0.5581
CRP, C-reactive protein; OR, Odds ratio; PVT, portal vein thrombosis.
Fig. 2. The path of the parameter estimated over a grid of values for l.
According to previous studies, both the LASSO algorithm and the SVM algorithm were widely used in the studies of small sample data. The LASSO algorithm is superior to the traditional regression methods in the variable selection. However, the LASSO algorithm could not be used for nonlinear models. The combination of SVM algorithm and LASSO algorithm could be applied to deal with nonlinear problems. Thus, the present study combined the LASSO algorithm and the SVM algorithm to explore the clinical diagnosis model for acute symptomatic PVT. In the present study, almost half of patients with PVT had liver cirrhosis (a rate of 44.7%), whereas few of the control subjects had liver cirrhosis (a rate of 5.3%). Previous studies indicated that patients with liver cirrhosis are more prone to PVT,
Fig. 3. ROC curves are represented for LASSO-SVM model. LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristics; SVM, support vector machine.
with a PVT rate of approximately 1% in patients with compensated cirrhosis,17 and 8e25% in patients awaiting liver transplantation.18 The mechanisms included the following: (1) local factors: intrahepatic structural changes, increased portal venous resistance, slowed blood flow, lymphangitis around the portal vein; (2) systemic factors: anticoagulation, coagulation imbalance, promotes blood coagulation; (3) other congenital or acquired factors.19 Normally, no bleeding or thrombosis occurs when the blood coagulation function is still balanced.20 But when severe liver damage occurs, the production of thrombin is affected,21 causing a hypercoagulable state that eventually causes PVT. The present study showed that liver cirrhosis might be one of the main predictive factors of PVT. PVT has been thought to be a rare complication of splenectomy.22 Recent studies9,23 have found a higher incidence of PVT after splenectomy in patients with cirrhotic portal hypertension as imaging levels increase. The exact cause of PVT after splenectomy is unclear. The causes include blood flow in the portal vein after splenectomy changes, the hypercoagulable state, the blind end formation after splenic vein ligation,24 the pathological changes of local blood vessels, local inflammatory response, and unreasonable use of hemostatic drugs. In addition, PVT might also be caused by platelet elevation after splenectomy.25,26 PVT after other abdominal surgery except splenectomy rarely happens. Recently, reports showed PVT rates of 5%e16% and 1%e2% after liver transplantation and pancreaticoduodenectomy, respectively.27e29 Cases with thrombophilia have genetic or acquired defects in coagulation-related genes, leading
Volume
-, -
2019
to a higher tendency for thromboembolism.30 PVT and deep venous thrombosis of lower extremities, coronary heart disease, and cerebral infarction are all thrombotic diseases. We speculate that these thrombotic diseases may have the same pathological mechanism. Thus, the history of cardiovascular and cerebrovascular diseases was included as risk factors in the study. However, no significant differences in history of cardiovascular and cerebrovascular diseases were detected between PVT group and nonPVT group in the present study. CRP is an acute phase protein of hepatocyte synthesis, which is associated with inflammation. It is often used to diagnose and identify acute infectious diseases; monitor postoperative infections; observe the efficacy of antibiotics; and determine the course of disease and prognosis. In general, CRP begins to rise for several hours after the onset of inflammation, reaching a peak in about 48 hours. After the lesion subsides and the tissue structure and function recover, CRP can gradually drop to normal levels. Clinically, CRP is widely detected, and it is gratifying that radiotherapy and chemotherapy and corticosteroid treatment do not affect the results of CRP. CRP can be a good predictor when judging the extent of PVT thrombosis, the severity of symptoms, etc.31 PVT is one of the manifestations of venous thromboembolism events in visceral vessels (in rare or special areas). CRP could be elevated in PVT, or as an indicator of its severity. Recent studies suggest that the occurrence of venous thromboembolism is associated with infection, inflammation, and immunity.32 So we speculate that CRP may be included as a diagnostic predictor of PVT in this study. The increase in plasma concentration of D-dimer reflects the secondary fibrinolysis process in the body. Ddimer is a specific marker in the fibrinolysis process. An increase in D-dimer means the presence of secondary fibrinolysis in the body. It has been confirmed to be a valid predictor in the diagnosis of venous thromboembolism events.32 The level of D-dimer concentration can be used to assess the clinical severity of PVT and to predict the prognosis of PVT.33 In summary, we introduce machine learning algorithms into medical research and solve the problem. When the sample size is small, there are many independent variables. Eleven variables were selected by LASSO variable selection, which has been discussed in the previous researches. The results of the selected variables are basically consistent with the risk factors described in the previous literature. It indicates that the LASSO algorithm can be used to filter the variables for the PVT diagnostic model. The model obtained by the joint
A diagnostic prediction model of acute symptomatic PVT 5
method of LASSO screening variable output into SVM has very high sensitivity and specificity. Machine learning algorithms have the advantages of traditional statistical methods, such as handling a large number of high-dimensional data. The PVT diagnostic prediction model that we finally get is not static and can be regularly revised based on new data so that machine learning continues to evolve. There are some limitations in the present study. Firstly, the study is a retrospective study; data collected from the hospitalized medical records of patients might lack of some data, such as P-selectin, thrombus precursor protein, which were important indicators. Secondly, the sample size was limited.
CONCLUSIONS We developed a LASSO-SVM model to diagnose PVT. We demonstrated that the model achieved a sensitivity of 91.5% and a specificity of 100.0%.
SUPPLEMENTARY DATA Supplementary data related to this article can be found at https://doi.org/10.1016/j.avsg.2019.04. 037. REFERENCES 1. Ogren M, Bergqvist D, Bjorck M, et al. Portal vein thrombosis: prevalence, patient characteristics and lifetime risk: a population study based on 23,796 consecutive autopsies. World J Gastroenterol 2006;12:2115e9. 2. Qi X, Bai M, Yang Z, et al. Occlusive portal vein thrombosis as a new marker of decompensated cirrhosis. Med hypotheses 2011;76:522e6. 3. Chawla Y, Duseja A, Dhiman RK. Review article: the modern management of portal vein thrombosis. Aliment Pharmacol Ther 2009;30:881e94. 4. Chen H, Lv Y, Han G. Anticoagulation for portal vein thrombosis in liver cirrhosis: not only recanalize the portal vein. Gastroenterology 2018;154:758e9. 5. Sogaard KK, Astrup LB, Vilstrup H, et al. Portal vein thrombosis; risk factors, clinical presentation and treatment. BMC Gastroenterol 2007;7:34. 6. Hoekstra J, Janssen HL. Vascular liver disorders (II): portal vein thrombosis. Neth J Med 2009;67:46e53. 7. Ponziani FR, Zocco MA, Campanale C, et al. Portal vein thrombosis: insight into physiopathology, diagnosis, and treatment. World J Gastroenterol 2010;16:143e55. 8. Wang J, Chai JS, Zhang YM. Anticoagulation treatment of portal vein thrombosis in a patient with cirrhosis awaiting liver transplantation: a case report. Medicine 2018;97: e11183. 9. Wu S, Wu Z, Zhang X, et al. The incidence and risk factors of portal vein system thrombosis after splenectomy and pericardial devascularization. Turkish J Gastroenterol 2015;26: 423e8.
6 Liu et al.
10. Dai J, Qi X, Peng Y, et al. Association between D-dimer level and portal venous system thrombosis in liver cirrhosis: a retrospective observational study. Int J Clin Exp Med 2015;8:15296e301. 11. Tibshirani RJ. Regression shrinkage and selection via the LASSO. J R Stat Soc B J R Stat Soc 1996;58:267e88. 12. Cortes C, Vapnik V. Support-vector networks. Machine Learn 1995;20:273e97. 13. Chen Q, Wang S. Variable selection for multiply-imputed data with application to dioxin exposure study. Stat Med 2013;32:3646e59. 14. Tibshirani R. Regression shrinkage and selection via the lasso: a retrospective. J R Stat Soc 2011;73:273e82. 15. Guo P, Xu X, Huang B, et al. Blood lead levels and associated factors among children in Guiyu of China: a populationbased study. PloS one 2014;9:e105470. 16. Friedman J, Hastie T, Tibshirani R. Pathwise coordinate optimization. Ann Appl Stat 2007;1:302e32. 17. Okuda K, Ohnishi K, Kimura K, et al. Incidence of portal vein thrombosis in liver cirrhosis. An angiographic study in 708 patients. Gastroenterology 1985;89:279e86. 18. Francoz C, Belghiti J, Vilgrain V, et al. Splanchnic vein thrombosis in candidates for liver transplantation: usefulness of screening and anticoagulation. Gut 2005;54:691e7. 19. Danila M, Sporea I, Popescu A, et al. Portal vein thrombosis in liver cirrhosis - the added value of contrast enhanced ultrasonography. Med ultrasonography 2016;18:218e33. 20. Monroe DM, Hoffman M. The coagulation cascade in cirrhosis. Clin Liver Dis 2009;13:1e9. 21. Tripodi A, Primignani M, Chantarangkul V, et al. Thrombin generation in patients with cirrhosis: the role of platelets. Hepatology 2006;44:440e5. 22. Ellison EC, Fabri PJ. Complications of splenectomy. Etiology, prevention, and management. Surg Clin North America 1983;63:1313e30. 23. Kinjo N, Kawanaka H, Akahoshi T, et al. Risk factors for portal venous thrombosis after splenectomy in patients
Annals of Vascular Surgery
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
with cirrhosis and portal hypertension. Br J Surg 2010;97:910e6. Ikeda M, Sekimoto M, Takiguchi S, et al. High incidence of thrombosis of the portal venous system after laparoscopic splenectomy: a prospective study with contrast-enhanced CT scan. Ann Surg 2005;241:208e16. Francoz C, Valla D, Durand F. Portal vein thrombosis, cirrhosis, and liver transplantation. J Hepatol 2012;57: 203e12. Zocco MA, Di Stasio E, De Cristofaro R, et al. Thrombotic risk factors in patients with liver cirrhosis: correlation with MELD scoring system and portal vein thrombosis development. J Hepatol 2009;51:682e9. Smoot RL, Christein JD, Farnell MB. Durability of portal venous reconstruction following resection during pancreaticoduodenectomy. J Gastrointest Surg 2006;10:1371e5. Ansari D, Ansorge C, Andren-Sandberg A, et al. Portal venous system thrombosis after pancreatic resection. World J Surg 2013;37:179e84. Rodriguez-Castro KI, Porte RJ, Nadal E, et al. Management of nonneoplastic portal vein thrombosis in the setting of liver transplantation: a systematic review. Transplantation 2012;94:1145e53. Donadini MP, Dentali F, Ageno W. Splanchnic vein thrombosis: new risk factors and management. Thromb Res 2012;129(Suppl 1):S93e6. Lisman T, Kamphuisen PW, Northup PG, et al. Established and new-generation antithrombotic drugs in patients with cirrhosis - possibilities and caveats. J Hepatol 2013;59:358e66. Zhou L, Mao Y, Wang L, et al. Comparison of peripheral blood T lymphocyte immune function among venous thromboembolism patients with and without infection and patients with simple infection. Int J Clin Exp Med 2015;8: 6585e91. Wells PS, Anderson DR, Rodger M, et al. Evaluation of D-dimer in the diagnosis of suspected deep-vein thrombosis. New Engl J Med 2003;349:1227e35.