Memoirs

A predictive model of response to peginterferon ribavirin in chronic hepatitis C using classification and regression tree analysis

Description
A predictive model of response to peginterferon ribavirin in chronic hepatitis C using classification and regression tree analysis
Categories
Published
of 10
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  Original Article A predictive model of response to peginterferon ribavirinin chronic hepatitis C using classification and regressiontree analysis hepr_607 251..260 Masayuki Kurosaki, 1 Kotaro Matsunaga, 2 Itsuko Hirayama, 1 Tomohiro Tanaka, 1 Mitsuaki Sato, 1 Yutaka Yasui, 1 Nobuharu Tamaki, 1 Takanori Hosokawa, 1 Ken Ueda, 1 Kaoru Tsuchiya, 1 Hiroyuki Nakanishi, 1 Hiroki Ikeda, 1 Jun Itakura, 1 Yuka Takahashi, 1 Yasuhiro Asahina, 1 Megumu Higaki, 4 Nobuyuki Enomoto 3 and Namiki Izumi 1 1 Division of Gastroenterology and Hepatology and  2 Division of Pathology, Musashino Red Cross Hospital, Tokyo, 3 First Department of Internal Medicine, University of Yamanashi, Yamanashi, and  4 Department of Medical Science, Jikei Medical University, Tokyo, Japan Aim:  Early disappearance of serum hepatitis C virus (HCV)RNA is the prerequisite for achieving sustained virologicalresponse (SVR) in peg-interferon (PEG-IFN) plus ribavirin (RBV)therapy for chronic hepatitis C. This study aimed to developa decision tree model for the pre-treatment prediction ofresponse.  Methods:  Genotype1bchronichepatitisCtreatedwithPEG-IFNalpha-2bandRBVwerestudied.Predictivefactorsofrapidor complete early virological response (RVR/cEVR) wereexplored in 400 consecutive patients using a recursive parti-tioning analysis, referred to as classification and regressiontree (CART) and validated. Results:  CART analysis identified hepatic steatosis ( < 30%)as the first predictor of response followed by low-density-lipoprotein cholesterol (LDL-C) ( 3 100 mg/dL), age ( < 50 and < 60 years), blood sugar ( < 120 mg/dL), and gamma-glutamyltransferase (GGT) ( < 40 IU/L) and built decision treemodel. The model consisted of seven groups with variableresponse rates from low (15%) to high (77%). The reproducibil-ity of the model was confirmed by the independent validationgroup ( r   2 =  0.987). When reconstructed into three groups, therate of RVR/cEVR was 16% for low probability group, 46% forintermediate probability group and 75% for high probabilitygroup. Conclusions:  A decision tree model that includes hepaticsteatosis, LDL-C, age, blood sugar, and GGT may be useful forthe prediction of response before PEG-IFN plus RBV therapy,and has the potential to support clinical decisions in selectingpatients for therapy and may provide a rationale for treatingmetabolic factors to improve the efficacy of antiviral therapy. Key words:  data mining ,  decision tree ,  HCV , low-density-lipoprotein-cholesterol ,  steatosis INTRODUCTION C OMBINATION THERAPY WITH pegylated inter-feron (PEG-IFN) and ribavirin (RBV) is now recog-nized as a standard treatment for patients with chronic hepatitis C. 1 However, the rate of sustained virologicalresponse (SVR) to 48 weeks of PEG-IFN RBV combina-tion therapy is only 50% in patients with hepatitis C virus (HCV) genotype 1b and high HCV RNA titer, socalled difficult to treat chronic hepatitis C patients. 2,3  Within this difficult to treat group, the response to treat-ment sometimes can be highly heterogeneous for cases which are apparently equivalent in HCV RNA titer,making the prediction of response before treatment adifficult task. It has been suggested that early virologicalresponse (EVR), defined as either undetectable HCV RNAora2logdropinHCVRNAatweek12,isareliablemeans to predict SVR. 2,4 More recently, it has been sug-gested that patients with a rapid virological response(RVR: undetectable HCV RNA at week 4) and a com-plete EVR (cEVR: undetectable HCV RNA at week 12) Correspondence: Dr Namiki Izumi, Division of Gastroenterology andHepatology, Musashino Red Cross Hospital, 1-26-1 Kyonan-cho, Musashino-shi, Tokyo 180-8610, Japan. Email:nizumi@musashino.jrc.or.jpReceived 26 May 2009; revision 25 August 2009; accepted 26 August  2009. Hepatology Research  2010;  40 : 251–260 doi: 10.1111/j.1872-034X.2009.00607.x  ©  2010 The Japan Society of Hepatology   251  achieve high SVR rates, while patients with a partial EVR (pEVR: 2 log drop in HCV RNA but still detectable at  week 12) have lower rates of SVR. 5 Since PEG-IFN RBV combination therapy is costly and accompanied by potential adverse effects, the ability to predict the pos-sibility of RVR or cEVR before therapy and identifying curablepatientsmaysignificantlyinfluencetheselectionof patients for therapy. Moreover, identification of base-line predictors of poor response is particularly impor-tant to establish a rationale for identifying therapeutic targets to improve the efficacy of antiviral therapy.Data mining is a method of predictive analysis whichexplorestremendousvolumesofdatatodiscoverhiddenpatterns and relationships in highly complex datasetsand enables the development of predictive models. Theclassificationandregressiontree(CART)analysisisacorecomponentofthedecisiontreetoolfordataminingandpredictive modeling, 6 is deployed to decision makers in various fields of business, and currently is being used inthe area of biomedicine. 7–13  The results of CART analysisare presented as a decision tree, which is intuitive andfacilitates the allocation of patients into subgroups by followingtheflow-chartform. 14 CARThasbeenshowntobe competitive with other traditional statistical tech-niques such as logistic regression analysis. 15 In the present study, we used the CART analysis toexplore baseline predictors of response to PEG-IFN plusRBV therapy among clinical, biochemical, virologicaland histological pretreatment variables and to define apre-treatment algorithm to discriminate chronic hepati-tis C patients who are likely to respond to PEG-IFN plusRBV therapy. MATERIALS AND METHODSPatients  A   TOTAL OF 419 chronic hepatitis C patients weretreated with PEG-IFN alpha-2b and RBV at MusashinoRedCrossHospitalbetweenDecember2001and December 2007. Among them, 400 patients whofulfilled the following inclusion criteria were enrolled inthe present study. (i) infection by genotype 1b (ii) HCV RNA higher than 100 KIU/mL by quantitative PCR (Cobas Amplicor HCV Monitor, Roche Diagnostic systems, CA) which is usually used for the definition of high viral load in Japan (iii) lack of co-infection withhepatitis B virus or human immunodeficiency virus (iv)lack of other causes of liver disease such as autoimmunehepatitis, primary biliary cirrhosis, or alcohol intake of more than 20 g per day, and (v) having completed at least 12 weeks of therapy with an early virologicalresponse that could be evaluated. Patients received PEG-IFN alpha-2b (1.5 microgram/kg) subcutaneously every  week and were administered a weight adjusted doseof RBV (600 mg for   < 60 kg, 800 mg for 60–80 kg, and1000 mg for   > 80 kg) which is the recommended dosagein Japan. Data from two third of patients (269 patients) were used for the model building set and the remaining one third of patients (131 patients) were used as a vali-dation set. Consent in writing was obtained from eachpatient and the study protocol conformed to the ethicalguidelines of the 1975 Declaration of Helsinki and wasapproved by the institutional review committee. Laboratory tests Blood samples were obtained before therapy, and at least once every month during therapy and analyzed for hematologic tests, blood chemistries, and HCV RNA. Inthe present study, RVR and cEVR was defined as unde-tectable HCV RNA by qualitative PCR with a lower detection limit of 50 IU/mL (Amplicor, Roche Diagnos-tic systems, CA) at week 4 and 12, respectively. SVR wasdefined as undetectable HCV RNA at week 24 after thecompletion of therapy. Histological examination For all patients, liver biopsy specimens were obtainedbefore therapy and were evaluated independently by three pathologists who were blinded to the clinicaldetails. If there was a disagreement, the scores assignedby the majority of pathologists were used for the analy-sis. Fibrosis and activity were scored according to theMETAVIR scoring system. 16 Fibrosis was staged on ascale of 0–4: F0 (no fibrosis), F1 (mild fibrosis: portalfibrosis without septa), F2 (moderate fibrosis: few septa), F3 (severe fibrosis: numerous septa without cir-rhosis) and F4 (cirrhosis). Activity of necroinflamma-tion was graded on a scale of 0–3: A0 (no activity), A1(mild activity), A2 (moderate activity) and A3 (severeactivity). Percentage of steatosis was quantified by deter-mining the average proportion of hepatocytes affectedby steatosis and graded on a scale of 0–3: grade 0 (nosteatosis), grade 1 (0–9%), grade 2 (10–29%), andgrade 3 (over 30%) as we reported previously. 17 Database for analysis  A pretreatment database of 72 variables was createdcontaining histological findings (grade of fibrosis, activ-ity,andsteatosis),laboratorytestsincludingthequantity of HCV RNA by Cobas Amplicor, and clinical informa-tion (age, gender, body weight, and body mass index). 252  M. Kurosaki  et al. Hepatology Research  2010;  40 : 251–260 ©  2010 The Japan Society of Hepatology    The baseline characteristics and test results are listed in Table 1. The overall rate of RVR/cEVR was 43% in themodel building set and 48% in the validation set. There were no significant differences in the clinical back-grounds between these two groups. Hepatitis C viralmutations, such as mutations in interferon-sensitivity determining region or core amino acid residues 70 and91,werenotincludedinthepresentanalysis.Thedataset of laboratory tests was based on the digitized records inthishospital.Continuousdatawassplitintocategorizeddata by increment of 10; For example, age was catego-rized into  < 30, 30–39, 40–49, 50–59, 60–69, and  3 70. Statistical analysis Based on this database, the recursive partitioning analy-sis algorithm referred to as CART was implemented todefine meaningful subgroups of patients with respect to the possibility of achieving RVR/cEVR. The CART belongs to a family of nonparametric regressionmethods based on binary recursive partitioning of data. Thesoftwareautomaticallyexplorethedatatosearchfor optimal split variables, builds a decision tree structureand finally classifies all subjects into particular sub-groups that are homogeneous with respect to theoutcome of interest. 18 During the CART analysis, first,the entire study population, and thereafter, all newly defined subgroups, were investigated at every step of theanalysis to determine which variable at what cut-off point yielded the most significant division into twoprognosticsubgroupsthatwereashomogeneousaspos-siblewithrespecttoestimatesofRVR/cEVRpossibilities. This algorithm uses the impurity function (Gini crite-rion function) for splitting. 19  A restriction was imposedon the tree construction such that terminal subgroupsresulting from any given split must have at least 20patients. The CART procedure stopped when either noadditional significant variable was detected or when thesample size was below 20. The resulting final subgroups were most homogeneous with respect to the probability of achieving RVR/cEVR. For this analysis, data mining software Clementine version 12.0 (SPSS Inc, Chicago,IL) was utilized. SPSS 15.0 (SPSS Inc, Chicago, IL) wasused for logistic regression analysis. RESULTSFactors associated with RVR/cEVR bystandard statistical analysis  W  E FIRST ANALYZED 72 variables by univariateand multivariate logistic regression analysis tofind factors associated with RVR/cEVR (Table 2).Patients with RVR/cEVR were significantly younger thanthose without. Among histological findings, grade of steatosis and stage of fibrosis was significantly lower inRVR/cEVR. Among hematologic tests, hemoglobin andhematcrit was significantly higher in RVR/cEVR. Among blood chemistry tests, creatinine and low-density lipo-protein cholesterol (LDL-C) was significantly higher and gamma-glutamyltransferase (GGT), low-density-lipoprotein cholesterol (LDL-C), and blood sugar weresignificantly lower in RVR/cEVR. The level of HCV RNA  was significantly lower in RVR/cEVR. There were nosignificant differences in other tests.Multivariate logistic regression analysis was per-formed on age, fibrosis stage, steatosis, HCVRNA, crea-tinine, hemoglobin, GGT, LDL-C, and blood sugar:hematcrit was not included since it is closely associated with hemoglobin. On multivariate analysis, age, gradeof steatosis, level of HCV RNA, creatinine, hemoglobin,GGT, and LDL-cholesterol remained significant whereasstage of fibrosis, hemoglobin and blood sugar were not. The CART analysis  The CART analysis was carried out on the model build-ing set of 269 patients using the same variables aslogistic regression analysis. Figure 1 shows the resulting decision tree. The CART analysis automatically selectedfive predictive variables to produce a total of seven sub-groups of patients. The grade of steatosis was selected asthe variable of initial split with an optimal cut-off of 30%. The possibility of achieving RVR/cEVR was only 18% for patients with hepatic steatosis of 30% or morecompared to 47% for patients with hepatic steatosis of less than 30%. Among patients with hepatic steatosisof less than 30%, the level of serum LDL-C, with anoptimal cut-off of 100 mg/dL, was selected as the vari-able of second split. Patients with higher LDL-C levelhad the higher probability of RVR/cEVR (57% vs. 32%). Among patients with LDL-C of less than 100 mg/dL,age, with an optimal cut-off of 60, was selected as thethird variable of split. Younger patients had thehigher probability of RVR/cEVR (49% vs. 15%). Among patients younger than 60, the blood sugar, with anoptimal cut-off of 120 mg/dL, was selected as the forth variable of split. Patients with lower blood sugar levelhad the higher probability of RVR/cEVR (71% vs. 31%). Among patients with hepatic steatosis of less than 30%and LDL-C of 100 mg/dL or more, age, with an optimalcut-off of 50, was selected as the third variable of split, younger being the predictor of higher RVR/cEVR prob-ability (77% vs. 50%). Among patients older than 50, Hepatology Research  2010;  40 : 251–260 Prediction model of response to peg-IFN and RBV   253 ©  2010 The Japan Society of Hepatology    Table 1  Clinical characteristics of patientsModel set  n  =  269 Validation set  n  =  131 P  -valueSex (M/F) 127/142 55/76 0.325 Age (years) 57.7  1  10.1 57.6  1  10.0 0.932Body weight (kg) 59.6  1  11.0 57.5  1  9.5 0.094Body mass index (kg/m 2 ) 23.2  1  3.1 23.3  1  3.8 0.934 Total protein (g/dL) 7.6  1  0.5 7.7  1  0.6 0.558 Albumin (g/dL) 4.2  1  0.3 4.2  1  0.3 0.349Globulin (g/dL) 3.4  1  0.5 3.4  1  0.6 0.989 Aspartate aminotransferase (IU/L) 58.1  1  43.1 55.8  1  37.5 0.601 Alanine aminotransferase (IU/L) 70.9  1  49.2 66.4  1  52.6 0.462Gamma-glutamyltransferase (IU/L) 49.6  1  44.0 45.2  1  34.4 0.33Lactate dehydrogenase (IU/L) 289.3  1  112.3 301.5  1  109.3 0.417 Total bilirubin (mg/dL) 0.71  1  0.28 0.69  1  0.23 0.317Direct bilirubin (mg/dL) 0.23  1  0.12 0.25  1  0.10 0.147Indirect bilirubin (mg/dL) 0.48  1  0.21 0.44  1  0.16 0.064 Alkaline phosphatase (IU/L) 290.9  1  107.6 292.5  1  107.6 0.917Leucine aminopeptidase (IU/L) 64.3  1  14.3 65.5  1  12.3 0.543 Thymol turbidity test (KU) 7.1  1  3.4 8.0  1  3.7 0.062Zinc sulfate turbidity test (KU) 15.4  1  4.9 16.3  1  5.4 0.188Choline esterase (IU/L) 318.1  1  81.7 321.1  1  78.1 0.798 Ammonia (microg/dL) 39.7  1  20.2 45.0  1  15.6 0.668Blood sugar (mg/dL) 125.9  1  41.1 117.4  1  47.9 0.081Glycohemoglobin (%) 5.6  1  1.6 5.4  1  1.2 0.797 Total cholesterol (mg/dL) 170.8  1  33.9 175.6  1  36.8 0.170Low-density-lipoprotein-cholesterol (mg/dL) 96.5  1  25.2 100.9  1  28.5 0.153High-density-lipoprotein-cholesterol (mg/dL) 54.2  1  15.9 55.2  1  17.4 0.612 Triglyceride (mg/dL) 108.5  1  47.8 102.8  1  46.4 0.306Creatinine (mg/dL) 0.72  1  0.15 0.74  1  0.17 0.236Urea nitrogen (mg/dL) 14.1  1  3.4 14.9  1  3.9 0.123Uric acid (mg/dL) 5.3  1  1.2 5.2  1  1.2 0.715Sodium (mEq/L) 142.2  1  2.0 142.4  1  2.0 0.471Potassium (mEq/L) 4.3  1  0.3 4.3  1  0.4 0.578Chloride (mEq/L) 104.0  1  2.2 104.0  1  2.6 0.905Calcium (mg/dL) 9.1  1  0.4 9.2  1  0.4 0.479Phosphorus (mg/dL) 3.5  1  0.5 3.5  1  0.6 0.814Magnesium (mg/dL) 2.2  1  0.2 2.3  1  0.3 0.390 Amylase (IU/L) 178.7  1  125.8 175.1  1  133.1 0.118Creatine kinase (IU/L) 114.9  1  147.6 119.3  1  73.7 0.849Iron (microg/dL) 104.7  1  53.2 109  1  37 0.726Ferritin (ng/mL) 111.3  1  103.3 59.7  1  118.5 0.405C-reactive peptide (mg/dL) 0.2  1  1.1 0.1  1  0.1 0.586Immunoglobulin G (mg/dL) 1849  1  426 1988  1  525 0.129Immunoglobulin M (mg/dL) 141  1  69 205  1  106 0.200Immunoglobulin A (mg/dL) 323  1  675 291  1  81 0.784 Triiodothyronine (pg/mL) 2.3  1  0.3 2.2  1  0.3 0.358 Thyroxin (ng/dL) 0.9  1  0.1 0.9  1  0.1 0.872 Thyroid stimulating hormone (micro IU/mL) 1.8  1  1.4 1.7  1  0.7 0.939 White blood cell count (/microl) 5243  1  1591 5286  1  1101 0.843Segmented neutrophils (%) 55.4  1  10.8 57.0  1  10.0 0.297Band neutrophils (%) 1.5  1  1.6 0.5  1  0.6 0.250Eosinophils (%) 2.9  1  2.3 2.4  1  1.4 0.127 254  M. Kurosaki  et al. Hepatology Research  2010;  40 : 251–260 ©  2010 The Japan Society of Hepatology   the level of GGT, with an optimal cutoff of 40 U/L, werethen selected as the fourth level of split, low levels being the predictor of higher RVR/cEVR probability (60% vs.35%). All five factors selected as significant variables in theCART analysis were also significantly associated withRVR/cEVR by univariate analysis (Table 2). In addition,steatosis, LDL-C, age and GGT were also independently   Table 1  Continued Model set  n  =  269 Validation set  n  =  131 P  -valueBasophiles (%) 0.6  1  0.4 0.6  1  0.3 0.727Lymphocytes (%) 34.6  1  9.6 34.0  1  9.3 0.682Monocytes (%) 6.6  1  2.2 6.2  1  2.6 0.149Red blood cell count (10 4  /microl) 458  1  43 455  1  47 0.643Hemoglobin (g/dL) 14.4  1  1.5 14.5  1  1.5 0.618Hematcrit (%) 42.7  1  4.0 42.9  1  4.4 0.717Reticulocytes (%) 1.4  1  0.4 1.4  1  0.4 0.762Mean corpuscular volume (fL) 93.3  1  4.5 93.8  1  5.41 0.466Mean corpuscular hemoglobin concentration (pg) 31.5  1  1.9 31.7  1  2.3 0.583Mean corpuscular hemoglobin concentration (g/dL) 33.8  1  0.9 33.7  1  1.3 0.910Platelets (10 4  /microl) 16.8  1  5.4 16.3  1  4.5 0.480Prothrombin time (s) 11.7  1  1.2 11.7  1  0.9 0.762Prothrombin time (activity %) 104.6  1  14.4 102.6  1  14.8 0.363Prothrombin time (international normalized ratio) 1.0  1  0.1 1.0  1  0.1 0.387 Thrombin time (%) 97.2  1  31.3 109  1  31.5 0.231 Activated partial thromboplastin time (s) 29.7  1  4.4 29.1  1  2.7 0.260Hepaplastin test (%) 97.8  1  20.3 95.4  1  19.4 0.523Fibrinogen (%) 237  1  44 225  1  45 0.069Hepatitis C virus RNA ( < 850/ 3 850 KIU/mL) 130/139 70/61 0.394Histological grade of  Activity (A1/A2/A3) 138/107/24 62/55/14 0.714Fibrosis (F1/F2/F3/F4) 135/74/57/3 58/40/27/6 0.131Steatosis (0%/1–9%/10–29%/30%  ) 89/109/37/34 49/45/21/16 0.643Hepatitis C virus RNA negative at week 12 ( yes/no) 116/153 63/68 0.349  Table 2  Factors associated with rapid or complete early virological response by univariate and multivariate logistic regressionanalysisParameter Category Univariate MultivariateOdds 95% CI  P  -value Odds 95% CI  P  -value Age (years)  < 50 vs.  3 50 2.65 1.51–4.65  < 0.001 2.03 1.04–3.97 0.039Fibrosis stage F1-2 vs. F3-4 2.47 1.31–4.66 0.005 1.77 0.85–3.68 0.120Steatosis (%)  < 30 vs.  3 30 4.11 1.64–10.29 0.003 2.88 1.07–7.79 0.037Hepatitis C virus RNA (KIU/mL)  < 850 vs.  3 850 1.97 1.21–3.22 0.007 1.93 1.09–3.43 0.025Creatinine (mg/dL)  3 0.8 vs.  < 0.8 3.30 1.96–5.56  < 0.001 3.54 1.88–6.67  < 0.001Hemoglobin (g/dL)  3 14.5 vs.  < 14.5 1.76 1.08–2.87 0.023 1.38 0.74–2.57 0.320Hematcrit (%)  3 43 vs.  < 43 1.75 1.07–2.84 0.003Gamma-glutamyltransferase (IU/L)  < 40 vs.  3 40 2.06 1.26–3.37 0.004 2.45 1.32–4.56 0.005Low-density-lipid cholesterol (mg/dL)  3 100 vs.  < 100 2.71 1.61–4.55  < 0.001 2.21 1.21–4.06 0.010Blood sugar (mg/dL)  < 120 vs.  3 120 2.00 1.02–3.95 0.045 1.42 0.64–3.13 0.390 CI, confidence interval. Hepatology Research  2010;  40 : 251–260 Prediction model of response to peg-IFN and RBV   255 ©  2010 The Japan Society of Hepatology 
Search
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks