Sách Cancer clinical trials current and controversial issues in design and analysis ( PDFDrive com )

Một cuốn sách hay về thử nghiệm lâm sàng trong lĩnh vực ung thư. Sách gồm các phần: 1 Endpoints for Cancer Clinical Trials Stephen L. George, Xiaofei Wang, and Herbert Pang 1.1 Introduction 1.2 Overall Survival 1.3 Endpoints Based on Tumor Measurements 1.3.1 RECIST Criteria 1.3.2 Response Rate as Primary Endpoints 1.3.3 Tumor Response as Continuous Variable 1.4 Progressionfree Survival and Other Composite Endpoints 1.5 Surrogate Endpoints 1.5.1 Definition 1.5.2 Surrogate Endpoint Validation 1.5.3 Remaining Issues 1.6 Patientreported Outcomes 1.6.1 Patientreported Outcomes 1.6.2 Types of PRO for Treatment Comparisons 1.6.3 Health Status, Functional and Symptoms Outcomes 1.6.4 General and Cancerspecific Quality of Life Outcomes 1.6.5 Criteria Used for PRO Instruments Selection 1.6.6 Reliability 1.6.7 Validity1.6.8 Responsiveness of Instruments to Change 1.7 Promising New Approaches 1.7.1 Limitations of Traditional Endpoints 1.7.2 Pharmacokinetic and Pharmacodynamics Responses 1.7.3 Imaging Techniques 1.7.4 Immune Biomarkersbased Endpoints 1.7.5 Criteria for Evaluating Biomarkerbased Endpoints 1.8 Summary 2 Use of Historical Data Simon Wandel, Heinz Schmidli, and Beat Neuenschwander 2.1 Introduction 2.2 Overview of Approaches for Incorporating Historical Data 2.2.1 Introduction 2.2.2 Metaanalytic Approaches 2.2.3 Robust Metaanalyticpredictive Priors and Priordata Conflict 2.2.4 Prior Effective Sample Size 2.3 Applications 2.3.1 Application 1: A Randomized Phase II Trial Using Historical Control Data 2.3.2 Application 2: Design of a Japanese Dose Escalation Study Incorporating Data from Western Patients 2.3.3 Application 3: Noninferiority and Biosimilar Trials 2.4 Discussion 2.5 Appendix 3 Multiplicity Dong Xi, Ekkehard Glimm, and Frank Bretz 3.1 Introduction to Multiplicity Issues 3.1.1 Sources of Multiplicity 3.1.2 Types of Error Rates 3.1.3 Why Multiplicity Adjustment 3.1.4 A Motivating Example 3.2 Common Multiple Comparison Procedures 3.2.1 General Concepts3.2.2 Methods Based on Univariate pValues 3.2.2.1 Methods Based on the Bonferroni Test 3.2.2.2 Methods Based on the Simes Test 3.2.2.3 Numerical Illustration 3.2.3 Parametric Methods 3.2.3.1 Dunnett Test 3.2.3.2 Multiple Testing in Linear Models 3.3 Advanced Multiple Comparison Procedures 3.3.1 Graphical Approaches 3.3.2 Gatekeeping Procedures 3.3.3 Group Sequential Procedures 3.3.3.1 Group Sequential Procedures with Multiple Hypotheses 3.3.3.2 Group Sequential Procedures with a Timetoevent Endpoint 3.3.4 Adaptive Designs 3.4 Applications 3.4.1 Multiple Comparison Procedure in the BELLE2 Trial 3.4.2 Comparison with a Common Control in Timetoevent Trials 3.5 Concluding Remarks 4 Analysis of Safety Data Steven Snapinn and Qi Jiang 4.1 Introduction 4.2 Phase I Clinical Trials 4.2.1 Phase I Designs 4.3 Planning Safety Analyses 4.3.1 Events of Interest 4.3.2 The Statistical Analysis Plan (SAP) 4.3.3 The Program Safety Analysis Plan (PSAP) 4.3.4 Data Monitoring Committee (DMC) 4.4 Safety Signal Detection 4.4.1 Classifying Adverse Events 4.4.2 Statistical Methods for Late Phase Trials 4.4.3 Postmarketing Signal Detection 4.4.4 Singlearm Trials and Combination Studies 4.4.5 Safety Noninferiority Trials4.5 Collecting, Summarizing, and Displaying Safety Data 4.5.1 Data Collection 4.5.2 Reporting Safety Information 4.5.3 Graphical Approaches 4.6 Metaanalysis of Safety Data 4.7 Benefitrisk Analysis 4.8 Summary II Early Phase Clinical Trials 5 Development and Validation of Predictive Signatures Michael C. Sachs and Lisa M. McShane 5.1 Introduction 5.1.1 Prognostic and Predictive Omics Signatures 5.2 Signature Development 5.2.1 Assay Development and Validation 5.2.2 Statistical Development 5.2.3 Iteration and Refinement 5.2.4 Performance Metrics 5.2.5 Estimation of Performance Metrics 5.2.6 Computational Reproducibility 5.2.7 Practical Considerations 5.3 Clinical Utility Assessment 5.3.1 How Omics Signatures Are Used in Clinical Trials 5.3.2 Evaluating Clinical Utility 5.3.3 Power and Sample Size Considerations 5.4 Summary 6 Phase I Trials and Dosefinding Mark R. Conaway and Nolan A. Wages 6.1 Background 6.2 Methods for a Single Cytotoxic Agent 6.2.1 Rulebased Designs 6.2.1.1 The Standard or 3+3 Design 6.2.1.2 Storer’s 2s tage Designs6.2.1.3 Biased Coin Designs 6.2.2 Methods Based on Toxicity Probability Intervals 6.3 Modelbased Methods 6.3.1 The Continual Reassessment Method 6.3.2 Escalation with Overdose Control (EWOC) 6.3.3 EWOC and CRM 6.3.4 Bayesian 2Parameter Logistic Models 6.3.5 Which Method to Use? 6.4 Timetoevent Toxicity Outcomes 6.5 Ordinal Outcomes 6.5.1 Rulebased Methods 6.5.2 Modelbased Methods 6.5.3 Toxicity Scores 6.6 Dose Expansion Cohorts 6.7 Dosefinding Based on Safety and Efficacy 6.8 Combinations of Agents 6.8.1 Assumption of a Single Ordering 6.8.2 Specifying Multiple Possible Orderings 6.8.3 Use of More Flexible Models 6.8.4 Finding Multiple MTDCs 6.9 Patient Heterogeneity 6.10 Noncytotoxic Agents 6.10.1 Locating the OBD 6.11 Summary 7 Design and Analysis of Phase II Cancer Clinical Trials SinHo Jung 7.1 Introduction 7.2 Singlearm Phase II Trials 7.2.1 Optimal Twostage Designs 7.2.2 Estimation of Response Rate 7.2.3 Confidence Interval 7.2.4 pValue Calculation 7.3 Randomized Phase II Trials 7.3.1 Singlestage Design7.3.2 Twostage Design 7.3.2.1 Choice of a1 and a 7.3.2.2 Choice of n1 and n2 7.3.3 Numerical Studies 7.4 Discussion III Late Phase Clinical Trials 8 Sample Size for Survival Trials in Cancer Edward Lakatos 8.1 Introduction 8.2 Departures from Proportionality 8.2.1 Treatment Lag 8.2.2 Treatment Antilag 8.2.3 Both Lag and Antilag 8.2.4 Sample Size Implications 8.2.4.1 Implications for Treatment Lag — Real World Example 8.2.4.2 Exploring the Implications of Treatment Lag and Antilag in a Controlled Setting 8.2.4.3 Sample Size and Power Calculations 8.2.4.4 Treatment Antilag 8.3 Two Paradigms for Which Conventional Wisdom Fails 8.3.1 Eventdriven Trial 8.3.2 Groupsequential Sample Size Inflation Factor 8.4 Sample Size Reestimation and Futility 8.4.1 Estimating the Treatment Effect in a Trial with a Threshold Treatment Lag 8.4.2 Increasing the Sample Size When There Is a Treatment Lag 8.4.3 Interaction between Weighted Statistics and Nonproportional Hazards 8.4.4 Estimating the Treatment Effect in a Trial with a Threshold Treatment Antilag 8.4.5 Sample Size Reestimation in the Presence of Treatment Lag or Antilag: Concluding Remark 8.4.6 Conditional Power, Current Trends, and Nonproportional Hazards 8.5 How the Markov Model Works8.5.1 Introduction 8.5.2 The Exponential Model for Calculating Cumulative Survival Probabilities 8.5.3 The Lifetable Approach to Calculating Cumulative Survival Probabilities 8.5.4 The Markov Model Approach to Calculating Cumulative Survival Probabilities 8.5.4.1 2State Markov Model: At Risk, Failure 8.5.4.2 3State Markov Model: At Risk, Failure, Loss 8.5.4.3 4State Markov Model: At Risk, Failure, Loss, ODIS (Noncompliance) 8.5.5 Using the Markov Model to Calculate Sample Sizes for the Logrank Statistic 8.5.6 Speed and Accuracy 8.6 Discussion and Conclusions 9 Noninferiority Trials Rajeshwari Sridhara and Thomas Gwise 9.1 Introduction 9.2 Endpoint Selection 9.3 Methods for Evaluating the Active Control Effect and Selecting the Noninferiority Margin 9.3.1 Fixed Margin 9.3.2 Synthesis Approach 9.3.3 Bayesian Approach 9.3.4 Placebocontrolled Approach 9.4 Sample Size Determination 9.4.1 Ratio of Proportions 9.4.2 Survival Endpoints 9.5 Interim Monitoring and Analyses 9.6 Multiple Comparisons 9.6.1 Testing of Noninferiority to Superiority and Superiority to Noninferiority 9.7 Missing Data and Noncompliance 9.8 Statistical Inference and Reporting 9.9 Summary10 Quality of Life Diane Fairclough 10.1 Introduction 10.2 Measures of HRQoL 10.3 QOL as an Endpoint in Cancer Trials 10.4 Multiple Endpoints 10.4.1 Summary Measures and Statistics 10.4.2 Multiple Comparisons Adjustments and Gatekeeping Strategies 10.5 Informative Missing Data Due to Dropout 10.5.1 Methods to Be Avoided 10.5.2 Recommended Approach 10.5.3 Sensitivity Analyses 10.5.4 QOL after Death 10.5.5 QALYs and QTWiST 10.5.6 How Much Data Can Be Missing? 10.6 Sample Size or Power Estimation 10.7 Summary IV Personalized Medicine 11 Biomarkerbased Clinical Trials Edward L. Korn and Boris Freidlin 11.1 Introduction 11.2 Analytic Performance of a Biomarker 11.3 Prognostic and Predictive Biomarkers 11.4 Biomarkers in Phase I Trials 11.5 Biomarkers in Phase II Trials 11.5.1 Trials without a Control Arm 11.5.2 Randomized Screening Trials with a Control Arm 11.6 Biomarkers in Phase III Trials 11.6.1 Biomarkers with Compelling Credentials 11.6.2 Biomarkers with Strong Credentials 11.6.2.1 Subgroupspecific Testing Strategies 11.6.2.2 Biomarkerpositive and Overall Strategies 11.6.2.3 Marker Sequential Test Design11.6.2.4 Sample Size Considerations 11.6.3 Biomarkers with Weak Credentials 11.6.4 Interim Monitoring 11.6.5 Retrospective Biomarker Analysis of Phase III Trial Data 11.6.6 Biomarkerstrategy Designs 11.7 Summary 12 Adaptive Clinical Trial Designs in Oncology J. Jack Lee and Lorenzo Trippa 12.1 Introduction 12.2 History of Adaptive Designs 12.3 Bayesian Framework and Its Use in Clinical Trials 12.4 Adaptive Dosefinding Designs for Identifying Optimal Biologic Dose 12.5 Multistage Designs, Group Sequential Designs, Interim Analysis, Early Stopping for Toxicity, Efficacy, or Futility 12.6 Sample Size Reestimation 12.7 Adaptive Randomization, Individual Ethics versus Group Ethics 12.8 Seamless Designs 12.9 Biomarkerguided Adaptive Designs 12.10 Multiarm Adaptive Designs 12.11 Master Protocols, Umbrella Trials, Basket Trials, and Platformbased Designs 12.12 Examples of Trials with Adaptive Designs — Lessons for Design and Conduct 12.13 Software for Adaptive Designs 12.14 Discussion 13 Dynamic Treatment Regimes Marie Davidian, Anastasios (Butch) Tsiatis, and Eric Laber 13.1 Introduction 13.2 Characterization of Treatment Regimes 13.2.1 Decision Rules and Regimes 13.2.2 Classes of Treatment Regimes 13.3 Potential Outcomes Framework 13.3.1 Single Decision 13.3.2 Multiple Decisions13.4 Sequential, Multiple Assignment, Randomized Trials 13.4.1 Data for Studying Dynamic Treatment Regimes 13.4.2 Considerations for SMARTs 13.4.3 Inference on Embedded Regimes in a SMART 13.5 Thinking in Terms of Dynamic Treatment Regimes 13.6 Optimal Treatment Regimes for Personalized Medicine 13.6.1 Characterizing an Optimal Regime 13.6.2 Regressionbased Estimation of an Optimal Regime 13.6.3 Alternative Methods 13.7 Discussion

Current and Controversial Issues in Design and Analysis

Chapman & Hall/CRC Biostatistics Series Chief

Editor-in-Shein-Chung Chow, Ph.D., Professor, Department of Biostatistics and

Bioinformatics, Duke University School of Medicine, Durham, North CarolinaSeries Editors

Byron Jones, Biometrical Fellow, Statistical Methodology, Integrated

Information Sciences, Novartis Pharma AG, Basel, Switzerland Jen-pei Liu,

Professor, Division of Biometry, Department of Agronomy, National Taiwan

University, Taipei, Taiwan Karl E Peace, Georgia Cancer Coalition,

Distinguished Cancer Scholar, Senior Research Scientist and Professor ofBiostatistics, Jiann-Ping Hsu College of Public Health, Georgia Southern

University, Statesboro, Georgia Bruce W Turnbull, Professor, School of

Operations Research and Industrial Engineering, Cornell University, Ithaca,New York Published Titles

Andreas Sashegyi, James Felli, and Rebecca Noel Benefit-Risk Assessment

Methods in Medical Product Development: Bridging Qualitative and Quantitative Assessments

Biosimilars: Design and Analysis of Follow-on Biologics

Shein-Chung Chow

Cancer Clinical Trials

Current and Controversial Issues in Design and Analysis

Except as permitted under U.S Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access

www.copyright.com ( http://www.copyright.com/ ) or contact the Copyright Clearance Center, Inc (CCC),

222 Rosewood Drive, Danvers, MA 01923, 978-750-8400 CCC is a not-for-profit organization that provides licenses and registration for a variety of users For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged.

To Ed Gehan and Marvin Zelen, who early in my career taught me much about the art and science of cancer biostatistics and served as role models while

– Herbert Pang

1.6.7 Validity

2.3.1 Application 1: A Randomized Phase II Trial Using HistoricalControl Data

2.3.2 Application 2: Design of a Japanese Dose Escalation StudyIncorporating Data from Western Patients

3.2.2.1 Methods Based on the Bonferroni Test3.2.2.2 Methods Based on the Simes Test3.2.2.3 Numerical Illustration

3.2.3 Parametric Methods

3.2.3.1 Dunnett Test3.2.3.2 Multiple Testing in Linear Models3.3 Advanced Multiple Comparison Procedures

3.3.1 Graphical Approaches

3.3.2 Gatekeeping Procedures

3.3.3 Group Sequential Procedures

3.3.3.1 Group Sequential Procedures with Multiple Hypotheses3.3.3.2 Group Sequential Procedures with a Time-to-event Endpoint3.3.4 Adaptive Designs

6.2.1.3 Biased Coin Designs6.2.2 Methods Based on Toxicity Probability Intervals

7.3.2.1 Choice of a1 and a 7.3.2.2 Choice of n1 and n2

a Controlled Setting8.2.4.3 Sample Size and Power Calculations8.2.4.4 Treatment Anti-lag

8.4.3 Interaction between Weighted Statistics and Non-proportional

Hazards8.4.4 Estimating the Treatment Effect in a Trial with a Threshold

Treatment Anti-lag8.4.5 Sample Size Re-estimation in the Presence of Treatment Lag orAnti-lag: Concluding Remark

8.4.6 Conditional Power, Current Trends, and Non-proportional Hazards8.5 How the Markov Model Works

8.5.2 The Exponential Model for Calculating Cumulative Survival

Probabilities8.5.3 The Life-table Approach to Calculating Cumulative Survival

Probabilities8.5.4 The Markov Model Approach to Calculating Cumulative SurvivalProbabilities

8.5.4.1 2-State Markov Model: At Risk, Failure8.5.4.2 3-State Markov Model: At Risk, Failure, Loss8.5.4.3 4-State Markov Model: At Risk, Failure, Loss, ODIS (Non-

compliance)8.5.5 Using the Markov Model to Calculate Sample Sizes for the Log-rankStatistic

9.6.1 Testing of Non-inferiority to Superiority and Superiority to Non-9.7 Missing Data and Non-compliance

9.8 Statistical Inference and Reporting

9.9 Summary

12.12 Examples of Trials with Adaptive Designs — Lessons for Design andConduct

13.4.3 Inference on Embedded Regimes in a SMART13.5 Thinking in Terms of Dynamic Treatment Regimes

13.6 Optimal Treatment Regimes for Personalized Medicine13.6.1 Characterizing an Optimal Regime

13.6.2 Regression-based Estimation of an Optimal Regime13.6.3 Alternative Methods

13.7 Discussion

Index

Preface

There are many important and controversial topics in the design and analysis ofcancer clinical trials, including adaptive approaches, biomarker-based trials, anddynamic treatment regime trials This book provides readers with a currentunderstanding of the critical issues in these and other topics with state-of-the-artapproaches Each chapter is written by well-known statisticians from academicinstitutions, regulatory agencies (FDA), the National Cancer Institute, or thepharmaceutical industry, all with extensive experience in cancer clinical trials.Examples are taken from actual cancer clinical trials The topics covered are:Endpoints for cancer clinical trials (Chapter 1)

We would like to acknowledge the partial support from the National CancerInstitute of the National Institutes of Health under award number CA142538.Stephen L George

Xiaofei Wang

Herbert Pang

in cancer throughout his career He has served on and chaired data monitoringcommittees for treatment and prevention trials in cancer and other diseases Dr.George is a Fellow of the American Statistical Association and of the Society forClinical Trials; served as Treasurer and a member of the Executive Committee

of the International Biometric Society for eight years; is a past President for theSociety for Clinical Trials; and served for four years as the biostatistician for theOncologic Drugs Advisory Committee for the Food and Drug Administration

Xiaofei Wang, Ph.D., obtained his Ph.D in Biostatistics from the University of

North Carolina at Chapel Hill He is currently an Associate Professor ofBiostatistics and Bioinformatics at Duke University School of Medicine, amember of Duke Cancer Institute (DCI), and the Director of Statistics of theAlliance Statistics and Data Center The Alliance for Clinical Trials in Oncology

is part of the NCI’s Clinical Trials Network (NCTN) He has been involved indesign and analysis of cancer clinical trials and translational studies in the pasttwelve years at Alliance, CALGB, and DCI He is an associate editor forStatistics in Biopharmaceutical Research and has served on special emphasispanels for NIH and FDA grants His methodology research is focused on thedevelopment of novel designs and methods for biomarker-integrated clinicalstudies, and methods for analyzing patient data from multiple sources

Herbert Pang, Ph.D., obtained his Ph.D in Biostatistics from Yale University

and B.A in Mathematics and Computer Science from the University of Oxford

He is an assistant professor at the School of Public Health, Li Ka Shing Faculty

of Medicine (LKSFM), the University of Hong Kong Dr Pang now holds anadjunct faculty position in the Department of Biostatistics and Bioinformatics at

Duke University He has been involved in the design, monitoring, and analysis

of cancer clinical trials, translational, and big data-omics research in cancer forthe CALGB, Alliance, Duke Cancer Institute, and LKSFM He served on theeditorial board of the Journal of Clinical Oncology Dr Pang received the YaleGraduate Fellowship, travel award/grant from the NIH, American Associationfor Cancer Research, Burroughs Wellcome Fund, and American StatisticalAssociation, as well as the US Chinese Anti-Cancer Association-Asian Fund forCancer Research Scholar Award

Biometric Research Branch Division of Cancer Treatment and DiagnosisNational Cancer Institute Bethesda, Maryland

Oncology Biometrics and Data Management Novartis Pharma AGBasel, Switzerland

General Issues

In cancer trials, the most common practice is to specify a single primaryendpoint and a primary trial objective based on this endpoint, used for setting thekey design and analysis specifications Other endpoints and objectives areusually relegated to secondary or exploratory roles However, an increasinglycommon practice is to define co-primary endpoints and objectives, necessitatingappropriate statistical adjustments for the resultant multiplicity [96].

Examples of endpoints commonly used in cancer clinical trials includeoverall survival, generally agreed to be the gold standard efficacy endpoint;tumor response rate or other endpoints based on tumor measurements; compositeendpoints such as progression-free survival and similar endpoints combiningindividual endpoints that may serve as surrogate endpoints for overall survival;patient-reported endpoints such as quality of life; and promising new approaches

for defining endpoints including pharmacokinetic and pharmacodynamicsresponses, imaging techniques, and biomarker-based endpoints These endpointsare summarized in Table 1.1 and are all discussed in more detail in the followingsections of this chapter.

The particular endpoints chosen for a clinical trial will depend on manyfactors including the phase of the trial, the cost and feasibility of assessing theendpoint, the follow-up studies planned, and other factors A simple statement of

an endpoint to be used is not sufficient to define the specific aspect of theendpoint and the specific objectives that will be addressed in the trial Forexample, comparing treatments via an endpoint such as overall survival or othertime-to-event endpoints can be assessed in terms of the entire survivaldistribution, hazard ratios, median survival time differences, survivalprobabilities at some prespecified time point (e.g., 5 years), or other measures.The exact specification of the endpoint is an important detail and will determinekey aspects of both design and analysis of the trial

TABLE 1.1: Endpoints in Cancer Clinical Trials

There have been recent efforts by regulatory authorities and others to definethe appropriate endpoints for trials of cancer treatment, both in general and forspecific diseases For example, the US Food and Drug Administration (FDA)

has issued general guidelines for industry on endpoints appropriate for use whenseeking regulatory approval for marketing of cancer drugs and biologics [98].Specific FDA guidance documents for non-small cell lung cancer (NSCLC) andfor imaging endpoints have also been published [100, 101] Similarly, aEuropean project entitled Definition for the Assessment of Time-to-eventEndpoints in CANcer trials (DATECAN) aims to provide recommendations fortime-to-event endpoints used in cancer clinical trials [10] DATECAN guidelinesfor GIST tumors and pancreatic cancer have been recently published [9, 13].Recommendations for hepatocellular carcinoma [63] and others have beenpublished.

1.2 Overall Survival

Overall survival (OS), the time from randomization (for randomized trials) orfrom trial registration (for non-randomized trials) to death from any cause, isgenerally acknowledged as the gold standard endpoint for cancer clinical trials[76] It reflects an obviously important clinical outcome, is easy to measure, anddoes not suffer from potential ascertainment or other biases existing for otherendpoints However, trials with objectives related to OS are generally quite largeand may require a lengthy follow-up period Surrogate endpoints for overallsurvival, discussed in detail below, are often used to reduce the size and duration

of a planned trial

In addition, there are other limitations and drawbacks in analysis andinterpretation when using OS as a primary endpoint in cancer trials Forexample, Phase I and II clinical trials are relatively small trials designed todetermine an appropriate dose or schedule for an agent or to detect someminimal activity and, with rare exceptions, OS would in general not be anappropriate primary endpoint for such trials Even for phase III cancer trials, inwhich OS is nearly always an important endpoint, there are difficulties Mostcancer therapies are given over a prolonged period of time, thereby increasingthe probability of non-adherence to the assigned treatment Even with excellentadherence to the originally assigned treatment regimen, cancer patients typicallymove through various disease states prior to death (e.g., patients may experience

a disease recurrence or disease progression), often requiring additionaltreatments at each change in state These treatments are often difficult orimpossible to specify fully in advance and in many settings (e.g., breast cancer)

there are effective post-trial therapies that complicate the interpretation of OS[86] Thus, a clinical trial designed to compare the OS of two initial treatmentsbecomes increasingly difficult to interpret if there are intervening additionaltreatments given prior to death An extreme example occurs when crossover orswitch to the alternative treatment is allowed at progression or relapse [110].

on the WHO criteria [109], the first internationally recognized criteria forassessment of solid tumor response Unlike the WHO criteria, which measurestumor responses based on 2D imaging, RECIST uses 1D imaging assessment todefine response To evaluate tumor changes per RECIST, tumor lesions areclassified as being target or non-target prior to study entry, and lesionscharacteristics (e.g., location), measures and the method used to assess thelesions are recorded At prespecified time intervals, pre-identified lesions arerepeatedly measured and any new identified lesions are also evaluated Theoverall tumor response at each time point is defined as follow:

No evidence of progression in any of the non-target lesions diagnosed atbaseline

Table 1.2 summarizes how overall response is determined based on tumorchanges in target, non-target, and new lesions at the patient level Best overallresponse (BOR) for a patient is defined as the best response evaluated from thestart of the treatment until disease progression/recurrence BOR represents thebest response level achieved among all overall responses with the rule that CR isbetter than PR and PR is better than SD Study endpoints to measure drugactivities can be defined on the occurrence and the durance of certain responsetypes Objective response rate (ORR) denotes the proportion of patients with atleast one CR/PR Disease control rate (DCR) is the proportion of patients with atleast one CR/PR/SD Response duration (DR) is the time from first assessment

of CR/PR until date of progression or last tumor assessment Per RECIST 1.1,when response rate and its confidence interval are reported for a clinical trial, allpatients included in a clinical trial should be assigned one of the followingcategories: 1) CR, 2) PR, 3) SD, 4) PD, 5) early death from malignant disease ortoxicity or other cause, and 6) unknown (not assessable, insufficient data) All ofthe patients who met eligibility criteria should be included in the denominatorwhen calculating the response rate Patients in response categories 4–6 should beconsidered as disease progression Supplementary analyses may be performed onvarious subsets of patients, e.g., excluding those with protocol deviations orearly death For most solid tumors, endpoints, such as progression-free survival(PFS) or time to progression (TTP), are also based on assessments of tumorchanges per RECIST For example, TTP is the time from the start of thetreatment until tumor progression (PD) with other events, including death,treated as “censored.”

TABLE 1.2: Evaluation of Best Overall Response (BOR)

In many diseases and treatment settings, there exists a strong associationbetween tumor response and progression (or survival) It is also commonlybelieved that drugs that induce tumor response are biologically active and maylead to improved survival or decreased symptoms Response rate can be assessedusing smaller trials in less time because of a bigger effect size and shorterrequired follow-up as compared to PFS and OS For these reasons, response rate

is frequently used as the primary endpoint for phase II trials with both arm and randomized designs The hope is that the trials with response rate as theprimary endpoint will lead to an early decision of whether a drug is promisingenough to warrant further investigation in phase III trials

single-The RECIST criteria for tumor response were designed primarily to assesscytotoxic agents and the appropriateness of evaluating tumor response to targetagents and immunotherapy via RECIST has been challenged In many recentclinical trials that involve molecularly targeted agents, tumor shrinkage hasfailed to translate into clear clinical benefit of patient symptom and survival.One possible reason is that the early activity of target agents, such as the EGFR-inhibitors and the VEGF-inhibitors, that occurs before tumor shrinkage mightnot be measured appropriately with conventional criteria such as RECIST [68].There are also cases of cytostatic agents that significantly prolong PFS and OSbut don’t significantly change the size of the tumor In evaluatingimmunotherapeutic agents, such as ipilimumab, there is evidence that durablemodest regressions or prolonged disease stability may be signalling a sustainedimprovement of survival, but tumor response per RECIST does not considerthese aspects of tumor response As a result, modified immune-response-relatedcriteria were proposed to optimize the assessment of tumor response or

progression to immunotherapy [71, 108] For these and other reasons, the use ofresponse rate to measure tumor response in phase II trials is becoming drug-and/or disease-specific Meanwhile, more randomized phase II trials are nowconducted with PFS or OS as primary endpoints with inflated type I error [85] tocontrol the size of the trial Alternative strategies, such as designs utilizing RR as

a co-primary endpoint with PFS [84] and randomized discontinuation designs[79], have also been proposed to address the problem of a weak connectionbetween response rate and gold standard endpoints such as PFS and OS

Response rate is rarely chosen as the primary endpoint in phase III trials Thecurrent stance of FDA is that cancer drug approval should be based on moredirect evidence of clinical benefit, such as improvements in overall survival(OS), health-related quality of life, tumor-related symptoms, and/or physicalfunctioning [76] However, exceptions do exist For example, response rate hasbeen used as the primary endpoint for FDA accelerated approval trials, wherelimited accrual is anticipated for rare diseases Complete response (CR) has alsoled to regular approval of drugs for treating acute leukemia [25], becauseresponse of significant magnitude and duration has been shown to be associatedwith longer OS in this disease

1.3.3 Tumor Response as Continuous Variable

Concerns over the high failure rate in phase III trials has led to pursuingalternatives to RECIST response as a phase II endpoint One criticism of usingresponse rate per RECIST as the phase II primary endpoint is the arbitrariness ofclassifying patients into four categories: CR, PR, SD, and PD For example,when a patient experiences a 19% versus 21% increase in disease, this indicates

a difference between SD and PD per RECIST Similarly, 29% versus 31%decrease in disease signals a difference between SD and PR In reality, suchsmall changes in SLD do not constitute a substantial change in the patient’sdisease state The arbitrariness of response classification underscores theinherent difficulty in using categorical measures to summarize tumor changes fordrug effect evaluation in clinical trials

Important information is potentially ignored when the continuum of tumorchange is categorized into groups [61] The use of tumor response as acontinuous variable has been hypothesized to be more informative thancategorical variables [3, 4] Waterfall and spider plots are data visualizationtechniques that display tumor shrinkage as observed % change of SLD frombaseline This continuous ‘measure’ provides a visual indication of the treatment

effect in reducing the tumor burden [32] As an example, CALGB 30704 is arandomized phase II trial comparing pemetrexed, sunitinib, or their combination

as second-line chemotherapy for advanced non-small cell lung cancer (NSCLC)[49] Tumor burden, measured as the SLD of target lesions, was collected for allrandomized patients within 30 days of registration and then every 2 cyclesduring protocol treatments and every 6 weeks until progression Waterfall plotand spider plot are used to display the percent change of tumor burden duringtreatment relative to baseline for 39 patients on pemetrexed and 37 patients onsunitinib As illustrated in Figure 1.1, each vertical line of a water plot representsthe tumor shrinkage of each patient receiving the treatment One advantage ofwaterfall plot is its capability of showing tumor measures for patients within theRECIST categories In Figure 1.2, a spider plot represents tumor percent growth

of these patients relative to baseline Waterfall plot and spider plot have becomeimportant tools to display graphically the continuous percent change of tumormeasurements The continuous measures can be analyzed quantitatively by

summarizing the mean (SD) and compared between groups using a t or

Wilcoxon test Waterfall plots and spider plots have been used effectively toshow the benefit of some treatments, such as sorafenib in renal cell carcinoma[81] and erlotinib in NSCLC [90]

Treating tumor changes as a continuous variable is also considered tocorrelate better with patient survival Unfortunately, studies so far have shown

no improvement in the usefulness of assessing tumor response as a continuousvariable as compared with RECIST measures [3, 4, 32, 64, 89] In renal cell

cancer, Stein et al [94] used mathematical models to calculate constants thatdescribe the exponential decrease and growth of the tumor burden for each

patient treated on a phase III trial comparing sunitinib vs interferon-α They

found that the median tumor growth constant of patients receiving sunitinib was

significantly lower than for those receiving interferon-α, which is consistent with

survival benefit of sunitinib over interferon found in the phase III trial [70] Theinvestigators suggested that calculation of a tumor growth constant could be aneffective surrogate endpoint for overall survival Tumor growth modeling hasalso been studied in lung cancer [103] and has led to the proposal of arandomized trial that uses early tumor growth, rather than PFS, as a primaryendpoint [55] The utility of analyzing tumor growth as a continuous variable orthe estimation of parameters of tumor growth model has yet to prove its value