Special issue introduction: Statistical Methods in Precision Medicine: Diagnostic, Prognostic, Predictive and Therapeutic

Dear Readers,

We are delighted to offer this special issue of Biostatistics & Epidemiology on ‘Statistical Methods in Precision Medicine: Diagnostic, Prognostic, Predictive and Therapeutic.’ Precision medicine, often referred to as personalized medicine, has a relatively short history but presents great opportunities and challenges. As former US Health and Human Services Secretary Michael Leavitt said in a 2007 meeting of the Personalized Medicine Coalition, advances in science and technology present an unprecedented ‘opportunity to bring health care to a new level of effectiveness and safety’ [1]. In particular, recent advances have been made in omics-based in vitro measurements [2–4], quantitative imaging biomarkers [5], artificial intelligence/machine learning [6], and electronic health record keeping [7]. These advances and others have led to a surge in medical research activity into personalized medicine, which has been described as ‘providing the right drug for the right patient at the right time’ [8]. As a result, the potential has never been greater to obtain powerful information for individualizing medical decision making, including but not limited to information on diagnosis, prognosis, and treatment selection, and for predicting dose, monitoring disease, modifying behavior, and aiding the development of a tailored therapy, that is, a drug or a medical device [9, 10].

The recognition that advances in science, technology, mathematics, and data collection could revolutionize healthcare has led to many important government initiatives. In 2015, the US launched the Precision Medicine Initiative (PMI), with the mission ‘to enable a new era of medicine through research, technology, and policies that empower patients, researchers, and providers to work together toward development of individualized care.’ This announcement was followed by the 21st Century Cures Act [11], which provided funding for PMI to drive research into the genetic, lifestyle and environmental variations of disease. Prior to PMI, the US Food and Drug Administration (FDA) had already made personalized medicine a top priority, issuing the discussion paper Paving the Way for Personalized Medicine: FDA’s Role in a New Era of Medical Product Development [12]. The FDA and the National Institutes of Health (NIH) published a working glossary of terminology for Biomarkers, EndpointS, and other Tools (BEST) [13]. The European Union Council [14] provided discussions on personalized medicine, including a formal definition. The European Medicines Agency (EMA) provided a perspective on pharmacogenomic information in drug labeling [15]. The first goal of EMA’s vision of Regulatory Science Strategy to 2025 [16] is ‘Catalysing the integration of science and technology in medicines development,’ under which the first core recommendation is to ‘support developments in precision medicine, biomarkers and omics’. These are just a few selected examples of regulatory efforts being made across the globe to facilitate the promise of precision medicine.

Many of the success stories in precision medicine have involved the discovery of a single biomarker or a set of biomarker variants understood to have pharmacokinetic effects (drug metabolism) or pharmacodynamic effects (drug target) [17]. Genetic polymorphisms in cytochrome P450 (CYP) enzymes (e.g. CYP 2D6 and 2C19) are well-understood to be predictive of metabolic phenotype (poor, intermediate, extensive, or ultrarapid metabolizer) for a class of therapies metabolized primarily by their gene products [18–22]. Metastatic colorectal cancer patients with epidermal growth factor receptor (EGFR)-positive tumors have been demonstrated to benefit from treatment with monoclonal antibodies (mAb) cetuximab and panitumumab if their tumor is KRAS wildtype [23–25]. Breast cancer patients with EGFR-positive tumors are more likely to respond to treatment with the mAb’s trastuzumab and pertuzumab if their tumors are positive for ERBB2, formerly known as HER2-neu [26]. PD-L1 expression is associated with differential efficacy of nivolumab on overall survival in non-squamous cell lung cancer patients [27].

In lieu of identifying biomarkers well-understood to be involved in a biological pathway, efforts have been made to build predictive models from high-throughput screening of large numbers of analytes (e.g. single nucleotide polymorphisms, gene expressions, or protein expressions) using advanced technologies (e.g. microarrays, next-generation sequencing, or mass spectrometry) and large-scale data analysis (e.g. genome-wide association study or polygenic risk score modeling). In untreated breast cancer patients, two well-known prognostic signatures based on gene expressions are Oncotype Dx and Mammaprint [28]. Drier and Domany [29] found that the only biological process or pathway significantly associated with the two signatures was cell proliferation, a process whose relevance was already well-known before gene-expression profiling. Recent advances in artificial intelligence/machine learning (AI/ML) have increased the potential to build clinically meaningful predictive models from large numbers of analytes based on big data sets. Though interpretability and explainability remain an issue [30], some AI/ML models have found acceptance. QuantX is a machine-learning-based, computer-aided diagnostic software device applied to magnetic resonance image data to obtain information (quantitative imaging risk score and component radiomic features) intended to assist radiologists in the assessment and diagnosis of breast abnormalities [31]. ContaCT is a clinical decision support, triage software device that uses an artificial intelligence algorithm to analyze computed tomography images of the brain and send a text notification to a neurovascular specialist if a suspected large vessel blockage has been identified [32, 33].

The development, translation, adoption, and implementation of precision medicine products into clinical care face regulatory, intellectual property, and reimbursement challenges, and, just as significantly, limitations in human ability to interpret and clinically apply omics and other complex data [34, 35]. The discovery and acceptance of important predictive biomarkers represent itself as one of the biggest challenges in personalized medicine. Tang and Pennello [36] obtained 21,471 hits in a PubMed search of the term ‘prognostic marker’ during the period January 2005 to July 2012, yet at the time only a handful of prognostic biomarkers had actually been widely accepted by the clinical community and even fewer were covered by private or public health plans such as the US Centers for Medicare & Medicaid Services [37]. The difficulty in translating biomarker research from bench to bedside is sometimes attributable to unfocused biomarker discovery studies, no firm demarcation between discovery and validation studies, or poor study design, conduct, analysis, or reporting. Some dramatic examples of unreproducible research (e.g. see [38]) prompted the Institute of Medicine to issue the report Evolution of Translational Omics: Lessons Learned and the Path Forward [3]. Challenges with the development, robustness, and clinical utility of omics-based tests for prognosis or therapy selection are reviewed in McShane and Polley [39].

Though the discovery of important predictive biomarkers is largely a biological or medical problem, statistics plays a vital role in all areas of precision medicine, ranging from study design to analysis. This special issue that you are about to enjoy focuses on the statistical issues in precision medicine. It emanated from the highly successful 2018 Medical Device Statistical Issues Conference on ‘Precision Medicine: Statistical Challenges and Opportunities’ hosted by the American Statistical Association (ASA) and Advanced Medical Technology Association (AdvaMed), at Washington DC, May 8 2018. The conference organizers were Ruey Dempsey, Rachel Johnson, and Cindi Brooks of AdvaMed, Xiting (Cindy) Yang, Gene Pennello, Martin Ho, and Norberto Pantoja-Galicia of FDA, Victoria Petrides (Abbott), Songbai Wang (Johnson & Johnson), Alicia Toledano (Biostatistics Consulting, LLC), Danping Liu (National Cancer Institute), Larry Tang (George Mason University), Zheyu Wang (Johns Hopkins University), and Xiaohua (Andrew) Zhou (Peking University). After the conference, Dr Zhou, the Editor-in-Chief of Biostatistics & Epidemiology: China Region, approached the ASA Medical Device and Diagnostics (MDD) section about publishing a special issue on precision medicine. In all, 11 papers have been contributed to the special issue by conference speakers and panelists but also others.

The 2018 conference was ‘Day Zero’ to the 11th Annual FDA/AdvaMed Medical Devices & Diagnostics Statistical Issues Conference. Traditionally, the ASA Medical Device and Diagnostics (MDD) Section and AdvaMed co-sponsor a ‘Day Zero’ conference that precedes the annual FDA/AdvaMed Medical Devices and Diagnostics Statistics Workshop. In prior years, the Day Zero conference had an industry focus and was in the format of short courses. However, MDD members, responding to a survey in late 2017, indicated a preference for an academically oriented statistics workshop on medical devices. In 2018, AdvaMed agreed to reformat the conference with a greater academic focus to better serve MDD section members from academia and FDA as well as industry. Thanks to the great efforts made by the ASA/MDD and AdvaMed organizers, the 2018 Day Zero on precision medicine was highly successful, with over 40 attendees, a well-received program, and substantive discussion of the presented material. The keynote address was delivered by Richard Simon, the long-time director of the Biometric Research Program of the National Cancer Institute (NCI) who recently retired from the agency. In his keynote address, entitled ‘Personalized Oncology: New Paradigms for Development’, Dr Simon focused on the hidden objective of clinical trials: to estimate the study population for whom the benefits of the treatment under investigation outweigh its risks. Many consider this problem in subset identification to be the key objective of precision medicine and perhaps the key biostatistical challenge of our time. In a recent publication, Dr Simon describes his proposal to treat the problem as one of classification, not multiple hypothesis testing [40].

In the first session, Sheng Luo (Duke University) presented on ‘Dynamic Risk Assessment and Prediction,’ with panelists Gregory Campbell (GCStat Consulting, LLC) and Songbai Wang (Johnson & Johnson) providing discussion. In the second session, Hormuzd Katki (NCI) presented on ‘Decision Analysis, Risk-stratification of Diagnostic Test Performance,’ with panelists Scott Evans (George Washington University) and Gene Pennello (FDA) providing discussion. In the third session, Zheyu Wang (Johns Hopkins University) presented on ‘Prediction and Diagnostic Accuracy without a Gold Standard,’ with panelists Zhen Chen (National Institute of Child Health and Human Development), Ao Yuan (Georgetown University), and Roseann White (Duke Clinical Research Institute in 2018, currently in Natera) providing discussion. The audience was highly engaged, participating in all of the discussion sessions.

Biomarker tests have been said to be ‘the key to unlocking precision medicine’ [41]. Many biomarker tests are intended as companion diagnostics, which are defined as medical devices that provide information that is ‘essential for the safe and effective use of a corresponding therapeutic product’ [42, 43]. In this special issue, Gregory Campbell provides an introduction to companion diagnostics for therapeutic and diagnostic statisticians and reviews some of the statistical challenges, including biomarker development, diagnostic performance, misclassification, prospective-retrospective validation, bridging studies, missing data, follow-on diagnostics and complex signatures. Songbai Wang and Richard Simon propose a bivariate Bayesian framework for simultaneous evaluation of two candidate companion diagnostic assays in a new drug clinical trial.

Because biomarkers and other diagnostic information (e.g. patient characteristics, disease characteristics, imaging, and diagnostic software output) are used to facilitate precision medicine, many papers in this special issue are devoted to diagnostic development and validation to evaluate binary diagnostic tests and risk predictions, Hormuzd Katki introduces the ‘number needed to test’ to identify one more subject with disease than by randomly declaring a subject positive for disease with the same probability as the test. Gene Pennello develops a framework for determining clinically meaningful performance goals for sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio based on desired risk stratification. Through extensive simulation, Zhen Chen and Soutik Ghosal investigate parametric models of placement values, which are attractive for covariate-adjusted evaluation of diagnostic information with receiver operating characteristic (ROC) curve analysis. Zhuang Miao, Liansheng (Larry) Tang, and Ao Yuan introduce the multiple outputation method to account for clustered data in non-parametric ROC evaluation. Norberto Pantoja-Galicia, Olivia Okereke, Deborah Blacker, and Rebecca Betensky consider global scale concordance measures that summarize sequences of time-dependent area under the ROC curve (AUC) over time. Xuehan Ren, Jue Wang, and Sheng Luo propose dynamic prediction of recurrent cardiovascular disease events with a Bayesian joint model of longitudinal and recurrent event data. Focusing on diagnostic imaging data from multi-reader multi-case studies, Stephen Hillis compares the advantages and disadvantages of the Obuchowski-Rockette and Gallas methods for making inferences on reader-averaged AUC.

A common challenge in diagnostic test evaluation is obtaining the reference standard (gold standard or ground truth) diagnosis on study subjects. In many cases, the reference standard is either imperfect or not available on some or even all subjects, due to its cost, invasiveness, or lack of existence. Zheyu Wang reviews the developments and debates on latent variable modeling in diagnostic studies when there is no gold standard. Danping Liu, Ashok Chaurasia, and Zheyu Wang consider selection and combination of biomarkers to improve diagnostic accuracy, extending imputation and reweighting techniques developed for ROC curve and AUC estimation to studies in which the gold standard is partially missing.

We hope that you enjoy this special issue. We are very grateful for the valuable contributions of the authors have made on contemporary issues in precision medicine. We are pleased to see their long-awaited papers now in print. We are indebted to the many colleagues who served as reviewers of the papers. Their careful reviews and constructive comments ensured that the research and its communication were of the highest quality and made this issue truly special. We are confident that the content therein will inspire further biostatistical and epidemiological research on precision medicine.