Organizational Improvements to Enhance Modern Clinical Epidemiology

Unprecedented investments in biological sciences and information technology are catalyzing revolutionary changes in clinical epidemiology and redefining human health and disease, as well as the approach to disease classification and treatment.¹ However, the potential of these innovations will only be realized within a biomedical enterprise that develops appropriate strategies for their use. Seven facets—traditional clinical epidemiology; biobanks; genomics and molecular profiling; imaging; bioinformatics; biostatistics; and decision making—are critical to this new approach but must be placed within a systems framework that integrates each discipline's contributions. Connecting these once-isolated spheres of endeavor would benefit clinicians and the public, if necessary preconditions are met.

Clinical disease states reflect complex interactions of genetic, environmental, and behavioral contributions, but tools for identifying and quantifying these factors simultaneously are only now being developed. The speed with which pragmatic effects on practice can be achieved now depends less on scientific progress than on the organized marshalling of these streams of activity, which ultimately may be the limiting factor in achieving “personalized medicine.”² In this Commentary, we describe actions essential for hastening this medical revolution.

Traditionally, clinical epidemiology describes characteristics of the whole human, human-environmental interactions, and outcomes over time related to those characteristics. Thus, lung cancer is a malignancy arising within the lung, whereas depression is defined by clusters of behavioral characteristics and self-expressed feelings. As assays and imaging advance, elements are added, so that lung cancer is classified by “stage” and depression in terms of “behavioral characteristics, severity, and biochemical status.” Still, the basic treatment approach rests on this construct. Radical change is imminent, however, as molecular pathways increasingly inform disease classification. Thus, divergent clinical trial results in seemingly similar populations remain common and may be due to participants with similar “phenotypes” who actually reside in different molecular and clinical clusters within that classical phenotype. If these clusters could be identified, prognosis could be estimated more accurately; more importantly, treatments could be better matched to patients, with higher probabilities of benefit and lower risks of toxicity.

Currently, however, a Tower of Babel in nomenclature and data standards is reinforced by fierce resistance of physicians to adopt common standards, slow diffusion of technology linking clinicians into coherent information networks, and many descriptive elements in medicine being subjective or requiring synthesis of diverse attributes into a gestalt. Only by agreeing on standardized classifications of clinical phenomena can the full power of new biological data be harnessed. Equally important is longitudinal measurement of outcomes.³ Preventive and medical decisions are based on predictions of future states; if these are not measured accurately, predictive rules permitting rational use of new knowledge cannot be computed.

Although the characteristics of prediction science have been described,^{3 - 4} few clinical leaders grasp this fundamental discipline. Academic medical centers should reward faculty for participating in this new, advanced discipline of clinical epidemiology, which, coupled with standards development,⁵ can accelerate understanding disease prevention and treatment. Furthermore, health systems should more fully engage in maintaining outcomes data.

Biospecimens collected by academic centers are in disarray.⁶ Samples are often stored in the laboratories of individuals or small groups of investigators and primitively categorized, with insufficient linkage to clinical data. Automated systems now exist that can store and retrieve biospecimens under optimal conditions with appropriate characterization. Ethical rules governing these repositories must be clarified and continuously updated as capacity for benefit or detriment changes. Such efforts are beyond most individuals' capabilities and require action at higher levels. Academic institutions and health systems must assign higher priority to biobanking coordination, adequately fund management and informatics support, and improve biobank effectiveness by developing common standards.

It is now feasible to perform whole-genome association studies, massive studies of genetic polymorphisms, gene expression profiling, and to measure thousands of proteins and metabolites simultaneously.⁷ While the science of fundamental measurements continues to advance, reproducible clusters of patterns of these myriad measures can be discerned within diseases that by the standards of conventional epidemiology are phenotypically identical.⁸ For example, a group of obese patients with like patterns of proteins and metabolites in their adipose tissue or blood will likely respond differently to intervention than those with different patterns. By discerning structure in patterns of gene, protein, or metabolite expression that provide snapshots of gene activity in a cell or tissue sample, subtle phenotypes can be distinguished in ways traditional methods cannot, thus transforming a major segment of biology from an observational molecular science to a data-intensive quantitative genomic science. These technologies are sufficiently mature to warrant broader availability among investigators through shared consortia within and among institutions, requiring that core laboratories that can generate high-dimensional data be made available to clinical investigators at costs that can be reduced by effective economy of scale.

With computed tomography, magnetic resonance imaging, positron emission tomography, and single-photon emission computed tomography, structural and functional aspects of human biology can be assayed directly with few invasive procedures. The exciting field of molecular probes can enable clinicians to determine whether a therapy is affecting the intended biological pathways early in the course of treatment and permit monitoring of activity in those pathways. Furthermore, genomic information and development of technology in animal models is driving a rapid evolution of molecular imaging to display and quantify molecular and cellular targets in vivo.

As with biobanks, most images are not stored or indexed in ways that facilitate research. While model systems have been developed through the Biomedical Informatics Research Network,⁹ they are not available for routine deployment, and many local solutions are not fully interoperable. Nevertheless, medical imaging has the advantage of DICOM (Digital Imaging and Communications in Medicine) standards, which provide uniform nomenclature and processes for handling images. Stored images provide biomarker information as do genomic technologies, but they will be useful only if the clinical and research communities focus on standards that allow interoperability of imaging data with other data about individuals and populations.

These advances are constrained by requirements for intensive data acquisition, storage, computational capacity, and standardized exchange.¹⁰ Health Level Seven has addressed standardization for laboratory values that can be measured numerically or as coded output. Unfortunately, such standards are only now emerging for clinical data, and considerable variation exists in the purposes, stated or perceived, for these efforts, including patient care, quality measurement, research, and improved administrative efficiency and billing. Effective cross-platform systems must be developed that expedite cooperation among industry, academia, and the National Institutes of Health (NIH) and embrace collaboration and work toward a unified concept of data sharing.

Similar to the 2 “translational blocks” identified by Sung et al,¹¹ 4 fundamental “bioinformatics blocks” must be overcome. First, there is a glut of basic biological data from assay and image analysis. Second, electronic health records (EHRs) must be developed and deployed. Third, research, clinical, and demographic data must be integrated into a common format. And fourth, clinical and molecular information must be linked so that analyses yielding both cross-sectional and longitudinal patterns relevant to disease subclassification can be performed in reproducible fashion.

If population characteristics are aggregated into repositories containing genes, proteins, metabolites, and images, converting that information into analyzable data sets presents formidable challenges. Each researcher tends to approach a disease from the perspective of a particular biological mechanism. Statistical modeling allows these views to be compared and integrated, revealing the whole problem. Even more challenging is the development of synthetic models from disparate data sources. The science of evaluating predictions and characterizing probabilistic statements in humans will require significant research. Advanced statistical concepts must be introduced to the medical community, given that ubiquitous P values and standard error bars will not suffice for complex new analyses.¹²

The revolution in clinical decision making afforded by genomics lies in increased resolution: the potential to place patients on a multidimensional risk spectrum based on individual molecular characteristics measured on a genomic scale. A major advantage of these enhancements is that innovations will be judged on their ability to measurably improve predictions about the human condition. In the case of therapeutic profiling, improvement in clinical outcomes provides a good measure. Given these challenges, the critical shortage of biostatisticians and deficits in quantitative skills among clinicians and researchers must be addressed by both academic institutions and continuing education for practitioners.

Producing a set of probabilities from a molecular analysis will not necessarily lead to better choices, and more information paradoxically can lead to bad decisions.¹³ Individuals make poor decisions when they misinterpret probabilistic statements about choices, perhaps because the presentation of options was misleading or the numeracy needed for informed decision making was lacking. Furthermore, even individuals capable of grasping quantitative or probabilistic arguments can be overwhelmed by emotions or time pressure. Marketing relies on positioning information to influence individuals to make particular choices; the complexities of these data increase the risk that inappropriate selection of parts of the data will drive poor decisions. Given that both medical practice and public health decision making are increasingly based on evaluation of data rather than subjective impressions, it is remarkable how little preparation clinicians receive in the fundamental biology, psychology, and sociology of medical decision making. A multidimensional understanding of human decision making should be stressed in professional education, and increased research funding will be necessary.

It is time for an operational approach consonant with the “big picture” described herein. An enhanced field of clinical epidemiology should incorporate shared models for health and disease, aggregated by multiple interdisciplinary investigative teams that develop “disease state models” that are shared without limitations. Thus, disease models can be developed by integrating fundamental knowledge with clinical and molecular databases and population records. As health systems implement reliable, interoperable EHRs and adopt centralized biobanking, they will be able to aggregate data in clinical and molecular data repositories that provide intensive cross-sectional information about populations and longitudinal data about clinical outcomes.

In the United States, key components of this approach are in place through the NIH Roadmap and Clinical and Translational Science Awards.¹⁴ The latter will eventually fund up to 60 academic health centers and associated clinical partners with significant infrastructure capacity in most of the areas above. These centers are bolstered by multiple grants and contracts that provide avenues for innovation in areas ranging from bioinformatics to community-based research networks.

Funding agencies must also anticipate the critical imperative for interdisciplinary work. Academic health systems should foster interdisciplinary focus among faculty to maximize the benefits of the coming revolution. Regulatory agencies and payers must also prepare for major shifts in assessing diagnostic and therapeutic technologies (and, subsequently, reimbursement).¹⁵ Effective planning (particularly the systems aspects of creating interdisciplinary hubs to support this new clinical epidemiology) will result in both an advanced clinical scientific discipline and more rapid diffusion of knowledge that improves health for individuals worldwide.