Reinventing Clinical Decision Support

In 2011, Kathy Halamka was diagnosed with Stage III breast cancer. Although news like this is never welcome, Kathy was fortunate in that she was married to a Harvard Medical School professor (the second author of this story) and had access to some of the best oncology services in the world. By the time Kathy’s tumor was detected, a sentinel node biopsy revealed that it had already spread to a few nearby lymph nodes. The malignancy was estrogen and progesterone positive but HER-2 negative, less than 5 cm in diameter, poorly differentiated, and fast-growing. On average, the 5-year relative survival rate for women like Kathy is 72%, meaning people who have the cancer are only about 72% as likely as people who do not have it to live for at least 5 years after being diagnosed.

The standard of care for cases like this is typically chemotherapy followed by mastectomy. But having access to digital resources such as the Shared Health Research Information Network (SHRINE), Informatics for Integrating Biology and the Bedside (i2b2), and Clinical Query 2 presented new options for Kathy and an opportunity to test a personalized medicine approach to healthcare. But the fact that too few patients and physicians have access to these sophisticated resources serves to emphasize the need to reinvent clinical decision support nationwide.

Tapping Sophisticated Databases

i2b2 is an open-source software platform that gives clinicians and researchers web-based access to a hospital’s electronic health records (EHRs), a resource that has the potential to locate treatment options not yet available in the medical literature or officially endorsed practice guidelines. i2b2 can be compared to an operating system on which applications such as Clinical Query 2 sit. Clinical Query 2 (screenshot below) consists of a website and database that let clinicians search patient records at Beth Israel Deaconess Medical Center. SHRINE is a network of computer systems that are affiliated with Harvard Medical School, giving users access to the EHRs of all of its affiliated hospitals, including Massachusetts General Hospital, Brigham and Women’s Hospital, and Dana-Farber Cancer Institute. With the assistance of these resources, it was possible to perform a very specific query about Kathy, a 50-year old Asian female with Stage III breast cancer. That query asked how many patients seen in all the Harvard-affiliated hospitals fit her profile. The system found more than 17,000 and provided the medications they received, their average white blood cell counts, their prognosis, and so on.

Clinical Query2 is used by clinicians at Beth Israel Deaconess Medical Center to help find individualized diagnostic and treatment options for patients at Harvard affiliated hospitals. (Image used with the permission of Beth Israel Deaconess Medical Center, Boston, Massachusetts.)

The results of that investigation concluded that stage III breast cancer is usually managed with doxorubicin (Adriamycin), cyclophosphamide (Cytoxan), and paclitaxel (Taxol). The query also pointed out that neuropathy is a common side effect of Taxol. Because Kathy is a visual artist, that consideration was important, as neuropathy could affect her fine motor skills. Further investigation found that there was only one clinical trial looking at the use of Taxol in this context, and it used a specific number of mg/kg body weight administered in 9 doses. There were no data to indicate that this was the optimal dosage regimen or if 3 doses or 11 doses would have resulted in better outcomes, both in terms of tumor shrinkage and adverse effects. That data prompted Kathy’s physicians to personalize her treatment by administering full protocols of Adriamycin and Cytoxan but only a half protocol of Taxol, giving her 5 doses rather than 9. The individualized approach caused her tumor to shrink and eventually disappear and resulted in minimal numbness in her hands and feet.

Clinical Decision Support Is Getting Smarter

While the digital tools used to help personalize Kathy Halamka’s cancer are impressive, they join a long list of clinical decision support platforms that are available to health professionals. These new resources can transform the practice of medicine, but only if clinicians are willing to recognize the need for such tools. That need can be summed up succinctly: “The complexity of medicine now exceeds the capacity of the human mind.” That observation, made by Ziad Obermeyer, MD, and Thomas H. Lee, MD, in the New England Journal of Medicine, emphasizes the fact that physicians and nurses, despite their years of education and clinical experience, have cognitive limitations. No single clinician can retain the petabytes of medical research and patient records now available in many clinical decision support systems. Nor can they be expected to see all the correlations and patterns required to make a fully informed diagnosis. No doubt these shortcomings are partially responsible for the disturbing number of misdiagnoses reported in the scientific literature.

The Institute of Medicine estimates that about 5% of US adults who obtain care in an outpatient setting are misdiagnosed every year. Postmortem studies have also found that diagnostic errors contribute to about 10% of patient deaths. Similarly, misdiagnosis likely contributes to up to 17% of adverse events in hospitals and are a major reason why so many malpractice claims must be paid out to injured patients.

The Merck Manual outlines several reasoning errors that physicians are prone to, including affective errors, confirmation bias, and availability errors. Clinicians fall victim to an affective error when they convince themselves that what they want to be true about a patient really is true. As the singer/songwriter Paul Simon once put it: “A man hears what he wants to hear and disregards the rest.” That single-mindedness is sometimes accompanied by a confirmation bias in which clinicians cherry-pick diagnostic observations that confirm their suspicions. Availability errors occur when physicians are continuously exposed to a large number of cases of a specific disorder and then begin to assume that same disorder exists in most of their other patients with similar signs and symptoms. This particular form of blindness makes it that much harder to detect rare disorders.

Knowledge- and Non-Knowledge based CDS Tools

While clinical decision support technology will not eliminate such biases, it challenges clinicians to think twice before jumping to conclusions by presenting them with other diagnostic possibilities that don’t readily come to mind. These digital tools come in a variety of types but can be divided into 2 broad categories: Knowledge-based and non-knowledge based. The former is made up of a database that includes scientific details on specific diseases, the best options for treatment, etc. The CDS tool typically includes decision trees with “If/Then” statements to guide the diagnostic process based on data from each patient’s medical record.

Non-knowledge based CDS systems rely on artificial intelligence (AI), machine learning, and neural networks to detect hidden patterns in patient data. Neural networks are software constructs that mimic the neurons and synapses found in the human brain. The artificial neurons, called nodes, are linked together in a network that can analyze input from medical images, for example, which in turn helps differentiate pathology from normal tissue. These deep-learning programs review thousands or millions of images, separating normal from abnormal findings. During the training phase, these programs make many mistakes, so they use a protocol called back propagation to review these miscalculations and gradually correct them, learning from its own errors without a programmer adding new code to the software.

UpToDate, available from Wolters Kluwer, is a good example of a knowledge-based CDS tool. This large online service is the equivalent of a massive medical textbook that covers numerous specialties—with an important difference: Textbooks are typically revised every 5 years, while the database is updated every few months. The company recently released a more sophisticated version of its service called UpToDate Advanced, which includes interactive algorithms that let clinicians take a more personalized approach to patient care. The advanced version of UpToDate also includes a tool to help providers interpret abnormal lab results. (pictured right courtesy Wolters Kluwer.)

VisualDx, another CDS system, provides point-of-care advice in primary care, emergency medicine, dermatology, and hospital medicine, with a strong emphasis on visual depictions of pathology. Unlike UpToDate, VisualDx includes a symptom and sign finder to help clinicians build a customized differential diagnosis. For instance, if the patient presents with anemia but also has seizures, the CDS tool will walk the user through a specific pathway, while a patient with anemia and hypotension will prompt the system to show a separate pathway and a different potential diagnosis. When the anemia/seizure option is chosen, VisualDx then presents the user with a list of possibilities, including chronic kidney disease, hemolytic uremic syndrome, and so on, using graphics and labels to walk the clinician through the next step in the diagnostic process. Including links to relevant articles to describe the condition, as well as ICD codes, drug reaction data, and proper tests to perform. It even provides links UpToDate and PubMed for clinicians who want to dive deeper.

ClinicalKey, like UpToDate, is a searchable database that addresses the needs of clinicians in a wide variety of specialties and in primary care practice. It relies on content from thousands of biomedical journals published by Elsevier, the company that makes ClinicalKey. One of the useful components of the database is its “smart” search function. As you begin to type a term into the search box, it lists numerous alternatives you may not have considered. For instance, as you begin typing the word diabetes, the service provides a drop-down menu that includes terms like diabetes-related complication, diabetic retinopathy, and diabetic ketoacidosis. Like several other CDS tools, ClinicalKey lets users integrate the service into an EHR system.

Elsevier also offers STATdx for Radiology, which contains thousands of images and complex diagnoses, helping specialists compare their patients’ finding to confirmed diagnoses of specific disorders. ExpertPath for Pathology is a similar service for pathologists.

While these CDS systems are impressive, they only scratch the surface. In recent years, researchers have demonstrated that machine-learning-enabled software can take the diagnostic process into previously uncharted territory. For example, Google scientists, working in conjunction with researchers at the University of Texas, Austin, used machine-learning algorithms to scan retinal images of patients suspected of having diabetic complications. When they compared computer-generated diagnoses of diabetic retinopathy to that performed by 54 experienced ophthalmologists and senior residents, they found the computer was more accurate than its human counterparts at detecting the disorder.

Researchers have also demonstrated that it’s possible to use neural networks that scan millions of skin lesions to differentiate between a normal mole and a malignant melanoma.

Implementing the latest CDS Tools

Although studies like these often generate headlines, commercially available CDS systems that take full advantage of AI are not quite ready for prime time. But that should not stop healthcare organizations from taking advantage of the sophisticated tools already available to the medical community. For instance, the SHRINE software mentioned above is freely available as open source, so it can be used by other medical centers to establish similar data sharing networks. In fact, several organizations have taken advantage of this opportunity. A project called SHRINE National, for instance, linked 8 medical centers to study co-existing disorders related to diabetes and autism. (Details on how to implement SHRINE are available here.)

Commercially available CDS systems can be plugged into a medical practice or hospital in several ways. UpToDate, for instance, offers single-user subscription options for medical professionals, residents, fellows, students, and patients. Multiuser subscriptions are available for hospitals, institutions, and group practices. Since the service is compatible with several EHR systems, it is also possible to fold the program into the same computer network that houses one’s patient records, reducing the need to log out of one electronic system and into another. Other vendors have similar implementation options, with pricing varying depending on specialty, the number of users, and whether an advanced informational option is chosen.

There’s little doubt that the advanced clinical decision support tools used to guide Kathy Halamka’s treatment have been invaluable. The individualized chemotherapy regimen minimized the impact of Taxol on her fine motor skills, her partial mastectomy was a success, and her osteopenia—an adverse effect of chemotherapy—has decreased. She continues to live an active, full life running an animal sanctuary in a Boston suburb. And as CDS tools like ClinicalQuery saturate the medical profession, it’s likely that the misdiagnosis statistics and their ugly effects will gradually shrink.

Paul Cerrato has more than 30 years of experience working in healthcare as a clinician, educator, and medical editor. He has written extensively on clinical medicine, electronic health records, protected health information security, practice management, and clinical decision support. He has served as editor of Information Week Healthcare, executive editor of Contemporary OB/GYN, senior editor of RN Magazine, and contributing writer/editor for the Yale University School of Medicine, the American Academy of Pediatrics, Information Week, Medscape, Healthcare Finance News, IMedicalapps.com, and Medpage Today. The Healthcare Information and Management Systems Society (HIMSS) has listed Mr. Cerrato as one of the most influential columnists in healthcare IT.John Halamka, MD, MS, is the international healthcare innovation professor at Harvard Medical School, chief information officer of the Beth Israel Deaconess System, and a practicing emergency physician. He strives to improve healthcare quality, safety, and efficiency for patients, providers, and payers throughout the world using information technology. He has written 5 books, several hundred articles, and the popular Geekdoctor blog.