By Deborah Borfitz
April 6, 2021 | Getting a diagnosis right when diseases share symptoms or co-occur due to advanced age or unhealthy habits may require a more holistic view of the situation than can be gleaned from the typical physician-patient encounter. Individuals might not remember all the illnesses they are being treated for, and doctors often have limited context for how diseases usually develop across populations and time, says Søren Brunak, professor at the Novo Nordisk Foundation Center for Protein Research at the University of Copenhagen.
Over the past few years, Brunak and his team have been developing disease trajectories—the sequential ordering of multiple diseases—using the experiences of millions of Danes to calculate the “most frequent highways” of disease progression. In envisioned real-world use, he likens the trajectories to Google Flights that has precomputed possible routes between two cities except the computation is a function of what happens to most patients with a certain disease, so doctors have a backdrop for evaluating individuals.
The end goal is a decision support tool embedded in electronic health record systems to reduce under-diagnosis and over-diagnosis of diseases, says Brunak, who has decades of experience correcting errors in complex databases. Medical misdiagnosis is a widespread, if under-discussed, issue in healthcare.
Denmark makes calculating disease trajectories possible because it has a national registry that follow its 5.6 million citizens over their lifetime, Brunak explains. “We now have more than 40 years of data from almost 10 million people [including those now dead].”
Brunak has co-authored numerous papers where the Danish National Patient Registry was tapped to systematically build disease trajectories from “statistically significant directional transitions from one disease to another,” which is sometimes also gender specific (e.g., breast cancer in men versus women). The control group are patients admitted the same week during the year, to “take seasonal variation out of the equation.”
In some cases, machine learning methods have been used to analyze the trajectories. But to identify potential incorrect or overlooked diagnoses, researchers needed only run-of-the-mill statistics to hunt for diagnoses that appeared in an “unusual context,” says Brunak.
Their initial analysis focused on chronic obstructive pulmonary disease (COPD), a common disease in Denmark because of the country’s high smoking prevalence. Comparing the unusual trajectories with the more frequent ones was a relatively simple exercise involving counting the number of identical diseases and quantifying the differences.
Results recently published in npj Digtal Medicine (DOI: 10.1038/s41746-021-00382-y). The study included a co-analysis of lung cancer.
An example of an unusual context would be an individual with diabetes and cardiovascular disease being diagnosed with COPD without any sort of prior respiratory issues, Brunak says. Lung cancer, on the other hand, “often comes more out of the blue” and may be less uncommon.
Based on the disease trajectories for 284,000 COPD patients between 1994 and 2015, approximately 69,000 disease progression paths were found to be “reasonably frequent” in that they were followed by at least 20 patients. The trajectories here involved three consecutive diseases, but “longer significant trajectories of additional disease are often found depending on the index disease and how systematic the disease development is,” says Brunak.
Among individuals who did not present with these time-ordered, common comorbidities, 9,597 were deemed unusual, he says. This group also died significantly earlier than COPD patients following a trajectory.
One subgroup comprising 2,185 patients was found to be at risk of misdiagnosed COPD and thought to instead have had lung cancer based on their lab test values and survival pattern. But only 10% of them had been so diagnosed, Brunak says, and only 4% had a lung function test to confirm a COPD diagnosis.
Researchers concluded that another subgroup with 2,368 patients were likely over-diagnosed with COPD in that they survived for more than 5.5 years post-diagnosis and had none of the typical complications of the disease, Brunak adds.
Disease Trajectory Browser
Although the algorithm was validated with data from COPD patients, it could be used for many other diseases to map typical trajectories and identify outliers, says Brunak. Once the mapping is done, “it only takes 10 seconds to match a single patient against everyone else.”
The chief requirement is having enough cases to analyze, which would exclude diseases that are rare or occur infrequently in the population under study. Rare diseases are in any case rarely over-diagnosed like COPD, he adds.
In a paper published last fall in Nature Communication (DOI: 10.1038/s41467-020-18682-4), University of Copenhagen researchers publicly released the Danish Disease Trajectory Browser based on data from the national registry from January 1994 to April 2018. The tool comprises electronic health data on 7.2 million Danes over the almost 25-year-long period, presented as summary statistics.
The browser enables users to search, filter, analyze, and visualize disease trajectories derived from statistically significant directional diagnosis pairs calculated from population-wide electronic health data, he says. The number of “frequent highways” will depend on the quantitative constraints on the trajectories when making the calculations (e.g., number of patients following a common progression path), including the “thresholds for relative risk of hopping from one disease to another” and prevalence of a disease in the population.
Outside research groups have been tapping the national patient registry for comparative data, he adds. “We do not necessarily expect all countries will have the exact same type of disease trajectories,” due to prevailing health habits and attitudes, but the way common conditions like cancer and COPD develop is likely to look similar place to place since those sorts of statistics get shared globally.
As the data populating the Disease Trajectory Browser stem from a complete population, they are probably less biased than other large data sets that focus on specific diseases, age groups, hospitals, or professions. In Denmark, like most Nordic countries, a social security certificate is virtually proof of one’s existence, says Brunak. “The number is used for everything—[including] your taxes, your insurances, and your diseases.”
This is what enabled researchers to integrate the disease registry with the register on causes of death and a database containing laboratory test results, he continues. The research team is also now working on integrating the Danish National Patient Registry with clinical trial registries. “Everything is tracked to the same number.”
Clinical trials comparing the use of the trajectories in a decision support tool are expected to happen at some point, says Brunak. Before that can happen, the software needs to be embedded in patient records systems (or something similar). And per recent regulations issued by the European Medicines Agency that cannot happen until the algorithm has been formally approved by regulators.
In the meantime, Brunak says, “we’re trying to convince people in a retrospective manner that… when you are over-diagnosed for something you might be under-diagnosed for something else.” As just demonstrated, getting it wrong can be deadly.