While clinical labs’ use of machine learning and artificial intelligence is perhaps most prominently associated with areas like pathology and microbiology testing, these tools are also seeing uptake for lab management applications like specimen routing and billing support.
These kinds of logistical applications have proved particularly relevant during the pandemic, helping laboratories handle dramatic increases in testing demand, but investments in this technology has at some labs long pre-dated COVID.
“We have been applying this kind of logistics optimization for a very long time,” said Lance Berberian, CIO at Labcorp, of his company’s use of AI in the management of its clinical testing business. “We routinely apply data science and artificial intelligence to our business problems.”
He cited AI as a key tool in managing its test capacity, ensuring the optimal distribution of instruments and supplies across the company’s different facilities and helping to route specimens as efficiently as possible.
“For everything from the reagent that goes into the instrument to the plastic pipette tips that get thrown away after a single use, you have to make sure you have all those consumables in the right place at the right time,” Berberian said.
Labcorp also uses machine learning to help optimize its lab staffing, he said.
“We have built very sophisticated machine learning models where we look at everything from the volume trends per location to the positivity rate” of a given test, he said. “We take all of that information and put it into machine learning models that allow us to make sure that our logistics flow of consumables, of labor, of the new instruments that we are bringing online, are placed in an optimized fashion.”
John Mooney, CIO at Opko Health’s BioReference Laboratories highlighted a similar use of machine learning at that company. The company runs a vast amount of instruments across multiple sites with only a small team overseeing the operations.
“We found that there are kind of certain times of the day and certain times of the week where demand is very, very consistent,” he said.
Taking advantage of this fact, Mooney and his team have incorporated pattern recognition into BioReference’s instruments to alert them to mismatches between the actual and anticipated test volumes.
“It says, at a given time, an instrument has to run around 100 tests,” he said. “Sometimes it runs 80 and sometimes it runs 120. But if it runs 12, there’s probably a problem. If it runs 12 six sessions in a row, there is probably a problem.
“We did this for all of our enterprise, stored permanently what the demand and output is in 15 minute increments based on the day of the week, and then we compare [expectations] against what is actually happening,” he said. “And that allows … a single person to manage an enterprise across 100 departments and thousands of instruments.”
Mooney said that, like Labcorp, BioReference’s machine learning and AI efforts predated the pandemic. (Quest Diagnostics and Sonic Healthcare declined to comment for this story.) He noted, though, that COVID testing was driving new implementations of this technology within the company.
For instance, BioReference has begun using machine learning to facilitate reading and quality control of point-of-care COVID tests, which Mooney said are in high demand among its corporate clients in particular.
“We have clients that are demanding, say, 3,000 tests in a six-hour window, so we are standing up huge operations with hundreds of people that are doing this,” he said.
These test results aren’t automatically integrated into the company’s information system, though, Mooney noted. “We can’t send in an order with the test. It doesn’t know who the patient is. We can’t get a result back with it.”
To address this, BioReference built a phone app that connects to its laboratory information system and allows the POC test user to enter their results, integrating them with the lab’s traditional data system. Additionally, they have developed a machine learning tool that analyzes pictures of completed tests and compares the pictured results against the patient-reported results.
After an individual takes a picture and uploads it, the image is sent to a machine learning service that reads the cartridge itself, Mooney said. “And it is flagging any discrepancy and saying we need this for additional review.”
Berberian said the pandemic has also driven new machine learning and AI applications at Labcorp, citing specifically work on building a COVID-19 registry that collects deidentified information on patients tested for SARS-CoV-2 that can be used for research into the disease.
An investigational review board has been designated for a project that Labcorp has with Ciox Healthsource in which it is collecting medical records from a cohort of patients including individuals who have tested positive for COVID-19.
These records, however, are typically in unstructured data formats like PDFs. Berberian and his team are using optical character recognition and natural language processing to convert this unstructured data to structured data that can be put into a database and manipulated, making it useful for researchers exploring, for instance, what aspects of patients’ medical records are correlated with severe disease or ventilator usage.
While the project is currently focused on COVID-19, Berberian said it has potentially wide-ranging applications beyond that particular use case.
“Medical records are very, very difficult to turn into actual usable data,” he said, noting that the same AI-driven process could be applied to large-scale studies of other diseases.
Berberian also said that the increasing focus within the lab industry on customer convenience and satisfaction also presented new opportunities for AI. Two years ago, Labcorp put in place at its roughly 2,000 specimen collection centers across the U.S. kiosks that allow customers to enter their insurance information by having their driver’s license and insurance card photographed, which lets them skip entering that information manually.
While it seems like a simple enough process, extracting the relevant information is actually quite complicated, especially when it comes to the insurance card, Berberian said, noting the thousands of different card formats.
“So how do you know where to get the important data off of it?” he said.
To address this challenge, the company built a convolutional neural network capable of pulling the required information off insurance card, and which, importantly, is able to learn to read new card formats as they are introduced by insurers.
This feature has proved a particular boon over the last year, he added. “Who wants to type on a screen and touch where other people have touched?”
One consumer-facing application BioReference is exploring is the use of machine learning to help predict a customer’s out-of-pocket costs for a particular test — especially for high-cost, high-complexity assays. It’s an area of growing relevance given that insurers are increasingly exposing lab customers to testing costs through copays and other mechanisms.
“If you’re going to do, for instance, a high-complexity genetic test that could cost lots of money, patients are price shopping,” Mooney said.
Calculating out-of-pocket costs is complicated, though, he noted, with the ultimate price for a patient dependent on factors including their insurance and how likely it is to reimburse and where they are at with their deductible.
Machine learning might help address this problem by training models on claims data for similar patients using similar tests with the same insurer that can then be used, along with information like deductible status, to predict an individual’s likely out-of-pocket payment.
Calling it a “tricky business,” Mooney added that if a test is going to cost thousands of dollars, patients want to know what the impact will be on their wallets, and healthcare providers are trying to help them with the answer.
“Not just, hey, here are what your options are, but this is what we think it is going to cost you,” he said.
This story first appeared in our sister publication, 360Dx, which provides in-depth coverage of in vitro diagnostics and the clinical lab market.