Advances in artificial intelligence (AI) and data analytics are changing the delivery of healthcare for the better. But integrating AI and machine-learning technology into routine practice in the clinical laboratory is not a simple endeavor.
Machine learning uses statistics to find patterns in extensive amounts of data. Machine learning can mine medical data to find cures for diseases, diagnose patients before they become ill, and better personalize diagnostics and therapeutics. Data analytics opportunities in the clinical lab abound and include diagnosis, prognosis, and test menu optimization.
But despite the promise of machine learning in clinical labs, there have been overexpectations. In addition, there are barriers to adoption, according to a session at the recent American Association for Clinical Chemistry (AACC) virtual meeting.
For example, laboratorians must determine which specific applications are best suited for AI and data analytics, and they must team with data experts and software engineers to come up with the best ways to develop algorithms that will improve diagnosis and treatment of specific conditions. To date, very few machine-learning tests are offered as part of regular diagnostic care.
First, lab management will have to determine whether the cost of developing and applying the appropriate algorithms and AI models fits in well with a lab's business model. Next, they will have to assess whether the laboratory's current infrastructure is built to make integration of AI and analytics feasible. That infrastructure might have to be updated to incorporate the needed processing capability and IT interconnectivity. There also is a need to maintain the integrity and reliability of developed AI algorithms, upgrading them from time to time to optimize their capabilities.
AI successes in lab medicine
But there have been some successes. Dr. Ulysses Balis of the department of pathology at the University of Michigan in Ann Arbor told AACC attendees how he and colleagues developed and implemented a machine-learning test that guides the treatment of inflammatory bowel disease (IBD) with azathioprine.
With an approximate cost of only $20 a month, azathioprine is much cheaper than other IBD medications, which can cost thousands of dollars a month. But azathioprine's dosage has to be tailored to the individual patient, making it difficult to prescribe. The university's test is actually a machine-learning algorithm called ThioMon that analyzes a patient's routine lab test results and indicates what the patient's dosage of azathioprine should be.
The scientists found that the test performs just as well as a colonoscopy, the gold standard for assessing an IBD patient's response to medication. The computational algorithm results in highly cost-effective and accurate testing, making it possible to treat IBD with thiopurine analogs, such as azathioprine. The algorithm has a 24-hour turnaround time, compared with a one- to two week waiting period for other tests, and is more accurate and less expensive than current testing options.
The algorithm is available as an internationally offered subscription service from University of Michigan's MLabs. As a reference lab, MLabs can offer the test to any hospital or clinic that can ship blood samples to the university health system. Currently, U.S. Food and Drug Administration (FDA) approval of laboratory-developed tests is not necessary, according to the university.
"The actual algorithm takes 2.1 seconds to run. With the use of web infrastructure (as a web-based report) and the use of e-Tubes (which involve generating test results at local labs and electronically transmitting the results elsewhere), the test can be made available on a global basis without needing to ship blood specimens," explained Balis.
ThioMon can also be integrated into a lab's laboratory information system.
Beyond IBD, clinical labs could use the algorithm to analyze routine lab test results and solve other patient care challenges. Balis indicated that there are additional diagnoses that can be extracted from routine lab values using machine-learning tools.
Generally, AI in the clinical lab has significant potential to improve a number of processes, including digital image analysis, chromatography, mass spectrometry, multianalyte predictions, quality assurance, disease prediction, and medication recommendations, according to Dr. Christopher McCudden, deputy chief medical and scientific officer of the Eastern Ontario Regional Laboratory Association in Ottawa, Ontario, Canada.
He told AACC attendees that AI technology usually does not replace the skilled human element but serves to augment human capabilities.
"The potential is to augment our ability and help us deal with shortages of lab technologists," he added.
As an example of using lab or demographic data to make predictions, McCudden explained how the myocardial ischemic injury index uses age, sex, and high-sensitivity cardiac troponin concentrations to predict an individual's risk for myocardial infarction. The developed algorithm was trained using data from about 3,000 patients, and was successfully tested on about 8,000 patients internationally, he indicated.
On another front, algorithms are able to identify various patterns in automated immunofluorescence. In hematology, machine learning has been used to differentiate different types of cells, he noted.
Four paths for AI in clinical labs
McCudden sees four potential paths for AI and laboratories. "We can be consumers of this information. We can be a source of data. We can be subject matter experts. We can create algorithms ourselves to solve our own problems."
McCudden provided some guidance for labs trying to develop AI algorithms. Labs should first define the specific problem to be solved. Then they should review the data on hand and determine if it's good quality data, or if it is missing needed specifics. "You have to ask yourself is an AI algorithm the solution to the problem, or would another approach work," he explained.
In addition, labs must consider how they would develop, or train, the algorithm, and then how the data and algorithm would be integrated and reported to users. Steps to maintain and improve the algorithm, such as with new data, also should be considered, he said.
"There is a collective anticipation over the potential of emerging technologies and computers to answer questions previously thought to be impossible, such as determining the difference between normal lymph node tissue and metastatic cancer. Algorithms can computationally and automatically answer questions with superhuman levels of accuracy. Yet medicine and healthcare trail other industries in their adoption of technology and the use of real world data," Dr. Jonathan Chen, PhD, assistant professor of medicine and biomedical informatics at the Stanford University Medical Center in Stanford, CA, explained to attendees in his presentation.
He asked, "What can we do with the masses of unlabeled data that are available? What if we could estimate the pretest probability of a diagnostic test, and how could we use that information in our approach to medical decision making?"
Chen indicated that test results can be predictable, even if the actual tests have not been performed, by moving available electronic medical data, such as from a patient's electronic health record testing history, into machine learning models to predict the results of future lab tests.
But, to this point, he's seen inflated expectations, disillusionment, and disappointment over what AI can actually do. Still, he believes there is significant potential for improving healthcare by using AI and related technology. By using every information and data source and combining the best data-driven machine-learning software with the best human clinical knowledge, it is possible to deliver better care, he indicated.
In response to the challenges facing labs, researchers can apply AI and machine-learning algorithms to their lab data to essentially create a "computational lab" that can convert primary lab information into medically actionable knowledge.