Also in This Issue
This Week in AI
Potentially deadly chemo: Why aren’t we testing before treating?
Using speech as a biomarker
A biomarker that only AI could see
An insightful series on wearable devices
This Week in AI
AI diagnoses lymphoma. . .
A deep learning model called the Lymphoma Artificial Reader System (LARS) was able to distinguish between positron emission tomography (PET) scans of patients with and without lymphoma with 87 - 90% accuracy. The model was trained on scans from the Memorial Sloan Kettering Cancer Center in New York City and tested on scans from both that center and from Austria’s Medical University of Vienna.
. . . but it isn’t a panacea for radiologists
AI assistance doesn’t always help radiologists make more accurate diagnoses, according to a study of 140 clinicians on 15 different diagnostic tasks involving chest X-rays. Less experienced radiologists were no more likely to improve with AI help than were more experienced ones. The only criterion that predictably made a difference was the quality of the AI model - the less accurate it was, the more mistakes clinicians made.
These results are at odds with several previous studies, but the authors suggest that their larger sample, wider range of subject ability, and broader slate of diagnostic tasks may make for a more accurate assessment.
Saving babies’ retinas with AI
In a recent study, all of the infants who had severe cases of retinopathy of prematurity (ROP) were accurately detected by a fully autonomous AI system. The system was trained on a dataset from the US and tested in two large, unrelated data sets, one from the US and one from India. While most cases of ROP don’t require treatment, severe cases make the condition the number-one cause of blindness in children worldwide. It’s particularly prevalent in low- and middle-income nations, where access to trained ophthalmologists can be spotty.
Three pieces of essential reading on medical AI for Dx folks
A Big Week in Medical AI (Eric Topol, Ground Truths blog) reviews some extraordinary recent advances
A trustworthy AI reality check (Frontiers in Digital Health) reveals the uneven transparency of algorithms in 14 radiology tools in current use
How to support the transition to AI-powered healthcare (Nature Medicine), while optimistic, pumps the brakes a little by highlighting what AI developers have to do to avoid obvious pitfalls - especially in assembling and maintaining relevant training data and carrying out clinical validation.
The ultimate non-invasive test: speech
A Google research team has published a draft paper on a novel tool that analyzes speech to detect infectious disease, beginning with COVID and tuberculosis (TB).
HeAR (Health Acoustic Representations - method reproduced below) is an AI application, of course, as all advanced tests that identify such complex patterns have to be. This one was trained on two-second snippets of speech to find diagnostically relevant patterns (listening for acoustic properties only - it's not what you say, it's how you say it).
Commentary: Results were not accurate enough to meet clinical needs: While a coin toss is 50:50 (useless for diagnosis), the best score for TB detection was 65:35, meaning the model was right ⅔ of the time. We wonder whether an improved and larger training set will show improved predictability. Still, a strong demonstration of future potential.
Wouldn’t you want to know if your chemo drug could kill you?
Fluorouracil (5-FU) is an inexpensive, standard drug used in the treatment of cancer, but about one in 1000 patients cannot metabolize it well - and in these patients, treatment at full chemotherapeutic levels can be fatal. A range of tests exist that each look for different sets of genetic variants that can cause the metabolism problem, but oncology guidelines don’t recommend testing before instituting treatment. Last Thursday, the FDA added new warnings to the drug’s label. Those warnings didn’t recommend testing either. Why? According to a KFF Health News article, the agency said “it could not endorse the 5-FU toxicity tests because it's never reviewed them.”
Commentary: Oncologists interviewed for this article noted that not all patients with these genetic variations have trouble metabolizing 5-FU. And different genetic variants tend to be present in different ethnic groups - and are diagnosed using different tests, so screening properly isn’t trivial. We hear that but are not convinced. This is a straightforward test - why is it not included with the standard CBC tests done before chemo starts? Will the state biomarker bills (now 15 states covering 47% of Americans) ensure payment for these tests.
Meanwhile, doctors in Europe - where testing is formally recommended - start patients with risky variants on a lower dose of the drug and see how they do. If they do well, the dose gets increased; if not, it’s stopped.
Food for Thought
Wearable digital health technologies - an insightful series
This week NEJM published the final installment of a series on how wearable technologies can help improve patient health. The series began in November with an editorial and a review of diabetes-management devices. December was depression, January was cardiovascular disease, and this week covers key issues involved in integrating wearable devices into clinical care. Issues that are highlighted include data management, privacy, equity of access, education of patients and their physicians, and - perhaps most important - integration into the US healthcare system.
These devices generate enormous amounts of data that could potentially be clinically relevant. What goes into the electronic health record (EHR) and what stays out? As the AI embedded in these devices becomes more sophisticated, the tools can provide wider and wider ranges of potentially actionable information. But no physician can read and absorb more than a few pages in the 2-4 minutes before seeing a patient.
Commentary: Wearable devices’ use and sophistication is growing dramatically and will continue to do so. In 2020, 84 million people in the US used a health app of some form meaning 30% of adults use their phone look to their phone for health information. Physicians need to acknowledge this and become the integrator of diagnostic insight provided by these devices. Sounds good but figuring out how to do that is not straightforward.
App reliability (which ones are accurate?)
Patient reliability (are they using app appropriately?)
Reference standards (how to compare to lab tests?)
Logistics (how does data actually transfer into chart / EMR?)
Physician payment (does MD get paid for the analysis of the wearable data?)
Patient Privacy
AI turns “junk DNA” into a potentially powerful biomarker
A new study uses AI to add another DNA feature to the pan-cancer biomarker portfolio: Alu elements. Also known as short interspersed elements (SINEs), Alu elements make up 11% of the human genome and used to be considered part of our “junk DNA.” However, it turns out that they have extremely complex and varied biological effects.
Using a machine-learning model, researchers were able to predict with 98.5% specificity whether people had any of 11 different types of cancer, based solely on changes in their Alu elements. When the analysis of Alu elements was combined with information about other DNA changes and eight known protein biomarkers, specificity went up to 98.9%. The biggest difference the model teased out was that patients with solid tumors have lower overall levels of one subfamily of Alu elements - a discovery that would not have been possible without the use of AI.
Commentary: In the Genomics 2.0 world, finding the mutations that mess up proteins isn’t enough (that was Genomics 1.0). At the 2.0 level, it’s about analyzing the non-coding 98% that controls when and where the genes should be activated. Curious about Genomics 3.0? It will involve multi-omics, in which we analyze 100% of the DNA plus all the other key factors that affect health and disease. AI will be the key to attaining that level.