Did AI get it wrong? Check the training set
Volume 10, Issue 11 | September 19, 2024
Also in This Issue
Even if concerns are valid, is Lancet slam of DTC warranted?
Can AI unlock the dx potential of metabolomics?
This Week in AI
Bird flu update: MO and MA
New and Noteworthy
This Week in AI
A cheaper way to evaluate DCIS
Ductal carcinoma in situ (DCIS) accounts for 25 percent of all breast-cancer diagnoses, and 30 to 50 percent of patients who have it go on to develop a highly invasive form of cancer. Tests that can determine whether a given patient falls into that category do exist, but they’re expensive. A Nature Communications paper documented an AI-based model that can do the job based on inexpensive, easily obtained chromatin-stained tissue samples. The technique takes into account both the proportion of different types of cells present in the sample and the way in which those cells are organized in the tissue.
Bird flu update: New info on Missouri case; MA is H5N1-free
Last week we reported that a person in Missouri had developed H5 influenza without any known animal contact. Officials have confirmed that this person did have H5N1 (last week they weren’t sure), but they still have no evidence of transmission to other people. One household contact of the infected person developed symptoms the same day but was never tested. If that second person had H5N1, it’s likely both people got it from the same source.
All 95 registered dairy farms in Massachusetts have tested negative for H5N1, based on bulk milk tank tests done at no cost to farmers. Massachusetts is the only state to have tested all of its dairy herds. A source at the lab where the milk is tested indicated that the state intends to continue testing once a month.
Food for Thought
To DTC or not to DTC? Lancet calls it “an industry built on fear”
A Lancet editorial this week is highly critical of direct-to-consumer (DTC) testing. The piece juxtaposes a $4.3 billion global market (forecast by NovaOneAdvisor.com to be $9.3 billion by 2033) with a standard litany of (mostly reasonable) concerns. Among the cited issues:
The unproven and often dubious claims of clinical value of many DTC tests, for example those that identify a genetic predisposition for a currently untreatable disease.
Diagnostics which skirt FDA approval requirements by being marketed as “lifestyle/wellness” tests.
Direct consumer access to a full suite of traditional biomarker tests taken out of any comprehensive clinical context.
The editorial indicts a “corporate playbook” marketing “false empowerment” and compares DTC tests with an idealized physician-driven model that’s integrated and wise. It clearly points out that most people’s real-world care experience with physicians and the traditional model are far from ideal - fragmented and inaccessible. (Some context commentary: The Lancet is an establishment British medical journal whose primary readers and contributors practice in the physician-driven National Health System.)
Commentary: The DTC train has left the station – it is not at full speed yet but it is unstoppable. Every medical care innovation has strengths and weaknesses, and frustrating as it is to see unbalanced editorials like this, several of the concerns raised are valid ones. As an industry, we need to protect against them. How? We believe that appropriate regulation and quality controls are foundational. If the federal government does not want to / does not have the resources to do it - should the “unregulated” DTC companies create their own quality code? If too many DTC tests are seen as inappropriate, it will negatively impact the quality tests. Education is another important baseline. Patients must feel comfortable and empowered to share DTC tests results with healthcare professionals and those providers need to know how to interpret (or debunk) them.
AI may unlock the diagnostic potential of metabolomics
Metabolomics is the study of all the small-molecule products that result from running a human body (see this review article, this chart reproduced here, and this review for details). Many of these individual products have long been used in diagnostics as part of routine blood tests. Less-invasive sample types also exist: urine and breath.
Until now, the clinical potential of analyzing the metabolome as a whole has been limited by two main factors. The first is knowledge: Linking complex metabolic profiles to disease has proven very challenging. The second is cost: This type of analysis requires high-cost equipment (primarily mass spectrometry) and highly skilled labor.
AI can help with the knowledge part by finding patterns that might otherwise be invisible or take an impossibly long time for a human to figure out. A recent Nature commentary reviews AI efforts to analyze complex data sets like these. On cost, a recent paper shows how cheaply acquired dried blood spots can be sent to a central lab to leverage the benefits of automation and scale.
Commentary: Rapid progress is being made in both metabolomics and microbiomics on both the knowledge and cost fronts, but in many ways these represent large areas of unexplored potential for disease diagnosis. May the progress continue and accelerate.
When AI gets it wrong, check the training set
When your AI’s training set isn’t diverse enough, your tool won’t work - no matter how good your code is. A study published in Frontiers in Digital Health is a case in point.
The AI in question was designed to measure mental health by analyzing vocal patterns. How might this be possible? Well, people with anxiety typically speak in higher tones, while people with depression often speak at a lower volume and use a “flat” intonation with little variation.
Sounds reasonable, but the tool misclassified people - because the data set it was trained on didn’t include enough women and didn’t represent a wide enough array of ethnicities. One problem: on average, women speak in higher tones than men do, even when women aren’t anxious. Another issue: Anxiety doesn’t translate to higher pitch in all ethnic groups. As Dr. Theodora Chaspari, one of the study’s authors, pointed out: “If AI isn’t trained well, or doesn’t include enough representative data, it can propagate these human or societal biases.”
Commentary: AI is not and will never be perfect. Think about it as if it were a person with an amazing, photographic memory. It’s very helpful, but it does not replace your knowledge and experience. (Of course, we do need to recognize that each of us also has biases.)
We are not concerned that AI gets it wrong sometimes, especially in the current early innings. We believe that it will get better and better. But just as we must do when evaluating a human’s conclusions (including our own), we must always check for bias when dealing with AI.
Quick Hits
In the quest to know everything about one’s health, glucose meters are being used by people who are not diabetic. Why? The hope that detecting pre-diabetes will lead to better health, the desire to optimize fitness regimens, and just plain curiosity, to name a few reasons. Large glucose-meter companies have announced their intention to create separate brands to target this market.
In contrast to the situation with personalized medicine, clinicians may not need much of a push to embrace the idea of self-testing for human papilloma virus (HPV). In a new study, researchers found more than half of healthcare providers would definitely or possibly offer HPV self-collection to their patients if the FDA approved the procedure.