We should be very careful about what we call "multi-cancer" detection, lest we get lost in all the hype.
Most people will think a "multi-cancer test" tells you which cancer you have. Behind the marketing vagueness, it often gives you only a binary answer to one question: "Is any cancer present?"
The label "multi-cancer" describes how many cancer types a test can pick up. Whether it can say which one it found is a separate question, and news coverage often blurs the two.
This is part of why the hype around multi-cancer tests looks so good: it's far easier to report 90%+ accuracy on a binary test that only asks, "Do I have any cancer?"
So far, most multi-cancer tests, whether they use blood samples or detection dogs, are built around that binary question. Recent studies trained dogs on breath from people with many different cancers, often pairing the dogs with a statistical model, and reported high accuracy. A positive result from these tests says that cancer may be present. It does not say where, which is the first thing a patient and their doctor need to know.
In this post we explain why a true multi-label test is so hard to build, why it matters, and how General Sense is working toward one that does not require training a dog on each cancer.
Machine learning sorts answers into three shapes, depending on how many options a test weighs and how many it can report at once.
A binary test weighs one question and reports yes or no: is any cancer present?
A single-label test, also called multi-class, sits between the two: it picks exactly one option from a fixed list, such as lung, breast, colorectal or none. It works like radio buttons: choosing one clears the others, so more confidence in one option means less in the rest.
A multi-label test reports a separate yes or no for every item on its list. It works like checkboxes: any number can be ticked, including none, and the list can include things other than cancer.
The limit comes from how a dog reports what it smells.
A detection dog gives one trained behavior per sample, such as a sit, and the behavior means target or not target. As we explained in Why so few dogs can do the job, the alert does not say which substance the dog found or how much of it. The nose takes in far more. Dogs have 811 working olfactory receptor genes to our 396, and each sniff registers as a pattern across many receptor types.
Training for "any cancer" also works against telling cancers apart. Trainers reward the same alert for lung, breast and colorectal samples, so the dog learns what the cancers have in common and gets nothing for noticing what separates them. That suits a yes/no test and works against a which-cancer test.
The workarounds each add cost. One specialist dog per cancer means training time and samples grow with every cancer added. A different alert for each cancer takes longer to train and invites confusion between alerts. Pooling several dogs on each sample steadies the answer but adds no labels.
Sensors on the dog, such as cameras or EEG, measure its response more precisely. If that response was trained as yes or no, the measurement is still yes or no.
A multi-label test needs two things: training samples labeled by cancer type, and a readout that carries more than one bit. The readout has to capture the signal before the dog turns it into a decision, and that is what General Sense is building.
We think a which-cancer answer is possible because of what is in the breath itself. Different diseases change the body's chemistry in different ways, and those changes show up in exhaled air. In one study, an array of chemical sensors classified 17 diseases, several of them cancers, from the breath of 1,404 people with about 86% accuracy in blind tests, and having one disease did not hide another (Nakhleh et al. 2017). Dogs pick up these differences too. Dogs trained on ovarian cancer did not alert on other gynecologic cancers (Horvath et al. 2008), and dogs trained on lung cancer were not thrown off by COPD (Ehmann et al. 2012). The information that separates one disease from another is in the sample; the hard part is reading it out of the dog.
Our approach is to record from the dog's olfactory bulb, the first brain structure that processes smell, through a thin electrode array placed on its surface in a one-time surgery. Different odors produce different patterns of activity there before the dog decides anything. A model trained on those patterns could score each label separately from a single sniff, and the dog doesn't have to be aware at all. The yes/no answer comes along as one more label. Adding a cancer type takes labeled samples and a model update, instead of months of training dogs on a new alert.
Researchers have decoded odor identity from neural activity in mice. Doing it in working dogs, for cancer, is the problem we are working on, and we have not yet published clinical results. When we do, we will report them against the checklist below.
General Sense does not diagnose disease. A per-cancer signal is a screening result that tells a clinician which confirmatory test to order first. We notice; clinicians decide.
Ask these questions of any multi-cancer result, ours included.
