Beating Bayes
Why a very good cancer test can still fail when you test everyone, and what dogs have to do with it. I worked this out in pieces across LinkedIn and X over two years; this puts them in order. It may grow into an essay.
The problem
"Screaming for Screenings"
So goes a subtitle in the recent Apollo Hospitals report on the increasingly grim situation of cancer in India. We know 15lakh (1.5M) people will be diagnosed with cancer next year, almost certainly an underestimate, with EY India's 2022 cancer report estimating the actual numbers to be 3x the reported incidence. Indians are also being afflicted with cancer much earlier - a whole decade before Americans for breast cancer (median 52yrs vs 63yrs). From the EY report we also know 80% of Indians are diagnosed too late, in Stage 3/4, where treatment success odds plummet from highs of 80% to lows of 10%. This means that the counterfactual lives lost because of late detection is around 4M people - the population of the city of Pune.
The most galling part is that we don't need a miracle cancer cure to save these lives. We simply need to detect cancer earlier. But current cancer screening methods are broken - they're capital-intensive, uncomfortable, invasive, and not that accurate. And they need to happen for every organ, ever so often. Needless to say, this doesn't happen.
To solve for this, Apollo asks people to come into the hospital for regular full-body checkups - expensive imaging scans, batteries of test etc. Only the elite (India1A <5% of popl) can even imagine affording the time and money for this. What is really needed is a novel cancer screening solution - one that is ultra-affordable, super non-invasive, and as easy to do as breathing - a cancer screening solution that is built for the scale of 1B+ ppl.
Perhaps the solution has been lying right under our noses.
The maths
It started as a reply to Elad Gil:
@eladgil Because we aren't teaching the machines to smell the same way we teach them to see and speak - by learning from the best. We're building a scalable platform for reinforcement learning by canine feedback (RLCF) at @dognosis so we can give AI a canine nose
and became a thread:
First @patrickc now @eladgil, seems like we're no longer blinded by vision and are finally paying more attention to the power of olfaction and the fact that "everything", including most diseases, has a scent and that dogs - not machines - are the ones that can detect it x.com/…
Many folks were kind enough to tag @dognosis as the company training dogs to detect cancer at scale, and I'm glad the scent is getting out there. With the right engineering around them, dogs already have the raw capabilities to solve for a big missing piece in early detection -
Beating Bayes law. Because it doesn't matter how good your test is, given cancer's prevalence (~1/100 for all types at a given point in time-space), the outcomes of any test are overdetermined by the low prior of having cancer in the first place.
This is the fundamental flaw holding back liquid biopsies like GRAIL, Exact etc etc. You can turn the knob on specificity as high as physically possible - sacrificing sensitivity in the process - and you'll still end up w an avalanche of false positives if you test "everyone"
The solution is to not test everyone. You want to push your prior of 1/100 to as low as possible, what the FDA calls "clinical enrichment". We already do this w current single organ screenings (much worse priors that multi-cancer) - age-thresholds, smoking history and so on
But you can still get breast cancer before you turn 40, in fact it's more common now than ever, and you can get lung cancer without ever having smoked before. Cancer may be correlated with all kinds of risk-factors but ultimately it's a random cascade of mutations gone awry
The best way to increase the priors of a cohort having cancer is to first test for cancer using an orthogonal mechanism that optimizes for sensitivity, user-interface, and unit-economics.
Why focus on the above 3? So you don't miss any cancers (esp early stage when it matters most), you get everyone on-board (no dealing with poop), and you can achieve population-scale without breaking a payor's pocket. And of everything we have at our disposal in the 21st century
olfactory tech is the one that excels at the trifecta of sensitivity, UI, and unit-economics. VOCs are the earliest signatures of a tumour's aberrant metabolism, can be captured from a few minutes of breathing, and can require just a few additional seconds of sensor time
and when your sensor is a friendly lab called Jessie who doesn't charge you an exorbitant salary, the unit economics looks very good indeed
while at @dognosis we believe that dog noses powered by engineering can already solve for early detection, no machine noses needed, we also believe we have the keys to teach AI how to smell. But unleashing the dogs to pass on the olfactory baton is a trail is for another day
So what
A year later, replying to coverage in India Today:
Thanks to India Today and Sumi Dutta for the coverage and Drs Kamal Saini and Ravi Mehrotra MD for the important notes in the conversation.
20+ years of 50+ studies have shown that dogs can detect human diseases. The questions people have always asked are - a) how can we deploy this ability in clinically relevant scenarios?, b) how do we trust they can do this in a standardized way?, c) how can we scale this ability?
Itamar Bitan and I have spent most of our waking hours over the past 5 years on answering these questions. We're now blessed to have an amazing team at Dognosis of 50 humans and 15 dogs to continue these efforts. None of these questions are intractable, and almost all of them are less about the dogs, and more about the humans. Choosing the right clinical application - pre-screening multi-cancer risk from breath, building statistical and AI tools so we can read the dogs with consistency, and automating the manual (human) read-and-write — all of these address these often-cited issues, to unlock a solution that solves problems that current approaches continue to struggle with. Our Journal of Clinical Oncology paper already shows the foundation for how this can be done.
I appreciate the good Doctor's note that canine olfaction could "one day" be part of the solution to democrazting early detection. I disagree that such a day is far away. There are real challenges ahead of us, yes, but we are confident in solving them within the coming months, so that we can provide real clinical value to individuals and communities. Ultimately, our collective goal should be to refuse the status-quo of today, where lakhs of Indians, and millions of individuals world-wide, are detected too late to be given fighting odds for curative treatment.
The Journal of Clinical Oncology paper is where this stops being an argument. More on Dognosis.