On September 4, Novartis reported two results from the same trial.
The drug lowered the marker it was built to lower. The trial missed its goal.
The Signal
A marker is a number doctors can measure today that stands in for something that takes years to show up. Cholesterol stands in for heart attacks. The bet behind every marker is simple: move the number and you move the outcome.
Lp(a) is one of those numbers. It is a particle in the blood, mostly inherited, and high levels are linked to heart disease. Novartis and Ionis built a drug, pelacarsen, to bring it down.
They tested it on 8,323 people who already had cardiovascular disease and high Lp(a). Some got the drug, some got a placebo. The drug did its job on the number; Ionis said Lp(a) fell substantially.
Then they counted what the number is supposed to predict: heart attacks, strokes, cardiovascular deaths and emergency procedures on blocked heart arteries. The trial could not show fewer of them than the placebo. Novartis wrote that the findings “did not demonstrate” reduced cardiovascular risk.
Some markers have earned that trust. Lowering LDL cholesterol has decades of trials showing fewer heart attacks. For Lp(a), that proof does not exist yet.
Nineteen days later, the same bet reached something you can buy. Galleri is a $949 blood test from Grail that looks for signs of many cancers at once. Here the number is a cancer found early. What it is supposed to predict is fewer people dying of cancer.
Galleri is sold today without FDA approval. On September 23, a panel of FDA advisers voted 10 to 0 that there is enough evidence it is safe. On whether there is enough evidence it works, the vote was 6 to 4.
Its biggest trial, run in England, enrolled more than 140,000 people. It compared people who got three yearly tests with people who got usual care. The goal was fewer cancers caught late, at stage III or IV, among twelve deadly types. Those late diagnoses came out about the same in both groups.
One narrower measure, stage IV alone, fell 14%. It was not the question the trial was built to answer. The trial has no data yet on deaths. Follow-up continues.
I have spent 25 years around technology sold on the number that could move by the next quarter. Health has the same temptation with higher stakes. The marker moves in weeks. The outcome takes years.
The question I run everything through is how this serves life. In health that question has a literal answer: how many people avoided the heart attack.
Pelacarsen moved the number. The trial could not show it moved the count.
The Application
More Lp(a) trials are still running. The National Lipid Association lists four other outcome trials, from Amgen and Lilly, testing whether their own Lp(a) drugs reduce cardiovascular events. They count the thing pelacarsen’s trial could not show.
What Galleri finds. In the FDA’s figures for the final version of the test, it caught about a third of the cancers diagnosed within a year. For early-stage cancers, it caught fewer than one in four. When it flagged cancer, it was right at least two times in three. The proposed label says a negative result does not rule out cancer, and that people should keep up their usual screening.
Gut bacteria and machine learning. A paper posted in September by one researcher, not yet peer reviewed, trained models to spot cancer from gut bacteria in stool samples. On samples held back from the studies it learned from, the best model scored 0.77, on a scale where 0.5 is a coin flip and 1.0 is perfect. On newer studies it had never seen, published from 2023 on, it scored 0.60. The author warns that testing a model on the same studies it learned from makes it look better than it is. That score on familiar data is a marker too: it stands in for how the tool will do with real patients.
The Noise
“Lower your biological age.”
Biological age tests read chemical tags on your DNA and turn them into a number they call your biological age. The algorithms that do it are called clocks.
On August 21, Nature Medicine published a large check of whether anything moves those clocks. Several of its authors work for or consult for TruDiagnostic, a company that sells these tests.
Of about 50 interventions, 13 lowered the clocks by the authors’ strict count. Five studies of senolytics, drugs meant to clear out aging cells, sent the clocks in different directions. The study did not measure whether anyone got healthier.
The authors write that nobody has defined how much a clock has to move to matter for health.
The Question
Think of the health number you check on yourself, the one that makes you feel better when it improves.
Has anyone ever shown that moving it changes what actually happens to you?
Now What?
The same question, five moves, for when you are the one buying health tests or programs for other people.
Ask every health vendor which outcome changed. A better score is a marker. So is a cancer found earlier. Ask what happened to people, such as fewer heart attacks or fewer deaths, in a named trial with a comparison group. If the evidence stops at the marker, say so in the decision memo.
For AI health tools, ask how they did on data they never saw. Ask for results from hospitals or studies the model was not trained on. The gut bacteria model went from 0.77 to 0.60 when it met newer studies. A vendor that only has results on familiar data has not answered the question.
Before offering a multi-cancer test as a benefit, budget the follow-up. Every positive result needs a diagnostic workup, and someone pays for it. In the FDA’s figures, up to a third of positives did not lead to a cancer diagnosis within a year. Ask the vendor how many positives to expect per thousand people like yours, and decide who pays before the first result comes back.
Write the negative result down carefully. In those same figures, about two of every three cancers diagnosed within the year were missed by the test. Use the label’s own words on negative results in what you send employees: a negative does not rule out cancer, and usual screening continues.
For biological age claims, ask two questions. Did several clocks agree in the same study? Did a second study of the same intervention find the same thing? Those are the two tests the Nature Medicine authors propose. Then ask whether the vendor sells the test it is quoting.
What I’m Watching
Post-quantum encryption arriving by update. Oracle released Java 27 on September 15. Java now offers, by default, encryption designed to resist a future quantum computer. Apps that upgrade and keep the defaults start offering it on modern encrypted connections.
Retailers choosing which AI agents get in. Meta launched Muse in the US on September 8. It is a personal AI agent people use for tasks that include shopping. Amazon asked Meta to take Amazon out of Muse, then blocked Muse on its site from September 20. Shopify partnered with Meta, so Muse users can pay with Shop Pay in Shopify stores.
Firm power from plants that already exist. Google signed a 22-year deal for up to half the output of Finland’s Loviisa nuclear plant, power that runs around the clock in any weather. It is Google’s first agreement to help keep an existing nuclear plant running longer and get more power out of it.
This is the Javier D’Ovidio Newsletter. Emerging tech without the hype. Real signals for strategic decisions.
If this edition helped you see something you were not seeing before, forward it to one leader who needs to be in the conversation.
Subscribe at newsletter.javierdovidio.com so you catch the next edition. Follow Javier D’Ovidio on LinkedIn for daily signals on the convergence.
How I make this
The judgment in this newsletter is mine. Every edition is the output of a framework I have built over 25 years in technology. I map what is happening across five forces. Then I filter signal from noise, and run each thing through one question: how does this serve life.
I use AI as a tool in that process. It helps me sweep sources across the five forces and draft. It does not decide anything. I choose the topic. I verify every figure and claim against primary sources before it goes out. I call what is signal and what is noise, and that call rests on having watched several technology waves arrive and get dismissed. I edit every edition against my own writing standards. The framework and the final call are mine, and I am accountable for every claim here.
That is the point of this newsletter. AI is a tool a person uses. The person stays responsible for what it produces.