Iris Holm: You look like you've been chewing on something since you walked in.
Lila Soto: I have, yeah — I want to hand you a scenario and I want your first reaction, not your considered one.
Lila Soto: Woman, late forties, routine mammogram. They find microcalcifications — small cluster, nothing dramatic. Biopsy, lumpectomy, six weeks of radiation. Five years pass. She sees a new oncologist, who looks at the original report and says — kind of casually — the finding was DCIS, ductal carcinoma in situ, and probably would never have progressed. She did all of that for probably.
Iris Holm: First reaction? That's not an edge case. That's the twelve-point-six percent.
Lila Soto: Wait — say that number again for people just tuning in.
Iris Holm: Twelve point six. That's the pooled overdiagnosis incidence from a systematic review of screening mammography — women forty and older, confidence interval ten-point-one to fifteen-point-six. Roughly one in eight screened women gets diagnosed and treated for something that was never going to harm her.
Lila Soto: And that's — okay, the thing I want to make sure lands is what overdiagnosis actually means. Because it sounds like 'the test was wrong.' It's not that.
Iris Holm: Right. The pathology is real. The cancer is there. It's just — it would never have surfaced. Never caused symptoms. Never killed anyone. Detecting it doesn't help. Treating it causes harm.
Lila Soto: So the test worked. That's what makes this so strange to sit with. The screening mammography did exactly what it was built to do — and the woman still went through surgery and radiation she didn't need.
Iris Holm: That's the thesis. Not bad doctors. Not bad machines. Something in the logic of testing people who feel fine.
Lila Soto: Which makes me wonder — if the test working perfectly is still the problem, what would a solution even look like?
Iris Holm: A solution would need to solve for something we haven't named yet. It's not the test. The test finding the cancer is — that's exactly correct. The problem is underneath: there's a reservoir. A population-level pool of cancers that are biologically real, confirmed at pathology, and will never surface. You can't design a test sensitive enough to find cancer and too precise to find that cancer.
Lila Soto: Wait — so the reservoir just... exists? In people walking around right now?
Iris Holm: Yes. That's what makes this structural. Screening taps it. Every time.
Lila Soto: But how do we know those cancers would never have progressed? I mean — isn't that kind of impossible to prove from the inside of a single case?
Iris Holm: Lead-time bias. Look, the intuition is that catching cancer earlier means living longer. But if the clock starts earlier and still stops at the same time, you didn't extend survival. You extended the period of being a cancer patient. The five-year survival rate looks better because detection moved the start date. Time of death: unchanged.
Lila Soto: That's — yeah. That's genuinely uncomfortable. Because five-year survival is basically the metric everyone uses to measure whether screening works.
Iris Holm: Right. And this is where the historical line matters. Cervical cancer screening — mid-twentieth century — that was a real success. Genuine mortality reduction. And it became the template. Every program that came after inherited that logic. But when you apply it to prostate or breast cancer, the indolent disease reservoir is far larger. The model doesn't transfer.
Lila Soto: So we built the whole architecture on a case that kind of... happened to work, and then assumed it would generalize.
Iris Holm: Exactly. And low-dose CT for lung cancer — meta-analyses of randomised controlled trials show it meets all three structural conditions. The reservoir exists, LDCT detects it, and competing causes of death in the screened population reduce whatever net benefit you'd get from catching it. The same mechanism, again.
Lila Soto: Joseph Wu's framing is something I think about a lot. He argues that even if you prove screening reduces mortality at the population level, that doesn't actually solve the problem. Because you still can't tell, at the individual level, which detected disease is the dangerous kind.
Iris Holm: That's the incorrigible part. Some harms — you can engineer around. Better risk stratification, maybe. But the inability to distinguish harmful from harmless detected disease? That's not a design flaw. That's baked into the logic of testing people who feel nothing.
Lila Soto: Which means asking whether screening is effective — whether it saves lives — might be the wrong question to start with.
Iris Holm: And if the question is wrong, then so is every answer we've been building on top of it. Which brings you directly to the cascade.
Lila Soto: Walk me through it. Like — step by step, what actually happens after that one abnormal result.
Iris Holm: Diagnostic imaging first. Then biopsy. Then, depending on what pathology shows, surgery — maybe chemotherapy, maybe radiation. Each of those steps has its own independent risk profile. You don't just get the harm from the original unnecessary diagnosis. You get compounding harm from every intervention that follows it.
Lila Soto: So the false positive doesn't just mean a scare. It means you might actually end up under a knife.
Iris Holm: The false positive alone triggers workup. And even if the biopsy comes back and you're clear — that biopsy happened. That risk was real.
Lila Soto: Okay, the PSA story is — I mean, that's kind of where this gets impossible to look away from. Because with prostate cancer the overtreatment isn't a hypothetical. The harms are specific. Urinary incontinence. Erectile dysfunction. Those are guaranteed consequences of treatment, and the mortality benefit is still — what, contested?
Iris Holm: The U.S. Preventive Services Task Force contested its own guidance on PSA testing because the harm-benefit calculus is genuinely unfavorable for many men at the population level. That's a federal body saying — actually, we're not sure this helps.
Lila Soto: And the competing risk piece underneath that is — most men whose prostate cancer gets detected through PSA screening will die of something else first. The cancer was never going to be what got them.
Iris Holm: So the mortality benefit is uncertain. The treatment harm is not. That's not a tradeoff. That's a guaranteed loss for a speculative gain.
Lila Soto: Overtreatment. That's the word, right? Real procedures. Real side effects. Zero clinical benefit for that person.
Iris Holm: And that word — overtreatment — only lands if you're willing to say the detection itself was the problem. Which is uncomfortable because the detection was correct. Pathology confirmed it. The tumor was there.
Lila Soto: And the numbers underneath all of this — the meta-analyses, the I-squared figures — there's a version of this conversation we haven't gotten to yet where even the evidence base for evaluating these programs turns out to be broken in a different way.
Iris Holm: Broken in a different way — that's the right frame. The twelve-point-six percent figure. Everyone cites it. But the I-squared on that meta-analysis is ninety-nine-point-nine-nine-three percent.
Lila Soto: Wait — what does that number actually mean?
Iris Holm: I-squared measures heterogeneity across studies. A hundred percent means the studies are measuring completely different things. Ninety-nine-point-nine-nine-three is — statistically, that aggregate is noise. We're issuing population-level guidance to millions of women on a point estimate that the math itself says you shouldn't treat as a number.
Lila Soto: So the confidence interval — ten-point-one to fifteen-point-six — that's not a tight band. That's the studies wildly disagreeing with each other and us averaging across the disagreement.
Iris Holm: And publishing the average. As if it settles something.
Lila Soto: Okay but — I mean, why are the studies so different from each other? Is it different populations, different definitions of overdiagnosis, or — actually, is it that the trial designs themselves are asking different questions?
Iris Holm: That last one. That's the load-bearing issue. Screening trials evaluate whether the test detects disease. Not whether detecting it changes outcomes that matter. The question of clinical necessity — would this ever have harmed the patient — is often not built into the trial design at all.
Lila Soto: Huh. So we've been running trials for decades that are — kind of asking the wrong thing by design.
Iris Holm: And when disease definitions expand — eligibility broadens, more people get screened — those expansions don't require evaluation of overdiagnosis harms. And the panels making those calls frequently have conflicts of interest. So the framework never corrects.
Lila Soto: Thyroid screening is telling. Countries with intense thyroid cancer screening programs have much higher incidence rates. Same mortality. The detection went up, the deaths didn't go down. That's almost like a controlled proof of the whole argument.
Iris Holm: Right. And it's not mammography, it's not LDCT, it's not PSA — it's a different modality entirely, same structural outcome. Which tells you this isn't a mammography problem. It's a feature of testing asymptomatic populations for anything with an indolent reservoir.
Lila Soto: So risk stratification — the idea of distinguishing which findings are actually dangerous — that would be the fix, in theory. And it kind of exists. But the tools aren't there yet, which means a clinician can recognize the problem clearly and still have nothing to do about it on a Tuesday morning with a patient sitting across from them.
Iris Holm: And informed consent is supposed to fix that gap. That's the mitigation on the table. Tell the patient the overdiagnosis risk, hand them balanced information, let them decide. Except — communicating conditional probabilities to someone sitting in a clinic who just heard the word cancer? That's not a knowledge transfer. That's noise landing on a person in crisis.
Lila Soto: I mean, it almost functions as institutional protection more than — yeah. Like, we told you. So now the decision was yours.
Iris Holm: Distributes the liability downward. Doesn't change the decision architecture. The patient still can't weigh a probabilistic harm against the psychological relief of 'we caught it early.'
Lila Soto: Which is where liquid biopsy comes in, kind of — as the hopeful next chapter. Minimally invasive, biofluid biomarkers, earlier detection. And I keep wondering if that's actually a solution or if it just — oh, I don't know how to phrase this — if it just relocates the problem to an earlier address.
Iris Holm: It relocates it. A more sensitive test that detects abnormalities at even earlier stages expands the pool of screen-detected disease. It does not improve our ability to tell harmful from harmless. If anything, the findings are earlier and therefore harder to characterize clinically. The discrimination problem gets worse.
Lila Soto: And liquid biopsy has barely moved from research to clinical practice anyway. Like, a limited number of assays have actually crossed that line. So we're projecting hope onto a technology that hasn't proven it can carry it yet.
Iris Holm: Unless the field changes what it calls success. That's the real condition. Detection rates — that's what the current framework rewards. If you don't shift the measure toward clinical outcomes that account for overtreatment harm, every new technology inherits the same structural tension. Liquid biopsy, whatever comes after it — doesn't matter.
Lila Soto: That's — yeah. That's the part that doesn't resolve. Because changing what success means requires admitting that what we called success before wasn't. And there's a whole culture built around early detection as — almost a moral act. Being the kind of person who gets screened. That framing is so deep I'm not sure a better assay touches it.
Iris Holm: The technology problem might actually be the easier one.
Lila Soto: We started with that woman in her late forties — routine mammogram, probably fine, and then six weeks of radiation for something that probably would never have progressed. That same Tuesday morning still happens. And I think what sits with me is — she walked in doing everything right. By every cultural signal she'd been given. And the system worked exactly as designed, all the way through the treatment cascade. I don't know what you do with that except hold it.
