Your diagnostic numbers are being recalibrated: p-tau217, a lung-cancer index, patient labels, GxE

Alzheimer's blood test has a kidney problem

Plasma p-tau217 is the blood biomarker everyone wants for Alzheimer's screening, but in a preprint, the authors show it climbs as kidney function falls. Plasma p-tau217 was 36% higher in people with eGFR 60–90 mL/min/1.73 m2 than in those with eGFR ≥90, and 116% higher in people with eGFR below 60. Levels fell 9.5% for every 10 mL/min/1.73 m2 increase in eGFR. m:…9943

The clinical consequence is direct: current fixed diagnostic thresholds for p-tau217 do not account for kidney function, potentially producing false positives. A kidney-blind cutoff risks reading organ decline as brain disease — exactly the confound population-scale screening cannot afford. m:…9943

The dose-response shape is worth noting: p-tau217 falls 9.5% for every 10-unit increase in eGFR — a smooth gradient, not a cliff. That means the confound is not confined to diagnosed kidney disease; it operates across the normal-to-mildly-reduced range where a large share of screening-age adults sit. m:…9943

A cheap lung-cancer signal with a low-confidence stamp

In a preprint meta-analysis, a lower C-reactive protein–albumin–lymphocyte (CALLY) index is associated with poorer overall survival (hazard ratio 1.78) and poorer recurrence/progression-free survival (hazard ratio 1.60) in primary lung malignancies — the kind of cheap, routine-lab prognostic tool clinicians love. m:…5717

But the authors grade the certainty of evidence as low, driven by risk of bias and heterogeneity across studies. So the recalibration here cuts both ways: an inexpensive signal exists, and the meta-analysis is honest that the literature behind it is not yet strong enough to bank on. m:…5717

The hazard ratios are not small — 1.78 for overall survival is a meaningfully worse outlook for the low-CALLY group — which is exactly why the low-certainty grade matters: a tempting effect size on a thin evidence base is the classic combination that has burned the prognostic literature before. m:…5717

For finding undiagnosed patients, ask them

In a preprint on training case-finding models, self-reported diagnoses beat diagnosis codes. Models trained on self-report were more effective than code-trained models at identifying uncoded patients with chronic fatigue and fibromyalgia — for chronic fatigue, a self-report-trained summary-count score reached an AUROC of 0.74 versus 0.67 for the code-trained version. m:…6118

The uncomfortable implication, and this is our reading of the direction of the result: for conditions medicine routinely undercodes — chronic fatigue and fibromyalgia — the medical record may be the noisy label and the patient survey the cleaner one. Elastic net models trained on self-report ranked uncoded self-reporters higher but showed reduced discrimination for later clinical codes — labels, it turns out, are the whole game. m:…6118

There is a subtlety in the elastic net result: self-report-trained models ranked uncoded self-reporters higher than summary-count scores but showed reduced discrimination for later clinical codes. Better at finding the hidden cases, worse at matching the eventual paperwork — which is what you would expect if the paperwork is the lagging, noisy part. m:…6118

Half the gene-environment literature might be a math artifact

In a preprint introducing REGENA, the authors disentangle phenotype scale-induced GxE from biological effect heterogeneity — and find that approximately 50% of trait–environment pairs with significant GxE in UK Biobank data are scale-dependent, an artifact of phenotype scale rather than biology. Modeling the distinction improves environment-specific polygenic prediction by 14% on average. m:…7748

The discovery payoff: environment-stratified GWAS detects 174 loci missed by pooled GWAS, with 67 showing opposite-signed effects across strata — effects that cancel out when you average over environments. It is a recalibration of the GxE literature itself, and of what 'missing heritability' might be hiding. m:…7748

The 67 opposite-signed loci are the sharpest detail: effects that point one way in one environment and the other way in another cancel out in pooled analysis. Our read: pooled GWAS does not just miss these loci, it actively hides them — the signal averages to zero by construction. m:…7748