How we know what we show

Evidence, graded honestly

Not all evidence is equal, and NephTox does not pretend otherwise. References are typed by study design (meta-analysis, RCT, cohort, case series, pharmacovigilance, mechanistic), and pharmacovigilance signals (FAERS reporting odds ratios) are labeled as reflecting disproportionate reporting — not incidence or proven causation. Where the renal signal for a newer agent is extrapolated from its class rather than established directly, that is stated.

Explore the full bibliography on the evidence base, see the field's evolution on the insights hub, or read the guidelines the atlas is aligned to.

FAERS, reporting odds ratios, and why they diverge from the literature

The FDA Adverse Event Reporting System is a spontaneous database: clinicians, manufacturers and patients file reports voluntarily. It has a numerator and no denominator. Nobody knows how many people took the drug, so FAERS can never yield an incidence, a risk or a rate — and this atlas never prints one from it.

What it can yield is disproportionality. A reporting odds ratio asks a comparative question: among all reports naming this drug, is this kidney phenotype mentioned more often than it is in reports about every other drug? An ROR of 3 means the phenotype is reported about three times as often as background — not that the drug triples anyone's risk. Every ROR here carries a 95% confidence interval, and a signal counts as significant only when that interval's lower bound clears 1. Below a floor of 50 total reports the atlas shows counts and refuses to divide at all: a death share of 3/7 renders “42.9%”, which reads like a rate when the true interval runs from roughly 10% to 82%.

The arithmetic is a four-cell table, and it is worth seeing because the fourth cell is where most misreadings come from. For one drug and one kidney phenotype: a is the reports naming both; b the drug's other reports; c that phenotype reported for every other drug in the database; and d everything else FAERS holds. The ratio is (a × d) / (b × c), with a 95% interval from exp(ln ROR ± 1.96 × √(1/a + 1/b + 1/c + 1/d)). Every cell must be populated or no ratio is published at all.

That third cell is the comparator, and it is the whole database — not other cancer drugs, and not patients with the same disease. So a phenotype that is simply common in advanced cancer is disproportionate for a great many anti-cancer agents at once, which is why hyponatraemia and haematuria produce broad signals that say more about the population than about any drug in it.

Reporting is also biased in ways that have nothing to do with the drug. A known toxicity gets reported more once it is known. Patients on anti-cancer therapy are volume-depleted, receive contrast and antibiotics, and have a disease that itself injures the kidney — none of which FAERS can adjust for. And the comparator is all other drugs, so a phenotype common across oncology inflates for the entire catalog at once.

This is why the atlas treats the two sources asymmetrically. The literature-documented lesion is what the atlas asserts; FAERS corroborates it or is silent, and silence is usually a property of the reporting vocabulary rather than evidence against the lesion. Naming acute tubular necrosis or interstitial nephritis on a report requires a biopsy, so real cases are filed as generic acute kidney injury and the specific term never appears. Running the other way, a phenotype is sometimes queried with terms that are merely consistent with it — haematuria for haemorrhagic cystitis, hyponatraemia for SIADH — both of which are common in advanced cancer for reasons that are not the drug. The insights hub reports that decomposition directly rather than a bare agreement score, because most apparent disagreement is one of those two vocabulary effects and only the remainder is a real open question.

The second effect used to be stated rather than measured, and that was a real weakness: when a phenotype is queried as a single OR over both kinds of term, the result cannot say which half carried the signal. So each phenotype's terms are classified — does this term name the lesion, or is it merely consistent with it? — and where the set mixes the two, the naming terms are asked a second time on their own, against their own comparator pool. Haemorrhagic cystitis is asked without haematuria; SIADH without hyponatraemia; nephrotic syndrome and the named glomerulonephritides without proteinuria. Where the narrow query still signals, the reporting really is about the lesion. Where it goes flat, the broad signal was the background all along, and the atlas says so instead of leaving the reader to guess.

Two term sets deliberately have no naming half. Crystal nephropathy is queried through stones and obstruction, which in this population are frequently the tumour; and the electrolyte node asserts a renal handling defect, which a reported potassium or calcium value does not establish — hypercalcaemia in advanced cancer is bone metastasis or PTHrP far more often than a tubular defect. Only hypomagnesaemia points reliably enough at the tubule to count as naming it. Signals resting on the rest are reported as what they are: real reporting, not attributable to the lesion.

The residue that survives all of this — a naming term disproportionate, no profile in the clinical family documenting the lesion, no biopsy requirement explaining the term away — is not left standing as a permanent mystery. In August 2026 every such cell was adjudicated against PubMed, one by one. Three outcomes: where the literature turned out to document the association, it entered the drug's profile with its citations; where the treated population, a co-administered drug, or class-level rather than agent-level literature explained the reporting, the cell is marked explained and counts alongside the not-attributable band; and where searching found nothing credible either way, the cell stays a lead — an open one, but a reviewed one. A cell shown as unreviewed is a signal that appeared after that review: the next worklist, not a claim.

Nothing derived from FAERS ever sets a drug's injury signature. A reporting signal with no literature behind it is a lead for review — adjudicated when a reviewer has searched it, unreviewed until then — and never a lesion the atlas claims.

CTCAE — grading severity, which is not the same as naming a lesion

Where this atlas states a grade, it grades against CTCAE v5.0, the NCI's Common Terminology Criteria for Adverse Events. For acute kidney injury that is a creatinine ladder: grade 1 Creatinine increase of >0.3 mg/dL; creatinine 1.5-2.0x above baseline · grade 2 Creatinine 2.0-3.0x above baseline · grade 3 Creatinine >3.0x baseline or >4.0 mg/dL; hospitalization indicated · grade 4 Life-threatening consequences; dialysis indicated. SITC's checkpoint-inhibitor toxicity guideline directs that immune-related AKI be graded by CTCAE, which is what makes it the right shared vocabulary rather than any single society's own bands.

It is held in one place for a reason. The creatinine bands had been restated by hand in five, and had drifted: one profile put grade 2 at 1.5–3× baseline, which pulls grade 1 patients into “hold the drug and start steroids”, and two tools had dropped grade 3's absolute >4.0 mg/dL limb, so a patient at 2.5× baseline with a creatinine of 4.5 mg/dL graded 2 instead of 3 — the difference between “rechallenge may be reasonable” and “generally not advised”. CTCAE is published standard vocabulary rather than a claim this atlas makes, so it carries no PMID of its own.

The limit worth stating plainly: a CTCAE grade measures how far the creatinine moved, not what happened to the kidney. Grade 3 AKI is the same grade whether the lesion is acute tubular necrosis, interstitial nephritis, a thrombotic microangiopathy or obstruction — and the same number can be reached by a drug that never touched the glomerulus, which is exactly what the pseudo-AKI agents do when they block tubular creatinine secretion without changing filtration.

Which is the thread running through this whole section. The atlas carries three vocabularies answering three different questions — the literature names the mechanism, MedDRA and FAERS record what somebody reported, CTCAE measures how severe the number got. They are not three measurements of one quantity and cannot be averaged into one. Where they diverge the atlas says which question each was answering rather than picking a winner.

The immune-related adverse event atlas

The irAE atlas covers immune checkpoint inhibitors only, and is sourced differently from the rest of the site. Each card is anchored on the society chapters this atlas holds for that specific organ — ASCO, ESMO and SITC, though not every organ has all three, and each page names its own — and the supporting sentence is stored beside each dose, threshold and rate verbatim, so a claim about the lung is never carried on a sentence about the kidney. Where a card also shows a position from another kind of source — a single center's protocol, or a grading vocabulary such as CTCAE — it is labeled as such beside the position, because those do not carry a society guideline's weight.

Two things it does that a conventional reference does not. Where a guideline states nothing, the card says so instead of filling the gap: a rung with no sourced corticosteroid dose reports that the dose is not carried, and each entry lists what it deliberately omits and why. And where two societies genuinely disagree — the grade-2 pneumonitis dose, how long a colitis taper should run, whether to scope at grade 2 — both positions are shown side by side rather than one being chosen silently.

NephTox is medical-education content — not medical advice. It supports, but does not replace, the judgment of a qualified clinician and confirmation against current primary sources.