Autopsy-confirmed registry pairing antemortem clinical diagnoses against postmortem findings — Goldman Class I/II major error classification
When a patient dies, the clinical record closes with a "final" diagnosis — the physician's best synthesis of history, exam, imaging, and labs. The autopsy asks a blunt question of that synthesis: was it right? Diagnostic discrepancy registries systematically pair antemortem clinical diagnoses against postmortem pathological findings, generating the single most rigorous form of diagnostic error surveillance available in medicine — one uncorrupted by the clinician's own recall or documentation bias.
Every method for measuring diagnostic error has a weakness. Malpractice claims capture only a tiny, biased slice of harm. Voluntary incident reporting is subject to underreporting and hindsight bias. Chart-review studies (e.g., using trigger tools) rely on the same documentation that may already encode the error. The autopsy is different: it is an independent, structural, anatomic ground truth that does not depend on what the treating team believed or wrote down.
A pathologist dissecting a body has no stake in confirming the clinical diagnosis and directly visualizes disease that imaging, biopsy, and clinical reasoning could only infer. This is why autopsy-discrepancy studies, despite their declining numbers, remain the backbone of what we know about the true rate of missed and wrong diagnoses in hospitalized patients — figures no other surveillance method can produce with comparable rigor.
The registry captures three linked data objects per case: the clinical (antemortem) diagnosis at time of death, the autopsy (postmortem) diagnosis after full pathological workup, and a discrepancy classification connecting the two. Over a cohort, these triplets generate a discrepancy rate — the fraction of autopsied deaths in which the clinical team's leading diagnosis and the anatomic truth disagreed.
Goldman et al. (NEJM 1983) reviewed 100 sequential autopsies at two teaching hospitals and found major diagnostic errors (Class I) in 10% of cases — diagnoses that, if made antemortem, would likely have changed management or prolonged survival. This single paper defined the classification scheme still used in discrepancy registries today.
The autopsy rate for non-forensic hospital deaths in the United States has collapsed from roughly 50% in the 1960s to below 10% today — and under 5% at many institutions. Multiple forces drove this decline: the 1971 removal of the Joint Commission's 20%-autopsy-rate accreditation requirement, the rise of "confident" imaging technology (CT, MRI) perceived as making autopsy redundant, cost pressure with no direct reimbursement for the procedure, and cultural/religious reluctance among families.
This creates a troubling epistemic paradox for the registry: as autopsy volume falls, so does statistical power and case-mix representativeness, precisely as the technology environment changes fastest and precisely when we most need contemporary discrepancy data. Selection bias grows too — the cases that still get autopsied tend to be diagnostically puzzling, sudden, or medicolegally sensitive deaths, not a random sample of hospital mortality. A modern registry must explicitly model and report this selection bias rather than ignore it.
At case intake, the registry logs a structured minimum dataset per decedent:
• Antemortem clinical diagnosis(es) — primary cause of death as documented by the treating team, ICD-coded • Clinical context — admission diagnosis, length of stay, ICU status, code status, diagnostic workup performed • Consent pathway — hospital (non-forensic) autopsy requires next-of-kin consent; medicolegal (coroner/ME) autopsy is mandated for specific death circumstances • Time-to-autopsy — interval between death and postmortem exam (affects tissue quality and interpretability) • Requesting clinician's stated diagnostic uncertainty, if documented
This intake record becomes the left-hand "clinical card" in the paired-comparison ledger — later matched against the pathologist's independent findings without the pathologist first reading the clinical chart in detail, a deliberate blinding step some rigorous registries use to reduce anchoring on the clinical impression.
A complete autopsy is a systematic, multi-hour anatomic and histological investigation: external examination, evisceration, organ-by-organ gross dissection, targeted histology, and — when indicated — toxicology, microbiology, and postmortem imaging. The output is the "autopsy-confirmed diagnosis": the cause and mechanism of death established by direct anatomic evidence rather than clinical inference.
A hospital (consented) autopsy follows a standardized protocol:
1. External examination — body habitus, surgical scars, lines/tubes in place, signs of trauma, skin findings 2. Evisceration — organs removed en bloc or individually (Rokitansky vs. Virchow technique) for systematic dissection 3. Gross organ examination — each organ weighed, sectioned, and inspected: heart (coronary arteries opened in cross-section every 3mm, valve and chamber inspection), lungs (parenchyma, vasculature, embolism check), brain (often fixed in formalin 2 weeks before sectioning), GI tract, kidneys, liver, spleen, endocrine organs 4. Histology sampling — representative sections from every organ plus any grossly abnormal tissue, processed to slides and reviewed microscopically 5. Ancillary studies as indicated — postmortem toxicology (blood, vitreous humor, liver), microbiology cultures, postmortem CT/MRI (increasingly used to guide dissection or as adjunct/alternative in some centers), genetic testing for sudden cardiac death syndromes
The pathologist synthesizes all of this into a final anatomic diagnosis list, ranked by causal contribution to death — the artifact that becomes the "autopsy card" in the registry's paired comparison.
Several structural reasons explain why even well-run clinical services generate discordant diagnoses relative to autopsy:
• Occult processes with nonspecific presentation — disseminated malignancy, fungal or mycobacterial infection, and small/silent pulmonary emboli frequently present with vague symptoms (fatigue, low-grade fever, mild dyspnea) that mimic more common conditions • Imaging sensitivity limits — CT pulmonary angiography misses small subsegmental emboli; echocardiography can miss vegetations <2-3mm or right-sided endocarditis; even MRI has false-negative rates for early ischemic stroke and small dissections • Competing diagnosis anchoring — once a plausible diagnosis (e.g., "sepsis of unknown source", "COPD exacerbation") is established, subsequent findings are often assimilated into that framework rather than triggering diagnostic reconsideration (a cognitive bias called anchoring/premature closure) • Terminal-phase confounding — in critically ill, mechanically ventilated, sedated patients, exam findings that would normally localize a new problem (focal neuro deficit, abdominal tenderness) are masked or unobtainable • Truncated workup — a patient who dies rapidly may not survive long enough for confirmatory testing (biopsy, culture results, specialist consultation) to return before death
The autopsy examination directly visualizes the anatomic substrate that these clinical inference chains could only estimate — which is exactly why postmortem findings so often diverge from the documented antemortem impression, and why unsuspected findings appear in a substantial minority of every autopsy series ever published.
Landmark autopsy series spanning five decades — Goldman 1983, Cameron & McGoogan 1981, Sonderegger-Iseli 2000, Roulson 2005, Shojania 2003 meta-analysis, and Winters 2012 ICU meta-analysis — consistently find major (Class I/II) discrepancy rates in the 10–25% range, despite dramatic advances in imaging, laboratory medicine, and molecular diagnostics across that period.
Because fewer than 1 in 10 non-forensic hospital deaths are autopsied in most modern health systems, the denominator of any discrepancy registry is not a random sample — it is a self-selected group. Families and clinicians are more likely to request or consent to autopsy when:
• The cause of death is genuinely unclear or clinically surprising • Death occurred unexpectedly in an otherwise stable-seeming patient • There is a medicolegal, insurance, or family concern about care quality • The case has research or educational value (transplant recipients, novel therapies, teaching hospitals)
This "diagnostic puzzle" selection almost certainly inflates the measured discrepancy rate relative to the true rate across all hospital deaths — a limitation every registry must report transparently, typically by comparing autopsied vs. non-autopsied death characteristics (age, service, admitting diagnosis, length of stay) to quantify how representative the autopsied subset is.
The heart of the registry is a simple but rigorously executed operation: place the antemortem clinical diagnosis card next to the postmortem autopsy diagnosis card for the same patient, and adjudicate agreement. This pairing is deceptively hard to do well — it requires structured criteria, ideally blinded adjudication, and explicit handling of partial matches, multiple competing diagnoses, and cause-of-death chains rather than single labels.
A rigorous paired comparison follows a defined sequence:
1. Extract the clinical cause-of-death chain — as documented on the death certificate and final progress note: immediate cause (1a), antecedent cause (1b), underlying cause (1c), and contributing conditions (Part II) 2. Extract the autopsy cause-of-death chain — the pathologist's independent formulation, ideally drafted before extensive review of the clinical chart to reduce anchoring 3. Map each clinical condition to its closest autopsy counterpart — using standardized terminology (ICD-10 or SNOMED CT mapping) to avoid disagreements that are merely lexical 4. Classify the match at each chain position: concordant (same diagnosis, right organ system and etiology), partially discordant (right organ/syndrome, wrong specific etiology — e.g., "pneumonia" correct but organism wrong), or fully discordant (different disease process entirely) 5. Roll up to a case-level discrepancy verdict — driven primarily by concordance of the immediate and underlying cause of death, since errors in minor contributing conditions are weighted less heavily
Because single-reviewer adjudication is subject to idiosyncratic judgment, higher-rigor registries use two independent reviewers (often one clinician, one pathologist) with a third adjudicator for disagreement — analogous to systematic-review methodology in evidence synthesis.
A subtle but important methodological issue: if the pathologist reads the full clinical chart before performing the autopsy, their gross and histologic search pattern can be unconsciously steered toward confirming the clinical impression (a form of diagnosis momentum crossing disciplines). Conversely, a pathologist with zero clinical context risks missing correlations that focus the dissection efficiently (e.g., not knowing about a heart murmur might mean less careful valve inspection).
High-rigor discrepancy registries address this with staged information release: the pathologist receives only demographic data and the immediate circumstances of death before dissection, performs the gross exam and forms a preliminary anatomic impression, and only then reviews the full clinical chart before finalizing the report — allowing genuine independent anatomic assessment while still benefiting from clinical correlation for final synthesis. This staged-blinding protocol is a marker of registry quality and should be reported in registry methods sections, though it remains a minority practice due to workflow burden.
Studies that explicitly compared blinded versus non-blinded autopsy diagnosis generation found blinded pathologists identified 15–20% more discordant cases than pathologists who reviewed the full chart first — evidence that the pairing methodology itself measurably shapes the discrepancy rate a registry reports.
Real cases rarely present as clean binary "agree/disagree" pairs. Common complex scenarios the registry must encode explicitly:
• Right syndrome, wrong etiology: clinical diagnosis "septic shock," autopsy reveals fungal (not bacterial) source — same clinical syndrome, different causative organism, different indicated antimicrobial therapy → typically scored as major discrepancy given treatment implications • Right organ, wrong process: clinical "pneumonia," autopsy reveals pulmonary infarction from silent PE — same anatomic location, entirely different disease mechanism and treatment → major discrepancy • Multiple simultaneous processes: patient with both correctly diagnosed CHF and an unsuspected occult malignancy contributing to death — registry logs both a concordant element and a discordant element within the same case, rolling up to an overall "partial discrepancy" case classification • Timing-sensitive discordance: correct diagnosis made, but so late in the clinical course that it could not plausibly have changed outcome — scored differently (Class II, not Class I) precisely because timing, not diagnostic accuracy per se, is the operative variable
Encoding this complexity, rather than forcing every case into a single binary label, is what separates a research-grade discrepancy registry from a simple tally.
Not every diagnostic discrepancy carries equal weight. A missed diagnosis of a slow-growing benign process found incidentally at autopsy is categorically different from a missed aortic dissection that, if caught, would have gone straight to the OR. The Goldman classification (1983), still the field's reference framework, sorts discrepancies by whether recognition would plausibly have changed management or survival — turning a simple concordance tally into an actionable severity-weighted signal.
Goldman et al. (NEJM 1983) proposed a four-tier severity schema, still the backbone of modern discrepancy classification (later refined by Battle 1987 and the Goldman-Battle-Podelski framework):
Class I — Major diagnostic error, missed diagnosis directly caused death or would likely have changed management/survival if recognized antemortem, AND is unrelated to the terminal/admitting condition (a genuinely missed, distinct disease process). Example: unsuspected aortic dissection in a patient treated for presumed MI.
Class II — Major diagnostic error with the same clinical significance as Class I, but which would likely NOT have changed the outcome even if diagnosed correctly (e.g., disease too advanced, patient too unstable for any intervention, or condition untreatable regardless of timing). Example: disseminated malignancy discovered at autopsy in a patient who died of multi-organ failure — earlier diagnosis would not have altered the terminal course.
Class III — Missed diagnosis related to the terminal disease process but of comparatively minor clinical significance — would not have changed management even under ideal conditions. Example: an additional small pulmonary embolus found incidentally alongside a correctly diagnosed and treated massive PE.
Class IV — Other discrepancies: incidental findings unrelated to the cause of death, minor errors in secondary/contributing diagnoses, coding or terminology mismatches without clinical import.
The registry's headline metric — "major discrepancy rate" — is conventionally defined as Class I + Class II combined, since both represent genuinely missed major diagnoses; Class I is the operationally most important subset because it identifies preventable-in-principle harm.
The Class I / Class II split is what makes the registry actionable rather than merely descriptive. A high Class I rate signals a genuine, addressable failure of the diagnostic process — the right information may have been available but wasn't synthesized correctly, or a key test wasn't ordered, or a finding was misattributed. These are the cases root-cause-analysis and M&M conference should prioritize, because a hypothetical "what could we have done differently" has real counterfactual weight.
A high Class II rate, by contrast, often reflects the fundamental limits of medicine against advanced or rapidly fatal disease rather than a correctable process failure — although it still has value for epidemiological surveillance (e.g., tracking how often specific occult conditions like disseminated fungal infection or silent malignancy appear) and for calibrating clinician and public expectations about diagnostic uncertainty in critical illness.
Separating the two prevents two symmetric errors: treating every discrepancy as a preventable error (demoralizing and imprecise), or dismissing all discrepancies as unavoidable (which would blunt genuine quality-improvement signal).
A 2002/2003 systematic review by Shojania, Burton, McDonald & Goldman — the same Goldman — pooled autopsy discrepancy studies from 1966–2002 and found the major error (Class I) rate had declined gradually over decades (roughly 10% per decade reduction), attributed to improved imaging and diagnostics — but cautioned that shrinking, non-representative autopsy samples make trend interpretation increasingly uncertain.
Since 1983, several groups have adapted the core Goldman logic for modern settings:
• ICU-specific schema (Combes 2004, Winters 2012 meta-analysis): critical-care discrepancy studies pool Class I rates of 8% and combined major discrepancy (Class I+II) of ~23% across mixed ICU populations, reflecting the diagnostic difficulty of critically ill, sedated, multi-organ-failure patients • Autopsy Confirmation Score adaptations: some registries now add a structured confidence rating on the strength of anatomic evidence supporting each discrepancy classification, since not every gross/histologic finding is equally unambiguous • Weighted severity scoring: rather than a categorical I–IV label, some modern registries assign a continuous severity score (0–100) incorporating estimated potential-years-of-life-lost and treatability, allowing finer statistical modeling of discrepancy burden across a cohort • Root-cause linkage: contemporary registries increasingly link each Class I case to a contributing-factor taxonomy (cognitive bias type, system/communication failure, test not ordered, result not followed up) — connecting the discrepancy registry to root-cause-analysis and cognitive-debiasing programs elsewhere in the diagnostic-safety literature.
Despite these refinements, the essential Goldman I–IV logic — discordant, related to death, and would/would-not have changed outcome — remains the dominant classification vocabulary in the field more than four decades after publication.
A single autopsy discrepancy is a case report. A thousand of them, filed consistently over years with stable classification criteria, become an epidemiological instrument capable of tracking whether diagnostic accuracy is improving, whether specific disease categories are chronically under-recognized, and whether a health system's diagnostic performance is drifting — the registry's ultimate purpose.
Each adjudicated case is filed into the permanent ledger with a structured record:
• Unique case ID, de-identified patient demographics, admitting service, length of stay • Full clinical diagnosis chain and full autopsy diagnosis chain (structured, coded terms) • Discrepancy classification (concordant / Class I / Class II / Class III / Class IV) with adjudicator identifiers • Contributing factor tags, where root-cause analysis was performed (cognitive, system, communication — see the companion Diagnostic Error Root-Cause Analysis registry) • Time-stamps: admission, death, autopsy performed, final report issued — enabling turnaround-time and timeliness analysis
This structured ledger is what enables the registry to answer questions no single case can: is the major discrepancy rate for stroke different from that for sepsis? Has the rate for pulmonary embolism changed since CT-angiography became routine? Does discrepancy rate vary by time of admission, service, or resident-vs-attending primary team?
The trend line plotted across the cohort — concordance rate (1 − major discrepancy rate) over time or over accumulating case count — is the registry's central analytic output. Three confounders must be actively managed to interpret it correctly:
1. Declining autopsy volume changes the denominator's composition over time (see Stage 2) — a rising trend line might reflect genuinely better diagnostics, or might simply reflect that only easier, less puzzling cases are still being autopsied as volume falls. Registries mitigate this by stratifying trend analysis by admitting service, case complexity score, or requesting-clinician stated uncertainty, so like is compared to like across periods.
2. Case-mix drift — as medicine advances, the mix of diseases causing hospital death shifts (e.g., relatively more deaths now occur in older, multimorbid, oncology and transplant patients than in the 1960s cohort Goldman studied) — a shift in what people die of will move the discrepancy rate independent of diagnostic skill.
3. Classification drift — if adjudication criteria or adjudicator training changes over the registry's life, an apparent trend may be an artifact of measurement rather than a real change in diagnostic accuracy; version-controlling the classification rubric and periodically re-adjudicating a sample of historical cases with current criteria is standard registry hygiene.
With these controls in place, the pooled literature does show a real, if modest, secular decline in major diagnostic error at autopsy over the second half of the 20th century — improved imaging, laboratory medicine, and specialist availability appear to have genuinely reduced the classic "missed MI" and "missed PE" errors that dominated Goldman's original 1983 cohort, even as new categories of diagnostic difficulty (multidrug-resistant infection, complex oncologic disease, polypharmacy in the very old) have emerged to partially offset the gain.
Because autopsy rates below roughly 20-30% are widely considered too low and too selection-biased to support institution-level benchmarking, many contemporary quality bodies argue for restoring minimum autopsy rate targets specifically to keep the discrepancy registry's statistical signal interpretable — a direct methodological argument for autopsy's continued clinical relevance in the era of advanced imaging.
A mature discrepancy registry closes the loop back into clinical practice through several channels:
• Morbidity & Mortality (M&M) conference case selection — Class I discrepancies are natural candidates for structured case review and teaching • Diagnostic calibration feedback — aggregated, de-identified discrepancy patterns by service or diagnosis category inform curriculum and checklist design (e.g., a persistently high missed-PE rate on a given service triggers a targeted diagnostic pathway review) • Autopsy-rate advocacy — registries with declining case volume often use their own data to make the institutional case for restoring consented autopsy rates, since the registry cannot function without a representative denominator • Cross-linkage with other diagnostic-safety instruments — root-cause taxonomies, cognitive-bias debiasing checklists, and closed-loop test-result-follow-up systems all draw on discrepancy-registry findings to prioritize where limited quality-improvement effort should be spent
The registry, in this sense, is not an archive of past failure but an operating instrument — its principal value lies in the discipline of forcing every autopsied death to answer one uncomfortable, structured question: did we get it right, and if not, could we have?