{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreib3rpfcjkprclenlcfrmy3owmtjxkd43ya5wvd2nc5xic7k7yxjde",
    "uri": "at://did:plc:wnd7xrumusq5uayjfi2pgfno/app.bsky.feed.post/3melibjq5hb52"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreiejjst4k2dixfhvvdjfg6q3pm45e36ns5z4o7dvowoi2ynl732bcu"
    },
    "mimeType": "binary/octet-stream",
    "size": 337118
  },
  "description": "🧬 Delphi-2M AI Predicts 1,000+ Diseases from 1.9B Health Records — But Eurocentric Bias Risks Global Equity\n\n1,000+ diseases predicted from 1.9B health records — enough to map a lifetime of risk for every Dane. 🧬 This isn’t just prediction—it’s preemptive care at scale. But who gets left out when the data is 95% European? Danish patients benefit today—what about the rest of the world?\n\nThe University of Potsdam-led consortium fed 1.9 billion structured EHR tokens—every Danish hospital admissio",
  "path": "/2026-02-11-161577247516580589774083647526905962962/",
  "publishedAt": "2026-02-11T12:31:57.000Z",
  "site": "https://espresso.cafecito.tech",
  "textContent": "## 🧬 Delphi-2M AI Predicts 1,000+ Diseases from 1.9B Health Records — But Eurocentric Bias Risks Global Equity\n\n> 1,000+ diseases predicted from 1.9B health records — enough to map a lifetime of risk for every Dane. 🧬 This isn’t just prediction—it’s preemptive care at scale. But who gets left out when the data is 95% European? Danish patients benefit today—what about the rest of the world?\n\nThe University of Potsdam-led consortium fed 1.9 billion structured EHR tokens—every Danish hospital admission since 1977—into a 2.1-billion-parameter transformer.\nOutput: probability curves for 1,061 ICD-10 codes, plus mortality, for any citizen at any future year.\nInternal test AUC ranges 0.87–0.93 across circulatory, neoplastic and rare congenital blocks; calibration slope 1.02 on 400 k UK-Biobank subjects.\nNo hand-crafted features; the model learns temporal disease trajectories end-to-end, then emits synthetic patient histories for counterfactual “what-if” queries.\n\n### What Makes Longitudinal JEPA Better Than Fine-Tuned BERT for Health Records?\n\nDelphi-2M replaces next-token prediction with a Joint-Embedding Predictive Architecture:\n(1) embed the past five visits, (2) predict embeddings of the next 12 months, (3) decode to ICD codes.\nThis halves parameter count versus BERT-style MLM and keeps 4 k-token context—enough for ten-year histories—within 24 GB GPU RAM.\nQuantization to INT8 costs only 0.4 % AUC, letting Danish regions run inference on a single A100 node for 5.8 M citizens.\n\n### Where Are the Bias Traps in a 50-Year Nordic Dataset?\n\nTop 20 predicted risks over-index for celiac, multiple sclerosis and Nordic-type diabetes; prevalence of sarcoidosis and sickle-cell remain under-represented.\nFix: consortium ships recalibration layers—two 256-neuron feed-forward blocks—that re-weight logits using local epidemiology tables.\nExternal validation on Framingham (23 % non-European ancestry) drops AUC by ≤0.015 after recalibration, within statistical tie of Danish hold-out.\n\n### Can Hospitals Deploy Delphi-2M Without Violating EU AI Act?\n\nModel classifies as “high-risk” under Annex III because it influences individual treatment paths.\nMandatory deliverables: CE-marked technical documentation, risk-management file, human-oversight API that surfaces top-five latent features driving each prediction.\nPilot in Central Denmark Region adds a 1.2 s latency budget; clinician override rate held at 11 %, matching existing clinical-decision-support norms.\n\n### Will Predictive-Analytics Revenue Match the 24 % CAGR Hype?\n\nBudget impact model: flagging 2 % of 40- to 70-year-olds for intensified hypertension control avoids 1,330 myocardial-infarction admissions yearly in Denmark.\nNet savings: €42 M annually; licensing fee ceiling €4.2 M gives region ten-month payback.\nScaled to NHS England (population 56 M) the same math yields £410 M avoidable cost, supporting a £40 M licence—still inside Office-of-Life-Sciences guidance of ≤10 % captured savings.",
  "title": "1,000+ Diseases Predicted from 1.9B Records — Denmark Leads AI Health Care, But Global Bias Looms",
  "updatedAt": "2026-02-11T12:31:57.000Z"
}