AI in Healthcare: Mayo's AI sees pancreatic cancer 16 months early  May 13th 2026

Watch & Listen First

• WBUR — "Introducing AI in the Doctor's Office" (May 5, 2026) · A four-part series anchored at Mass General Brigham and Boston-area primary care, the kind of grounded clinician-and-patient reporting that's been missing from the discourse. Listen
• WBUR — "Many people now trust AI with their feelings" (May 7, 2026) · Boston therapists react to actual ChatGPT/Claude/Gemini mental health transcripts; the companion interactive lets you score the bots against clinician feedback. Listen

Key Takeaways

• Pancreatic screening just got plausible. Mayo's REDMOD flagged 73% of prediagnostic cancers ~16 months out on CTs read as normal — but 19% FPR rules out general screening; rollout has to target elevated-risk cohorts.
• CMS made AI a chronic-care vendor. ACCESS launches July 5 with 150 participants and a PMPM thin enough that the math only closes with heavy automation — the first Medicare model designed around AI economics.
• Utah held the line on autonomous prescribing. Medical Licensing Board's suspension request denied; Doctronic continues in phase one with physician review. The state-level AI-clinician licensing test case.
• Open-weight clinical AI moved up the stack. Google's MedGemma 1.5 adds 3D CT/MRI and whole-slide pathology to a 4B open model, +47% macro-F1 on WSI vs MedGemma 1.
• The benchmark gap is the story. Nature Medicine: LLM users identified correct conditions in <34.5% of real cases despite near-perfect USMLE scores. Drive that into every procurement conversation.

The Big Story

Mayo Clinic's REDMOD detects pancreatic cancer up to 3 years before diagnosis · May 11, 2026 · Mayo Clinic News Network
→ The radiomics-based detection model analyzed ~2,000 CTs — scans from patients later diagnosed with PDAC, all originally read as normal — and flagged 73% of prediagnostic cancers at a median 16 months pre-diagnosis, roughly double unaided specialist sensitivity. The catch: 81/430 controls flagged, a 19% false-positive rate that rules out general-population screening. The clinical pathway is the prospective AI-PACED study targeting elevated-risk cohorts (new-onset diabetes after 50, familial risk), where pretest probability raises PPV enough to justify EUS workup. The likely regulatory route is De Novo — there's no predicate device for asymptomatic PDAC detection on incidental CTs.

Also This Week

CMS launches 10-year ACCESS model to pay AI for chronic-care outcomes · May 12, 2026 · TechCrunch
→ 150 orgs accepted into a model paying only on measured outcomes (BP, A1c, pain, depression); the PMPM is thin enough that the math requires automation — a de facto reimbursement code for chronic-care agents.

Utah Medical Board's Doctronic suspension request denied · May 11, 2026 · STAT
→ 47 states have 250+ clinical-AI bills in flight; STAT's framework — licensing AI as a clinician with malpractice carriage — is the most concrete policy text health system boards can react to today.

Google ships MedGemma 1.5 with 3D radiology and whole-slide pathology · May 5, 2026 · Google Research
→ Open-weight 4B model now handles CT/MRI volumes, WSI histopathology, anatomical localization, and multi-timepoint comparison; Taiwan's NHIA already runs it on 30,000+ pathology reports for lung-surgery planning.

Cleveland Clinic uses LLM screening to speed Phase 3 polycythemia vera enrollment · May 12, 2026 · Cleveland Clinic ConsultQD
→ Synapsis AI identified eligible candidates faster than conventional prescreening — the next operational case for ambient + agentic AI is trial recruitment, not just documentation.

Digital health Q1 funding hits $4B with average deal size at a 4-year high · April 30, 2026 · Fierce Healthcare
→ 12 megadeals of $100M+ drove a $36.7M average check — capital concentrating in fewer companies with deployed AI revenue, not seed models.

From the Lab

"AI Prognosis: When the path to good AI is littered with bad data" · STAT
→ The underappreciated procurement question: what training corpus does your scribe, your decision-support agent, and your chronic-care bot actually run on? STAT argues the next round of evaluation has to look upstream at data provenance, not just downstream at AUROCs — directly relevant to any health system standing up an AI governance committee.

Worth Reading

• AI chatbots are becoming mental health tools before they are ready — Leading models still miss indirect risk signals (food, withdrawal, hopelessness escalation), and 28% of under-30s have already used a chatbot for mental health.
• Health system AI adoption surges in 2026 with execs reporting increased ROI — 75% of US health systems now run one-plus AI platform and over half of those tracking ROI report 2x; the inflection from pilots to enterprise is real, with 2027 procurement implications.

The week early detection got cheap and prescribing got political — and CMS, of all agencies, wrote the first real AI reimbursement code.