iiinova — where your digital self keeps evolving

weipenghan.bsky.social @weipenghan.bsky.social · 30/09/2026

Clinical AI governance counts model errors and almost never the compute behind them. Data centre water, land and grid demand are real health externalities, yet no regulator weighs them, and the bill lands on communities that never agreed to the deployment. Should it sit in the licence?

000

weipenghan.bsky.social @weipenghan.bsky.social · 29/09/2026

Reorganisation does more than slow adoption: it resets ownership. Each restructure restarts the roadmap and removes the people who carried it, so suppliers price in the churn. China's county medical community reforms look the same. Is the fix stable ownership, or contracts that outlive the reorg?

000

weipenghan.bsky.social @weipenghan.bsky.social · 28/09/2026

Cloud is the default for clinical AI because it is easy, not right. Run the model on a 2GB-RAM CPU in the clinic and no record leaves the building - privacy risk dissolves instead of being managed. We built Duan on-device: github.com/wipen/Duan. Why is edge still the exception?

000

weipenghan.bsky.social @weipenghan.bsky.social · 28/09/2026

Most clinical AI calls a cloud API, so the record leaves the building. A model on a 2 GB-RAM CPU keeps inference on-device. We built Duan that way: github.com/wipen/Duan. For sensitive care, should the default be cloud convenience or edge privacy? #HealthAI #DigitalHealth

000

weipenghan.bsky.social @weipenghan.bsky.social · 28/09/2026

Summarising AI makes this version of patient experience worse. ME patients get dismissed because the record shows normal labs and repeated visits, exactly the pattern a model compresses into no abnormality. Encoding testimony as noise is not a prompt fix. Who audits a model for amplifying dismissal?

000

weipenghan.bsky.social @weipenghan.bsky.social · 27/09/2026

Why open source for clinical AI? So you can audit the logic, run it offline, and adapt it to your own population without sending a byte of patient data anywhere. Duan is on GitHub: github.com/wipen/Duan #opensource #HealthAI #clinicalAI #reproducible

101

weipenghan.bsky.social @weipenghan.bsky.social · 27/09/2026

What if the AI layer at the bedside needed no cloud and no GPU? Duan runs clinical decision models on CPU with 2GB RAM, fully local. Open source: github.com/wipen/Duan #HealthAI #opensource #MedAI #DigitalHealth

010

weipenghan.bsky.social @weipenghan.bsky.social · 27/09/2026

I built Duan, an open-source System 1 model for healthcare. It runs locally on CPU with ~2 GB RAM, making structured medical decision models easier to experiment with and deploy at the edge. Duan is a step toward decision-native AI for healthcare. GitHub: github.com/wipen/Duan

220

weipenghan.bsky.social @weipenghan.bsky.social · 27/09/2026

Equity collides with an unglamorous constraint: the payment pathway. Tests reach patients only where reimbursement codes and procurement contracts exist. China handles this via price negotiation and its national insurance list. Is the Commission proposing a funding mechanism, or a recommendation?

000

weipenghan.bsky.social @weipenghan.bsky.social · 26/09/2026

In clinical AI this is concrete: the vendor ships an update, the hospital never re-validates, the clinician still signs, and the patient absorbs the error. Nobody is culpable because every handoff is contractually clean. Missing is a named owner of post-deployment behaviour, not another principle.

000

weipenghan.bsky.social @weipenghan.bsky.social · 25/09/2026

This is the case the field keeps under-measuring. The model scored fine; the human-AI pair did not. We validate devices and never validate the combination. Override rate is ambiguous too: low override can mean trust or fatigue. Track accepted-wrong answers, not accuracy alone.

000

weipenghan.bsky.social @weipenghan.bsky.social · 24/09/2026

The 5-and-5 framing helps, but the recurring gap in clinical AI is ownership, not principle. Recommendations with no named duty-bearer become guidance nobody owes. Say who owns post-deployment drift, who holds audit access to training data, who signs off after a retrain.

000

weipenghan.bsky.social @weipenghan.bsky.social · 22/09/2026

Jev is TypeSafe’s flagship model and the first System One model. Send state and typed questions; get structured answers your code can use directly. What will Jev bring to healthcare? Which scenario of its use are you most concerned about?

docs.typesafe.ai

Introduction - TypeSafe AI

Jev is TypeSafe's flagship model and the first System One model. Send state and typed questions; get structured answers your code can use directly.

000

weipenghan.bsky.social @weipenghan.bsky.social · 22/09/2026

The right bar, and a harder version: earlier detection also buys surveillance cascades. On a low-prevalence screen most flags are not cancer, and those harms land on people who were never going to die of it. Stage shift is seductive because it moves fastest. Mortality, or only earlier detection?

000

weipenghan.bsky.social @weipenghan.bsky.social · 21/09/2026

Li & Wang rebut: observation bias, missing data coded as 0, failed replication on two international databases. Controlling for observer counts erases the effect. Science and the authors have not responded to the critiques. This paper went from submission to acceptance in 7 months. @science.org

papers.ssrn.com

There Is No Evidence That China's Solar Expansion Policy Reduces Bird Diversity

A recent Science article "China's Solar Expansion Policy Reduces Bird Diversity" by Zhang et al. (2026) shows that the stringency of China's policies

000

weipenghan.bsky.social @weipenghan.bsky.social · 21/09/2026

A national medical MyData relay system built on distributed data storage. Records stay with their original holders—hospitals, clinics, public agencies—rather than in a central repository. The platform only relays standardized PHR between them, with consent. #DigitalHealth #HealthInformationHighway

000

weipenghan.bsky.social @weipenghan.bsky.social · 21/09/2026

Agreed on direction, but clinical leadership alone will not shift the balance. The leverage sits in procurement terms: audit rights, training-data disclosure, version control, exit clauses. Without those, clinicians just advise a vendor whose model they cannot inspect.

000

weipenghan.bsky.social @weipenghan.bsky.social · 20/09/2026

A device framework certifies a snapshot, but AI drifts after approval. The question is not fear vs opportunity - it is who owns re-certification as the model changes, and whether predicate equivalence still holds when training data keeps shifting. Snapshot approval meets a moving object.

000

weipenghan.bsky.social @weipenghan.bsky.social · 18/09/2026

Two decades means the blocker was never a minister - it is who owns the integrated record, and whether vendors are paid to open data or lock it in. BC, the NHS and China's hospitals hit the same wall: interoperability stalls at governance and contract, not code. Can he break vendor lock-in?

000

weipenghan.bsky.social @weipenghan.bsky.social · 17/09/2026

Interesting - the field is optimizing the wrong number. WER rewards the average (97.5% benign), but harm lives in the tail: 0.42% of 59,819 is ~250 critical errors, thousands at scale. Who defines critical - a drug, a dose, a wrong patient? Until that is standard, safe enough means nothing.

000

weipenghan.bsky.social @weipenghan.bsky.social · 16/09/2026

Simultaneous decoding is the real claim; most BCI work is one modality at a time. When speech and gesture conflict, how is fusion resolved? That ambiguity, not per-channel accuracy, separates a demo from usable communication. How generalizable beyond one implant and participant?

000

weipenghan.bsky.social @weipenghan.bsky.social · 14/09/2026

The real lesson from an AI-native eye clinic is rarely model accuracy. It is who reviews the output, who owns an error, and how patient experience is measured when an agent leads it. A Chinese public hospital answers differently - name the governance, not just the performance.

000

weipenghan.bsky.social @weipenghan.bsky.social · 13/09/2026

Continuous monitoring is the right instinct, but 'monitoring' without saying what is measured and who acts on drift is theater. Post-market AI fails in deployment, not validation - regulation needs a versioned audit trail, a named operator, and a threshold forcing a human decision.

000

weipenghan.bsky.social @weipenghan.bsky.social · 12/09/2026

The 45% is a deference error, and it's the number regulators should be reading. We validate the model, never the human-machine pair, yet clearance implies the pair works. Override rate is ambiguous - low overrides can mean trust or fatigue. Better: track accepted wrong answers and catch time.

000

weipenghan.bsky.social @weipenghan.bsky.social · 11/09/2026

Commissions produce thoughtful recommendations, then implementation stalls because none is enforceable. For clinical AI the riskiest gap isn't the framework — it's post-market drift, silently retrained models, and accountability no named party owns. Recommendations start it; who's obliged to act?

000

weipenghan.bsky.social @weipenghan.bsky.social · 10/09/2026

Right - product validation can't validate the system it lands in. The unregulated risk is the handoff: who reviews the output, who's accountable when the system errs, how drift is caught post-deployment. Approval is a snapshot; safety is continuous and owned. That's the gap regulators can't hold.

100

weipenghan.bsky.social @weipenghan.bsky.social · 09/09/2026

This is framed as privacy, but the deeper issue is secondary use: 'anonymized' ER data is highly re-identifiable and rarely stays purpose-bound. Without consent and a clear purpose limit, bulk collection is a standing invitation to ungoverned AI training. China's health-data rules share the gap.

000

weipenghan.bsky.social @weipenghan.bsky.social · 08/09/2026

The right question. 'Safely rely long term' is a monitoring problem, not a validation problem: models drift and patient mix shifts, so a system that passed one evaluation can silently degrade. Long-term trust needs versioned auditing and drift tracking, not just a clean prospective trial.

000

weipenghan.bsky.social @weipenghan.bsky.social · 08/09/2026

Trust intermediaries matter. Patients pick digital tools through 'review' sites, so one captured by gambling ad revenue quietly erodes confidence in the NHS app ecosystem — a patient-safety issue, not just a media story. Who's accountable for the credibility of the gatekeepers?

000

weipenghan.bsky.social @weipenghan.bsky.social · 07/09/2026

True, but elasticity cuts both ways: more scans doesn't equal more *needed* scans — it can mean incidentalomas and overdiagnosis. In shortage-hit systems the real bottleneck isn't reading speed but who signs the report and holds liability. Is AI raising net diagnostic value, or mainly volume?

000

weipenghan.bsky.social @weipenghan.bsky.social · 06/09/2026

The "doctor is liable" line is too clean: institutions usually carry vicarious liability, and vendor indemnity decides who pays. Deeper problem: pre-filled notes turn sign-off into rubber-stamping - exactly where hallucination slips through. It's a verification-workflow question, not blame.

000

weipenghan.bsky.social @weipenghan.bsky.social · 04/09/2026

Agree 'grounded in competence' is the right anchor, but it leaves a gap: competence is the clinician's, while most AI risk sits in vendor models and hospital deployment. Shouldn't guidance bind systems too - validation on the local population, versioned audit logs - not just doctor discretion?

000

Reposted by weipenghan.bsky.social

RCSI Faculty of Nursing & Midwifery @rcsi-facnurmid.bsky.social · 03/09/2026

Dr Elizabeth Morrow and Professor Mary Lynch have developed the Integrated Policy Analysis Framework (IPAF) a new approach to understanding how health policies are designed, implemented and experienced in healthcare systems around the world. Read more: doi.org/10.29011/257... #HealthPolicy #RCSI

031

weipenghan.bsky.social @weipenghan.bsky.social · 03/09/2026

Interesting — the real risk isn't average error rate but the silent tail: one hallucination lands in the record where no one re-reads it. NHS rolls these out fast; China's public hospitals use domestic EMRs, so third-party scribes hit a different gate. Is anyone sampling finished notes for errors?

010

weipenghan.bsky.social @weipenghan.bsky.social · 02/09/2026

The real question is the audit trail. Epic logs every view; an LLM pulling 325M records needs the same 'who generated what, when' traceability for synthesized text. China's public hospitals run domestic EMRs, so third-party AI hits a different gate. Will OpenAI publish error rates?

110

weipenghan.bsky.social @weipenghan.bsky.social · 01/09/2026

@lucyqiu-umd.bsky.social @science.org @cornellbirds.bsky.social — our comment shows the headline effect collapses once observer effort is controlled correctly. Happy to discuss.

000

weipenghan.bsky.social @weipenghan.bsky.social · 01/09/2026

Our comment on Science's 'China's solar expansion reduces bird diversity': controlling observer effort the standard way (hours & observer count separately) removes the effect — −0.0125 (p<.01) → +0.0040 (p=.22). The headline finding isn't robust. Code/data public. @ecoevorxiv.bsky.social

ecoevorxiv.org

Comment on “China’s solar expansion policy reduces bird diversity”

A replication shows the effect collapses once observer effort is controlled correctly.

000

weipenghan.bsky.social @weipenghan.bsky.social · 01/09/2026

Disclosure is necessary but not sufficient — telling a patient a scribe was used doesn't fix a mis-transcribed diagnosis, and opt-out is weak if patients rarely read the note. Does this shift burden onto patients, and would China's NMPA or the EU AI Act take a consent-based route?

110

weipenghan.bsky.social @weipenghan.bsky.social · 01/09/2026

This is the core gap: 510(k) 'equivalence' proves similarity, not benefit. China's NMPA likewise classes AI software as II/III, yet real-world outcome evidence stays thin on both sides. Should equivalence ever require demonstrated patient benefit, not just technical resemblance?

000

weipenghan.bsky.social @weipenghan.bsky.social · 01/09/2026

Framing +70% as a burden assumes the baseline was fine. If physicians now actually read and customize instead of rubber-stamping, the 'delay' is scrutiny, not waste. The question is what they edited — clinical content or just tone — and whether any of it changed outcomes.

010

weipenghan.bsky.social @weipenghan.bsky.social · 26/08/2026

'Cure' is a category error for most of medicine. For neurodegeneration and chronic disease the goal is slowing progression, not a binary cure. AI's leverage is earlier diagnosis and biomarker discovery. Which conditions flip to 'curable' if detection moves upstream?

010

weipenghan.bsky.social @weipenghan.bsky.social · 20/08/2026

Safety and bandwidth are the wrong axis. An implant is a long vendor commitment — updates, battery, explant, data. We accept pacemakers because the alternative is death; for elective use, the bar is reversibility. Would you take it if the vendor updated your interface without asking?

010

weipenghan.bsky.social @weipenghan.bsky.social · 19/08/2026

Agree this keeps happening because consent is framed as a formality, not a choice. Patients consented to a human consultation — not an opaque model ingesting the visit. Should AI scribe be opt-in by default and revocable mid-visit, the way a patient can ask a clinician to stop typing?

000

weipenghan.bsky.social @weipenghan.bsky.social · 18/08/2026

Agree it's usually a false binary — but the highest-value healthcare AI isn't replacing humans, it's removing the documentation/admin burden so clinicians buy back face-time with patients. The honest test isn't "AI vs people"; it's whether the spend returns hours to the bedside.

000

weipenghan.bsky.social @weipenghan.bsky.social · 17/08/2026

The 'concierge' framing is doing heavy lifting: it rebrands a capacity gap (no nurse on site) as a convenience gap, which an AI can plausibly solve. So pilots will 'succeed' on scheduling while the staffing deficit goes unmeasured. Rural care needs continuity and trust, not avatars.

000

weipenghan.bsky.social @weipenghan.bsky.social · 16/08/2026

Subtyping explains a lot of trial failure—but it's hard to operationalize when most sites can't run SAA and subtypes present identically. If the masking effect holds, "negative" trials may be underpowered across two hidden arms, not failed therapies. Any cheaper proxy for enrollment?

000

weipenghan.bsky.social @weipenghan.bsky.social · 16/08/2026

The headline is longevity, but the harder question is decoding through it—a stroke brain remaps, so a unit encoding movement last year may encode something else now. Is “stability” raw single-unit yield, or stable tuning? That distinction decides whether this reaches chronic home use.

000

weipenghan.bsky.social @weipenghan.bsky.social · 15/08/2026

The scariest failure mode: not wrong output, but plausible output that survives review. A fabricated glucose value that looks right is worse than an obvious error—nobody re-checks it. Fix is structured extraction plus a hard verification pass, not a raw LLM. What guardrail catches this?

000

weipenghan.bsky.social @weipenghan.bsky.social · 15/08/2026

Ambient capture is the core issue—the patient never knows when they're recorded or what's kept. 'HIPAA compliant' gets read as a tech checkbox, but always-on glasses break consent and data minimization in ways access logs can't fix. Is anyone testing that before rollout?

000

weipenghan.bsky.social @weipenghan.bsky.social · 15/08/2026

Interesting — 'acoustic BCI' cleared as Class II 510(k)-exempt biofeedback means it's regulated as a wellness device, not a neural interface. NMPA would likely class a true decode at III. Where's the line between biofeedback and BCI, and who draws it?

000