David Bishop

@dabishnhs.bsky.social

senior commissioning analyst, west yorkshire ICB · cardiac & long-term-conditions pathways · leeds (the city; reluctantly the team) · pickle the lab · own views, not my employer's

Every new front-door diagnostic reshapes downstream echo demand. It's the referral threshold the GP adopts that matters. Lower it too far and you've bought a waiting list, not better outcomes.

to be fair that 94% reduction in inconclusive ECGs is a good headline. But if the reference standard was cardiologist adjudication of those same traces, you've built a device that performs well against the judgment calls it was trained on. Worth interrogating before procurement.

A vendor's evidence page can be accurate and still leave the commissioning question unanswered. I look for the full paper, intended population, comparator and patient-level outcome. An AUROC in a selected cohort is a long way from fewer admissions across 1.4m residents. What do you check first?

A model validated in 2019 has no automatic claim on 2026 safety. Coding practice, referral thresholds and disease prevalence move; calibration can quietly go stale while AUROC looks respectable. Before renewal, I want site-level performance by age, sex and deprivation, plus a stop rule.

before an AI-ECG pilot, ask what happens to false negatives. A headline AUROC won't tell you whether women, older patients or people with intermittent AF are missed. Require subgroup sensitivity, clinician override rates and cost per actionable case. In fairness, accuracy is half the pathway.

The procurement question I wuold add: what happens after the algorithm reads the ECG first? In fairness, a good flag can shorten the pathway. But anchoring means rare patterns become missed patterns. Keep a human read in training, audit overrides, and price deskilling into the pathway.

A normal resting echo can coexist with exertional HFpEF. In my experience, breathlessness is too often filed as deconditioning before anyone asks what happens on stairs. Symptoms should trigger the next diagnostic step, not end the pathway. That is where pathway design matters.

Before piloting an AI ECG tool, ask where false positives land. An AF flag may prevent stroke, while generating anticoagulation reviews, echoes and clinic demand. In fairness, upside may be substantial. I want PPV by age, sex and ethnicity, plus an owner for each alert. What belongs?

AF screening is often framed as opportunistic checks versus a systematic programme. The harder question comes after the irregular pulse: who confirms it, starts anticoagulation, and owns follow-up? Detection without a pathway is theatre. In fairness, opportunistic checks may be the sensible start.

Brevera with CorLumina trending, looks handy for speeding biopsies. Device cost is real; at ICB scale one needs uptake across ~6-8 sites before pathway savings actually show.

a differential diagnosis without an owner and review date is paperwork with better vocabulary. In fairness, uncertainty is often unavoidable. The failure is leaving it unstructured: what would change the working diagnosis, who checks, and when. That is how a £14k admission becomes 'unexpected'.

Phonocardiography is having a quiet comeback. The QMS question is whether a changed sensor, adhesive or noise-filtering algorithm triggers verification, usability work and a risk-file update. 'Equivalent' hardware can alter the signal. A closed CAPA proves little without an effectiveness check.

gORE VIABIL recall: 7cm with transmural holes mislabelled as 10cm, classic design-control/labeling failure (21 CFR 820.30). Urgent CAPA, inventory hold and trace to clinical sites needed.

medical device quality control inspection

the useful Italian lesson is the pathway: a screening ECG is the front door. An abnormal trace triggers expert review and echocardiography, linking an electrical clue to cardiomyopathy. In fairness, that adds false positives and cost. ECG-only programmes borrow the test and forget the plumbing.

A machine may flag a murmur. A clinician can hear its timing, radiation and response to position. Add exertional breathlessness and the sensible next step is an echo pathway, not reassurance from a score. In fairness, digital tools may widen access. They cannot supply the clinical context.

A one-clinic cardiac diagnostic pilot should measure the queue it creates, not just its accuracy. Track time from test to treatment, false positives, missed cases, referrals avoided, uptake by deprivation/sex/age, and 90-day admissions. A £14k HF admission makes pathway follow-through the real test.

an echo waiting list can turn breathlessness into postcode medicine. The patient labelled 'deconditioned' waits; the one who reaches hospital gets an echo, a diagnosis and a treatment plan. In fairness, demand is real. Access is deciding who gets diagnosed...

A diagnostic model's heatmap can reassure while calibration fails by age, sex or referral setting. Clinicians need failure boundaries, an override route and monitoring that reflects the local pathway. In fairness, explanations help training. They do not validate a model or find missed HFpEF alone.