Series: 'Medical desertification and care wandering' — Article 5, Part 4/5
How to scale industrially without turning predictive prevention into an error machine, a cognitive overload generator, and an iatrogenic factory? Industrialising in healthcare is not neutral: at small scale, human adjustment and local attention compensate for flaws; at large scale, anything not structurally designed becomes a systemic risk.
Every digital health catastrophe begins with the same illusion: what works in a pilot works at scale. Therac-25 shows that a rare bug multiplied by volume becomes a lethal statistical certainty. The NHS National Programme for IT (NPfIT, over £10 billion) shows that organisational complexity exceeding governance capacity collapses. HealthCare.gov shows that insufficient industrial capacity is not a technical detail but a major political and social risk.
The experience of large-scale EHR deployments (Epic Systems) reveals a specific failure: systems were designed to be deployed, not to be cognitively sustainable at tens of millions of patients. Technical industrialisation succeeded. Clinical industrialisation — maintaining usability and relevance at scale — failed.
The projection for 12 million PREDICARE patients is sobering: 3.6 million alerts per day for 60,000 physicians, or 60 alerts per physician daily — the critical threshold exceeded from year 1 without trajectory-based filtering via the digital twin. PREDICARE precisely transforms a threshold-based monitoring logic into a trajectory-based monitoring logic, where it is not the variation that triggers the alert but the probability of it leading to an avoidable adverse event.
The article establishes non-negotiable boundaries: industrialise collection, algorithmic processing, and traceability, but protect final clinical decisions, patient communication, and individual benefit/risk assessment. The clinician remains the decision author, never the validator of an automatic recommendation.
Industrial interoperability requires mandatory FHIR R4, documented APIs, third-party testing, and graduated sanctions. Change management is not a marginal accompaniment — it is an infrastructure to be funded at face value, on par with servers and algorithms. Article originally published on LinkedIn (in French).
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.