Article — Position paper · ○ Open access

From planned to predictive maintenance — Part 1: a possible architecture

Series: 'Medical desertification and care wandering' — Article 3 (Part 1)/5

Jérôme Vetillard · · LinkedIn Pulse · 12 pages · 1 min read
🇫🇷 Lire en français ↓ Download PDF

Series “Medical abandonment and clinical drift” — Article 3, Part 1

Modern medicine rests on a reassuring but false idea: if something goes wrong, the patient will seek care. This is false for chronic diseases. The current system operates on a planned-maintenance logic that industry abandoned 20 years ago because it was structurally unable to manage the dynamic complexity of critical systems.

The calendar paradigm and its 8,756 invisible hours

Over 99 % of a chronic patient’s time occurs outside any medical presence. Eight annual 30-minute consultations yield 4 hours of contact out of 8,760 hours. It is within those invisible 8,756 hours that progressive glycaemic drift, silent fluid retention, and worsening hypertension develop. The “come back in 3 months” model is a calendar-based maintenance approach applied to a system whose degradation is continuous and non-linear.

What industry solved 20 years ago

Aviation, energy and rail solved this problem by combining continuous sensors, dynamic digital models, and predictive algorithms, reducing unplanned failures by 30 to 50 %. The question stands: why does a jet engine benefit from continuous predictive monitoring, but not a high-risk polypathological patient?

A five-layer architecture

The proposed system relies on a layered functional architecture. The IoMT acquisition layer collects weak signals via simple connected medical sensors. The companion app acts as a silent and reassuring sentinel — it humanises technology to drive ecosystem adoption. The ML models and digital twin layer uses multiple specialised AIs (Agentic AI), each responsible for a specific task. The decision and alert system generates qualified alerts under human clinical governance. The professional dashboard provides a 360° patient view with pre-synthesised data and prioritised alerts.

The LLM does not decide, it explains

A central architectural choice is deliberate: the LLM does not intervene in the predictive core. Specialised models understand clinical state; the LLM makes the system comprehensible and usable by humans. This separation is not a technical compromise — it is a clinical safety position. The generative model serves explainability, not decision-making. Article originally published on LinkedIn (in French).

Read the document