Intermittent vigilance against longitudinal chronicity, clinical vigilance density as structuring operator, and the conditions of a French institutionalization before default private standardization
The public question is no longer whether a predictive vigilance infrastructure will emerge in the French health system. It is under which doctrine, which institutional architecture, and which contestability regime it will emerge. The absence of public experimentation does not suspend the transformation. It lets the lock-in conditions form under existing law, by accumulation of sensors, cloud infrastructures, models, platforms, and economic vehicles carried by private actors. This article demonstrates, in five hierarchized steps, why the status quo has become an implicit decision with documented cost, and qualifies PREDICARE as an institutional trial protocol allowing France to explicitly test the sustainability conditions before asymmetric industrial lock-in.
This article does not address AI in healthcare. It does not address platforms. It does not address medical automation. It addresses the vigilance regime within which the French health system observes the trajectories of its chronic patients. Four structural determinants converge: declining medical demography standardized by care needs (from 331 to 312 physicians per 100,000 inhabitants between 2012 and 2021), activity-based pricing that mechanically rewards acute hospitalization and penalizes prevention, professional cognitive load exceeding longitudinal information-processing capacity, organizational fragmentation depriving any single actor of an integrated trajectory view. To these are added 13.8 million patients in long-term illness status as of January 1, 2024, and 6.7 million French citizens without a declared primary care physician. A crucial analytical distinction is required. Human continuity, the relational quality exercised by caregivers over time, is not observational continuity, an informational property of the surveillance device. The consultation as a unit of observation remains discrete regardless of relational quality. The relative collapse of clinical vigilance density on longitudinal chronic trajectories is not an individual failure of caregivers. It is the structurally expected effect of a regime we have not refounded.
The direct cost is measurable. The Weissman method places the number of annually avoidable hospital stays in France at 742,474, for an associated envelope between 2 and 3 billion euros for strictly avoidable stays alone. The PREDICARE Medico-Economic Note retains a widened range of 510,000 to 742,000 potentially avoidable hospitalizations and approximately 4 million potentially avoidable emergency-room visits. But direct cost does not exhaust the stake. The absence of a framed public experimentation lets the conditions of a private lock-in form under existing law, through the convergence of three signal families, distinct but cumulative. First family, consumer-grade sensors (Apple Watch, Fitbit, continuous glucose monitors, oximeters) that install themselves into lives without any specific French medical authorization procedure. Second family, the cloud platforms and infrastructures whose de facto interoperability standards (HL7 FHIR, SNOMED CT, LOINC) have international governance forms whose trajectory is not arbitrated in French arenas. Third family, insurers building their own actuarial evaluation devices for AI health risk and instrumented adherence, backed by corporate-wellness platforms with premium modulation. No isolated signal is demonstrative; their juxtaposition is. Non-decision is not prudence. It is a default decision whose historical trajectory is known.
Observation is not supervision. Supervision is not intervention. Intervention is not clinical improvement. The complete chain alone is pertinent, and its rupture at any single point reduces the whole to a technological figure without clinical reach. The single objective is to increase clinical vigilance density without increasing net cognitive load. The minimal architecture reduces to four cumulative blocks: observe and qualify (entry into the partial-continuous regime, with enforceable opposable qualification of the model in the sense of the AI-governance pentalogy); prioritize and intervene (triage along a temporal triptych of actionability, without which the model predicts into the void); trace and contest (a quaternary grammar of true alert, false alert, true silence, false silence, algorithmic silence becoming opposable in a continuous regime); evaluate-redistribute-govern-audit (four cumulative functions to be rigorously distinguished, the order govern then audit being an institutionalization requirement). Four transverse conditions deploy: the threshold as the central institutional act (six distinct thresholds to be posed as political acts), the six structurally incompatible temporalities of the clinical system, the non-negotiable institutional separation of critical functions, operational contestability as constitutive. The complete matrix reads as ten cumulative links: observe, qualify, prioritize, intervene, trace, contest, evaluate, redistribute, govern, audit. The failure of a single link breaks the chain. The confusion of two compromises the institutional legitimacy of the whole.
PREDICARE is not evidence. PREDICARE is a machine for producing admissible evidence. This precision fixes the precise epistemic status of the device. PREDICARE does not claim to demonstrate ex ante the conclusions of the experimentation; it claims to produce, under methodologically defensible and institutionally governed conditions, the evidentiary material from which French institutionalization may be deliberated. Clinical perimeter on fragile chronic patients with instrumentable decompensation signatures (heart failure, COPD, imbalanced diabetes). Design as prospective cohort with comparison arm or documented quasi-experimental setting, observation duration of 24 to 36 months, longitudinal follow-up by IoMT instrumentation. Five authentic limits structure the experimentation: clinical effect not yet generalizable at national scale, economic model to be built within the experimentation rather than as its precondition, city-hospital interoperability as the cardinal operational risk, net cognitive load to be contained, territorial equity to be treated institutionally. Nine operational indicators set the trial: signal-decision delay, decision-intervention delay, rate of alerts handled within the useful window, rate of events without prior alert, documented review of algorithmic silences, net clinical load, cost per avoided decompensation and per stabilized trajectory, impact on non-equipped patients, effectively exercised patient contestation capacity. Silence becomes both measurable as event and qualifiable as judgment, which renders it institutionally opposable.
Clinical vigilance density on longitudinal chronic trajectories will not be substantially increased by adding consultations, nor by injecting resources into an unchanged frame. It will be increased only by a change of vigilance regime, and only on condition that this change does not increase the net cognitive load of the system. The required political gesture is therefore not to authorize or prohibit a predictive vigilance infrastructure. It is to pose, upstream of any deployment, the explicit governance frame within which that infrastructure will be bounded: thresholds, uses, prioritizations, state doctrine. Without this frame, the technical device will deploy regardless. It will simply deploy without mandate or transparency, and its historical trajectory will be that of actuarial surveillance rather than that of a sustainable medical vigilance. The public choice is no longer between prudence and innovation. It is between two forms of institutionalization. The first is explicit, experimental, contestable, bounded by a public doctrine. The second is implicit, industrial, progressive, carried by the actors who already possess the sensors, the data, the models, the infrastructures, and the economic incentives. PREDICARE does not claim to close this choice. It makes it, at last, instructable. Pharmacovigilance took three decades to institute itself as a continuous regime after thalidomide. The 2026-2030 window is not a horizon. It is an institutional calendar.
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.