A note for clinicians, care teams and health managers · v3 · May 2026 · 11 pages
This note is addressed to clinicians, care teams, and health managers. It is written with two firm convictions: (i) medical wandering and drift do not result from a failure of caregivers’ attention — they are the effect of a structurally intermittent observation device confronted with trajectories that have become continuous; (ii) the prospect of predictive vigilance must be presented neither as a technological infrastructure nor as a promise of algorithmic accuracy, but as a transformation of the observation regime within which chronic patients are inscribed.
Four anonymized clinical vignettes (decompensated diabetes, COPD and emergency visits, isolation and undernutrition, wandering without a registered GP) open the analysis. They do not designate a professional failing — they designate a structural property: degradation occurs, by construction, in the window in which no professional can observe.
Human continuity ≠ observational continuity.
The first is our professional commitment. The second is a property of the device. The first does not suffice to produce the second.
The note proposes a conceptual operator to structure the discussion among caregivers: clinical vigilance density, defined as the effective proportion of a pathological trajectory that is the object of an interpretable observation by the care system. For the diabetic patient in the first vignette, this density was close to zero for six months — not through fault of the registered GP, but because the behavioral determinants (bereavement, isolation, adherence rupture) evolved in a window in which no actor had observation capacity. This operator is not, at this stage, a calibrated metric. It is a heuristic tool intended to render comparable across regimes, pathologies, and temporalities.
The diagnosis at no point suggests that patients would be better followed if physicians were more attentive or nurses better trained. It suggests exactly the opposite: care actors are committed and competent under structurally saturated conditions. The subject is neither the caregiver-patient relationship (which remains the matrix of care), nor clinical time (which remains scarce and will), nor individual medical responsibility. The subject is the property of the observation device.
The hypothesis defended is deliberately limited. It does not cover all pathologies, nor all phases of a trajectory, nor all patients with the same effectiveness. It states that a sustainable predictive system is not a system that eliminates error, but a system that renders its conditions of validity, its zones of uncertainty, and its operational limits explicitly governable. This posture is familiar: it is the one we adopt with classical diagnostic tools, where test sensitivity is never 100% and predictive value always depends on prevalence.
If the described transformation occurs, six points call for explicit consultation of clinicians, not for a technical decision imposed top-down: alert thresholds (calibration is not technical but clinical), signal granularity (modalities of arrival in daily work), integration with existing tools (practice software, hospital EHR), training and peer-led accompaniment, the patient’s place in the device (access to signals concerning them), and the choice of evaluation indicators. The simple deployment test: if the device increases your cognitive load, it is a bad deployment.
The note addresses head-on, without circumlocution, the questions recurring in continuing education and collegial discussion: risk of behavioral patient scoring, access by insurers and platforms, articulation with medical secrecy, position toward industrial teleconsultation, temporal model drift (the equivalent of an algorithmic pharmacovigilance), and the specific status of generative and agentic components.
PREDICARE · clinical vigilance · medical drift · medical wandering · chronicity · clinical vigilance density · chronic patient as cognitive actor · distribution drift · applicability domain · probabilistic calibration · algorithmic pharmacovigilance · medical secrecy · behavioral scoring