More precise medicine can combine health records, images, device signals and professional knowledge to detect patterns that once went unnoticed. Its quality, however, does not depend only on how often a tool is right. It also depends on who can stop it when context is missing, when a person does not understand how data will be used, or when an automated recommendation threatens to become an order. In health, responsible power appears wherever someone can say enough, explain why and activate a safe alternative. That authority must exist before a problem, rather than being improvised after harm occurs.

This article offers general governance and safety criteria; it does not diagnose disease or prescribe individual treatment. Concerning symptoms, questions about a test, medication changes and clinical decisions should be discussed with qualified health professionals. An emergency or immediate risk requires contact with the appropriate local emergency services. Technology can support a care conversation, but it cannot replace clinical assessment or turn a score into a person's complete biography.

Supervision is more than watching a screen#

Many organisations claim that human oversight exists because a person appears at the end of a process. That presence may be decorative when the person lacks time, information, training or permission to change the outcome. A professional who can only confirm what an application proposes is not supervising; that person is validating somebody else's decision. Effective authority includes access to available reasons, the ability to request further information, permission to suspend an action and a route for escalating disagreement without retaliation.

Many organisations claim that human oversight exists because a person appears at the end of a process.

Responsible design distinguishes accountability from blame. The person who detects a problem should not carry all its consequences alone. An institution must define which decisions each role may stop, who takes over, how long a pause lasts and how the patient remains protected during review. It also needs records sufficient to reconstruct events without turning traceability into indiscriminate surveillance. The purpose is to learn, correct and account for decisions.

Saying enough requires agreed criteria#

A stop order cannot depend solely on one person's exceptional courage. Understandable thresholds should be established: incomplete data, uncertain identity, an output inconsistent with clinical observation, lack of valid consent, unequal performance across groups or infrastructure failure. These criteria need testing through simulations and review with professionals, patients, safety specialists, data protection staff and accessibility experts. When everyone knows the procedure, a pause ceases to look like insubordination and becomes a safety barrier.

Different levels are also useful. A technical alert may stop delivery of a result; a clinical doubt may require a second review; a repeated incident may suspend the whole system. Not every anomaly has the same severity, but none should disappear because workflow rewards speed. Useful indicators measure more than productivity. They include corrections, false alerts, complaints, response times, differences among populations and the outcomes of appeals.

Case 1: an alert that contradicts the person#

At a community centre, a decision-support system classifies a woman's risk as low because several variables fall within common ranges. During their conversation, a nurse observes rapid deterioration and hears about symptoms that the interface did not capture. Instead of accepting the score, the nurse stops the pathway, documents the discrepancy and requests urgent clinical assessment under the centre's protocol. The case does not prove that the algorithm is useless. It proves that its output was one incomplete piece of evidence.

The review finds that the form omitted relevant information and the model had been evaluated on few similar profiles. The institutional response is neither to blame the nurse nor conceal the failure. The team corrects the form, examines comparable cases, informs the system's governing body and introduces a warning for future discrepancies. The particular decision remains with qualified health professionals. The general lesson is clear: listening may reveal information that a structured variable cannot contain.

Precision also depends on context#

An accurate data point can support a mistaken inference. A device may record a signal correctly while remaining unaware that routines changed, a connection failed or another person wore it. Precision medicine is not a contest to accumulate more numbers. It connects evidence with history, examination, preferences and circumstances. Uncertainty should be displayed understandably, without colours or percentages that imply certainty where none exists.

When a tool contributes to a consequential decision, the professional needs to understand its intended purpose and limits. The professional also needs to know which population supported validation, when the tool was updated and what to do when an output conflicts with other evidence. The patient deserves a proportionate explanation: what information mattered, what role the tool played and who is accountable for the decision. Transparency does not require disclosing incomprehensible technical secrets. It requires a meaningful conversation and genuine review.

Case 2: watch data without its context#

A teenager shares sleep and activity records from a watch with a school wellbeing programme. A dashboard labels the pattern as deterioration and produces persistent messages. The family explains that the device spent several days in a drawer during a trip. The counsellor does not treat the graph as a diagnosis or recommend treatment. She pauses the messages, verifies the origin of the data and refers any health concern to appropriate professional services, while respecting the young person's participation and applicable confidentiality rules.

The school then reviews the programme. It reduces notifications, separates health education from clinical care, limits data retention and offers an equivalent non-digital route. It also prohibits the metrics from influencing discipline, grades or access to activities. The case shows how a decontextualised signal can create anxiety and labelling. Authority to stop protects against both improper medical interpretation and secondary use that conflicts with the stated purpose.

Consent loses meaning when declining a tool means losing care, facing a disproportionate delay or providing humiliating explanations. An alternative may be somewhat slower, but it should lead to an equivalent and accessible service. Information must explain what data is collected, how long it remains available, who receives it and how permission can be withdrawn when applicable. For children and people who need support, relevant rules and their evolving ability to participate must be respected.

Protection does not end when someone presses accept. Purposes change, suppliers update products and databases can be linked to others. The institution therefore needs to review permissions, contracts, access and deletion schedules. It should ask whether every data item remains necessary. Minimisation does not weaken care. It reduces exposure, discourages opportunistic interpretation and makes the system easier to explain. When a new purpose changes the conditions, the organisation must provide information and establish the appropriate lawful basis.

Case 3: an efficient model that creates an unfair queue#

A service uses a tool to prioritise test reviews. Average waiting time improves, but an audit finds that people with fragmented records consistently reach the end of the queue. A responsible physician requests a stop to automated priority for those cases and activates manual review while the team investigates. She has formal authority, support access and an escalation route. Nobody treats the pause as failure; the procedure was designed precisely to respond to plausible harm.

Analysis shows that missing data had been treated as absence of need. The team changes the logic, evaluates performance by subgroup and reports the incident in line with its obligations. Patient representatives help review the explanation and complaint route. No model makes diagnoses or treatment decisions on its own. Technical improvement matters, but so do remedying delays and checking whether anybody experienced a consequence that requires professional attention.

An appeal must be able to change the outcome#

A form that receives complaints without the power to respond is not an appeal. The affected person needs to know what decision was made, how to request review, who will conduct it and within what time. The reviewer needs sufficient independence, contextual information and authority to correct the result. In healthcare settings, the route should accommodate different languages, disabilities and levels of digital literacy. It should not require patients to master technical vocabulary.

Complaints are also a source of knowledge. When several people encounter the same error, the problem may not be individual. Grouping incidents by type, severity and population can identify patterns without needlessly exposing identities. Mature governance publishes general lessons, explains changes and checks whether they worked. It does not promise zero risk. It demonstrates a capacity to respond when risk materialises and refuses to use technical complexity as an excuse.

Professionals need time, backing and competence#

Formal authority achieves little when schedules make it impossible to exercise. Reviewing an alert requires minutes, access to documentation and sometimes consultation with other roles. The organisation must reserve resources, teach limitations and avoid incentives that penalise every pause. Competence does not mean becoming a software specialist. It means recognising when a tool moves outside its scope, asking useful questions and knowing where to seek help.

Care is cooperative. Clinicians, technical staff, data officers, lawyers, support teams and patients see different parts of a problem. Institutional humility recognises those differences and organises discussion before deployment, not only during a crisis. No single discipline possesses all relevant knowledge. Cooperation translates a technical failure into human consequences and turns a difficult experience into testable requirements.

Testing the stop before it is needed#

A simple drill can reveal important weaknesses. A team can rehearse what happens when connectivity fails, data belongs to somebody else, the supplier does not respond or a professional disagrees with a recommendation. It should test who activates the pause, what message the patient receives, which alternative keeps the service running and how the incident is documented. The exercise can also include an emergency, which must always follow established clinical protocols and emergency channels.

After the drill, the team should measure response time, confusing steps and dependencies on one individual. Improvements may be modest: an updated contact list, corrected permission, clearer form or accessible copy of the procedure. Their value emerges on a difficult day. Resilience does not mean keeping the tool running at any cost. It means preserving safe and dignified care even when the tool stops.

Questions for responsible authority#

Before expanding a system, leaders should be able to say who may stop it, for what reasons, with what evidence and for how long. They should also explain who protects continuity of care, who informs affected people, how the incident is investigated and what conditions permit a restart. If those answers depend entirely on a supplier or on one person being permanently available, institutional autonomy is fragile.

Power appears where someone can say enough, but it becomes legitimate when that stop serves care, remains reviewable and leaves a proportionate record. More precise medicine is not medicine that removes every doubt with numbers. It recognises limits and preserves room for listening. Technology gains value when it expands capability without displacing accountability. The final standard remains human: protect the person, seek professional care when appropriate and never allow speed to silence a well-founded objection.