The exception does not break the rule: it reveals where the rule was drawn, whom it represents, and which reality it left outside. In medicine that aims to become more precise without ceasing to listen, a discordant result should not immediately become a label attached to a person. It may point to incomplete data, a poorly adjusted instrument, an underrepresented population, a communication barrier, or a change that deserves professional attention. Listening to the exception means investigating methodically, keeping alternative explanations open, and protecting the person who will bear the consequences of the decision.
This article provides general criteria about digital health, quality, equity, and care. It does not diagnose disease or prescribe individual treatment. A device reading, an algorithmic score, or any of the fictional cases described here cannot determine what is happening to a particular person. New, persistent, or worsening symptoms, concerning results, questions about medication, and clinical decisions should be discussed with qualified health professionals. If there is an emergency or an immediate risk, the appropriate local emergency service should be contacted.
The exception contains information about the system#
Protocols are necessary because they organise work, reduce avoidable variation, and support action under pressure. Yet every protocol rests on assumptions: a measurement method, a reference population, available resources, and an error margin judged acceptable. When one person does not fit, it is tempting to attribute the problem to that person. The safer approach first asks whether the rule anticipated reasonable diversity in bodies, ages, languages, abilities, social settings, and clinical trajectories.
Protocols are necessary because they organise work, reduce avoidable variation, and support action under pressure.
An exception does not prove by itself that the system is wrong. Nor should it be discarded as noise. It is an observation that needs context. The team can verify data provenance, instrument calibration, recording time and conditions, earlier values, missing information, and the person’s reported experience. This sequence separates signal from interpretation and interpretation from decision. Prudence does not mean paralysis; it means knowing which uncertainty remains before acting.
Clinical and technological precision depends on acknowledging limits. A model may perform well on average but fail more often in one subgroup; an interface may work in testing but become inaccessible with pain, fatigue, or low vision; a questionnaire may lose nuance in translation. A singular case can expose those mechanisms. When institutions record exceptions respectfully, look for patterns, and correct design, the burden of learning no longer falls solely on the affected person.
Listening does not compete with measurement#
Professional listening is not a sentimental concession against rigour. It supplies information that a sensor or administrative record does not contain: what change the person notices, when it occurs, how it affects daily life, which barriers exist, and which outcome matters. That account also has limitations and needs corroboration, but excluding it produces false exactness. Precise medicine combines reliable measurements, clinical knowledge, informed preferences, and follow-up instead of granting absolute authority to one source.
Language matters. Presenting a prediction as certainty may cause fear, resignation, or unnecessary decisions. Presenting it as worthless may waste a valid warning. Appropriate communication explains what the tool measures, the population in which it was studied, which data are absent, what errors are possible, and who will review the result. It also provides time for questions and an accessible correction route. A comprehensible explanation is part of safety, especially when a decision affects access to care.
Disagreement should also be documented. If a professional believes that an automated recommendation does not fit the observed situation, that professional should be able to record the concern, obtain review, and act without blind administrative pressure. If a person finds an error in the record or does not understand a score, a channel staffed by someone with authority is needed. Human oversight is real only when the reviewer has information, time, and power to change the expected course.
Three questions before treating a case as an anomaly#
The first question is whether the data describe what they are assumed to describe. Identity, unit, date, device, software version, measurement conditions, and missing fields can alter meaning. The second question is whether the rule was validated in a population sufficiently similar to the one in which it is used. The third is what harm each possible error would cause. Requesting one additional check does not have the same consequence as delaying necessary care or closing access to a service.
These questions should be answered before blame is assigned. In overstretched settings, an exception may look like an interruption, but hiding it shifts cost into later visits, complaints, and lost trust. A short, structured review can prevent harm: confirm the record, hear the discordance, compare other evidence, and name the decision owner. When doubt remains, the responsible route is professional assessment or an appropriate test, not a household interpretation of the algorithm.
Case 1: the watch that conflicts with experience#
In this fictional case, Elena receives several night-time alerts from a consumer watch. She feels well during the day, but the app displays an alarming colour without explaining signal quality. The system might reflect a relevant event, movement, poor contact, inappropriate configuration, or another cause. None of these possibilities can be resolved by looking at the screen alone. The useful exception is not an assumed disease; it is the discordance between the reading, its context, and the warning’s design.
Elena keeps the readings and notes when they occurred, without changing medication or starting treatment independently. She checks the device documentation and seeks guidance from a health professional, who can consider history, symptoms, and whether confirmation is needed. If severe symptoms or sudden deterioration appeared, she would seek emergency care. For the manufacturer, the case raises different tasks: display signal quality, avoid diagnostic language, support data export, and explain clearly when an alert warrants consultation.
Case 2: a score that penalises the barrier#
In the second fictional case, Samira uses a portal to prepare for an appointment. The questionnaire was written in a language she understands only partly and assumes stable internet access. Several answers remain blank, and the system interprets missingness as low engagement, pushing her request down the queue. The classification does not measure need alone: it incorporates a barrier created by the service itself. Treating Samira as an individual exception would hide that others facing similar conditions may leave before being recorded.
The centre reviews the case with language support, restores the request under clinical access criteria, and analyses abandonment by language, age, disability, and channel. It offers telephone and face-to-face routes without penalty, corrects the handling of missing values, and assigns an appeal owner. The action neither presumes a diagnosis nor guarantees clinical priority; it ensures that professionals assess need with sufficient information. Equity becomes a testable property of the pathway rather than a general statement.
Case 3: the average that conceals a different recovery#
In the third fictional case, Jordi completes a digital follow-up scale after a procedure. His average improves, but an everyday task he considers essential remains difficult. The screen summarises progress in green and offers no way to explain the difference. The aggregate may be correct and still insufficient. The exception shows that the outcome selected by the system does not fully match the outcome relevant to this person.
Jordi reports the difficulty at the scheduled review. The team compares the trajectory, checks for possible warning signs, and discusses suitable goals and options without allowing the app to decide. The service adds a short narrative field and reviews whether the scale performs similarly across groups. If severe pain, sudden worsening, or other concerning signs arose, the general advice would be to seek professional or urgent care according to severity, not wait for the dashboard colour to change.
Health data: control, quality, and purpose#
Health data are particularly sensitive because they may reveal conditions, habits, and relationships. Collecting more does not automatically mean caring better. Every project should define a purpose, minimise information, control access, set retention periods, and enable correction. It should also distinguish care use from secondary use for research, management, or innovation. Where consent is the applicable basis, it should be understandable rather than a generic gateway to every future reuse.
The European Commission’s European Health Data Space addresses access, control, secure exchange, and interoperability through gradual implementation. The Spanish Data Protection Agency gathers specific criteria concerning health information. These institutional frameworks show that an exception can originate in inaccurate or decontextualised data and that the ability to know, correct, and limit uses is part of care.
Quality requires preserving provenance, units, date, modifications, and missingness. A zero must not be confused with an unrecorded value; copying between systems should not erase context; a correction should propagate traceably. Teams need audits that ask who is excluded, not merely overall accuracy rates. If exceptions concentrate in one group, the issue is no longer singular and calls for changes in design, resources, and governance.
Governing artificial intelligence in health#
The World Health Organization’s guidance on ethics and governance of artificial intelligence for health places autonomy, well-being, transparency, accountability, inclusion, and sustainability at the centre. Its practical lesson is that technical performance is insufficient. Effects on professionals and communities must be assessed, identifiable owners must remain accountable, and complaint and remedy mechanisms should exist before deployment.
The WHO’s guidance on large multi-modal models in health extends attention to generative systems that can produce text, images, and other outputs. These tools may support tasks, but they can also invent plausible content, omit context, or reproduce bias. They should therefore not replace clinical assessment or be used to obtain personalised diagnoses or treatments without qualified professionals and appropriate safety processes.
A responsible organisation defines purpose, population, clinical owner, version, maintenance, withdrawal thresholds, and contingency plan. It tests the system in ordinary situations and out-of-average cases. It monitors false positives, false negatives, delays, abandonment, and differences across groups. It also asks whether claimed savings add invisible labour for patients or staff. When outcomes deteriorate or context changes, the organisation must be able to limit, suspend, or replace the tool.
Turning the exception into learning and care#
A review meeting can follow a simple sequence: describe what happened without blame, separate data from inference, listen to the affected person, reconstruct the pathway, identify the failed assumption, and agree on a correction with an owner and date. The organisation then checks whether the change works in other cases. Anonymising or limiting shared information prevents learning from becoming unnecessary exposure. Transparency should protect people while making mechanisms visible.
For patients and families, participation means being able to ask what a result means, what uncertainty it contains, who reviews it, and how information can be corrected. They should not bear the burden of proving that a system fails. For professionals, listening to exceptions requires institutional permission to depart from a recommendation when evidence and judgement support doing so. For suppliers, it entails designing logs, explanations, and incident channels that permit investigation without relying on informal memory.
The rule improves when it learns from what it failed to anticipate. Precise medicine does not promise to remove all uncertainty; it makes uncertainty visible, distributes it fairly, and prevents an automated output from being mistaken for a complete decision. An exception handled rigorously may expose faulty measurement, a repeated barrier, or an ignored human priority. Its value lies not in romanticising the singular, but in turning it into an opportunity to measure better, listen sooner, and refer to professional care whenever an individual’s actual health is at stake.
