Inequality often hides inside the average. A health system can report an overall improvement in response time, accuracy, or satisfaction while leaving behind people who live far away, speak another language, have a disability, work in insecure conditions, or barely appear in the data. Precision medicine promises to recognise relevant differences, but it deserves that name only when scientific knowledge, responsible technology, and clinical listening operate together. An average describes a group; by itself, it does not say who received help, who waited too long, or which barrier turned an available service into an inaccessible one.
The average is useful for orientation, yet it becomes dangerous when it stands in for reality. Two hospitals can have the same mean waiting time even though one responds fairly consistently and the other alternates very fast service with extreme delay. A tool can perform well overall and fail more often for an underrepresented subgroup. A remote programme can reduce travel for many people while creating a new burden for those without connectivity, privacy at home, or digital support. Inequality does not always appear as explicit exclusion; it often looks like variation, missing data, or a statistical exception.
This perspective does not turn every difference into an injustice or permit an automatic claim about cause. Age, clinical situation, social context, accessibility, record quality, and organisational decisions may interact. Responsible analysis develops hypotheses, seeks disaggregated evidence, and listens to professionals and patients before reaching conclusions. It also provides no individual diagnosis or treatment: any specific concern about symptoms, tests, or medication requires assessment by a qualified healthcare professional, and signs of an emergency require contact with the appropriate emergency services.
An average can inform and erase#
A mean compresses different values into one figure. That operation makes periods and institutions easier to compare, but it can conceal the full distribution. The median, extremes, spread, and proportion of people crossing a meaningful threshold all matter. Relevant characteristics should also be examined where justified, with privacy protections and without creating groups so small that individuals become identifiable. The purpose is not to collect every sensitive detail. It is to test whether benefit and harm are distributed with reasonable fairness.
A mean compresses different values into one figure. That operation makes periods and institutions easier to compare, but it can conceal the full distribution.
The same standard applies to missing data. If a characteristic is not recorded because asking is difficult, the gap may be concentrated among the very people facing the greatest barriers. If someone leaves a digital pathway before completion, a system may count only those who reach the end and confuse success among participants with universal access. Absence needs a label and an explanation. A responsible dashboard shows how many cases were excluded from a calculation, why they were excluded, and what uncertainty that omission creates.
Comparison also requires a suitable reference. An aggregate outcome may improve because the composition of the population has changed rather than because care is better. Comparable cohorts, changes over time, local conditions, and outcomes that matter to people therefore need attention. Statistical precision does not replace clinical judgement or lived experience. The former signals patterns; the latter helps an institution understand what those patterns mean and which organisational response might correct them without causing a new harm.
Precision means recognising relevant differences#
More precise medicine does not treat every person as an isolated data point or assume that more data always creates better care. It recognises that certain biological, social, and environmental differences may alter risk, access, or response, and it requires evidence that a distinction is relevant. It also resists turning historical correlations into individual destinies. A postcode may reflect accumulated disadvantage, but careless use could perpetuate lower investment where support is most needed. A tool should assist a reasoned decision, not replace it with an opaque label.
Quality begins before a model is built. Participant selection, sensor calibration, the method used to record symptoms, and the availability of translation determine who is represented. External validation, monitoring in real use, and a safe route for reporting incidents follow. Performance should be reviewed across groups relevant to the intended purpose, with adequate sample sizes and explicit limitations. When evidence cannot support a reliable conclusion, the system should express uncertainty and favour human review.
The World Health Organization places autonomy, well-being, transparency, accountability, inclusiveness, and sustainability among the principles for artificial intelligence in health. That framework shifts attention from whether a technology looks impressive to whether it works safely and fairly in real conditions. FDA guidance on age-, race-, and ethnicity-specific data in medical-device studies reinforces the value of representation and transparent reporting of performance. None of these principles authorises individual clinical advice outside a competent care relationship.
Case 1: an alert that performed well except where it mattered most#
A fictional hospital introduced an alert to detect deterioration and prioritise review. The initial report showed high sensitivity overall. Months later, the safety team examined false negatives and found a persistent difference among patients whose data came from wards with older monitoring equipment. The tool had not discovered a clinical characteristic of that group; it was receiving lower-quality signals. The average had blended two infrastructures and turned a material inequality into what looked like medical variation.
The prudent response was neither to assign blame nor to remove all automated support immediately. The hospital reviewed sensors, recording protocols, staffing by shift, and performance by location. It added signal-quality warnings and required clinical review when essential data were missing. Leadership set a harm threshold that would pause the alert. The case shows that equity also depends on maintenance, procurement, and organisation. A system designed to support staff cannot be evaluated separately from the environment supplying its calculations.
Case 2: the remote appointment that improved average attendance#
A fictional primary-care centre offered video follow-up and reduced the mean number of missed appointments. The result looked excellent until analysts separated the data by channel and reason for dropout. Older people living alone, patients with hearing difficulties, and workers without a private space completed fewer sessions. People who left during setup did not even appear as missed appointments because registration began only after they entered the virtual room. The metric rewarded a doorway that did not count those unable to cross it.
The centre retained the remote channel but added telephone access, in-person support, captions, a preliminary technical check, and an easy choice of format. It recorded failed attempts without automatically interpreting them as lack of interest and consulted patient groups and reception staff. Improvement did not require forcing everyone through the same route; it required equivalent outcomes through different routes. When a technical obstacle prevents access to necessary care, a health service should provide another channel and direct the person to qualified staff.
Case 3: a precise study with an excessively narrow sample#
A fictional consortium developed a wearable device and reported very high overall performance. Its validation sample, however, consisted mainly of younger volunteers with stable connectivity and experience using apps. During a pilot deployment, invalid readings were more common among people with limited mobility and users unable to hold the device in the indicated position. The problem did not prove that the device was useless; it showed that the broad claim exceeded the available evidence.
The consortium expanded validation, adapted the instructions, tested physical supports, and published subgroup results with uncertainty intervals. It also created a process ensuring that an invalid reading led to professional assessment rather than false reassurance. The central lesson was institutional: representativeness is not a formality added after development but part of design. Commercial communication had to distinguish what was demonstrated, what was probable, and what remained unknown.
Listening is a method of measurement#
Individual stories do not replace quantitative evaluation, but they reveal mechanisms that an average cannot name. One complaint may expose an ambiguous translation; repeated calls may identify an impossible opening time; observation in a consultation may show that a screen demands more attention than it saves. Listening becomes rigorous when an accessible channel, categories for analysing patterns, accountable decision-makers, and feedback to contributors all exist. Collecting testimony without changing anything merely transfers more work to affected people.
Participation should begin before a tool is selected. Patients, clinical professionals, administrative staff, accessibility specialists, data-protection officers, and maintenance workers understand different parts of the pathway. Their involvement helps define which outcome deserves measurement and which harm should be unacceptable. It also reveals displaced costs: minutes that disappear from a central report because a family member, nurse, or community organisation absorbs them. Genuine efficiency considers total time and the distribution of effort.
Listening includes acknowledging disagreement. A person may decline secondary use of health data or prefer face-to-face care without being uninformed. Autonomy requires understandable information, meaningful choices, and proportionate consequences. The European Health Data Space establishes a framework for access to, control of, and secure exchange of electronic health data; its gradual application requires attention to privacy, interoperability, and rights. A technically connected architecture becomes humane only when people can understand who has access, for what purpose, and how to request a correction.
An evaluation that makes distribution visible#
A responsible protocol defines the intended population, relevant outcomes, justified subgroups, minimum data quality, and pause criteria in advance. It then combines aggregate and distributed measures: false positives and negatives, dropouts, extreme delays, incidents, accessibility, workload for professionals, and patient-reported outcomes. Differences require statistical support and contextual knowledge. A small table cannot establish causation, but a repeated signal should not be dismissed merely because the overall average remains favourable.
Review needs a schedule and an accountable owner. Systems change when a model is updated, a sensor is replaced, the population shifts, or staff adapt their practice. Previous performance is no guarantee of future results. Versions should be documented, traceability preserved, independent audit enabled, and routes of redress provided. If disproportionate harm emerges, the organisation must be able to restrict, correct, or withdraw the tool while preserving continuity of care.
Privacy and equity are not opposing goals. Evaluating differences may require sensitive data, but the work must apply minimisation, clear purpose, security, limited retention, and controlled access. Analysis protected by an independent unit or techniques that lower identification risk may sometimes be preferable. Missing safeguards cannot be repaired by collecting more information. Trust depends on explaining the use and making responsibility verifiable.
From finding a gap to making a concrete repair#
When a gap appears, publishing the difference is only a beginning. The organisation must investigate its mechanism, listen to those bearing the consequence, and assign resources. Repair may require accessibility, translation, changes to opening hours, updated equipment, training, additional staff, or redesigned criteria. It may also require contacting affected people, explaining the failure, and offering professional review. The response should be proportionate and should not force each patient to prove separately a systemic problem that the institution already knows exists.
Mature transparency avoids two extremes: presenting technology as neutral and dismissing it because it has any limitation. Every intervention has a context in which evidence supports its use. A useful report explains purpose, population, data, group-level performance, uncertainty, oversight, and the complaint route. Communication must distinguish a support tool from a clinical decision. General information never replaces assessment by a healthcare professional; a personal concern belongs with the care team, and an emergency belongs with local urgent services.
Precision medicine that keeps listening always asks who is missing from the figure. It examines both the average outcome and its distribution, treats missing data as meaningful, and considers each exception as possible information about the system. This discipline does not promise that technology will eliminate all inequality. It does allow gaps to be found earlier, discussed honestly, and addressed responsibly. Success does not mean giving every person an identical process. It means preventing an avoidable difference from deciding who receives safe, understandable, and dignified care.
