Digital health and AI in healthcare

Digital health covers technologies used to support health and care. Artificial intelligence, or AI, is one group of methods within this wider field. Benefits depend on evidence, inclusive design, safe deployment and clear accountability.

Digital health is broader than AI Digital health includes electronic records, online booking, video consultations, messaging, remote monitoring, health apps and clinical software. These tools do not all use artificial intelligence. AI is a broad label for computer systems that produce outputs such as classifications, predictions, recommendations or generated content. Many current systems learn statistical patterns from data. Some combine learned methods with programmed rules. Generative AI creates new text, images, audio or other content from patterns learned during training. It can produce fluent answers without understanding a person or situation as a clinician does. Digital tools may improve communication, access, monitoring or administrative work. These benefits are possibilities, not guarantees. A tool can also add workload, delay care or create new barriers when it is poorly designed or introduced without support. A tool must be judged for a specific purpose The intended purpose describes what a tool is designed to do, who should use it and in which setting. It also includes the population, information used and meaning of the output. A system tested for organising messages is not automatically suitable for diagnosing illness. A tool validated for one population or service may perform differently elsewhere. Clinical validation examines whether a tool performs well enough for its intended purpose. Evaluation should consider accuracy, missed cases, false alarms, patient outcomes, workflow effects and possible harms. Average performance can hide important differences between groups. Local testing and monitoring matter because people, equipment, workflows and disease patterns can differ from the original study. Performance may also change after software updates or changes in the data entering the system. Safety therefore needs continuing monitoring rather than one check before launch. Some software with an intended medical purpose is regulated as a medical device. Other health related software does not meet that definition. Regulation is an important safeguard, but it does not make a product perfect or suitable for every task. Safety and accountability are shared Safe digital care depends on several groups. Developers and manufacturers are responsible for design, testing, documentation, updates and responding to product problems. Health and care organisations are responsible for selecting appropriate tools, assessing local risks, training staff and monitoring use. In England, national clinical safety standards apply to relevant health IT manufacturers and deploying organisations. Healthcare professionals remain responsible for the decisions they take when using a tool. They should understand its purpose and limitations, work within their competence and report concerns. Responsibility should not be placed on one clinician alone. Suppliers, organisations, professionals and regulators have different duties across the technology's lifecycle. Human oversight must be meaningful. The reviewer needs enough training, information, time and authority to question or reject an output. Adding a person who can only approve the result does not create effective oversight. Automation bias means trusting an automated output too readily. It can cause someone to overlook contradictory information or accept a confident recommendation without enough checking. The level of human review should match the possible harm. A low risk administrative task may need different safeguards from software influencing diagnosis or treatment. Generative AI can sound convincing and still be wrong Generative AI can produce inaccurate statements, invented references or missing details that sound plausible. This is often called a hallucination. A polished answer is not evidence that the information is correct. Important outputs need checking against reliable sources, the original record and the clinical context. General purpose AI systems are not automatically validated clinical tools. A service may be useful for drafting or explanation while remaining unsuitable for diagnosis, prescribing or urgent triage. An AI answer should not be used to rule out a serious condition. Severe, rapidly worsening or time critical symptoms need the appropriate urgent or emergency service rather than further chatbot discussion. Bias can enter at several stages Algorithmic bias does not arise only from unrepresentative training data. It can enter through the problem chosen, how outcomes are labelled, what information is measured and which errors are treated as acceptable. Bias can also appear during deployment. A tool may perform differently when used with new equipment, languages, populations or workflows. Fair evaluation examines performance across relevant groups and settings. It also asks whether the tool changes access, workload or outcomes in ways that widen existing inequalities. Digital access must remain inclusive Digital exclusion can involve lacking a suitable device, affordable data, reliable internet, digital skills or confidence. Language, disability, health literacy and inaccessible design can also create barriers. A digital route may help some people while disadvantaging others. NHS England advises that digital approaches should complement suitable non digital support rather than quietly remove it. Inclusive design involves people who will use the service. It also provides accessible information, assistance and alternative routes when needed. Health data needs careful protection Health information is sensitive. Data protection requires lawful, fair and transparent use, clear purposes, data minimisation, accuracy, security and accountability. Privacy, confidentiality and cybersecurity are related but different. A system can have strong encryption while still using more personal data than necessary or producing unsafe clinical outputs. Do not assume that a public AI service is approved for confidential health information. Use authorised secure systems and follow the healthcare organisation's instructions before entering identifiable information. People should receive proportionate information when AI meaningfully affects their care. This may include the tool's role, important limitations, human involvement and how to question or correct an error. The central question is specific AI is not simply good or bad. The useful question is whether a particular tool is safe, effective, fair and appropriate for a defined purpose. Good deployment combines evidence, regulation where applicable, clinical risk management, data protection, inclusive design, training and ongoing monitoring. Human judgement remains important, but it must sit within a safe system rather than carry the entire burden.

Do not judge healthcare AI by the label alone. Judge the specific tool, intended purpose, evidence, population, setting, safeguards, data use and accountability.