Key points
- WHO's classification (2nd edition, 2023) groups digital health interventions by their main user, namely persons, healthcare providers, health management and support personnel, and data services.
- Going through all 32 categories shows possible AI uses well beyond clinical care, in supply chains, financing, surveillance, vital registration and data management.
- Some categories are mostly a job for conventional software, and only a few AI uses are established. Most are emerging or exploratory.
Discussions of AI in health often focus on one product at a time, such as a chest X-ray reader or a chatbot. WHO’s Classification of digital interventions, services and applications in health offers a way to look at the whole health system instead. It is a shared language for describing what digital technology does in health, and it is organised by who uses each intervention.
The classification was not written as a framework for AI. The second edition, published in 2023, does name a few AI related functions directly, such as simulated human-like conversations (1.6.2), generative AI for tailored content creation (2.5.6) and automated analysis of data to generate new information or predictions (4.1.4). This piece goes through every category and asks a simple question. Where could AI play a part, and where is the work mostly done by conventional software?
How to read the tables
Each table follows one of WHO’s four user groups, with the WHO category number and name as they appear in the classification. For each category, the table gives possible AI uses, the kind of AI involved, and a rough status.
- In use in some settings means the use exists in real services somewhere. It does not mean the use has been shown to improve health outcomes.
- Emerging means there are pilots, products or research, but use is not widespread.
- Limited AI role means the category is mainly served by conventional software, rules or standards, and AI adds little.
It is worth being clear about what counts as AI here. Sending a scheduled reminder, applying a clinical checklist or running a fixed rule is conventional software, even when it is automated. The tables list a use as AI only when it involves models that learn from data, such as prediction, computer vision, speech recognition or language models.
Persons
| WHO category | Possible AI uses | Kind of AI | Status |
|---|---|---|---|
| 1.1 Targeted communication to persons | Predicting who is likely to miss a visit, so that reminders can be prioritised. Drafting messages in local languages | Prediction, language models | Emerging |
| 1.2 Untargeted communication to persons | Drafting and translating health promotion content, for review by staff | Language models | Emerging |
| 1.3 Person to person communication | Helping moderators of peer support groups spot harmful content or misinformation | Language models | Emerging |
| 1.4 Personal health tracking | Wearables that notify users of a possible irregular heart rhythm | Prediction from sensor data | In use in some settings |
| 1.5 Person based reporting | Sorting and summarising reports and feedback sent in by the public | Language models | Emerging |
| 1.6 On demand communication with persons | Chatbots and voice assistants that answer health questions (1.6.2) | Language models, speech | In use in some settings |
| 1.7 Person-centred financial transactions | Flagging unusual transactions for review | Anomaly detection | Emerging |
| 1.8 Person-centred consent management | Recording and managing consent | Conventional software | Limited AI role |
WHO itself launched S.A.R.A.H., a prototype digital health promoter powered by generative AI, in 2024. Chatbots can answer questions at any hour and in many languages, but a fluent answer is not necessarily a correct one, which is why content review and clear limits matter.
Healthcare providers
| WHO category | Possible AI uses | Kind of AI | Status |
|---|---|---|---|
| 2.1 Identification and registration of persons | Biometric matching, such as face or fingerprint, to verify identity | Computer vision | In use in some settings |
| 2.2 Person-centred health records | Drafting clinical notes from consultations, and summarising long records | Speech recognition, language models | In use in some settings |
| 2.3 Healthcare provider decision support | Screening persons by risk (2.3.3), such as computer-aided detection of TB on chest X-rays, and risk prediction models | Computer vision, prediction | In use in some settings |
| 2.4 Telemedicine | Pre-screening images sent for remote review, and flagging remote monitoring readings that need attention | Computer vision, prediction | Emerging |
| 2.5 Healthcare provider communication | Generative AI for tailored content creation (2.5.6), such as drafting guidance or summaries for staff | Language models | Emerging |
| 2.6 Referral coordination | Forecasting demand for emergency transport | Prediction | Emerging |
| 2.7 Scheduling and activity planning for healthcare providers | Identifying persons in need of services (2.7.1), such as those at risk of dropping out of care | Prediction | Emerging |
| 2.8 Healthcare provider training | Simulated patients for practice, and feedback on written answers | Language models | Emerging |
| 2.9 Prescription and medication management | Finding possible adverse drug effects in reports and notes (2.9.3) | Language models | Emerging |
| 2.10 Laboratory and diagnostics imaging management | Reading retinal images, pathology slides and radiology scans | Computer vision | In use in some settings |
| 2.11 Healthcare provider financial transactions | Suggesting billing codes from clinical notes | Language models | Emerging |
This group has the clearest examples of established AI. In 2021, WHO recommended that computer-aided detection software may be used in place of human readers to interpret chest X-rays when screening people aged 15 and over for TB. In 2018, the US Food and Drug Administration authorised a device that uses AI to detect diabetic retinopathy without a specialist reading the image. Regulatory authorisation and guideline recommendations are about specific uses and conditions, and do not by themselves show better outcomes everywhere a tool is used. Much of what is called decision support, such as prompts and checklists (2.3.1 and 2.3.2), is rules based and does not involve AI at all.
Health management and support personnel
| WHO category | Possible AI uses | Kind of AI | Status |
|---|---|---|---|
| 3.1 Human resource management | Forecasting staffing needs, often done with conventional statistics | Prediction | Emerging |
| 3.2 Supply chain management | Forecasting demand for medicines and commodities | Prediction | Emerging |
| 3.3 Public health event notification | Scanning news and other open sources for signals of public health events | Language models | In use in some settings |
| 3.4 Civil registration and vital statistics | Assigning probable causes of death from verbal autopsy interviews | Statistical algorithms, increasingly language models | In use in some settings |
| 3.5 Health system financial management | Flagging unusual insurance claims for human review | Anomaly detection | Emerging |
| 3.6 Equipment and asset management | Predicting when equipment needs maintenance, from sensor data | Prediction | Emerging |
| 3.7 Facility management | Checking facility lists and locations against imagery and other data | Computer vision | Emerging |
| 3.8 Person-centred health certificate management | Issuing, validating and revoking certificates | Conventional software | Limited AI role |
WHO’s Epidemic Intelligence from Open Sources (EIOS) initiative brings together information from news and other open sources to support the early detection of public health events, with automated processing to help analysts filter large volumes of text. For verbal autopsy, computer algorithms that assign a probable cause of death have been used for many years. Many of them are statistical methods rather than AI in the modern sense, which is a good reminder that not every automated analysis is AI. Forecasting in this group is similar. A good forecast may come from a simple statistical model.
Data services
| WHO category | Possible AI uses | Kind of AI | Status |
|---|---|---|---|
| 4.1 Data management | Automated analysis of data to generate new information or predictions (4.1.4), such as spotting unusual values in routine reports | Prediction, anomaly detection | Emerging |
| 4.2 Data coding | Turning free text and scanned forms into structured data (4.2.1), matching duplicate records (4.2.2), and suggesting disease or cause of death codes (4.2.3) | Language models, computer vision, prediction | Emerging |
| 4.3 Geospatial information management | Estimating where people and settlements are from satellite imagery (4.3.3) | Computer vision, prediction | In use in some settings |
| 4.4 Data exchange and interoperability | Suggesting matches between local codes and standard terminologies | Language models | Limited AI role |
| 4.5 Data governance compliance | Finding and removing personal details in clinical text, as part of data privacy protection (4.5.2) | Language models | Emerging |
Most of this group is about getting data right, and much of it is done with conventional software and agreed standards. Interoperability in particular depends on standards and governance rather than on AI. Where AI does help, it is often in the slow, practical work of turning unstructured data into something usable.
What the mapping shows
Three points stand out.
- AI uses appear across the whole classification. They go well beyond clinical care, into supply chains, financing, surveillance, vital registration and data management.
- Established uses are few. Most of the uses in the tables are emerging or exploratory, and being in use somewhere is not the same as being shown to improve health.
- Some categories need little AI. Consent, certificates, interoperability and many scheduling and decision support tools are largely served by conventional software, rules and standards.
About this mapping
This is my own mapping of possible AI uses against the WHO framework. It is not an official WHO taxonomy, not a systematic review and not an assessment of whether any use is effective. The WHO categories and numbers are taken from the second edition of the classification. The AI uses and status labels are my reading of the general picture, and they will change as evidence develops.
Views are my own.
Sources
- WHO. Classification of digital interventions, services and applications in health. A shared language to describe the uses of digital technology for health, 2nd edition. 2023
- WHO. Consolidated guidelines on tuberculosis. Module 2, screening. Systematic screening for tuberculosis disease. 2021
- US Food and Drug Administration. De Novo authorisation of an AI-based device for detecting diabetic retinopathy. April 2018
- WHO. S.A.R.A.H., a Smart AI Resource Assistant for Health. 2024
- WHO. Epidemic Intelligence from Open Sources (EIOS)
- WHO. Verbal autopsy standards, the 2022 WHO verbal autopsy instrument