Text-based triage tools help people find the right level of care. For skin concerns, adding a photo can provide important information that questions alone cannot capture.
AI symptom checkers are becoming a common first step in digital healthcare. Patients answer questions about their symptoms, and the tool suggests what might be happening and where to seek help. This could mean self-care, a pharmacist, a GP or an emergency department.
These tools play an important role. Research suggests they are generally better at guiding people to the right level of care than identifying the exact condition. That is a valuable function for any healthcare service.
Skin concerns, however, present a particular challenge. Many skin conditions are recognised by their appearance. A written description can explain symptoms, but it cannot show the whole picture. Adding image-based AI does not replace text-based triage. It completes it.
What text-based triage tools are Text-based triage tools come in three main forms:
Consumer apps, including versions licensed to insurers and healthcare systems.
API-based tools integrated into other services, such as telehealth apps and national health advice lines.
Tools used within hospitals, including emergency departments.
These tools follow a similar process. They ask patients questions about their symptoms and medical history. They then provide possible conditions and recommend what the patient should do next. The recommendation often includes a level of urgency.
Some tools are regulated medical devices rather than general wellness apps, certified under the EU Medical Device Regulation for defined purposes.
The evidence suggests that text-based triage can support healthcare services when used for its intended purpose.port healthcare services when used for its intended purpose.
Research on symptom checkers shows mixed results. However, one finding is consistent. These tools tend to perform better at assessing urgency than at identifying the correct condition.
Triage is the main strength. A systematic review of 10 studies found that triage advice was correct in 48.8% to 90.1% of cases. The first condition suggested was correct in only 19% to 37.9% of cases. For a tool designed to guide patients towards appropriate care, the ability to assess urgency is particularly important.Improving care decisions at national scale. A study of Australia’s national virtual triage service found that the proportion of users planning to seek appropriate, lower acuity care increased from 21.3% before triage to 52.9% afterwards. The number of people unsure where to seek care fell from 670,502 to 2,557 (Gellert et al., 2026 ). However, the study measured users’ stated intentions, not their subsequent actions.Influence the actual behaviour. A study in Portugal’s largest private health network examined 382 patients with non-emergency complaints. The proportion whose planned next step was considered clinically appropriate was 29.8%. After using the tool, 64.4% took an action rated clinically appropriate. Researchers assessed actual behaviour mainly through health records (Cotte et al., NEJM AI, 2026 ).Support in the emergency departments. In a university hospital in Barcelona, the condition eventually identified by the treating physician appeared among the tool’s top three suggestions in 75.4% of 622 cases. It was the first suggestion in 42.9% of cases (Moreno Barriga et al., 2017 ).
These findings show why text-based triage remains useful. It helps patients understand where to seek care and supports decisions about urgency. The challenge is not that these tools have no value. It is that some complaints require information that text alone cannot provide.
Skin conditions are a clear example of this limitation. Researchers tested eight symptom checkers using 15 rash cases. Across 120 entries, the tools identified the correct condition in 37 cases, or 30.8%. Individual tools achieved results ranging from 13.3% to 73.3% (Berry et al., 2023 ).
The researchers identified two possible reasons for the results. The tools collected limited information about patients’ histories, and their algorithms covered too few skin conditions. Better questions and broader coverage could help address these issues.
However, another limitation remains. Skin conditions are largely visual.
Clinicians assess features such as shape, colour, texture, borders and distribution. A description such as “a red, itchy patch” could fit many different conditions. Most people also lack the medical vocabulary needed to describe subtle visual differences.
Asking more questions can help narrow the possibilities. But questions cannot show the actual appearance of a rash.
General-purpose chatbots face a similar challenge. Autoderm discusses this issue in its article on why AI chatbots struggle with skin conditions.
Skin conditions are not a minor category of healthcare. They account for a substantial share of everyday primary care.
In Finland , skin conditions were the main reason for 7.8% of GP visits in public primary care between 2015 and 2019. This represented almost 1.5 million consultations (Salava et al., 2021 ).In the UK , a one-month audit of 2,175 GP consultations found that 11.31% concerned dermatological conditions. Around one-third of these consultations, or 32.5%, involved conditions that could have been self-managed (Dolfini and Patel, 2020; conference abstract ).
These figures highlight an important opportunity for digital healthcare. Skin complaints are common, and many do not require urgent medical attention. Some may be managed through appropriate self-care, while others need a GP or specialist.
This is exactly the type of situation in which digital triage can help. Yet skin complaints are also a category where written questions have clear limitations. Adding an image step to an existing digital care pathway could help address this gap. It would allow the service to gather visual information without losing the benefits of its existing triage process.
Image-based AI can analyse visual information that text-based tools cannot capture. Research suggests that combining photographs with clinical history can improve the assessment of skin concerns.
Performance on common skin conditions. A study examined an AI system that combined photographs with basic clinical history. It covered 26 common conditions that accounted for 80% of skin cases seen in primary care. The system achieved 66% top-1 accuracy. The corresponding figures were 63% for dermatologists, 44% for primary care physicians and 40% for nurse practitioners (Liu et al., Nature Medicine, 2020 ).Support for non-specialists. Another study involved 40 clinicians and 1,048 cases. With AI support, agreement with a panel of dermatologists increased from 48% to 58% among primary care physicians. For nurse practitioners, agreement rose from 46% to 58% (Jain et al., JAMA Network Open, 2021 ).Skin tone remains an important consideration. A study involving 389 dermatologists and 459 primary care physicians from 39 countries found that AI support increased accuracy by more than 33% for both groups. However, it also widened the difference in general practitioners’ accuracy between lighter and darker skin (Groh et al., Nature Medicine, 2024 ).Autoderm in primary care. In a prospective study of 100 patients in central Catalonia, GPs reported that Autoderm’s suggestions helped them consider possible conditions in 92% of cases. They also reported that a teledermatology referral might have been avoided in 34% of cases (Escalé-Besa et al., Scientific Reports, 2023 ; see Autoderm’s write-up of the study).
The first two studies combined photographs with clinical history. These findings support the potential value of image-based information in primary care. They should be understood in the context of the studies and their specific settings.
Text-based triage and image-based AI have different strengths.
Text-based triage gathers medical history, identifies warning signs, assesses urgency and recommends where to seek care.Image-based AI provides visual information and returns informational condition suggestions based on a photograph
The two approaches can work together in a straightforward process.
The patient answers the existing text-based questions and identifies a skin concern.
The service requests a photograph and provides clear instructions on lighting, distance and focus.
An image service returns informational condition suggestions, which are presented to the user.
The triage tool assesses urgency from the patient’s answers, including symptom duration and warning signs, and recommends the next step. The image suggestions are shown to the user alongside that recommendation.
Each component retains its role. The text-based triage tool remains responsible for assessing urgency and guiding the patient to the next step. The image component adds visual information that questions cannot provide.
Image quality also matters. Poor lighting, an unclear focus or an unsuitable distance can affect the usefulness of a photograph. Clear instructions should therefore be part of the process from the beginning. Further information is available in how image quality shapes AI dermatology accuracy.
For teams developing or selecting a digital care pathway, four questions deserve early attention:
When should the service request a photograph, and which skin complaints should trigger this step?
How should image-based suggestions inform the triage process, and who is responsible for the final routing decision?
Has the image component been tested across different skin tones and with photographs similar to those users are likely to submit?
Are the intended uses of both components suitable for the combined workflow?
Answering these questions helps teams integrate image AI without disrupting the existing triage process.
Where Autoderm fits
Autoderm is an image API that can sit alongside a text-based triage engine in this kind of workflow. It returns informational condition suggestions across skin concerns of every kind from a smartphone photograph. It is designed to integrate into existing services. This allows healthcare providers to add an image step without replacing their existing text-based process.
Autoderm is already used by telemedicine services and online pharmacies. More information is available on its telemedicine and online pharmacy pages, and on its clinical evidence page .
Text-based triage remains valuable for assessing urgency and guiding patients towards appropriate care. For skin concerns, adding image-based AI can provide visual information that questions alone cannot capture. The aim is not to replace text-based triage. It is to strengthen the existing process by combining symptom information with visual information.
Text triage has made the digital front door work for many complaints. For skin, it needs a camera.
References
Wallace W et al. npj Digital Medicine, 2022. doi.org/10.1038/s41746-022-00667-w
Gellert GA et al. Mayo Clinic Proceedings: Digital Health, 2026. doi.org/10.1016/j.mcpdig.2025.100331
Cotte F et al. NEJM AI, 2026. doi.org/10.1056/AIoa2500833
Moreno Barriga E et al. Emergencias, 2017. pubmed.ncbi.nlm.nih.gov/29188913
Berry NA et al. Journal of the American Academy of Dermatology, 2023. doi.org/10.1016/j.jaad.2022.06.034
Salava A et al. Skin Health and Disease, 2021. doi.org/10.1002/ski2.53
Dolfini L, Patel Y. British Journal of General Practice, 2020 (conference abstract). bjgp.org/content/70/suppl_1/bjgp20X711053
Liu Y et al. Nature Medicine, 2020. doi.org/10.1038/s41591-020-0842-3
Jain A et al. JAMA Network Open, 2021. doi.org/10.1001/jamanetworkopen.2021.7249
Groh M et al. Nature Medicine, 2024. doi.org/10.1038/s41591-023-02728-3
Escalé-Besa A et al. Scientific Reports (Nature Portfolio), 2023. doi.org/10.1038/s41598-023-31340-1
