An AI is only as good as the decisions it helps clinicians make.
We evaluate Legit.Health through development testing and clinical studies. Together, they examine how the software performs, how clinicians use its results and how it supports diagnosis, severity assessment and referral decisions.
Our clinical validation programme includes more than ten studies across different conditions and populations. This page explains the methods and highlights selected findings. The Clinical Validation Studies in the Instructions for Use provide the study catalogue and identify the clinical benefits each study supports.
Two different numbers
Model performance and clinical performance answer different questions.
These evaluations complement each other: one measures the software’s output, while the other examines its use in a clinical task.
The algorithm analyses images, and its output is compared with a reference diagnosis. Sensitivity, specificity and area under the curve (AUC) describe different aspects of its performance on that set of cases.
In our reader studies, the same clinicians assess the same cases without and with Legit.Health. Comparing their answers measures the change in diagnostic accuracy for the cases and reading conditions studied.
Development test sets and clinical studies can contain different conditions, populations and reference diagnoses. Their results are reported separately, with the study and endpoint identified, rather than averaged into a single accuracy figure.
Why it matters here
Legit.Health supports the decision; the clinician makes it.
Legit.Health returns a ranked list of the conditions a photograph may show, each with its probability, together with the referral and malignancy indexes and severity scores. The clinician reads that next to everything the photograph cannot show: the history, the examination, the medication, the previous visits. The decision is theirs.
Top-1 measures how often the correct condition appears first; Top-3 and Top-5 count whether it appears among the first three or five. These describe the ranked differential. Reader studies then assess how clinicians’ answers change when they use the report.
Published clinician results
Getting the first differential diagnosis right is a demanding task for clinicians too. In a teledermatology study of 963 cases, primary-care physicians achieved 44% accuracy for their first diagnosis (Top-1), and dermatologists 63%, against a dermatologist panel's reference diagnosis. Liu et al., Nature Medicine (2020).
In a cohort of 144 cases from a prospective study of suspected skin neoplasms at South Korean hospitals, attending dermatologists achieved 61.8% Top-1, 69.4% Top-2 and 71.5% Top-3 accuracy. Kim et al., PLOS ONE (2022).
These results provide context for diagnostic accuracy: conditions and diagnostic reference standards differ from IDEI_2023, which evaluated histopathology-confirmed pigmented lesions with nevus subtypes grouped together. They are not a direct comparison with Legit.Health.
How the studies are built
The same clinicians, the same cases, with and without.
Our reader studies use a multi-reader multi-case design: several clinicians assess a shared set of cases.
- The cases: Photographs are selected for the study’s clinical question, with a reference diagnosis for each case.
- The first read, without Legit.Health: Each clinician records their diagnosis for every case, on their own.
- The second read, with Legit.Health: The same clinician sees the same case again with the Legit.Health report beside it, and records their diagnosis again.
- The comparison: Each clinician’s answers are compared case by case between the two readings. The study protocol defines the endpoints and analysis before the results are evaluated.
Using photographs lets every clinician assess the same cases. Prospective clinical investigations and studies of referrals complement these image-reading tasks.
What the studies found
Clinicians named the right condition more often with Legit.Health.
Every study below compared the clinicians using Legit.Health with the same clinicians reading the same cases without it.
The specialty analysis in MAN_2025 is exploratory.
Primary care physicians
BI_2024Generalised pustular psoriasis and other conditions+20 percentage points 49% to 69%PH_2024Skin lesions in primary care+18 percentage points 64% to 82%SAN_2024A wide range of skin conditions+27 percentage points 63% to 90%MAN_2025Dark skin, Fitzpatrick V and VI+28 percentage points 37% to 65%
Dermatologists
In every study shown that compared the two professional groups, the gain was larger for primary care physicians. The charts show first-ranked diagnostic accuracy; gains are expressed in percentage points.
The studies
BI_2024, generalised pustular psoriasis: Evaluated recognition of this rare disease alongside conditions with similar presentations. See the study.SAN_2024, a wide range of conditions: Primary care physicians and dermatologists, on cases from the common to the rare. See the study.MAN_2025, dark skin: Every case was a Fitzpatrick type V or VI presentation, the skin types most datasets under-represent.PH_2024, primary care: General practitioners, on the skin lesions they see every week. See the study.
Skin cancer
Tested on real patients.
Clinical investigations assess the malignancy score and the ranked differential in defined patient populations. These examples address different clinical questions.
- Pigmented lesions confirmed by histopathology: In the prospective arm of
IDEI_2023, Legit.Health’s malignancy score achieved an AUC of 96.69% (89.00%-100.00%) for distinguishing malignant from benign pigmented lesions, using histopathology as the reference. The interval shown is the 95% confidence interval. This estimate comes from a small prospective sample, with independent confirmation planned. See the study. - Basal cell and squamous cell carcinoma of the head and neck: A prospective study at a head and neck surgery clinic evaluated recognition of these cancers in the differential diagnosis, using histopathology as the reference. Published in the European Archives of Oto-Rhino-Laryngology. See the study.
AUC describes how well the scores distinguish malignant from benign lesions across decision thresholds. It is not the percentage of cancers detected at a particular cutoff.
Choosing a dermatology AI
Four questions worth asking any vendor.
These are the questions we ask of our own evidence, and they apply to any tool.
- Was it measured with clinicians in the loop? A sensitivity on a test set says how the software behaves on its own. Ask for the result of clinicians with and without it, on the same cases.
- Which clinicians? Ask for results in the professional groups who will use the tool, such as primary care physicians or dermatologists.
- Which conditions? Ask whether the evidence covers the conditions your patients bring, or only one of them, such as suspected skin cancer.
- Is it certified for that use? The intended purpose on the certificate says what the evidence had to prove. Legit.Health's is set out on the certification page.
Planifions unedémonstration en direct
Legit.Health est la solution d'intelligence artificielle tout-en-un avec des performances ultra élevées, des fonctionnalités exclusives et un soutien clinique.
