Validation Research

ScanSkinAI Accuracy Report 2026

Independent clinical validation of AI skin screening across 80+ conditions and all six Fitzpatrick skin types.

Published: 2026-07-06 · Last updated: 2026-07-06

Reviewed by Dr. Anand S. Urhekar (independent clinical reviewer)

Key findings

Overall clinically acceptable accuracy: 95.3% across 80 skin conditions.
By dimension (Tier 2, clinically acceptable): diagnostic 95%, clinical explanation 93%, triage 98%.
Under-triage rate — the critical safety metric — was 2% (2 out of 100 cases).
Independently reviewed by Dr. Anand S. Urhekar, MD (25+ years of clinical experience), under a locked protocol (IVY-CLIN-PROTO-001) with the reviewer blinded to internal metrics and performance targets.
Stratified across all six Fitzpatrick skin types (I–VI), with the sample weighted toward Types III–IV to reflect Asia-Pacific user demographics.
Independent audit dataset: 100 anonymised real-world clinical cases, reviewed 9–10 March 2026.

What "accuracy" means in this report

"Clinically acceptable accuracy" is the percentage of cases where the AI output would not lead to patient harm or a materially incorrect clinical pathway. It is the standard metric used in the medical AI validation literature for triage and screening tools. It captures both exact agreement with the reviewing clinician and clinically safe partial agreement — for example, over-triaging a benign lesion to further review is counted as acceptable, whereas under-triaging a concerning lesion to reassurance is not.

It is not the same thing as sensitivity or specificity for a single disease. Sensitivity and specificity are per-condition metrics used in single-disease detection studies (for example, melanoma versus benign naevus). The 95.3% figure reported here is measured across 80 different skin conditions and reflects overall screening safety and agreement with an independent dermatologist, not per-disease detection performance.

Concretely, this figure describes how often ScanSkinAI produced an output a clinician would consider a safe and reasonable first-pass screening response — across a diverse set of conditions and skin types. It does not tell you the probability that any individual scan result is correct for you personally, and it is not a substitute for clinical assessment.

How ScanSkinAI was validated

Three-tier clinical assessment framework

Each case in the audit was scored on three tiers, mirroring the reporting structure recommended in the medical AI validation literature:

Tier 1 — Strict Concordance Rate. The percentage of cases where the AI output received the highest level of agreement from the independent reviewer, with no qualifications. This is the most conservative measure and counts only exact agreement. Formula: strict rate = (top-category count / total cases) × 100.
Tier 2 — Clinically Acceptable Rate. The percentage of cases where the AI output was clinically acceptable, meaning it would not lead to patient harm or a materially incorrect care pathway. This tier includes both exact agreement and clinically safe partial agreement. Formula: acceptable rate = ((top-category + partial-category) / total cases) × 100. 95.3% is the Tier 2 figure.
Tier 3 — Critical Failure Rate. The percentage of cases where the AI output represented a clinically significant error — one that could lead to misdirection, misdiagnosis, or an unsafe care pathway. Reporting this metric alongside the others is a transparency requirement, not a marketing choice.

Dataset

The validation dataset is a 100-case independent audit set composed of anonymised real-world production scans spanning 80 skin conditions and Fitzpatrick skin types I–VI. Cases were sampled by stratified random sampling across condition category and skin type, and were not used in training or fine-tuning of the model. The audit was retrospective in design, reflecting real-world performance rather than a controlled trial setting, and was conducted 9–10 March 2026 under protocol IVY-CLIN-PROTO-001, which was locked before any case review commenced.

Independent review process

Case-by-case review was conducted by Dr. Anand S. Urhekar, MD, Section Head of Dermatology at M.P. Shah Hospital in Nairobi, with over 25 years of clinical experience across India, South Africa (UN), Micronesia, Uzbekistan, and Kenya, and prior consultant dermatology roles including 8 years at Aga Khan Hospital Kisumu. Qualifications include MD (Kharkov State Medical University, Ukraine), DVD Masters in Dermatology (M.R. Medical College, India), FAM, and PGDCC. Medical Council of India registration No. 15334. Dr. Urhekar was independently contracted, was not involved in developing the AI model, and holds no equity, advisory, employment, or financial interest in Ivy AI Solutions. He was blinded to internal confidence scores and performance targets, and evaluated only the user-facing AI output for each case.

Scoring rubric

Each case was scored on three dimensions with three rating levels per dimension:

Diagnostic Concordance: Agree · Partially Agree · Disagree.
Clinical Explanation: Accurate · Minor Issues · Inaccurate.
Triage Appropriateness: Appropriate · Over-triaged · Under-triaged.

Conditions covered

ScanSkinAI screens for 80 visible skin conditions using a two-tier vision–language architecture (a DINOv2-Base Vision Transformer visual encoder combined with a large language model classification head). Coverage spans 30 flagship AI classifier classes plus 50 additional high-frequency conditions curated for outpatient dermatology in humid, high-density, high-UV environments typical of Hong Kong and the broader Asia-Pacific region. All 80 conditions are mapped to ICD-10 code families in the underlying dossier (IVY-CLIN-007). Full architectural detail is proprietary; what matters for this report is that the same model configuration was audited end-to-end.

Condition coverage by clinical category, showing total conditions and flagship AI classifier classes.
Category	Total conditions	Flagship (★)
Neoplastic	12	5
Infectious	19	11
Inflammatory	16	4
Pigmented	17	6
General Dermatology	16	4
Total	80	30

Limitations

Company-sponsored. The audit was commissioned and funded by Ivy AI Solutions Limited. To mitigate bias, the protocol was locked before review and the reviewer was independently contracted with no financial interest in the company.
Single reviewer. Phase 1 used one independent dermatologist, so inter-rater reliability cannot be assessed. Phase 2 expands to two or more reviewers.
Sample size. A 100-case audit limits the statistical power of subgroup analyses (for example, per-condition or per-Fitzpatrick-type breakdowns).
Retrospective design. The audit reviews historical production scans rather than prospective cases; it reflects real-world use, not a controlled trial.
Performance depends on image quality. Poor lighting, motion blur, focus problems, and heavy makeup or occlusion reduce reliability. Some cases in the audit had suboptimal image quality noted by the reviewer, which affects both AI and expert assessment.
ScanSkinAI does not cover every possible skin condition. Rare presentations, deep or subsurface pathology, and conditions requiring dermoscopy, histology, or imaging beyond visible-light photography are outside its scope.
The report describes screening and triage performance. It does not describe diagnostic performance for any single disease.

Results

Headline result

95.3% clinically acceptable accuracy

Across 80 skin conditions, 100-case independent dermatologist audit.

Three-tier performance summary

All three reporting tiers are shown together for full transparency. Each serves a different audience: Tier 1 is the most conservative measure (regulatory and peer review), Tier 2 is the standard metric in medical AI validation (partner and commercial contexts), and Tier 3 is the critical failure rate (clinical governance).

Three-tier performance summary across diagnostic concordance, clinical explanation, and triage appropriateness.
Dimension	Strict rate (Tier 1)	Clinically acceptable (Tier 2)	Critical failure (Tier 3)
Diagnostic concordance	91%	95%	5%
Clinical explanation	87%	93%	7%
Triage appropriateness	90%	98%	2% (under-triage)
Overall (mean)	89.3%	95.3%	4.7%

Absolute counts across the 100 cases: Diagnostic — Agree 91 / Partially Agree 4 / Disagree 5. Explanation — Accurate 87 / Minor Issues 6 / Inaccurate 7. Triage — Appropriate 90 / Over-triaged 8 / Under-triaged 2.

Triage safety analysis

Triage is the most safety-critical dimension. Over-triage and under-triage are not equivalent errors and are always reported separately.

Over-triage — 8 cases. The AI recommended a higher urgency than clinically necessary. This may cause an unnecessary specialist referral but poses no patient safety risk; in a screening context it is the preferred error direction.
Under-triage — 2 cases. The AI recommended a lower urgency than appropriate. This is the single most important safety metric. Both cases were documented in detail by the independent reviewer: one urticaria presentation classified as granuloma annulare (noted as requiring emergency attention due to angioedema risk), and one case where the reviewer favoured scalp actinic keratosis over the AI's pediculosis capitis classification.

The 2% under-triage rate and 98% clinically acceptable triage rate confirm that ScanSkinAI consistently errs on the side of caution — the appropriate default behaviour for a screening tool.

Fitzpatrick-stratified results

ScanSkinAI is validated on all six Fitzpatrick skin types. The Phase 1 sample is weighted toward Types III–IV (51.3% combined) to reflect the Asia-Pacific user demographics. Per-skin-type accuracy figures are being finalised for publication when the stratified analysis is complete; the overall 95.3% clinically acceptable accuracy is measured across all six types combined.

Fitzpatrick skin type coverage in the ScanSkinAI validation dataset.
Fitzpatrick skin type	Description	Cases (n)	% of sample	Clinically acceptable accuracy
Type I	Very fair skin, always burns	7	8.8%	To be published
Type II	Fair skin, burns easily	14	17.5%	To be published
Type III	Medium skin, sometimes burns	21	26.3%	To be published
Type IV	Olive / moderate brown, rarely burns	20	25.0%	To be published
Type V	Dark brown, very rarely burns	13	16.3%	To be published
Type VI	Very dark / black, never burns	5	6.3%	To be published
Total		80	100%

Most consumer AI skin tools do not publish skin-tone-stratified performance data. Publishing per-type figures — including where they are lower than the aggregate — is part of the commitment of this report.

Roadmap — ongoing clinical evidence programme

This Phase 1 audit is the baseline for a quarterly, repeatable evidence generation cycle with increasing sample sizes and reviewer panels.

Clinical validation roadmap across Phase 1 through ongoing monitoring.
Phase	Sample	Timeline	Key enhancement
Phase 1	100 cases	Q1 2026 — complete	Single reviewer; establish baseline
Phase 2	300–500 cases	Q3 2026	Expand to 2+ reviewers for inter-rater reliability
Phase 3	500–1,000 cases	Q4 2026 – Q1 2027	External CRO or academic institution for independent validation
Ongoing	1,000+ per cycle	2027+	Continuous monitoring; multi-site, multi-reviewer

Regulatory positioning

ScanSkinAI is positioned as a wellness/wellbeing screening tool by default and is not a diagnostic device. For compliance-sensitive partners, UKCA Class I medical device quality standards are applied. The platform operates under ISO 27001 (information security) and ISO 13485 (medical device quality management) certified processes.

How this compares with published research

Peer-reviewed meta-analyses of AI in dermatology consistently report that convolutional and transformer-based models can achieve accuracy comparable to board-certified dermatologists in controlled image-classification tasks. See, for example, Liu et al.'s 2019 systematic review of deep learning versus healthcare professionals in medical imaging (PubMed 32908386), and Esteva et al., "Dermatologist-level classification of skin cancer with deep neural networks", Nature, 2017 (PubMed 28117445).

These studies focus primarily on single-disease detection (typically melanoma versus benign naevus) in curated benchmark image sets. The ScanSkinAI report differs in scope: it measures clinically acceptable screening performance across 80 different conditions in a broader triage context, rather than single-disease sensitivity on a curated benchmark. The two kinds of result answer different questions and should not be compared as if they were the same metric.

The wider literature has also documented a persistent gap in performance and dataset representation for darker skin tones. See Adamson & Smith, "Machine learning and health care disparities in dermatology", JAMA Dermatology, 2018 (PubMed 30073279). Reporting Fitzpatrick-stratified results — even when incomplete — is a direct response to that gap.

What this does and does not mean

ScanSkinAI is a screening and monitoring tool. These results describe screening performance, not diagnostic performance.
ScanSkinAI does not diagnose skin cancer or any other disease. It does not replace a dermatologist, dermoscopy, biopsy, or clinical examination.
A high aggregate accuracy figure does not mean any individual result is certain. Users should always seek professional medical advice for concerning, changing, bleeding, non-healing, or otherwise worrying skin changes — regardless of what the AI output says.

Frequently asked questions

How accurate is ScanSkinAI?

In an independent 100-case dermatologist audit, ScanSkinAI achieved 95.3% clinically acceptable accuracy across 80 skin conditions. Clinically acceptable accuracy is the standard metric used in medical AI validation literature and describes screening performance, not diagnostic performance.

Who validated ScanSkinAI's accuracy?

The audit was reviewed independently by Dr. Anand S. Urhekar, MD, a board-certified physician with more than 25 years of clinical experience. Dr. Urhekar reviewed the AI's outputs against ground-truth clinical assessments case by case.

Does ScanSkinAI work on darker skin tones?

Yes. ScanSkinAI is validated across all six Fitzpatrick skin types (I to VI). The Phase 1 audit was stratified as Type I: 7 cases (8.8%), Type II: 14 (17.5%), Type III: 21 (26.3%), Type IV: 20 (25.0%), Type V: 13 (16.3%), and Type VI: 5 (6.3%) — deliberately weighted toward Types III–IV to reflect the Asia-Pacific user demographics. The overall 95.3% clinically acceptable accuracy is measured across all six types combined, not on lighter skin only.

What does 'clinically acceptable accuracy' mean?

It is the percentage of cases where the AI output would not lead to patient harm or a materially incorrect clinical pathway. It is the standard metric in published medical AI validation studies and is more meaningful for a triage and screening tool than raw single-disease sensitivity, because it captures both exact agreement and clinically safe partial agreement (for example, over-triaging rather than under-triaging).

Does a high accuracy mean ScanSkinAI can diagnose skin cancer?

No. ScanSkinAI is a screening and monitoring tool. It does not diagnose skin cancer or any other disease and does not replace a dermatologist, dermoscopy, biopsy, or clinical examination. Any concerning, changing, bleeding, or non-healing skin change should be assessed by a qualified clinician.

Why is under-triage the key safety metric?

Over-triage and under-triage are not equivalent clinical errors. Over-triage recommends a higher urgency than necessary — inconvenient but safe. Under-triage recommends a lower urgency than appropriate — the only triage error that can delay needed care. In the Phase 1 audit, ScanSkinAI produced 8 over-triage cases (conservative behaviour) and only 2 under-triage cases out of 100, for a 2% under-triage rate. Both under-triage cases were documented in detail by the independent reviewer.

How often is this report updated?

This page is refreshed as validation continues and as new audits are completed. The last-updated date at the top of the page reflects the most recent revision.