Clinical evidence portfolio

1. Executive summary

Autoderm is an AI-powered dermatology decision support API that analyses smartphone images to screen for 70+ skin conditions, returning results in under 1 second. The software has been researched since 2018 and has processed over 2 million API calls across deployments with Boots UK, DocMorris Germany, myGP UK, and others. This document summarises Autoderm’s complete clinical evidence portfolio across nine categories: peer-reviewed performance studies, white paper performance studies, exploratory diversity evaluations, human-AI reader studies, real-world safety data, algorithm progression, demographic evidence, regulatory compliance, and evidence strength assessment.

Intended use

Intended users

Overall assessment

Autoderm’s clinical evidence portfolio is sufficient to transition from MDD Class I to MDR Class IIa. The combination of five peer-reviewed publications across five countries, five white papers providing reader studies, standalone performance validation, bias evaluation, and real-world deployment data, 2M+ API calls with zero adverse events, and FDA Breakthrough Device designation places Autoderm among the best-evidenced AI dermatology devices at Class IIa level.This document summarises Autoderm’s complete clinical evidence portfolio across nine categories: peer-reviewed performance studies, white paper performance studies, exploratory diversity evaluations, human-AI reader studies, real-world safety data, algorithm progression, demographic evidence, regulatory compliance, and evidence strength assessment.

Regulatory Note on White Papers

The Boots UK GP Reader Study (listed as white paper #6 below) is the same study as the Boots UK PMCF study described in Section 9.2. It is listed in both contexts because it serves dual purposes: as standalone clinical evidence and as a formal PMCF obligation under the Quality Management System.

2. Peer-reviewed publications

The Boots UK GP Reader Study (listed as white paper #6 below) is the same study as the Boots UK PMCF study described in Section 9.2. It is listed in both contexts because it serves dual purposes: as standalone clinical evidence and as a formal PMCF obligation under the Quality Management System.

2.1 Standalone AI performance studies

3. Exploratory diversity evaluation

The following study was conducted as an external exploratory evaluation on Fitzpatrick Skin Type VI (African dark skin) populations using V0.1 of the algorithm. It is classified separately from the peer-reviewed performance studies due to important methodological differences in how accuracy was measured.

Methodological note on kamulegeya et al. 2023

This study used Autoderm’s distribution scores (per-condition similarity scores that do not sum to 100%) as the basis for ranking diagnoses. The scoring system recorded the rank position (1–5) of the correct diagnosis in the top-5 output list. The resulting 17% figure is therefore a top-5 retrieval rate, the proportion of cases where the correct diagnosis appeared anywhere in the ranked output, not a standard sensitivity or specificity metric. This methodology differs fundamentally from the top-1/top-3 sensitivity metrics used in Zhu 2023, Lu 2022, and the accuracy metrics in Zaar 2020. Figures from this study should not be directly compared to those from other studies in this portfolio without this caveat clearly stated. The paper’s reference to ‘69.9% on Caucasian skin’ was Autoderm’s own stated training performance figure used as a benchmark, it was not a finding of this study.

Strategic Value

Despite these methodological limitations, this study holds significant strategic and regulatory value. Together with Zaar 2020, it provides the only direct same-model (V0.1) cross-population comparison in the portfolio: 56.4% top-5 retrieval (Sweden, FST I-II) versus 17% top-5 retrieval (Uganda, FST VI). This contrast directly motivated the Fitzpatrick17K bias evaluation and subsequent dataset diversification in V2.x, forming a compelling ‘identified problem → implemented solution → validated improvement’ narrative for regulators and enterprise partners.

3.5 Standalone AI performance: White papers

Why White Papers?

The pace of AI model development frequently outstrips the peer-review publication cycle, which typically takes 12–24 months from submission to publication. By the time a study is published, the model under evaluation may already have been superseded by a newer version with improved performance. Autoderm therefore supplements its peer-reviewed evidence base with white papers that provide timely performance data on current or recent model versions, while peer-reviewed publications for the same studies are pursued in parallel where appropriate.

4. Human-AI reader studies

Three reader studies demonstrate that clinicians and students using Autoderm as a decision support tool consistently outperform those working without AI assistance. All three studies show the same directional result: AI + clinician > clinician alone.

4. Human-AI reader studies

Three reader studies demonstrate that clinicians and students using Autoderm as a decision support tool consistently outperform those working without AI assistance. All three studies show the same directional result: AI + clinician > clinician alone.

5. Aggregate performance summary

The table below synthesises key performance metrics across all comparable studies. Note that direct numerical comparisons should be interpreted with caution given differences in study design, population, conditions evaluated, and model versions tested.

Regulatory framing: Escalé-Besa 2023 dual analysis

The Escalé-Besa 2023 study contains two complementary analyses that serve distinct regulatory arguments. The all-patients analysis (n=100) demonstrates real-world robustness: even when presented with 18 conditions outside its trained scope, Autoderm maintained 98% specificity and did not generate dangerous false positives. The in-scope analysis (n=82) reflects true clinical performance within the intended use, conditions the algorithm was designed to screen, where Autoderm matched GPs at Top-3 (AI 75% vs GP 76%) and matched dermatologists at Top-3 when assessed at Top-5 (AI 89% vs dermatologist 90%). For FDA and MDR submissions, the n=82 analysis is the primary clinical performance benchmark; the all-100 analysis supports the safety and robustness argument. Both were conducted on V2.0; V2.3.0 is expected to perform better.

6. Real-world safety record

Autoderm has operated continuously since 2018. Post-market surveillance data represents the strongest real-world safety evidence among comparable AI dermatology devices at MDR Class IIa level.

The device is intended for use as a clinical decision support tool (CDST) under healthcare worker oversight, not as a standalone diagnostic system. All output is presented as a ranked list of possible skin conditions rather than definitive diagnoses, mitigating risk of over-reliance. Residual risks include false negatives for serious conditions (e.g. melanoma ~5% miss rate in top-5) and image quality dependency, both addressed in the Instructions for Use (IFU V3.11).

7. Algorithm progression

Autoderm’s algorithm has undergone documented continuous development from V0.1 (2017 MVP) to V2.3.0 (current production version). Each major version introduced expanded training data, improved condition coverage, and measurable accuracy gains validated by external studies.

Training data note: The current V2.3.0 model is trained on 150,000+ dermatologist-annotated images from a proprietary dataset of 3M+ real-world smartphone images (100K new images per month). Annotations follow a consensus protocol requiring agreement from at least 2 of 3 independent board-certified dermatologists per image. Expanded training with ~220,680 images including augmentation has also been completed.

8. Demographic & diversity evidence

Autoderm’s portfolio includes validation data across five Fitzpatrick skin type categories (FST I-II, III-IV, VI) and six geographies (Sweden, China, Spain, UK, Uganda, USA), providing unusually broad demographic coverage for a Class IIa SaMD.

Outstanding gap: An absolute per-FST-type sensitivity/specificity table for V2.3.0 remains desirable for the CER but is not a blocking requirement given the Fitzpatrick17K evaluation already completed.

9. Post-market clinical follow-up (PMCF)

Autoderm has an established PMCF programme under its ISO 13485:2021 Quality Management System, comprising three completed formal PMCF studies, an ongoing PMS vigilance monitoring framework, and a pipeline of additional planned studies.

9.1 PMCF Study 1 - Visiba Care (September 2021)

Performance Results (Real-World, n=1,092)

Notable per-condition performance (Top-5 recall)

Performance context

Pre-V0.1 closed-set testing (Visiba 2019 contract) documented 49.3% top-1, 70.1% top-3, 81.7% top-5. The real-world V2.0 figures (~42% top-1, ~75% top-5) reflect open-set clinical conditions where the correct diagnosis may fall outside the 44 evaluated conditions. The difference between closed-set and real-world performance is well-characterised in the AI/ML literature and expected.

Calibration analysis

A calibration plot comparing raw model output against a calibrated version demonstrated that raw model outputs over-predict (observed probability consistently exceeds predicted probability). This analysis formally supports the classification of Autoderm’s output percentages as distribution scores across ranked skin disease classes, not calibrated probability estimates. This framing is critical for both MDR CER accuracy and FDA De Novo CDS classification.

Healthcare professional feedback

GPs using Autoderm within Visiba Care reported satisfaction with diagnostic accuracy. No complaints about missing skin diseases were received. GPs requested addition of 7 new conditions (predominantly paediatric): Chickenpox (B01.9), Scarlet fever (A38.9), Hand/foot/mouth disease (B08.4), Measles (B05.9), Viral exanthema (L27.1), Fifth disease (B08.3), and Erysipelas (A46.9).

PMCF conclusion

The clinical data confirmed the overall benefit-risk profile of Autoderm V2.0 was satisfactory. No patient safety concerns were identified. No changes to the PMCF plan were deemed necessary.

9.2 PMCF study 2: Boots UK GP survey (November 2024)

Performance results

User acceptance (Study group)

User acceptance (Study group)

The study confirms Autoderm’s intended use as a clinical decision support tool improves primary care diagnostic performance and workflow efficiency in real-world UK general practice settings.

9.3 Post-market surveillance report: myGP UK (May 2024)

Disease Distribution (May 2024)

Top conditions flagged: Seborrhoeic Keratosis (350 cases), Actinic Keratosis (258), Lentigo (236), Atypical Melanocytic Nevus (205), Basal Cell Carcinoma (202), Dermatofibroma (192). Malignant tumours comprised 8.3% of all results.

Melanoma screening

Autoderm flagged 79 cases as possible melanoma during May 2024 (approximately 2–3 per day). With approximately 46 melanomas diagnosed daily in the UK, this suggests approximately 5% of UK melanoma cases may have been first signposted via the myGP app, demonstrating the potential role of AI screening in raising early awareness at population scale.

Key findings

The disease distribution in the myGP consumer population (35.2% inflammatory, 32.2% benign tumours, 22.9% infectious, 8.3% malignant, 1.4% genital) is consistent with expected population-level skin disease epidemiology. The over-representation of benign tumours such as seborrhoeic keratosis reflects a population of health-conscious users seeking reassurance, directly supporting Autoderm’s intended use as a signposting tool that builds patient awareness and directs them to appropriate care.

9.4 User testimonials

9.5 PMS vigilance monitoring framework

Adverse event database searches

Section 11 of the CER summarises adverse event searches across: MHRA Medical Device Alerts and Field Safety Notices (UK); BfArM (Germany); FDA MAUDE database (US); and ClinicalTrials.gov. No alerts related to the safety of Autoderm were identified across any database.

9.6 Planned PMCF studies

10. Regulatory status & certifications

11. Evidence strength assessment

Overall Assessment: SUFFICIENT for MDR Class IIa transition, subject to addressing documented gaps.

The two remaining critical gaps for MDR submission are: (1) intended purpose language reconciliation between IFU V3.11 and the 2022 CER, a documentation decision rather than an evidence deficiency; and (2) the equivalence pathway pivot, Autoderm should proceed on the basis of its own clinical data rather than the weak SkinVision equivalence claim in the 2022 CER. Both gaps are on track for resolution by June 2026 per the MDR Evidence Mapping roadmap.

Appendix: Full publication & study references