Methodology for the Digital Skin Health Insights Report 2026

Research objectives

1. Describe the distribution of available risk bands across approved scan records.
2. Describe the most frequent preliminary AI-supported output categories within the medical skin-module subset.
3. Describe the scale of lower-frequency output labels grouped into the long-tail category.
4. Explain the limits of interpreting scan-level AI outputs as health or population statistics.
5. Provide transparent information about data scope, privacy and methodological constraints.

No hypotheses were specified and no inferential analyses were performed.

Data source and coverage

The source is routinely collected ScanSkinAI production scan history, summarised into an approved aggregate extract. Data coverage runs from 1 January 2026 to 28 July 2026, extracted on 29 July 2026.

The analysis covers eligible production outputs included in the approved aggregate extract. Model-version-level results are not reported publicly.

Inclusion and exclusion

Included

• Scan records included in the approved aggregate production extract
• Records within the verified data-coverage period
• Risk-band records included in the 50,265-scan risk analysis
• Medical skin-module records included in the 45,270-scan output analysis

Potentially excluded where verified

• Records outside the approved reporting scope
• Non-medical or other module records from the output-category analysis
• Records unavailable in the approved aggregate extract

AI-output definitions

• AI output groups are preliminary model-generated labels.
• They are not necessarily user-selected concerns.
• They are not confirmed by biopsy, pathology or dermatologist review within this analysis.
• A label ending in “indication” signals a preliminary concern output, never a confirmed diagnosis.
• A “normal skin” output does not rule out disease.
• “Not a skin image / unclassified” represents an image without a reportable skin category.
• Approximately 430 lower-frequency labels were grouped into “all other labels”.

Statistical methods

Only descriptive statistics were used: counts, percentages, ordered category tables, bar charts, a stacked risk-band chart, and approximate grouped totals where explicitly marked.

Percentages are calculated as the number of scan records in a category divided by the relevant analysis denominator, multiplied by 100. Risk-band percentages use 50,265 scans; medical skin-module AI-output percentages use 45,270 scans.

Rounding

• Percentages are displayed to one decimal place.
• Approximate source values retain the approximation symbol (≈).
• Displayed percentages may not sum to exactly 100% because of rounding.
• No exact count is derived from a rounded percentage where the original count was not approved.

Methods not used

• Hypothesis tests
• P-values
• Confidence intervals
• Regression
• Causal modelling
• Accuracy calculations
• Sensitivity or specificity calculations
• Population weighting
• Imputation
• Age standardisation
• Geographic adjustment

Handling the long tail

• Approximately 430 distinct lower-frequency labels were combined into one group.
• Combined count: approximately 11,900 scans.
• Combined share: approximately 26% of the medical skin-module dataset.
• Grouping improves readability and avoids an excessively long public table.
• It reduces category-level detail.
• The grouped result should not be interpreted as one clinical category.
• Approximate values are reproduced as supplied and are not presented as exact.

Bias and limitations

1. Selection bias

   Scans arise from people who chose to use a consumer skin app; the distribution reflects that group, not a general population.

2. Self-selection into platform use

   Users often scan because something worries them, which can raise the share of concern-related outputs.

3. Repeated scans

   One person may contribute several scans, so frequent conditions or persistent worries can be over-represented.

4. Unknown unique-user denominator

   Because the number of people is unknown, no per-person rate can be derived from any figure here.

5. Unknown clinical confirmation

   No biopsy, pathology or dermatologist confirmation was linked to these records within this analysis.

6. AI-category misclassification

   Preliminary labels can be wrong in either direction; counts should be read as output categories, not conditions.

7. Model and taxonomy changes

   If more than one product or model version is represented, label availability and thresholds may have shifted during the period.

8. Different analysis subsets

   The two datasets have different denominators, so cross-table arithmetic would produce invalid results.

9. Missing risk-band values

   147 records (0.3%) had no reportable band and are shown separately rather than redistributed.

10. Unclassified images

    305 module scans (0.7%) could not be assigned a reportable skin category, slightly reducing interpretable output volume.

11. Approximate long-tail values

    The grouped long-tail count and share are approximate as supplied and must not be treated as exact.

12. Absence of demographics

    Age, sex, skin type and other characteristics were not approved for public disclosure, so no subgroup interpretation is possible.

13. Absence of geographic information

    No country or region breakdown is published, so findings cannot be attributed to any market.

14. Absence of time-series information

    No monthly or seasonal series is published, so no trend or change over time may be inferred.

15. Absence of clinical outcomes

    Whether any user sought care, or what a clinician concluded, is unknown.

16. Absence of follow-up outside the platform

    Outcomes occurring outside ScanSkinAI are not observable in this dataset.

17. No general-population denominator

    Without a population base, no prevalence or incidence estimate can be produced.

18. No adjustment for campaigns or product changes

    Marketing activity, partner launches and interface changes may influence scan volume and category mix; no adjustment was applied.

19. Possible technical retries or duplicates

    If deduplication rules have not been independently verified, retries could inflate counts marginally.

20. Company ownership

    ScanSkinAI owns the platform and performed the analysis, so the report should not be read as an independent evaluation.

Reproducibility and data availability

The underlying row-level dataset is not publicly available because it may contain sensitive health-related platform information, images, contractual information and proprietary system data.

Publicly available:

• The aggregate tables shown in the report
• This public methodology
• Definitions
• Version history
• The correction process

Version history

Corrections

Substantive corrections increment the version, are explained in this table, preserve the previous change record and update the modification date. Material results are never silently replaced. Requests: info@scanskinai.com.