Equitable Dermatology, the open science face of Programme 01
A Biopratiq open science programme building PCA driven diagnostic AI that works for every human regardless of skin tone, geography, or access to specialist care.
The Representation Check
One question, one answer, and no account. It reports on a published literature, not on you.
Take the Check ›Dermatological AI has a systemic bias problem built into its foundations and the consequences fall hardest on the communities already least served by healthcare systems.
Less than 10% of dermatology training image datasets adequately represent Fitzpatrick types V and VI, the skin tones of the majority of the global population.
1 dermatologist per 1 million people in many low- and middle-income regions, versus 1 per 30,000 in high-income countries. AI should bridge this gap not widen it.
Delayed diagnosis of pigmented lesions including melanoma in darker skin types is measurably linked to worse clinical outcomes. This is a bias problem with life-or-death stakes.
A photograph has to get through the epidermis twice
Move through Fitzpatrick I to VI
Light goes in, reflects off whatever you are trying to see, and comes back out. Melanin sits in the epidermis, on both legs of that journey, and absorbs part of it each way. This is not a flaw in anyone's skin. It is the physics of taking a photograph, and it is the reason a camera based model gets quietly worse as skin gets darker.
Photons completing the round trip through the epidermis.
A sensor measuring at the skin surface never makes the round trip.
At Fitzpatrick III, roughly 7 of every 12 photons complete the journey. The rest are absorbed in the epidermis before they reach the thing you were trying to photograph, or on the way back out. A model reading that photograph is not being given less accurate information. It is being given less information.
This is the argument for hardware. You cannot post process your way out of a signal that never arrived.
The majority of clinical training datasets are composed predominantly of lighter Fitzpatrick skin types I to III, leaving people with darker skin tones underserved and at risk.
Conditions like melanoma, lupus, and eczema present differently across the full spectrum of human skin. A model trained without this diversity does not fail gracefully. It fails silently, and those failures disproportionately affect communities already facing the greatest barriers to specialist care.
This is not a minor statistical artefact. It is a structural failure at the foundation of modern dermatological AI and it requires a structural response, not a post-hoc patch.
Access is the second axis of inequity. Dermatologists are concentrated in high-income urban centres. In large parts of sub-Saharan Africa, South Asia, and Latin America, the ratio of specialist to patient is a fraction of what clinical guidelines recommend.
For the billions of people who will never see a dermatologist, AI is not a luxury it is the only realistic path to any expert-level skin assessment. That makes the quality and fairness of dermatological AI a public health issue of the first order.
We are building for these contexts first, not as an afterthought.
A PCA-driven multimodal diagnostic architecture designed from first principles to be skin-tone invariant with equity embedded in the mathematics, not bolted on at the end.
We use Principal Component Analysis across the full Fitzpatrick spectrum during feature extraction, ensuring that learned representations capture biologically meaningful variance not spectral artefacts of skin tone.
Rather than retrofitting bias corrections onto existing models, we embed equity into the mathematical structure of the model itself. Fairness is a design principle, not a compliance checkbox.
The model is trained via a federated learning framework, enabling institutions in low- and middle-income countries to contribute training signal without centralising sensitive patient data.
Privacy is architectural, not an afterthought. Data stays local only gradients are shared. This enables participation from institutions that could not transfer patient data under local regulations, dramatically expanding our training distribution.
Dermoscopy, clinical photography, and structured symptom data. Designed for low cost hardware.
Skin-tone stratified PCA normalisation removes spectral bias before any classification occurs.
Federated model with uncertainty quantification the model knows what it does not know.
Offline-capable output designed for non-specialist operators in low-bandwidth environments.
Three non-negotiable commitments that define how we build, deploy, and share everything we create.
All model weights, training protocols, and validation datasets generated by this programme will be published under open-access licences. We believe equitable dermatology requires a commons not proprietary moats.
Peer-reviewed publication and pre-print release are built into our research roadmap at every milestone not as optional extras.
Our deployment architecture prioritises low-resource clinical settings from day one. The model is optimised for constrained compute, intermittent connectivity, and non-specialist operators.
We are seeking clinical validation partners in sub-Saharan Africa and South Asia, and we intend to validate there before any high income market release. Partners will be named here once agreements are signed.
Fairness is a loss function term, not a post-hoc audit. We minimise diagnostic performance disparity across skin tone subgroups during training.
A model that is 95% accurate on average but 70% accurate on darker skin is not 95% accurate it is inequitable. We build to close that gap to zero.
Whether you are a clinical institution, a researcher, a potential federated partner, or a funder there are meaningful ways to contribute to this programme. All research outputs are open access.
