Page 163 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

diseases and those that are in the state of outbreaks to provide everyone with a plausible
solution to diagnosis, regardless of race or region. By proactively creating skin-tone-inclusive
diagnostic frameworks for underrepresented and data-biased conditions, EquiDermAI aims
to serve as a critical resource for global health equity and preparedness. 4.1-Healthcare

3 Use case requirements
• REQ-01: It is critical that the diagnostic framework integrates deep-generative
augmentation techniques to expand datasets with diverse skin tones synthetically,
ensuring balanced model performance across all racial and ethnic groups.
• REQ-02: It is critical that the system supports lightweight, quantised AI models optimised
for deployment on low-power edge devices to enable accessibility in remote and
resource-constrained regions.
• REQ-03: It is critical that the system include tools to simulate how emerging diseases
(e.g., monkeypox, measles) present across different skin tones, supporting rapid, inclusive
diagnosis during public health emergencies.
• REQ-04: It is critical that the framework ensures privacy by performing all inference locally
on-device, with no data transmission or centralised aggregation, thereby containing
sensitive health data entirely within the user’s hardware and mitigating exposure risks
without requiring encryption or differential privacy layers.
• REQ-05: It is expected that all model training and evaluation will incorporate fairness
metrics—including subgroup Receiver Operating Characteristic-Area Under the Curve
(ROC-AUC), demographic parity difference, and equality of opportunity, with deployment
thresholds enforcing less than 5% performance deviation across stratified skin tone
groups.

4 Sequence diagram

Figure 22�1: Our System sequence diagram

127