In the pursuit of global health equity, artificial intelligence (AI) stands at a transformative crossroads. While AI-powered diagnostics have shown immense promise in high-income healthcare systems, the question remains: how can we design these tools to deliver impact where resources are limited and clinical needs are greatest?

Building diagnostic AI for low- and middle-income countries (LMICs) is no longer just an ethical imperative; rather it’s a strategic frontier for global medicine. From domain adaptation and computational efficiency to local partnerships and hardware design, innovation must go beyond performance metrics to prioritize accessibility, robustness, and relevance.

Training AI When Data Is Scarce

A foundational challenge in LMIC settings is the lack of large, well-annotated datasets. AI models developed on high-income populations often struggle to generalize to different demographics, imaging protocols, or disease presentations found elsewhere.

Domain adaptation techniques which fine-tune existing models on small target datasets, and few-shot learning which enables models to learn from very limited examples are increasingly being explored to bridge this gap. These techniques help tailor diagnostic algorithms to local populations without requiring large-scale retraining or data collection, enabling faster and more inclusive deployment.

Infrastructure Matters: Cloud vs. Edge Computing

A major consideration for diagnostic AI in LMICs is infrastructure variability. Some regions have intermittent internet access or limited bandwidth, making cloud-based diagnostics infeasible or insecure for patient data.

In response, innovators are turning to edge computing, where AI models are embedded directly into portable or offline-capable devices such as smartphones, USB sticks, or point-of-care microscopes. These lightweight deployments ensure that diagnostic insights can be delivered on-site, without reliance on remote servers.

By combining on-device inference with intermittent cloud syncing, hybrid models can also maintain up-to-date performance while functioning in disconnected environments.

Mobile and Lightweight AI Tools

Deploying AI in rural clinics, community hospitals, or mobile screening units requires hardware-aware AI models. These models must:

• Run efficiently on low-power devices (e.g., smartphones or Raspberry Pi boards)
• Tolerate noisy or low-resolution data inputs
• Operate with minimal user interaction or training

Projects such as AI-powered smartphone-based microscopy or offline pathology slide classification tools are already proving that mobile-first design can make advanced diagnostics viable in areas with few specialists.

From Co-Development to Capacity Building

Beyond technology, success hinges on deep collaboration with local institutions, healthcare workers, and NGOs. These partnerships ensure that AI tools:

• Are clinically and culturally adapted to local needs
• Support, rather than disrupt, existing clinical workflows
• Contribute to local research capacity and technical upskilling

Initiatives that pair global AI teams with regional medical institutions not only improve model relevance but also help establish a sustainable ecosystem for innovation, where local stakeholders are not just end-users but co-creators.

Toward Equitable Diagnostics

Designing diagnostic AI for LMICs is not about creating “lighter” versions of tools meant for wealthy systems. It’s about rethinking performance, interoperability, and human impact from the ground up. It’s about ensuring that the benefits of AI in healthcare are not gated by geography, bandwidth, or data privilege.

The shift toward inclusive diagnostics marks a critical step toward global health equity, one where AI doesn’t just predict disease, but helps close the gap between need and care.

From Local Deployment to Global Inclusion

If AI is to fulfill its promise in healthcare, it must not stop at technological innovation, rather it must drive representation, inclusion, and equity. The success of AI-powered diagnostics in low-resource settings is not just a technical challenge; it’s a global justice issue.

That’s why efforts like the Remaining84 campaign are critical.

Only 16% of data used in cancer research and diagnostics comes from non-Western, non-white, and non-urban populations. This means 84% of the world is underrepresented in the development of AI-driven health solutions, despite often facing the highest disease burden and lowest access to care.

PAICON’s Remaining84 campaign is a call to address this imbalance by building inclusive datasets, fostering partnerships in LMICs, and deploying AI tools where they can have the widest and most equitable impact.

🔗 Explore more: www.remaining84.com

References

Ngiam KY, Khor IW. Big data and machine learning algorithms for health-care delivery. Lancet Oncol. 2019;20(5):e262–e273. https://doi.org/10.1016/S1470-2045(19)30149-4

Rajpurkar P, Chen E, Banerjee O, Topol EJ. AI in health and medicine. Nat Med. 2022;28:31–38. https://doi.org/10.1038/s41591-021-01614-0

Howard FM, Dolezal J, Kochanny S, et al. The impact of site-specific digital histology variability on deep learning model accuracy and bias. Nat Commun. 2021;12:4423. https://doi.org/10.1038/s41467-021-24698-1

Xu, J., Glicksberg, B.S., Su, C. et al. Federated Learning for Healthcare Informatics. J Healthc Inform Res 5, 1–19 (2021). https://doi.org/10.1007/s41666-020-00082-4