Aletheia: an offline diagnostic tool for clinical decision-making in sub-Saharan Africa
Researchers described Aletheia, an offline system for differential diagnosis built on a fine-tuned Qwen2.5-3B model, intended for hospitals in sub-Saharan Africa with limited connectivity and hardware.
Researchers introduced Aletheia, a clinical decision support system (differential diagnosis) designed as an offline-first solution for healthcare in sub-Saharan Africa. The motivation is a shortage of specialists in the region, where the doctor-to-patient ratio in some rural areas falls below 1:25 000, while existing AI diagnostic tools typically require a reliable internet connection and powerful hardware, which is impractical for healthcare workers in district hospitals and health centers.
Aletheia is built on Qwen2.5-3B-Instruct, fine-tuned using QLoRA (Quantised Low-Rank Adaptation) on a curated dataset of 27 000 clinical reasoning samples covering 50 diseases with higher prevalence in East Africa. According to the authors, the system achieves diagnostic accuracy of 80.0 % on the first attempt (Top-1) and 100.0 % within the first three attempts (Top-3), BERTScore-F1 of 0.909 and METEOR of 0.467 across ten categories of clinical cases. A calibration error (Expected Calibration Error) of 0.275 is also reported.
According to the authors, the system meets the memory limit of 7 168 MB for the Africa Deep Tech Challenge 2026 (ADTC 2026) competition, with peak RAM usage during inference of approximately 3 630 MB on a standardized test laptop. From this, the authors conclude that deploying clinical reasoning based on large language models at the primary care level in resource-constrained settings without cloud infrastructure is feasible.
The source text is an abstract of a scientific paper on arXiv and contains no information about availability, licensing, pricing or planned deployment of the system in practice.
Why it matters
The study provides a concrete example of how a small language model with modest computational requirements, fine-tuned for a narrow domain, can eliminate the need for a cloud connection where both infrastructure and specialists are lacking – according to the authors, the system runs within the memory budget of a typical laptop. For developers of similar tools, the combination of techniques (QLoRA, quantization, a small base model) and the reported calibration error of 0.275 are particularly relevant, suggesting that the confidence of the model in its predictions is not always reliable and needs to be taken into account when designing the clinical interface.
Relevant practical impact
What this means
For a business
The demonstration shows that a small model (3B parameters) fine-tuned using QLoRA on specialized clinical data can run offline within the memory budget of a typical laptop (approximately 3.6 GB RAM), which is relevant to companies developing edge AI or specialized diagnostic tools for settings without reliable connectivity.
Development More business impacts →Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.