EVENT DOSSIER
On September 7, 2026, an offline first-visit clinical decision-making system named Aletheia was approved at Africa Deep Technology Challenge 2026 (ADTC 2026). Built based on Qwen2.5-3B-Instruct, this system is designed for low-resource medical environments in sub-Saharan Africa. It was fine-tuned using quantitative low-rank adaptation techniques on a dataset containing 27,000 samples, covering 50 diseases with high incidence rates in East Africa. Evaluation results showed that its Top-1 diagnostic accuracy was 80% across ten representative clinical case categories, and the Top-3 accuracy was 100%. The expected calibration error was 0.275. The system’s peak推理 memory usage was approximately 3630 MB, meeting the challenge’s memory budget limit of 7168 MB, demonstrating the ability to deploy clinical reasoning based on large language models in resource-constrained environments without cloud infrastructure.
Aletheia 是一款专为撒哈拉以南非洲低资源医疗环境设计的离线首诊临床决策支持系统，旨在解决当地专家稀缺问题。该系统基于 Qwen2.5-3B-Instruct 构建，利用量化低秩适应（QLoRA）技术在包含 27,000 个样本的 curated 数据集上进行微调，涵盖东非洲高发病率的 50 种疾病。评估显示，其在十个代表性临床案例类别上的 Top-1 诊断准确率为 80%，Top-3 准确率为 100%；系统预期校准误差（ECE）为 0.275，并在标准化基准笔记本电脑上以约 3630 MB 的峰值推理内存通过了非洲深度技术挑战 2026（ADTC 2026）7168 MB 的内存预算限制。这些结果表明，无需云端基础设施即可在资源受限环境中部署基于大语言模型的临床推理。