From 0 to 193 Logins in 88 Days: Evidence‑Driven AI Empowers 5 Million Chinese Doctors

Chinese physicians work under extreme pressure, handling millions of consultations annually, yet they struggle to keep up with rapidly evolving medical evidence. Conventional AI models often hallucinate—fabricating references or providing incorrect DOI links—making them unsafe for clinical decision‑making.

A recent study published on medRxiv showed that adding Retrieval‑Augmented Generation (RAG) to a large language model increased the unsupported‑claim (hallucination) rate from a baseline of 5.0% to a staggering 43.6%, an 8.7‑fold rise in factual errors.

To address this, Alibaba Health introduced Hydrogen Ion , a medical AI product built on a four‑layer evidence architecture:

Layer 1 – Evidence Understanding: Medical texts are transformed into structured, traceable evidence units using the PICO framework (Population, Intervention, Comparison, Outcome) and the GRADE system for confidence grading.

Layer 2 – Structured Retrieval: Instead of keyword matching, the system searches for structured PICO elements, turning queries like “Can ibuprofen lower fever faster than acetaminophen in children?” into a formal evidence question, dramatically reducing retrieval failures.

Layer 3 – Model Alignment: Reward modeling and rubric‑based scoring teach the AI to generate low‑hallucination, source‑cited answers that meet clinical quality standards.

Layer 4 – Expert‑in‑the‑Loop: Over 300 senior physicians continuously review, score, and correct AI outputs, feeding back into the training loop to improve evidence extraction, retrieval logic, and rubric strictness.

Hydrogen Ion integrates massive data sources—including PubMed, Google Scholar, thousands of Chinese and international journals, 30 000+ clinical guidelines, and more than 60 000 drug monographs—ensuring up‑to‑date, authoritative references. In internal testing, a senior emergency physician queried the system for a precise ticagrelor dose adjustment; the AI returned the correct 180 mg loading dose and 90 mg bid maintenance dose, linking directly to the 2025 Chinese Society of Cardiology guideline PDF.

Speed is critical in emergency settings. The same physician reported that the AI produced a complete, citation‑rich answer in just three seconds, compared to the 10‑20 minutes required to switch between PubMed, guideline PDFs, and drug labels.

Beyond speed, the system emphasizes timeliness and authority: each answer is tagged with the publication year (e.g., 2025) and the evidence level (e.g., RCT, meta‑analysis), allowing doctors to verify that the recommendation is both recent and high‑quality.

Hydrogen Ion’s performance metrics—193 logins over 88 days and a reported 2‑3× lower hallucination rate than domestic competitors—demonstrate the practical impact of the evidence‑centric design. The product’s success illustrates that in medical AI, the true competitive moat lies not in model size but in the end‑to‑end pipeline that transforms high‑grade evidence into trustworthy clinical answers.