Article review by: Matt Martin August 2026
AI4Doctor: Clinical large language model centered on electronic medical records.
Paper: Zhuang, Y., Wang, B., Yin, C., Zhang, J. et al. "Clinical large language model centered on electronic medical records." npj Digital Medicine 9, 511 (2026). https://doi.org/10.1038/s41746-026-02509-5. Chinese PLA General Hospital, Beijing Institute of Technology, and NOVA Medical School Lisboa.
This paper introduces AI4Doctor, a 32 billion parameter Chinese language clinical LLM built on the premise that existing medical LLMs over rely on literature and exam style training (Med PaLM 2, GPT 4V Medical, HuatuoGPT and similar) and under use the richer, patient specific reasoning signal locked inside electronic medical records (EMRs). Their fix is an EMR centred training pipeline applied to more than three million de identified, multi centre inpatient records spanning 46 departments at the PLA General Hospital.
The pipeline has four stages. First, continued pretraining on the de identified EMR corpus plus medical literature and textbooks, using a novel "Special Temporal Grid" representation that encodes multi resolution timing, trends and missingness across a patient's hospitalisation (admission through discharge) rather than relying on fixed interval resampling. Second, curriculum based instruction fine tuning organised via a "task cube" crossing NLP ability, specialty and clinical application (discharge summaries, admission records, treatment plans, and so on). Third, clinical alignment via Odds Ratio Preference Optimisation (ORPO), a lightweight reinforcement style method that folds in doctor expert preferences and EMR derived signals (diagnostic priors, risk thresholds) at roughly one tenth the compute cost of full RLHF. Fourth, a hybrid retrieval and inference layer that pulls similar EMR cases and guideline snippets at query time across major clinical domains.
On the results side: senior physicians scored AI4Doctor at 80.52/100 on complex clinical questions across specialties, ahead of Qwen 34B (74.03), HuatuoGPT 34B (73.07), and notably ahead of the senior physicians' own average performance (72.99) and a 6.5 point margin over the next best system. On medical documentation generation (diagnosis plans, admission and postoperative records, discharge summaries) it again led with an average of 67.73. On the public CMMLU Med and C Eval Med benchmarks it scored 81.20 and 70.39, beating every other tested model bar GPT 4, and its smaller 13B variant outperformed larger 34B rivals, which the authors read as evidence that EMR grounded training beats raw parameter scaling. Retrieval augmentation added a further 10.4% lift in consistency ("hit" rate) and 6.0% in fluency; ORPO alignment improved all eight measured clinical quality dimensions (comprehensiveness, risk/side effect awareness, professionalism, accuracy, and so on) by 3 to 10%.
Alongside the model, the authors release 301MedQA, a new benchmark of 2,277 EMR grounded clinical questions (1,200 foundational queries, 176 clinician written and 601 LLM expanded advanced clinical queries, and 300 real world clinic questions) built from the same multi centre EMR repository, intended as a closed book test of episode grounded reasoning rather than exam style recall. On the CMB Clin benchmark, which mixes physician, nurse, pharmacist, technician and graduate exam question sets, AI4Doctor's 55.70 average trailed only GPT 4 (59.46) and beat every other tested model, general or medical, open or closed source, including HuatuoGPT and ChatGLM2 6B by a wide margin.
The authors flag two honest limitations: performance drops on rare conditions and on cases that exceed the model's context window, and they propose knowledge graph reranking, sliding window decoding and multimodal (image plus text) extensions as future work. The study had institutional ethics approval, used retrospective de identified data, and neither the underlying data nor a public model release is available; code is available from the corresponding author on reasonable request.
Is this approach being explored by others?
This is a genuinely active line of research right now, not a one off. Below are two papers that independently converge on the same core idea AI4Doctor is built on, EMR grounded fine tuning of an LLM, optionally layered with retrieval, rather than training only on literature and exam questions.
1. NYUTron (Jiang et al., Nature, 2023)
"Health system-scale language models are all-purpose prediction engines." Jiang, L.Y., Liu, X.C., Nejatian, N.P. et al. Nature 619, 357–362 (2023). https://doi.org/10.1038/s41586-023-06160-y (NYU Langone Health / NYU Center for Data Science; PMID 37286606, according to PubMed)
This is arguably the landmark precedent for AI4Doctor's central thesis. NYU's team pretrained a clinical language model (NYUTron) directly on unstructured hospital EHR notes, rather than general medical text, then fine tuned it across five clinical/operational prediction tasks (30 day readmission, in hospital mortality, comorbidity index, length of stay, insurance denial). It beat traditional structured data models by 5.4 to 14.7 AUC points and was deployed in a prospective single arm trial. The philosophy is identical to AI4Doctor's, that a hospital's own EMR corpus is a better training substrate than public literature, but the application differs: NYUTron targets structured outcome prediction, while AI4Doctor targets open ended clinical dialogue and documentation. It's a strong signal that "train on your own hospital's EMR" is an established, published, and clinically deployed strategy, not something AI4Doctor invented from scratch, though AI4Doctor's temporal grid encoding and instruction tuned conversational format go further.
2. AIDx (Alwakeel, Scientific Reports, 2026)
"AIDx: a locally deployable AI system for physician clinical decision support." Alwakeel, A. Scientific Reports 16 (2026). https://doi.org/10.1038/s41598-026-47470-1 (independent developer; PMID 42277077, according to PubMed)
This one is architecturally the closest match I found. AIDx's core component, AIDx Copilot, is an LLM fine tuned on de identified EHR data and optionally grounded with retrieval augmented generation from open medical references, evaluated on the MultiMedQA suite. The ablation results mirror AI4Doctor's Figure 5 almost exactly in shape: EHR fine tuning drove the large majority of the performance gain (+17.8 percentage points over the base model), while adding retrieval contributed a much smaller additional lift (+0.4 points on average). That's the same pattern AI4Doctor reports, fine tuning on EMR data does the heavy lifting, retrieval is a smaller but real add on for grounding and consistency. AIDx is a smaller, single author, benchmark only study (no clinician evaluation or real world deployment), so it's not in the same weight class as AI4Doctor's multi centre, physician graded evaluation, but its independent arrival at "fine tune on EMR, add RAG as a secondary layer" supports that this is becoming a converging recipe rather than one team's idiosyncratic design choice.
A third data point worth knowing about: OncoPainBot (Liu, Hu, Li et al., npj Digital Medicine, 2026, https://doi.org/10.1038/s41746-026-02362-6), published in the same journal as AI4Doctor, applies a comparable EMR plus RAG recipe to a narrower problem (cancer pain assessment and management), using a multi agent LLM framework validated against 516 real world EMRs. It's narrower in scope than AI4Doctor but reinforces that npj Digital Medicine is currently publishing multiple EMR grounded, retrieval augmented clinical LLM papers in short succession, so the field is clearly moving this direction in parallel, not just at AI4Doctor's institution.
My take: AI4Doctor is not proposing a wholly novel paradigm, EMR centric pretraining and RAG layered fine tuning both have independent precedent (NYUTron for the former, AIDx for the pattern of "fine tuning dominates, RAG adds a smaller lift"). What looks more genuinely novel here is the scale and rigour of the evaluation (senior physicians across specialties, head to head against practising clinicians, a released 2,277 question benchmark) and the Special Temporal Grid method for handling multi resolution, missing longitudinal EMR data, which I didn't see replicated in either comparator. I'd put that novelty claim at around 70% confidence since I haven't done an exhaustive literature search specifically for STG style temporal encoders; if that distinction matters for how you're using this paper, it's worth a more targeted search.
Sources: