You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Terms of access — Vivax Clinical World Model (RAIL-M-NC-Research licence)
- Access is granted solely for non-commercial academic or scientific research, teaching, method development, reproducibility checks and privacy/safety audits of the model itself. Any commercial use, including commercial research and development, requires a separate written agreement with the Licensor.
- No clinical use. The model, derivatives and outputs may not be used to inform, support, recommend or make decisions about the diagnosis, prognosis, monitoring, treatment, medication, dosing, triage or discharge of any person, nor as a component of any clinical decision support, monitoring or alerting system, in any care setting.
- Data protection. The model was trained on pseudonymized intensive-care records under institutional approval. You will not attempt to re-identify individuals, to extract training records or memorized sequences, or to link outputs to identifiable persons, and you will report suspected memorization or privacy findings to the Licensor before publication.
- Redistribution and derivatives must carry this licence unchanged, bind recipients to the same restrictions, retain the model card's limitation and provenance statements, and identify changes. Restrictions may not be removed by relicensing.
- Attribution. Publications and software must cite the model card and accompanying paper, state the version used (SHA-256 manifest) and that the model is not clinically validated.
- The model is provided "as is" without warranty; the Licensor is not liable for damages arising from its use. Nothing here is medical advice.
- Access terminates automatically on breach. The Licensor may revoke access and may publish updated versions under their own terms.
- Governing law: Republic of Türkiye. The full licence text is in LICENSE.md and prevails over this summary.
By requesting access you confirm that you have read LICENSE.md and agree to these terms.
Log in or Sign Up to review the conditions and access this model content.
Vivax Clinical World Model — Phase A research freeze (corrected-v5 ensemble)
Frozen configuration acudx-wm-phase-a-20261021: three-seed Mamba ensemble, 8.19 M
parameters per member, trained on deduplicated v5 features of a single-institution adult
ICU cohort. This is a research freeze, not a clinically validated model. It is
released so that the method, its known defects and the later corrected release (v6) can be
inspected and reproduced. Do not use it for patient care.
Repository:
getVivax/cwm-icu-v5. Hashes below were verified against the uploaded files on 2026-11-06.
What this model is
A probabilistic forecaster of observed adult-ICU physiology under documented care. Given an episode prefix (hourly-binned observations, plans and administrations with availability timing, plus static context), it forecasts observed values of 21 physiological concepts at 1, 4, 12, 24, 48 and 72 hours as predictive samples. The primary forecast is HOLD-LAST (hold the last complete hour's care vector; historical insulin compartment decays; no future doses supplied). OBSERVED-CONDITIONAL additionally uses recorded future care and answers a different question. Neither is a learned care policy.
Architecture (per member): four Mamba layers (h = 512, state 16, conv 4, expand 2);
256-d observation encoder over values, masks and time-since-availability; stochastic
latent state of nine organ/care blocks × 8 + 64 residual dims (136 total); linear_h
decoder with organ masking; learned concept-specific carry decay; a bounded exclusive
insulin mechanism (fixed τ = 2.5 h, strength 0.35 in transformed-glucose units, learned
modulation bounded in [1/3, 3]). Ensemble = equal mixture of member draws (300 draws).
A selected GCS integer/spread postprocessor (horizon scales 1/1/1/1/1.1/1.1, support 3–15)
is part of the frozen inference configuration. Structured A2 pathways are disabled.
Files
| File | Content | SHA-256 |
|---|---|---|
ssm_v6_s0/best.pt |
seed 0 state dict, step 24,000 | 92cee369b0d11ad00ddc148dfa0a83e87e243cbc25f943ab64914e2e08e1569d |
ssm_v6_s0/config.json |
frozen architecture/training config | 710c041a9c3756514f4ce6cb1d1b3b88f098022436e1c2210c1c19971ef76406 |
ssm_v6_s0/norm_stats.parquet |
frozen TRAIN normalization per channel | 035e11e663620aff0e5c35f01cccc235f6f032ee67db5335810017c0abe04ddd |
ssm_v6_s0/vocab.parquet |
134 concepts (name, kind, aggregate counts) | 106febe6712c86c49a1bc5ab49f92ef7de0345b175d8923557a84ac1dc0463c9 |
ssm_v6_s1/best.pt |
seed 1 state dict, step 25,000 | 74da1dfa31d4e7939cccef192be57bff15d6a187085be007930f48d1e96b7ba2 |
ssm_v6_s1/config.json |
frozen architecture/training config | 0a405a1313c9639950619f624148455be416d73377797ce9feb0df3965e62566 |
ssm_v6_s1/norm_stats.parquet |
frozen TRAIN normalization per channel | 035e11e663620aff0e5c35f01cccc235f6f032ee67db5335810017c0abe04ddd |
ssm_v6_s1/vocab.parquet |
134 concepts (name, kind, aggregate counts) | 106febe6712c86c49a1bc5ab49f92ef7de0345b175d8923557a84ac1dc0463c9 |
ssm_v6_s2/best.pt |
seed 2 state dict, step 19,000 | f899d702fdb76f4979cc4eb3e23efb9cc4cccfa178ca27245f78057f12273b69 |
ssm_v6_s2/config.json |
frozen architecture/training config | 97082fc768259b24fa42dd70b0b3a8a9a54b66726528e4f22a51a089d4f63ce0 |
ssm_v6_s2/norm_stats.parquet |
frozen TRAIN normalization per channel | 035e11e663620aff0e5c35f01cccc235f6f032ee67db5335810017c0abe04ddd |
ssm_v6_s2/vocab.parquet |
134 concepts (name, kind, aggregate counts) | 106febe6712c86c49a1bc5ab49f92ef7de0345b175d8923557a84ac1dc0463c9 |
frozen_phase_a.json |
complete frozen configuration and artifact hashes | 32f512eb12cf8cf39742e258519561915091e1966ba19f6987ff745d1d8911a6 |
LICENSE.md |
Vivax Clinical World Model RAIL-M-NC-Research licence v1.0 | ec462fcbf1835f0e3638808a27a278ac75ec420767d7e8045de172a72b133c8a |
Run directories are named ssm_v6_s* for historical reasons; the feature snapshot is
deduplicated v5. Each best.pt holds {"model": state_dict, "step": int, "val": {loss, nll, rec, kl, overshoot}} and nothing else (pre-release audit 2026-11-06: no
strings, identifiers, samples or paths; side files are concept-level aggregates only).
Training data (not released)
Pseudonymized HIS extract of one tertiary-hospital adult ICU service, Turkey. Frozen patient-disjoint temporal split: TRAIN 2015–2023 (45,479 episode metadata rows, 34,257 patients; 43,419 episodes with usable sequences), VAL 2024 (8,741 episodes, 7,106 patients; 8,445 usable), TEST reserved and closed. 37.3 M events, 134 concepts; 84 observation channels, 49 plan-class channels, 6 administration/intake channels. Raw data, ID maps and clinical notes never leave the institution and are not part of this release. Pseudonymization preserves clinical timing and is not a guarantee of anonymity; hence the licence's re-identification and extraction prohibitions.
Evaluation (development VAL, 1,500 episodes, 452,189 forecast slots, 1,021,239 targets)
Uncalibrated HOLD-LAST macro CRPS skill over 126 concept × horizon cells, equal weights, versus the persistence reference; 1,000 patient-bootstrap resamples for intervals.
| Model | Draws | HOLD-LAST skill | Coverage50 | Coverage90 |
|---|---|---|---|---|
| seed 0 | 100 | 0.1443 | 0.571 | 0.891 |
| seed 1 | 100 | 0.1399 | 0.582 | 0.894 |
| seed 2 | 100 | 0.1515 | 0.583 | 0.897 |
| pooled ensemble | 300 | 0.1592 | 0.591 | 0.909 |
| matched-count mixture | 100 | 0.1520 | 0.583 | 0.901 |
The prespecified Phase A1 gate (pooled skill ≥ +0.1469 and all insulin T1 criteria in every seed) passed. The 300-draw score carries a smaller finite-sample CRPS bias than 100-draw members; comparisons across draw counts must say so. Insulin T1 checks (242 isolated-dose anchors) show a negative 1/2/4 h glucose response with dose-response ρ ≈ 0.97 and ≈ 5–6 mg/dL per IU at 2 h: these are responses of the supplied mechanism, not treatment-effect estimates. Full tables, T2 ablations and T3 plan contrasts are in the repository model card.
Known defects and limitations (read before use)
- Single institution, one VAL year. No evidence of transportability across centers, years, units or practice changes. No held-out TEST comparison, prospective study or completed clinician adjudication.
- Outcome-label defect (post-ICU). A coordinator review found ICU deaths misclassified as ward deaths; all historical ward-death/composite claims are retired until outcome rule v1.2 is applied. The physiology forecaster does not use those labels, but ICU-boundary accuracy has not been independently revalidated.
- Training/evaluation timing mismatch. Training bins administrations by clinical time without the evaluation adapter's availability gate; random crops reset context; evaluation prefixes can exceed the loader's 14-day cap. Causal and long-context claims are limited.
- Entangled care and severity. The bounded insulin mechanism and supplied scenarios do not identify counterfactual responses, benefit or harm. Latent block names are supplied, not verified physiological states.
- Observed-only targets. Informative missingness, report-time proxies, copy-forward and clipping persist; charted MAP is excluded as an HIS artifact. Aggregate coverage can hide concept-, horizon- and subgroup-specific failures.
- Defects corrected after this freeze (v6 line). Later protocol work found: an unconstrained prefix-preparation cost class in the pipeline; fold-roster handling that must exclude a separate calibration role; ten included TRAIN episodes with zero-length ICU intervals; and the need for availability-gated training. These do not change the frozen v5 weights but motivate the forthcoming v6 release, which will be published alongside this one so that the correction history is inspectable.
Intended use
Non-commercial research only: method comparison, representation and mechanism studies,
reproducibility checks, privacy and safety audits of the model itself. Not for clinical
decision support, dosing, triage, prognosis for identifiable persons, or any deployment.
See LICENSE.md (RAIL-M-NC-Research, Attachment A).
How to load
import torch, json
cfg = json.load(open("ssm_v6_s0/config.json"))
ck = torch.load("ssm_v6_s0/best.pt", map_location="cpu")
state = ck["model"] # 65 tensors, 8,187,463 parameters
Model construction requires the acudx.worldmodel code at the pinned source archive
(hashes in frozen_phase_a.json; Torch 2.11, mamba-ssm 2.3.2.post1, causal-conv1d 1.7.0).
Inputs must follow the v5 feature contract (vocab.parquet, norm_stats.parquet, frozen
transforms and clipping). Code release and exact loading instructions will accompany the
paper preprint.
Citation
Paper in preparation. Until then cite this model card and the configuration tag
acudx-wm-phase-a-20261021.
Contact
Vivax — Hugging Face getVivax. Use the repository's discussion tab for access or licence questions, or the contact e-mail in the accompanying paper.