Downloads · 30 days
0
mayinai/Mayin-ICU-WM-v0.1
Mayin-ICU-WM-v0.1 is a machine learning model from mayinai. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Mayin-ICU-WM v0.1 is Mayin AI's first clinical world-model release for learning and forecasting longitudinal patient-state representations from electronic health records.
Downloads · 30 days
0
Access
Public
Updated Sep 3, 2026
Repo size
2.3 MB
Likes
1
Public
Click a slice to open those files.
.png1.8 MB · 99%
From the Hugging Face model README
Mayin-ICU-WM v0.1 is Mayin AI's first clinical world-model release for learning and forecasting longitudinal patient-state representations from electronic health records.
The v0.1 model establishes an ICU-focused latent trajectory modeling core based on the Clin-JEPA formulation, while the Mayin-ICU-WM roadmap extends the framework toward hospital-wide, irregularly sampled EHR trajectories and latent-to-text generation.
Availability. This Hugging Face repository is the public documentation home for Mayin-ICU-WM v0.1. Model weights are not distributed here. A credentialed release through PhysioNet is planned; the access link will be added after publication.
| Version | v0.1 |
| Status | Research release |
| Developed by | Mayin AI |
| Base encoder | Qwen3-8B |
| Domain | Intensive care and longitudinal electronic health records |
| Weights | Planned for a future credentialed PhysioNet release |
Mayin-ICU-WM v0.1 represents a longitudinal ICU stay as admission context followed by hourly patient states and clinical actions. Observations, interventions, and admission context are serialized as readable EHR text and mapped by a shared Qwen-based clinical encoder into a continuous patient-state space. An action-conditioned Transformer predicts future states directly in that latent space, enabling history representations and model-generated future representations to be used together in downstream studies.

Solid components constitute the Mayin-ICU-WM v0.1 release. Dashed components indicate extensions currently under development.
The Mayin-ICU-WM development roadmap expands the v0.1 core in four directions:
Mayin-ICU-WM v0.1 was trained on ICU trajectories derived from MIMIC-IV v3.1. The cohort contains 84,504 ICU stays from 64,879 patients. Train, validation, and test partitions were separated at the patient level, with 59,107, 12,615, and 12,782 stays respectively.
Clinical observations and interventions were organized into hourly windows and serialized into state and action text. No raw MIMIC-IV records, processed patient trajectories, patient-level embeddings, or patient-level predictions are included in this repository.
The v0.1 evaluation examines three complementary properties of the learned patient-state space: downstream predictive utility, stability during pretraining, and sensitivity to clinically meaningful changes in patient trajectories.
The downstream evaluation covers 15 binary prediction tasks in two complementary groups: seven early-event prediction tasks and eight stay-level outcome tasks. One-hidden-layer ReLU MLP probes were evaluated on two representations:
Ridge, LightGBM, LSTM, and TCN models trained on the corresponding raw clinical features provide reference points. Reported values are macro averages across tasks; task-level uncertainty was estimated with 500 stay-clustered bootstrap samples.

| Evaluation group | Representation | Mean AUROC | Mean AUPRC |
|---|---|---|---|
| Early-event prediction, 7 tasks | History only | 0.827 | 0.407 |
| Early-event prediction, 7 tasks | History + rollout | 0.835 | 0.413 |
| Stay-level outcomes, 8 tasks | History only | 0.868 | 0.521 |
| Stay-level outcomes, 8 tasks | History + rollout | 0.886 | 0.576 |
Rollout-enhanced representations improve the aggregate results in both evaluation groups. On the eight stay-level outcomes, Mayin-ICU-WM v0.1 reaches 0.886 mean AUROC and 0.576 mean AUPRC.

Mayin-ICU-WM v0.1 completed the full pretraining schedule without representation collapse. Its latent-state variability remained well above the collapse threshold at every logged evaluation and stabilized at z<sub>std</sub> = 0.2548, while the random-mask and no-warmup comparison variants collapsed early. These results demonstrate stable representation learning throughout the complete training run. For context, the comparison curves follow Figure 2b of Clin-JEPA.

Latent-geometry analysis on 100 held-out 72-hour ICU stays. (a) Individual trajectories and cohort centroids; (b) admission and hour-71 distributions, with arrows showing cohort-mean movement; (c) admission-to-end displacement for SOFA-stable and SOFA-rising stays.
Mayin-ICU-WM v0.1's latent space captures differences in clinical progression. Across 50 SOFA-stable and 50 SOFA-rising ICU stays, clinically stable trajectories remained closer to their admission state, while trajectories with worsening organ dysfunction moved farther over the 72-hour window. The difference was statistically significant (one-sided Mann–Whitney U = 1,670; p = 0.0019; Cohen's d = 0.571) and remained directionally consistent across nine tested UMAP settings.
The Mayin-ICU-WM model artifacts are derived from MIMIC-IV and are not hosted on Hugging Face. A credentialed PhysioNet release is planned. Until that release is published, this repository provides the model description, architecture, and aggregate evaluation results only.
Researchers seeking to work with MIMIC-IV must independently complete the applicable PhysioNet credentialing, training, and data-use requirements.
We thank the Clin-JEPA authors for developing and documenting the clinical joint-embedding predictive framework, the Qwen team for the base language model, and the MIMIC-IV and PhysioNet teams for the clinical research infrastructure that enabled this work.