README.md

---
library_name: keras
tags:
  - healthcare
  - icd-10
  - mortality-prediction
  - readmission-prediction
  - deepset
  - tabular
---

# NRD ICD-10 outcome models

DeepSet + (optional) Transformer models trained on the National Readmission
Database (NRD, 2016–2020) to predict two outcomes from up to 40 ICD-10
diagnosis codes plus demographics (AGE, FEMALE, PAY1, ZIPINC_QRTL).

| Subfolder            | Outcome  | Description                          |
|----------------------|----------|--------------------------------------|
| `mortality_30day/`   | `MOR30`  | 30-day post-discharge mortality      |
| `readmission_30day/` | `REA30`  | 30-day all-cause readmission         |

`encoders/` holds the fitted `LabelEncoder` (ICD-10 → integer IDs) and
`MinMaxScaler` (AGE) used at training time. Inputs must be encoded with the
**same** artifacts at inference, or predictions will be meaningless.

## Loading a model

The `.keras` files contain three custom serializable components
(`DeepSet`, `TransformerBlock`, `F2Score`) that must be importable (and
registered via `@tf.keras.utils.register_keras_serializable(package="Custom")`)
before `load_model`:

```python
import tensorflow as tf
from huggingface_hub import hf_hub_download
# Register your custom classes — see src/train/ in the source repo
from custom_layers import DeepSet, TransformerBlock, F2Score  # noqa: F401

path = hf_hub_download(
    repo_id="<user-or-org>/<repo-name>",
    filename="mortality_30day/mort_hypertrial_auc.keras",
)
model = tf.keras.models.load_model(path)
```

## Variants

Within each outcome subfolder, file suffixes denote the architecture:

- `_hypertrial_auc` — best model from the Keras-Tuner search (recommended)
- `_icd_only` — ICD codes only, no demographics (ablation)
- `_no_deepset` — flattened ICD input, no DeepSet aggregation (ablation)
- `_with_transformers` / `_transformer` — DeepSet + TransformerBlocks

## Data restrictions

NRD is a HCUP product distributed under a Data Use Agreement. These weights
do not contain individual records, but downstream users should be aware of
the source.