---
license: cc-by-4.0
pretty_name: spectf
task_categories:
- tabular-classification
tags:
- tabular
- biomedical
- benchmark
- TabBioMed
- clinical-&-healthcare
- binary
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: fold_0_train
    path: data/fold_0/train.parquet
  - split: fold_0_validation
    path: data/fold_0/val.parquet
  - split: fold_0_test
    path: data/fold_0/test.parquet
  - split: fold_1_train
    path: data/fold_1/train.parquet
  - split: fold_1_validation
    path: data/fold_1/val.parquet
  - split: fold_1_test
    path: data/fold_1/test.parquet
  - split: fold_2_train
    path: data/fold_2/train.parquet
  - split: fold_2_validation
    path: data/fold_2/val.parquet
  - split: fold_2_test
    path: data/fold_2/test.parquet
  - split: fold_3_train
    path: data/fold_3/train.parquet
  - split: fold_3_validation
    path: data/fold_3/val.parquet
  - split: fold_3_test
    path: data/fold_3/test.parquet
  - split: fold_4_train
    path: data/fold_4/train.parquet
  - split: fold_4_validation
    path: data/fold_4/val.parquet
  - split: fold_4_test
    path: data/fold_4/test.parquet
---

# spectf

Part of **TabBioMed** (*TabBioMed: A Large-Scale Benchmark for Biomedical Tabular Learning*), a benchmark of 94 curated biomedical
tabular datasets. Code, preprocessing and documentation: https://github.com/AI-sandbox/TabBioMed.

Normal (0; 55) versus abnormal (1; 212) cardiac SPECT, using 44 supplied rest/stress image summaries per patient. Remove 82 redundant PMLB copies; the 267 retained records exactly match the corrected UCI train/test union. Pool that cohort for benchmark CV rather than using the UCI 80/187 split. Use five stratified outer folds and three inner folds; fit imputation, encoding and scaling within training partitions.

| Domain | Task | Samples | Features | Classes |
|---|---|---|---|---|
| Clinical & Healthcare | binary | 267 | 44 | 2 |

**Splits.** 5 stored outer cross-validation folds (`data/fold_<i>/`), each with train/val/test files (fold 0: 186/27/54 rows). Every row appears in exactly one outer test fold. HPO in the paper uses inner folds drawn from each outer non-test pool.

## Files

- `data/**.parquet`: one row per sample. `row_idx` is the row index in the original
  TabBioMed array; `f_<i>[_<name>]` are the features (in order); `target` (or `target_<j>`
  for multi-target tasks) is the label. Categorical feature indices (0-based, among the
  features): `[]`. Missing values are stored as nulls.
- `metadata.json`: TabBioMed task metadata (task type, categorical indices, source, description).
- `feature_names.json`: original feature names, in column order.
- `sample_metadata.csv`: per-row provenance (identifiers, source annotations); not features.

```python
from datasets import load_dataset
ds = load_dataset("TabBioMed/spectf")
```

A Croissant (1.1) description with Responsible-AI fields is distributed with the
TabBioMed release.

## Source and license

Source: PMLB (spectf); verified against original UCI release

This derived dataset is distributed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Please also
cite and respect the terms of the original source.

## Citation

```bibtex
@inproceedings{tabbiomed2026,
  title     = {TabBioMed: A Large-Scale Benchmark for Biomedical Tabular Learning},
  author    = {Mateo Bernad{\'o}, Pau and Polapragada, Saivenkata Nagavyjayanthi and Bonet, David and Arbiol Rakuljic, Pol and Mogas Pladevall, Laia and Comajoan Cara, Mar{\c{c}}al and Gonzalez Ferrer, Jesus and Abante, Jordi and Mas Montserrat, Daniel and Haussler, David and Ioannidis, Alexander G.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}
```