| --- |
| license: pddl |
| license_name: physionet-credentialed-health-data-license-1.5.0 |
| license_link: https://physionet.org/content/mimiciv/view-license/3.1/ |
| language: |
| - en |
| tags: |
| - medical |
| - clinical |
| - ecg |
| - ehr |
| - multimodal |
| - question-answering |
| - prognosis |
| - mimic-iv |
| task_categories: |
| - question-answering |
| size_categories: |
| - 100K<n<1M |
| --- |
| |
| # UniPACT-MDS-ED Prompted Dataset |
|
|
|  |
|  |
|  |
|  |
|
|
|  |
|
|
| A multimodal prompt-formatted version of the **MDS-ED** prognostic benchmark (built on MIMIC-IV-ECG + MIMIC-IV), used to train and evaluate **UniPACT**. Each example pairs a raw 12-lead ECG waveform with a natural-language prompt that textualizes structured EHR (demographics, biometrics, vital signs) plus a binary clinical question (Yes/No). |
|
|
| 📄 **Paper:** Tang, Xia, Lu, Saeed. *UniPACT: A Multimodal Framework for Prognostic Question Answering on Raw ECG and Structured EHR.* ICASSP 2026 ([arXiv:2601.17916](https://arxiv.org/abs/2601.17916) · [IEEE Xplore](https://ieeexplore.ieee.org/document/11461333)) — read this for architecture, training recipe, ablations, and full results. |
|
|
| ## 🔑 Access |
|
|
| This dataset inherits the **PhysioNet Credentialed Health Data License** from MIMIC-IV. Before you can download: |
|
|
| 1. Create a [PhysioNet](https://physionet.org/) account |
| 2. Complete the required CITI "Data or Specimens Only Research" training |
| 3. Sign the MIMIC-IV data use agreement |
|
|
| Without credentialing, the files will not be accessible. |
|
|
| ## 📊 Task coverage |
|
|
| The dataset spans **1443 binary prognostic sub-tasks** across four families: |
|
|
| | Family | Sub-tasks | Description | |
| |---|---:|---| |
| | Diagnosis | 1428 | Disease-level binary diagnosis classes | |
| | Deterioration | 6 | Acute clinical deterioration outcomes | |
| | ICU admission | 2 | ICU-related admission outcomes | |
| | Mortality | 7 | Mortality at multiple time horizons | |
|
|
| Splits follow the official MDS-ED train / validation / test partition. |
|
|
| ## 🧪 Examples |
|
|
| Each row contains a raw ECG tensor (`(5000, 12)` float array · 10 s @ 500 Hz · 12 leads), a textualized EHR prompt, a binary clinical question, and a Yes/No label. One real sample from each task family: |
|
|
| **Deterioration** — `deterioration_severe_hypoxemia_362` |
| ```text |
| ECG: …/p16463772/s48979490/48979490 |
| EHR: 54 y/o BLACK AFRICAN AMERICAN female; BMI 29.8, weight 76.2; |
| temp 36.4, HR 60, RR 16.5, SpO₂ 100, BP 153/94, acuity 2.0. |
| Q: Will the patient "experience severe hypoxemia"? |
| A: Yes |
| ``` |
|
|
| **ICU admission** — `deterioration_icu_24h_7` |
| ```text |
| ECG: …/p18585855/s44869927/44869927 |
| EHR: 54 y/o WHITE female; BMI 28.9, weight 71.7, height 157.5; |
| temp 36.8, HR 77.5, RR 16.5, SpO₂ 99.5, BP 115.5/66.5, acuity 3.0. |
| Q: Will the patient "require ICU admission within the next 24 hours"? |
| A: Yes |
| ``` |
|
|
| **Mortality** — `deterioration_mortality_1d_3` |
| ```text |
| ECG: …/p13953606/s48605173/48605173 |
| EHR: 61 y/o WHITE male; BMI 28.7, weight 104.1, height 190.5; |
| temp 36.3, HR 101, RR 20, SpO₂ 99, BP 156/81, acuity 1.0. |
| Q: Will the patient "die within 24 hours"? |
| A: No |
| ``` |
|
|
| **Diagnosis** — `diagnoses_i10_0` |
| ```text |
| ECG: …/p11922120/s41055417/41055417 |
| EHR: 66 y/o WHITE male; BMI 31.7, weight 118.1; |
| temp 36.3, HR 81, RR 18, SpO₂ 99, BP 141/73, acuity 2.0. |
| Q: Will the patient be diagnosed with "Essential (primary) hypertension"? |
| A: No |
| ``` |
|
|
| --- |
|
|
| ## Reference |
|
|
| <details> |
| <summary><b>Citation</b></summary> |
|
|
| ```bibtex |
| @INPROCEEDINGS{11461333, |
| author = {Tang, Jialu and Xia, Tong and Lu, Yuan and Saeed, Aaqib}, |
| booktitle = {ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, |
| title = {UniPACT: A Multimodal Framework for Prognostic Question Answering on Raw ECG and Structured EHR}, |
| year = {2026}, |
| pages = {22537-22541}, |
| doi = {10.1109/ICASSP55912.2026.11461333} |
| } |
| ``` |
|
|
| Please also cite the underlying [MDS-ED benchmark](https://arxiv.org/abs/2407.17856) and [MIMIC-IV](https://physionet.org/content/mimiciv/) per their respective requirements. |
|
|
| </details> |
|
|
| <details> |
| <summary><b>Source data & construction</b></summary> |
|
|
| - **Source benchmark:** [MDS-ED](https://arxiv.org/abs/2407.17856) (Alcaraz et al., 2024) |
| - **Underlying clinical data:** [MIMIC-IV-ECG](https://physionet.org/content/mimic-iv-ecg/) and [MIMIC-IV](https://physionet.org/content/mimiciv/), collected at Beth Israel Deaconess Medical Center and de-identified under HIPAA Safe Harbor by the MIT Laboratory for Computational Physiology |
| - **EHR textualization:** 3 demographic + 3 biometric + 7 vital-parameter fields are inserted into fixed sentence templates |
| - **Labels:** derived programmatically from existing structured fields (ICD codes, ICU stay tables, mortality timestamps, MDS-ED deterioration outcomes) — no human re-annotation |
| - **ECG:** raw 12-lead waveforms kept in native form for waveform encoders; no conversion to text reports |
|
|
| See paper §2 for the full construction pipeline. |
|
|
| </details> |
