Ipds Eval

---
name: ipds-eval
description: Evaluates large language models' ability to support inpatient clinical decision-making by classifying patient cases into appropriate triage, diagnosis, and treatment pathways. It probes the models' clinical reasoning, diagnostic accuracy, and alignment with real-world physician judgments. Use when the user wants to benchmark on IPDS, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2503.13205
  bibtex_key: chen2025map
  confidence: high
---

# ipds-eval

> MAP: Evaluation and Multi-Agent Enhancement of Large Language Models for Inpatient Pathways — Chen et al. (2025) (arXiv:2503.13205, 2025)

## What this evaluates

Evaluates large language models' ability to support inpatient clinical decision-making by classifying patient cases into appropriate triage, diagnosis, and treatment pathways. It probes the models' clinical reasoning, diagnostic accuracy, and alignment with real-world physician judgments.

## Datasets

- **IPDS** — total 51274; splits: test (-1); repo https://github.com/franciszchen/MAP

## Metrics

- `accuracy` **(primary)** — range: percent
  - Proportion of correctly predicted clinical pathway labels (triage, diagnosis, or treatment) out of the total number of cases. Calculated as (number of correct predictions / total predictions) × 100.
- `intra-class correlation coefficient (ICC)` — range: [0, 1]
  - Statistical measure of inter-rater reliability used to quantify the agreement between model/clinician predictions and ground truth or among clinicians. Values range from 0 to 1, with higher values indicating stronger agreement.

## Input / output format

**Input**: Patient clinical cases/notes derived from MIMIC-IV, presented as structured or free-text clinical presentations requiring pathway classification.

**Output**: Classification labels for triage, diagnosis, and treatment pathways. For clinical validation, a multiple-choice format allowing the top 3 ranked diagnoses.

## Scoring recipe

```python
def calculate_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (correct / len(gold_labels)) * 100

def calculate_icc(predictions, gold_labels):
    # Uses standard ICC(2,1) or ICC(3,1) for absolute agreement
    # Implemented via scipy.stats or pingouin
    return pingouin.intraclass_corr(data=predictions, targets=gold_labels, raters=None)['ICC'].values[0]
```

## Common pitfalls

- MIMIC-IV data cannot be sent to external APIs (e.g., OpenAI, Google) due to privacy agreements, restricting evaluation to locally hosted models.
- LLM classification performance is sensitive to the temperature parameter, introducing randomness that must be controlled during evaluation.
- Clinical validation sample size is limited (e.g., 100 cases) due to the high time/energy cost of expert physician review.

## Evidence (verbatim from paper)

> MAP achieved an overall diagnosis accuracy of 78.10%, reflecting an 28.80% improvement over LLaMA3-8B, which had an accuracy of 49.30%. Notably, MAP outperformed the best specialized LLM, HuatuoGPT2-13B, by a 25.10% improvement in accuracy (i.e., 78.10% vs. 53.00%).

## Citation

```bibtex
@misc{chen2025map,
  title={MAP: Evaluation and Multi-Agent Enhancement of Large Language Models for Inpatient Pathways},
  author={Chen et al. (2025)},
  year={2025},
  note={arXiv:2503.13205}
}
```

- arXiv: 2503.13205