README.md

metadata

license: other
task_categories:
  - text-classification
  - question-answering
language:
  - en
tags:
  - medical
  - clinical-reasoning
  - pulmonology
  - safety-evaluation
  - benchmark
  - lung cancer
  - lungcancer
  - lungCancerScreaning
size_categories:
  - 1K<n<10K

PulmoBench: A Structured Clinical Reasoning Benchmark for Pulmonology

Overview

PulmoBench is a structured benchmark designed to evaluate Large Language Models (LLMs) in clinical risk stratification, safety-aware pulmonary reasoning, and zero-shot diagnostic generalization.

The benchmark features two evaluation tracks:

1. Core Benchmark (train, validation, test): Synthesized natural-language clinical vignettes with structured risk tiers, escalation flags, and systematic perturbations (lexical rephrasing, symptom omission, contradictory cues).
2. External Validation (External_validation): Real-world patient records deterministically mapped from the Biswas & Nath clinical dataset into natural language narrative vignettes for out-of-distribution evaluation.

---

Dataset Structure

1. Core Benchmark Splits (train, validation, test)

Field	Type	Description
case_id	int	Unique case identifier
age	int	Patient age in years
sex	str	Biological sex (male, female)
risk_tier	str	Ordinal risk category (low, medium, high)
risk_score	int	Normalized risk score (0–2)
urgency	str	Recommended triage urgency level (routine_followup, specialist_consult, urgent_ct_scan, immediate_referral)
gold_differential	list[str]	Ground truth differential diagnoses
escalation_required	bool	Safety flag denoting whether immediate medical escalation is required
vignette	str	Clinical patient narrative presented to the model
perturbation_type	str	Stress-test variant (original, lexical_rephrase, symptom_removed, contradictory_signal)

2. External Validation Split (External_validation)

Generated from raw tabular clinical records containing 13 clinical symptoms/lifestyle factors (smoking, yellow fingers, chronic disease, wheezing, chest pain, etc.) transformed into standard narrative vignettes.

Field	Type	Description
case_id	int	Original row index from source cohort
age	int	Patient age in years
sex	str	Biological sex (male, female)
vignette	str	Natural language case presentation constructed from reported symptoms
gold_lung_cancer	bool	Ground-truth diagnostic confirmation (True / False). Do not include in prompt input.

---

Loading the Dataset

  from datasets import load_dataset
  dataset = load_dataset("saibhossain/PulmoBench")
  print(dataset)

output :

  DatasetDict({
      train: Dataset({
          features: ['case_id', 'age', 'sex', 'risk_tier', 'risk_score', 'urgency', 'gold_differential', 'escalation_required', 'vignette', 'perturbation_type'],
          num_rows: 2544
      })
      validation: Dataset({
          features: ['case_id', 'age', 'sex', 'risk_tier', 'risk_score', 'urgency', 'gold_differential', 'escalation_required', 'vignette', 'perturbation_type'],
          num_rows: 544
      })
      test: Dataset({
          features: ['case_id', 'age', 'sex', 'risk_tier', 'risk_score', 'urgency', 'gold_differential', 'escalation_required', 'vignette', 'perturbation_type'],
          num_rows: 548
      })
  })

Loading External Validation

from datasets import load_dataset

ext_val = load_dataset(
    "saibhossain/PulmoBench", 
    data_files={"external_validation": "External_validation/lung_cancer_external.jsonl"}
)
print(ext_val)

--

Task Definitions & Evaluation

Given a clinical vignette, models are evaluated on:

1. Clinical Risk Stratification: Predicting risk_tier and triage urgency.
2. Safety Compliance: Correctly identifying escalation_required without hallucinating or under-triaging critical findings.
3. Robustness: Stability under adversarial perturbations (perturbation_type).
4. External Generalization: Zero-shot binary detection of lung cancer risk on the External_validation cohort.

Intended Use & Safety Notice

PulmoBench is curated strictly for academic research, benchmark evaluation, and robustness testing of LLMs in biomedical contexts.

Clinical Disclaimer: This benchmark is not intended for direct clinical deployment, diagnostic decision-making, or real-world patient management.

Citations & Acknowledgments

Source Dataset

If you use PulmoBench or its external validation set, please cite the underlying sources:

Derived from: [Lung Cancer Prediction](https://www.kaggle.com/datasets/thedevastator/cancer-patients-and-air-pollution-a-new-link)

@misc{mr__abhinaba_biswas_mr__akash_nath_2024,
    title={Lung Cancer Dataset},
    url={[https://www.kaggle.com/dsv/8795028](https://www.kaggle.com/dsv/8795028)},
    DOI={10.34740/KAGGLE/DSV/8795028},
    publisher={Kaggle},
    author={Mr. Abhinaba Biswas and Mr. Akash Nath},
    year={2024}
}

@misc{pulmobench2026,
  title={PulmoBench: A Structured Clinical Reasoning Benchmark for Pulmonology},
  author={Hossain, Saib},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{[https://huggingface.co/datasets/saibhossain/PulmoBench](https://huggingface.co/datasets/saibhossain/PulmoBench)}}
}

License

Other (specified in description)