clinical-nlp-fairness-auditor

Clinical NLP Fairness Auditor: DistilBERT Medical Specialty Classifier

Author: Steven Hill
Audit date: 2026-04-24 (corrected 2026-10-07)
Base model: distilbert-base-uncased
Task: Single-label text classification, 20 medical specialties
Repository: github.com/srhill12/clinical-nlp-fairness-auditor

This model was trained to be audited. It sorts clinical transcription notes by medical specialty, and the accompanying audit documents its performance disparities and ends in a do-not-deploy recommendation. The author trained the model and conducted the audit, so the audit is a self-audit, not an independent one.

This model is not intended for clinical or operational use.

---

Intended use

Intended: research on auditing clinical text classifiers, and study of the failure modes documented below.

Hypothetical use evaluated by the audit: administrative routing of transcribed notes to specialty work queues.

Out of scope: diagnosis or treatment decisions, emergency triage or patient prioritization, any production deployment, and any use with real patient data.

---

Training data

• Source: MTSamples sample transcription reports, via galileo-ai/medical_transcription_40. That dataset has no published license or dataset card.
• Size: 4,087 notes after keeping specialties with at least 50 examples
• Split: 70% train / 15% validation / 15% test, stratified (test set: 614 notes)
• Classes: Cardiovascular / Pulmonary; Consult - History and Phy.; Discharge Summary; ENT - Otolaryngology; Emergency Room Reports; Gastroenterology; General Medicine; Hematology - Oncology; Nephrology; Neurology; Neurosurgery; Obstetrics / Gynecology; Ophthalmology; Orthopedic; Pain Management; Pediatrics - Neonatal; Radiology; SOAP / Chart / Progress Notes; Surgery; Urology

Data limitations

• Sample transcriptions, not notes from a live EHR
• No patient demographic attributes, so protected-class fairness cannot be evaluated
• Class imbalance: Surgery is about 24% of the data
• Overlapping labels: MTSamples lists some notes under more than one specialty (8 test notes appear twice with different labels)
• 66.6% of test notes (409 of 614) exceed 512 tokens and are truncated

---

Training procedure

• Framework: HuggingFace Transformers, PyTorch
• Epochs: 3
• Batch size: 16
• Learning rate: 2e-5, warmup ratio 0.1
• Imbalance handling: inverse-frequency class weights in the cross-entropy loss
• Max input length: 512 tokens
• Hardware: NVIDIA T4 GPU (Google Colab), about 157 seconds

---

Evaluation results

Held-out test set, 614 notes.

Per-class performance

                               precision    recall  f1-score   support

   Cardiovascular / Pulmonary      0.544     0.740     0.627        50
   Consult - History and Phy.      0.417     0.072     0.123        69
            Discharge Summary      0.636     0.933     0.757        15
         ENT - Otolaryngology      0.310     0.750     0.439        12
       Emergency Room Reports      0.095     0.700     0.167        10
             Gastroenterology      0.432     0.533     0.478        30
             General Medicine      0.258     0.229     0.242        35
        Hematology - Oncology      0.000     0.000     0.000        12
                   Nephrology      0.381     0.727     0.500        11
                    Neurology      0.436     0.567     0.493        30
                 Neurosurgery      0.393     0.846     0.537        13
      Obstetrics / Gynecology      0.425     0.850     0.567        20
                Ophthalmology      0.435     0.909     0.588        11
                   Orthopedic      0.415     0.562     0.478        48
              Pain Management      0.875     0.875     0.875         8
        Pediatrics - Neonatal      0.000     0.000     0.000         9
                    Radiology      0.444     0.526     0.482        38
SOAP / Chart / Progress Notes      0.737     0.609     0.667        23
                      Surgery      0.619     0.088     0.154       148
                      Urology      0.548     0.773     0.642        22

                     accuracy                          0.419       614
                    macro avg      0.420     0.564     0.441       614
                 weighted avg      0.476     0.419     0.369       614

Subgroup performance disparities

Gaps are the highest group accuracy minus the lowest, in percentage points. They are differences, not ratios.

These are performance disparities across note categories, not protected-class fairness metrics.

---

Audit findings

Zero recall. Every Pediatrics-Neonatal note (9) and every Hematology-Oncology note (12) in the test set was misclassified.

Class-weight over-correction. Surgery, the largest class, has 8.8% recall (13 of 148; false negative rate 91.2%). It was predicted only 21 times, while Emergency Room Reports, with 10 test notes, was predicted 74 times. The two largest classes are heavily under-predicted and most small classes are over-predicted, which is consistent with the inverse-frequency weights over-correcting. An unweighted baseline was not retained, so this is not a controlled comparison.

Note length and truncation. Truncated notes score 35.9% vs 53.7% for untruncated notes. Truncation does not explain the length gap by itself: Q3 and Q4 are both fully truncated, yet Q4 outperforms Q3 by 14.8 points.

High miss rates. Five specialties have false negative rates of 70% or higher: Pediatrics - Neonatal, Hematology - Oncology, Consult - History and Phy., Surgery, and General Medicine.

Explainability. SHAP (PartitionExplainer) on four hand-picked notes, first 300 words each. In the misclassified Surgery example, a cardiac catheterization note predicted as Cardiovascular / Pulmonary, the top attributions are cardiac terms such as "coronary" and "catheterization." The model keys on organ-system vocabulary that crosses specialty boundaries. Four examples are illustrative, not systematic.

---

Deployment recommendation

Do not deploy. Before any deployment decision:

1. Zero-recall specialties resolved
2. Human review in place for high false-negative-rate specialties
3. Protected-class performance evaluated on data that includes demographic attributes
4. Independent audit sign-off

The absence of demographic data is a governance gap in its own right: a healthcare model that cannot be evaluated for protected-class fairness should not be put into use.

---

Governance documentation

The repository includes a scorecard that maps this project's work to NIST AI RMF 1.0 subcategories across GOVERN, MAP, MEASURE, and MANAGE, with a status for each. It is a mapping, not a claim of conformance. See outputs/nist_rmf_scorecard.md in the repository.

This card follows the structure proposed in Mitchell et al., "Model Cards for Model Reporting," Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT* '19), 2019 (doi:10.1145/3287560.3287596).

---

Corrections (2026-10-07)

An earlier version of this card reported that 20.7% of notes were truncated at 512 tokens. That figure could not be reproduced; the recomputed figure is 66.6%. It also attributed the note-length gap to truncation and said the model over-predicts high-frequency specialties. Neither is supported by the data. The repository's scorecard revision log lists all changes.

---

Citation

@misc{hill2026clinicalnlp,
  author = {Hill, Steven},
  title  = {Clinical NLP Fairness Auditor},
  year   = {2026},
  note   = {Audit of a DistilBERT medical specialty classifier},
  url    = {https://github.com/srhill12/clinical-nlp-fairness-auditor}
}