cihatyldz/sifahane-bert-turkish-medical

🏥 Şifahane — Turkish Medical Text Classification (BERT)

A fine-tuned Turkish BERT model for classifying patient complaints into 12 medical departments with high accuracy.

This model is part of the Şifahane project, a dual-model medical triage demo comparing a fine-tuned BERT classifier against a zero-shot LLM.

Model Overview

Property	Value
Base Model	dbmdz/bert-base-turkish-cased
Language	Turkish (tr)
Task	Multi-class Text Classification (12 classes)
Parameters	~111M
Inference Speed	~15–130 ms on CPU
License	Apache 2.0

Intended Use

Primary use case: Classifying Turkish patient complaints into the appropriate medical department for triage routing.

Input: Free-text patient complaint in Turkish.
Output: Predicted department label with confidence scores.

Example

Input:  "Göğsümde şiddetli ağrı var, sol koluma yayılıyor, terleme eşlik ediyor."
Output: Kardiyoloji (98.7%)

Supported Classes (12 Departments)

#	Department	Description
0	Dermatoloji	Skin conditions (eczema, fungal infections, allergies)
1	Endokrinoloji	Hormonal/metabolic disorders (diabetes, thyroid)
2	Gastroenteroloji	Digestive system (gastritis, liver, IBS)
3	Göğüs Hastalıkları	Respiratory (asthma, COPD, pneumonia)
4	Göz Hastalıkları	Eye conditions (conjunctivitis, glaucoma)
5	KBB	Ear, nose, throat (sinusitis, otitis, tonsillitis)
6	Kadın Doğum	Obstetrics & gynecology
7	Kardiyoloji	Cardiovascular (hypertension, arrhythmia, coronary)
8	Nöroloji	Neurological (migraine, epilepsy, stroke, vertigo)
9	Ortopedi	Musculoskeletal (disc herniation, meniscus, fractures)
10	Psikiyatri	Mental health (depression, anxiety, sleep disorders)
11	Üroloji	Urological (UTI, kidney stones, prostate)

Quick Start

With Transformers Pipeline

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="cihatyldz/sifahane-bert-turkish-medical",
    top_k=3,
)

result = classifier("Midemde şiddetli yanma var, 2 haftadır devam ediyor.")
print(result)
# [[{'label': 'Gastroenteroloji', 'score': 0.98}, ...]]

Manual Inference

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import torch.nn.functional as F

model_name = "cihatyldz/sifahane-bert-turkish-medical"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.eval()

text = "Başımın sol tarafında zonklayıcı ağrı var, ışığa hassaslaştım."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    probs = F.softmax(outputs.logits, dim=-1)[0]

pred_id = probs.argmax().item()
label = model.config.id2label[pred_id]
confidence = probs[pred_id].item()

print(f"Department: {label} ({confidence:.1%})")
# Department: Nöroloji (97.2%)

Training Details

Dataset

• Source: cihatyldz/sifahane-turkish-medical-complaints
• Size: ~6,000 synthetic Turkish patient complaints
• Generation Method: Template-based synthesis with randomized clinical parameters (duration, intensity, symptoms, body location)
• Coverage: 12 departments × 36 conditions × 3 urgency levels
• Split: 80% train / 10% validation / 10% test (stratified)

Training Configuration

Hyperparameter	Value
Base model	dbmdz/bert-base-turkish-cased
Epochs	4
Batch size	32
Learning rate	2e-5
Weight decay	0.01
Warmup ratio	0.1
Max sequence length	128
Optimizer	AdamW
LR scheduler	Linear with warmup
FP16	Enabled
Early stopping	Patience = 2 (metric: F1 macro)

Training Infrastructure

• Hardware: NVIDIA Tesla T4 (Google Colab)
• Training time: ~5–10 minutes
• Framework: Hugging Face Transformers + Trainer API

Evaluation Results

Evaluated on the held-out test split (~600 examples):

Metric	Score
Accuracy	~99.3%
F1 (macro)	~99.3%
F1 (weighted)	~99.3%

Note: High scores reflect the synthetic nature of the training data with clear template patterns. Real-world clinical text would yield lower but still useful performance. The model demonstrates strong pattern recognition within its training distribution.

Live Demo

Try the model in action at the Şifahane Space, where it runs side-by-side with a zero-shot LLM (Qwen2.5-7B):

🚀 Şifahane Demo

The demo compares two approaches:

	BERT Classifier (this model)	LLM (Qwen2.5-7B)
Speed	~15–130 ms (CPU)	~1–3 s (Inference API)
Approach	Fine-tuned, single-task	Zero-shot, multi-task
Flexibility	Fixed categories	Open-ended
Cost	Free (CPU)	Free (HF Inference API)

Project Portfolio

This model is part of a Turkish NLP portfolio demonstrating three different AI architectures across three domains:

Project	Domain	Architecture	Link
🪙 Akçe	Banking	Fine-tuned LLM (Generative)	Space
🐪 Kervan	Logistics	RAG (Retrieval + Generative)	Space
🏥 Şifahane	Healthcare	Dual-Model (Classifier vs LLM)	Space

Limitations & Ethical Considerations

• Not a medical device: This model is for research and educational purposes only. It should never be used for actual medical diagnosis or triage decisions.
• Synthetic training data: The model was trained on template-generated data, not real clinical records. Performance on real-world patient language may differ.
• Turkish only: The model is designed for Turkish text and will not perform well on other languages.
• Fixed taxonomy: The model only supports the 12 predefined departments. Complaints outside these categories may be misclassified.
• No urgency assessment: This model predicts department only. Urgency classification requires additional logic (see the Şifahane Space for a rule-based urgency module).

Citation

@misc{yildiz2025sifahane,
  author       = {Cihat Yıldız},
  title        = {Şifahane: Turkish Medical Text Classification with Fine-tuned BERT},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/cihatyldz/sifahane-bert-turkish-medical}
}

Contact

• Developer: Cihat Yıldız
• Hugging Face: @cihatyldz
• Demo: Şifahane Space