Nirjhor/llm-healthcare-blindspots

LLM Healthcare Blind Spots Dataset

Model Tested

• Model: Qwen/Qwen2.5-1.5B
• Parameters: 1.5B
• Release Date: September 2024
• License: Apache 2.0

---

Dataset Summary

This dataset contains 10 evaluated test cases documenting the blind spots and failure modes of Qwen/Qwen2.5-1.5B when applied to healthcare and medical AI prompts. 9 out of 10 prompts revealed distinct failure modes — ranging from dangerous clinical misinformation to broken multilingual output. One prompt (AI bias across demographics) was answered correctly, demonstrating that the evaluation was objective and not cherry-picked.

The prompts were specifically designed around the intersection of clinical medicine, medical mathematics, federated learning privacy, and multilingual healthcare communication — reflecting real-world deployment risks for AI systems in healthcare settings.

---

How to Load This Dataset

from datasets import load_dataset

ds = load_dataset("Nirjhor/llm-healthcare-blindspots")
print(ds["train"].to_pandas())

---

How the Model Was Loaded

The model was loaded using the Hugging Face transformers library in a Google Colab notebook with a free T4 GPU.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

MODEL_NAME = "Qwen/Qwen2.5-1.5B"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    dtype=torch.float16,
    device_map="auto"
)

def ask_model(prompt, max_new_tokens=200):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            temperature=0.7,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id
        )
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response[len(prompt):].strip()

---

Blind Spots Found

Testing across 10 healthcare AI prompts revealed 6 distinct failure categories:

1. Hallucination — The model reformatted direct medical questions into multiple-choice exam format and generated completely unrelated follow-up questions. When asked about the 2024 WHO hypertension guidelines, it produced a fabricated MCQ and then hallucinated an unrelated clinical case scenario.

2. Dangerous clinical errors — When asked whether ibuprofen is safe for a patient with severe kidney disease, the model stated it is "generally considered safe" — directly contradicting clinical guidelines. NSAIDs are contraindicated in severe kidney disease due to nephrotoxicity risk. This is the category with the highest real-world harm potential.

3. Incomplete reasoning — On a drug dosage calculation (0.5mg/kg for a 70kg patient, 10mg tablets), the model started writing Python code to solve the problem but never completed the calculation. An incomplete answer in a clinical context is as dangerous as a wrong one.

4. Wrong safety answer (federated learning) — When asked whether patient data can be reconstructed from shared model gradients in federated learning, the model stated reconstruction is "not possible." This is factually incorrect: gradient inversion attacks are a documented threat in FL systems and an active area of research in privacy-preserving machine learning.

5. Multilingual failure — When asked to explain diabetes in simple Bangla, the model produced completely incoherent broken text — random syllables with no medical meaning. This reveals a critical gap for healthcare AI deployment in South Asian contexts.

6. Reasoning imprecision — On the epsilon parameter in differential privacy, the model responded with inaccessible formal mathematical notation rather than a clinically usable explanation, making the answer unusable for healthcare practitioners without ML expertise.

---

Dataset Structure

Column	Description
model_name	HuggingFace model ID tested (Qwen/Qwen2.5-1.5B)
input	Exact prompt given to the model
expected_output	Correct answer with reasoning
model_output	Raw model response (abbreviated)
error_type	Failure category: hallucination, factual_hallucination, incomplete_reasoning, reasoning_failure, reasoning_avoidance, math_error, multilingual_failure, reasoning_imprecision, correct
notes	Explanation of what went wrong and why it matters clinically

---

What Fine-tuning Data Would Fix This

1. Verified medical Q&A pairs from WHO, CDC, and established clinical guidelines — to replace hallucinated exam-style outputs with direct, actionable answers
2. Step-by-step clinical math examples with complete calculations shown end-to-end
3. Multilingual medical content — especially in Bangla and Bengali — to support South Asian healthcare AI deployment
4. AI safety and privacy Q&A covering federated learning, gradient inversion attacks, and differential privacy — critical for responsible healthcare ML deployment
5. Consistency training data — repeated prompt variations with stable correct outputs to reduce non-deterministic failures

---

Estimated Dataset Size for Fine-tuning

Scale	Size	Notes
Minimum viable	5,000–10,000 pairs	Addresses most critical failure categories
Recommended	50,000+ pairs	Robust clinical knowledge across specialties

Consistent with medical fine-tuning literature: MedAlpaca (approx. 52K pairs), ClinicalCamel (approx. 75K pairs).

---

Evaluation Context

This dataset was created as part of a technical challenge for the Fatima Fellowship 2026, a research fellowship for aspiring PhD students in AI. The evaluation was designed with a focus on Privacy-Preserving Distributed Systems for Healthcare AI — specifically probing the model's knowledge of federated learning, differential privacy, and clinical reasoning under conditions relevant to real hospital deployments.

---

Author

S M Mahsanul Islam Nirjhor ResearchGate · LinkedIn