binary-classification-drug-reactions

binary-classification-drug-reactions is NeuroKnow AI's 184.4M-parameter DeBERTa-v3-based binary classifier for assessing the severity of adverse drug reactions from patient-reported narrative text. It classifies each input as Severe or Non-Severe for pharmacovigilance and drug-safety workflows.

What it does

The model classifies patient-reported drug-experience narratives into:

• Non-Severe
• Severe

It supports adverse drug reaction triage, pharmacovigilance analysis, drug-safety workflows, biomedical text classification, and research-oriented decision-support pipelines.

Model architecture

• Base architecture: microsoft/deberta-v3-base
• Transformers architecture: DebertaV2ForSequenceClassification
• Architecture family: DeBERTa-v3
• Task: binary sequence classification
• Classes: Non-Severe, Severe
• Transformer layers: 12
• Hidden size: 768
• Attention heads: 12
• Intermediate size: 3072
• Classifier output dimension: 2
• Parameter count: 184,423,682
• Model dtype: float32
• Tokenizer: SentencePiece Unigram via DebertaV2Tokenizer
• Vocabulary size: 128,100
• Application maximum sequence length: 256 tokens
• Maximum positional embeddings: 512

Training data

The model was trained on:

• 8,153 patient-reported drug-experience narratives
• Adverse drug reaction severity labels
• A near-balanced class distribution:
  • 53.4% Severe
  • 46.6% Non-Severe

Additional biomedical training

The model was further adapted on OpenMed/DDI-Corpus-Processed for biomedical domain adaptation. This additional training was used to strengthen:

• Drug-interaction representations
• Pharmacological safety understanding
• Biomedical domain adaptation
• Representations relevant to adverse drug reaction severity classification

This corpus was used as additional biomedical training and is not presented as a Severe/Non-Severe adverse drug reaction dataset.

Training configuration

learning_rate: 2e-5
optimizer: adafactor
batch_size: 16
gradient_accumulation: 4
effective_batch_size: 64
epochs: 8
early_stopping_patience: 3
warmup_ratio: 0.1
lr_scheduler: cosine
weight_decay: 0.01
max_seq_length: 256
cv_folds: 5

Validation metrics

Metric	Score
Macro F1	0.9162
Accuracy	95.52%

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "neuroknowai/binary-classification-drug-reactions"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

id2label = {
    0: "Non-Severe",
    1: "Severe",
}

def classify_adr(text):
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        max_length=256,
        padding=True,
    )

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

    class_id = int(probabilities.argmax())
    label = id2label[class_id]
    confidence = float(probabilities[class_id])

    return {
        "label": label,
        "confidence": round(confidence, 4),
    }

result = classify_adr(
    "I experienced severe insomnia, heart palpitations, and extreme anxiety "
    "after taking this medication for two weeks."
)

print(result)

Intended use

This model is intended for research, pharmacovigilance analysis, drug-safety workflows, ADR narrative classification, and decision-support pipelines.

It should not be used as a diagnostic model or as the sole basis for clinical decisions.

Organization

Developed and maintained by NeuroKnow AI.

Hugging Face:

neuroknowai/binary-classification-drug-reactions