Diagnosis Matching — Distilled ModernBERT

A binary classifier that determines whether two free-text diagnoses written in Russian refer to the same condition.

This model is a single-transformer distillation of a complex multi-component meta-model for diagnosis matching. The original system is an AutoGluon ensemble that aggregates 16 engineered features from four distinct sectors:

The full meta-model achieves F1 = 0.910, avg_prec = 0.948 on the held-out test set — but requires running 5 separate LLMs, 3 embedding models, and 6 handcrafted NLP features at inference time.

This distilled model compresses the entire pipeline into a single ModernBERT-base forward pass, replacing all 16 features and their upstream models with one 577 MB checkpoint.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("sm1rk/diagnosis-matching-distilled-model")
model = AutoModelForSequenceClassification.from_pretrained("sm1rk/diagnosis-matching-distilled-model")

inputs = tokenizer("атеросклероз сонных артерий", "гиперлипидемия", return_tensors="pt")
with torch.no_grad():
    prob = torch.softmax(model(**inputs).logits, dim=1)[0, 1].item()

print(f"P(match) = {prob:.4f}")

Model Details

Input: a pair of free-text diagnoses (diagnosis_1, diagnosis_2). Output: match probability P(match) ∈ [0, 1].

Intended Use

Matching semantically equivalent diagnoses expressed in different wording — e.g., clinical records, discharge summaries, billing codes. Not intended for clinical decision support.