ledphrey's picture
transfer from thankrandomness
3061f29 verified
import os
import torch
from datasets import load_dataset, DatasetDict
from transformers import AutoTokenizer, AutoModel
import chromadb
import gradio as gr
import numpy as np
from sklearn.metrics import precision_score, recall_score, f1_score
# Mean Pooling - Take attention mask into account for correct averaging
def meanpooling(output, mask):
embeddings = output[0] # First element of model_output contains all token embeddings
mask = mask.unsqueeze(-1).expand(embeddings.size()).float()
return torch.sum(embeddings * mask, 1) / torch.clamp(mask.sum(1), min=1e-9)
# Load the dataset
dataset = load_dataset("thankrandomness/mimic-iii")
# Split the dataset into train and validation sets
split_dataset = dataset['train'].train_test_split(test_size=0.2, seed=42)
dataset = DatasetDict({
'train': split_dataset['train'],
'validation': split_dataset['test']
})
# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("neuml/pubmedbert-base-embeddings-matryoshka")
model = AutoModel.from_pretrained("neuml/pubmedbert-base-embeddings-matryoshka")
# Function to normalize embeddings to unit vectors
def normalize_embedding(embedding):
norm = np.linalg.norm(embedding)
return (embedding / norm).tolist() if norm > 0 else embedding
# Function to embed and normalize text
def embed_text(text):
inputs = tokenizer(text, padding=True, truncation=True, max_length=512, return_tensors='pt')
with torch.no_grad():
output = model(**inputs)
embeddings = meanpooling(output, inputs['attention_mask'])
normalized_embeddings = normalize_embedding(embeddings.numpy())
return normalized_embeddings
# Initialize ChromaDB client
client = chromadb.Client()
collection = client.create_collection(name="pubmedbert_matryoshka_embeddings")
# Function to upsert data into ChromaDB
def upsert_data(dataset_split):
for i, row in enumerate(dataset_split):
for note in row['notes']:
text = note.get('text', '')
annotations_list = []
for annotation in note.get('annotations', []):
try:
code = annotation['code']
code_system = annotation['code_system']
description = annotation['description']
annotations_list.append({"code": code, "code_system": code_system, "description": description})
except KeyError as e:
print(f"Skipping annotation due to missing key: {e}")
if text and annotations_list:
embeddings = embed_text([text])[0]
# Upsert data, embeddings, and annotations into ChromaDB
for j, annotation in enumerate(annotations_list):
collection.upsert(
ids=[f"note_{note['note_id']}_{j}"],
embeddings=[embeddings],
metadatas=[annotation]
)
else:
print(f"Skipping note {note['note_id']} due to missing 'text' or 'annotations'")
# Upsert training data
upsert_data(dataset['train'])
# Define retrieval function with similarity threshold
def retrieve_relevant_text(input_text):
input_embedding = embed_text([input_text])[0]
results = collection.query(
query_embeddings=[input_embedding],
n_results=5,
include=["metadatas", "documents", "distances"]
)
output = []
#print("Retrieved items and their similarity scores:")
for metadata, distance in zip(results['metadatas'][0], results['distances'][0]):
#print(f"Code: {metadata['code']}, Similarity Score: {distance}")
#if distance <= similarity_threshold:
output.append({
"similarity_score": distance,
"code": metadata['code'],
"code_system": metadata['code_system'],
"description": metadata['description']
})
# if not output:
# print("No results met the similarity threshold.")
return output
# Evaluate retrieval efficiency on the validation/test set
def evaluate_efficiency(dataset_split):
y_true = []
y_pred = []
total_similarity = 0
total_items = 0
for i, row in enumerate(dataset_split):
for note in row['notes']:
text = note.get('text', '')
annotations_list = [annotation['code'] for annotation in note.get('annotations', []) if 'code' in annotation]
if text and annotations_list:
retrieved_results = retrieve_relevant_text(text)
retrieved_codes = [result['code'] for result in retrieved_results]
# Sum up similarity scores for average calculation
for result in retrieved_results:
total_similarity += result['similarity_score']
total_items += 1
# Ground truth
y_true.extend(annotations_list)
# Predictions (limit to length of true annotations to avoid mismatch)
y_pred.extend(retrieved_codes[:len(annotations_list)])
# for result in retrieved_results:
# print(f" Code: {result['code']}, Similarity Score: {result['similarity_score']:.2f}")
# Debugging output to check for mismatches and understand results
# print("Sample y_true:", y_true[:10])
# print("Sample y_pred:", y_pred[:10])
if total_items > 0:
avg_similarity = total_similarity / total_items
else:
avg_similarity = 0
if len(y_true) != len(y_pred):
min_length = min(len(y_true), len(y_pred))
y_true = y_true[:min_length]
y_pred = y_pred[:min_length]
# Calculate metrics
precision = precision_score(y_true, y_pred, average='macro', zero_division=0)
recall = recall_score(y_true, y_pred, average='macro', zero_division=0)
f1 = f1_score(y_true, y_pred, average='macro', zero_division=0)
return precision, recall, f1, avg_similarity
# Calculate retrieval efficiency metrics
precision, recall, f1, avg_similarity = evaluate_efficiency(dataset['validation'])
# Gradio interface
def gradio_interface(input_text):
results = retrieve_relevant_text(input_text)
formatted_results = [
f"Result {i + 1}:\n"
f"Similarity Score: {result['similarity_score']:.2f}\n"
f"Code: {result['code']}\n"
f"Code System: {result['code_system']}\n"
f"Description: {result['description']}\n"
"-------------------"
for i, result in enumerate(results)
]
return "\n".join(formatted_results)
# Display retrieval efficiency metrics
# metrics = f"Precision: {precision:.2f}, Recall: {recall:.2f}, F1 Score: {f1:.2f}"
metrics = f"Accuracy: {avg_similarity:.2f}"
with gr.Blocks() as interface:
gr.Markdown("# Automated Medical Coding POC")
# gr.Markdown(metrics)
with gr.Row():
with gr.Column():
text_input = gr.Textbox(label="Input Text")
submit_button = gr.Button("Submit")
with gr.Column():
text_output = gr.Textbox(label="Retrieved Results", lines=10)
submit_button.click(fn=gradio_interface, inputs=text_input, outputs=text_output)
interface.launch()