Teaching LLMs Clinical Reasoning for ICUs

# Illustrative sketch (not from the paper)
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
# Load ICU-REACT dataset (clinician‑in‑the‑loop, OMOP‑aligned)
train_dataset, test_dataset = load_icu_react(split="train"), load_icu_react(split="test")
# Initialize base model (e.g., Llama) and tokenizer
model_name = "meta-llama/Llama-2-70b"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_name)
# Apply LoRA for low‑rank adaptation (Clin‑REACT fine‑tuning)
model = apply_lora(model, rank=4)  # rank chosen as an example of LoRA config
# Simple training loop (supervised fine‑tuning on ICU‑REACT)
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
model.train()
for epoch in range(1):
    for batch in train_dataset:
        inputs = tokenizer(batch["question"], return_tensors="pt", truncation=True, padding=True)
        labels = tokenizer(batch["answer"], return_tensors="pt", truncation=True, padding=True).input_ids
        outputs = model(**inputs, labels=labels)
        loss = outputs.loss
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()
# Evaluate on ICU‑REACT test set and external benchmarks
results = evaluate_on_benchmarks(model, tokenizer, test_dataset, benchmarks=["SCT-Bench", "ER-Reason", "MedRBench", "VivaBench"])
print("Evaluation results:", results)

// Illustrative sketch (not from the paper)
const { AutoModelForCausalLM, AutoTokenizer } = require('@huggingface/transformers');
const torch = require('torch-js'); // placeholder for torch bindings
// Load ICU‑REACT dataset (clinician‑in‑the‑loop, OMOP‑aligned)
const { loadIcuReact } = require('./data_utils');
const trainDataset = loadIcuReact('train');
const testDataset = loadIcuReact('test');
// Initialize base model (e.g., Llama) and tokenizer
const modelName = 'meta-llama/Llama-2-70b';
let model = AutoModelForCausalLM.from_pretrained(modelName, { dtype: 'bfloat16' });
const tokenizer = AutoTokenizer.from_pretrained(modelName);
// Apply LoRA for low‑rank adaptation (Clin‑REACT fine‑tuning)
const applyLoRA = require('./lora_utils');
model = applyLoRA(model, { rank: 4 }); // rank is an example LoRA config
// Simple training loop (supervised fine‑tuning on ICU‑REACT)
const optimizer = new torch.optim.AdamW(model.parameters(), { lr: 1e-4 });
model.train();
(async () => {
  for (const batch of trainDataset) {
    const inputs = tokenizer(batch.question, { returnTensors: 'pt', truncation: true, padding: true });
    const labels = tokenizer(batch.answer, { returnTensors: 'pt', truncation: true, padding: true }).input_ids;
    const outputs = await model.forward({ ...inputs, labels });
    const loss = outputs.loss;
    loss.backward();
    optimizer.step();
    optimizer.zeroGrad();
  }
  // Evaluate on ICU‑REACT test set and external benchmarks
  const evaluate = require('./eval_utils');
  const results = await evaluate(model, tokenizer, testDataset, ['SCT-Bench', 'ER-Reason', 'MedRBench', 'VivaBench']);
  console.log('Evaluation results:', results);
})();