Getting Started with clinChat

library(clinChat)

Overview

clinChat lets you explore clinical datasets through natural language while keeping the data local. It registers tools that the language model can call to summarize, count, and group your data.

Example 1: Basic Clinical Trial Data

# Simulated clinical trial data
set.seed(42)
trial_data <- data.frame(
  subject_id = paste0("S", sprintf("%03d", 1:80)),
  age = sample(18:85, 80, replace = TRUE),
  sex = sample(c("Male", "Female"), 80, replace = TRUE),
  arm = sample(c("Drug A", "Drug B", "Placebo"), 80, replace = TRUE),
  response = sample(c("Responder", "Non-responder"), 80, replace = TRUE, prob = c(0.55, 0.45)),
  baseline_score = round(rnorm(80, 50, 12), 1)
)

# Create a chat object
chat <- create_clinical_chat(
  data = trial_data,
  system_prompt = "You are a clinical data analyst. Use the available tools to answer accurately."
)

# Example questions you can ask (requires API key)
# chat$chat("How many subjects are in each treatment arm?")
# chat$chat("What is the mean age by sex?")
# chat$chat("Summarize the dataset")

Example 2: Lab Results Dataset

lab_data <- data.frame(
  patient_id = 1:60,
  age = sample(25:75, 60, replace = TRUE),
  diagnosis = sample(c("Type 2 Diabetes", "Hypertension", "Healthy"), 60, replace = TRUE),
  hba1c = round(rnorm(60, 6.8, 1.4), 1),
  systolic_bp = round(rnorm(60, 132, 16))
)

chat2 <- create_clinical_chat(lab_data)

# chat2$chat("How many patients have Type 2 Diabetes and what is their average HbA1c?")
# chat2$chat("Describe the columns in this dataset")

Launching the Interactive App

# Launch a Shiny chat interface
run_clinchat_app(
  data = trial_data,
  title = "Clinical Trial Explorer",
  provider = "openai/gpt-4o-mini"   # or your preferred provider
)

Using the Tools Directly

# Inspect the tools that will be registered
tools <- clinical_tools(trial_data)
length(tools)   # should be 4

Privacy

All data queries happen locally via tools. Only the results of those tools (summaries, counts, aggregates) are passed to the language model. Raw patient-level rows are never sent by default.