README.md
In clinChat: Conversational Interfaces for Clinical Data with Large Language Models

clinChat

Conversational interfaces for clinical data using Large Language Models in R.

clinChat helps clinical researchers, biostatisticians, and data scientists explore structured clinical datasets through natural language chat interfaces. It is built on the modern R LLM stack (ellmer + shinychat) and prioritizes privacy: your data stays local; the model only sees tool results you explicitly allow.

Privacy-aware chat with clinical data frames
Pre-built tools for summarization, filtering, column description, and basic statistics
Ready-to-launch Shiny chat applications
Easy integration with ellmer providers (OpenAI, Anthropic, local models, etc.)
Designed for clinical and pharmaceutical data review workflows

# Install dependencies first
install.packages(c("shiny", "ellmer", "shinychat", "dplyr"))

# Then install this package (once on CRAN or from GitHub)
# install.packages("clinChat")
# or
# remotes::install_github("shekharchalwadigrepo-hub/clinChat")

library(clinChat)

# Example 1: Clinical trial data
set.seed(42)
trial_data <- data.frame(
  subject_id = paste0("S", 1:50),
  age = sample(20:80, 50, replace = TRUE),
  sex = sample(c("M", "F"), 50, replace = TRUE),
  treatment = sample(c("Drug A", "Placebo"), 50, replace = TRUE),
  response = sample(c("Yes", "No"), 50, replace = TRUE)
)

chat <- create_clinical_chat(trial_data)
# chat$chat("What is the response rate by treatment?")

# Example 2: Lab values
lab_data <- data.frame(
  patient_id = 1:40,
  age = sample(30:70, 40, replace = TRUE),
  diagnosis = sample(c("Diabetes", "Hypertension", "None"), 40, replace = TRUE),
  hba1c = round(rnorm(40, 7.0, 1.2), 1)
)

chat2 <- create_clinical_chat(lab_data)
# chat2$chat("How many patients have diabetes? What is their mean HbA1c?")

run_clinchat_app(trial_data, title = "Trial Data Chat")

Raw patient-level data is never sent to the LLM by default.
Tools return aggregated summaries, counts, or filtered results that you control.
Prefer local models (via Ollama or similar) for maximum privacy.
Always review tool outputs before relying on model responses in regulated environments.

This is an initial release designed to be CRAN-ready. Contributions and feedback are welcome.

MIT