N-of-1 Causal Lab: Personalized Precision Decision-Making via Bayesian Causal Inference

Core Introduction

N-of-1 Causal Lab is an end-to-end LLM-driven framework designed to convert natural language questions into Bayesian causal inference workflows. It supports N-of-1 level causal analysis on personal time-series data (health, learning, behavior, etc.) to help users make personalized scientific decisions. This project combines N-of-1 research methodology, Bayesian statistics, large language models, and GPU-accelerated computing, focusing on solving the problem where population statistical conclusions cannot directly guide individual decisions.

Background: The Conflict Between Population Statistics and Personal Decision-Making

Traditional data science and medical research rely on large-sample analysis to draw universal conclusions, but when it comes to personal decisions, population averages often fail (e.g., if a drug works for 80% of people, do you belong to the 20% where it doesn’t?). N-of-1 research focuses on the time-series data of a single individual, using systematic self-experiments and causal inference to identify interventions suitable for that individual. This project combines this concept with Bayesian statistics, LLM, and GPU acceleration to build an end-to-end personal causal inference platform.

Technical Workflow: From Natural Language Query to Causal Estimation

The core workflow of the system is: User's natural language query → Output causal effect estimation, divided into 5 stages:

1. Problem Understanding and DAG Construction: Parse the causal structure, build a Directed Acyclic Graph (DAG), and determine treatment/outcome variables;
2. Measurement Model Matching: Analyze user datasets (e.g., Apple Health), handle irregular timestamps and semantic heterogeneity;
3. Causal Identification Validation: Run the ID algorithm to verify effect identifiability; if unidentifiable, revise the DAG;
4. State Space Model and Estimation: Convert to a continuous-time State Space Model (SSM), use the Ornstein-Uhlenbeck process to describe variable dynamics, apply Poisson/Bernoulli/Beta likelihood functions based on data types (count/binary/proportion), and perform estimation via JAX GPU-accelerated MCMC;
5. Counterfactual Simulation: LLM simulates intervention scenarios on the fitted model and outputs causal effect results.

Data Sources: Covering Multi-Domain Personal Data

The project supports various data types obtained via DSAR:

• Health & Fitness: Apple Health, Oura Ring, 23andMe (genetics), Strava;
• Learning & Cognition: Anki, Duolingo, YouTube watch history;
• Mental Health & Behavior: Google Takeout, WhatsApp chat logs, ChatGPT/Claude logs. Cross-analysis is supported (e.g., sleep + learning + mood data) to discover hidden causal relationships.

Technical Highlights: Analysis of Key Features

1. User-Friendly: Automatically handles complex decisions like model selection and prior setting, with no technical burden;
2. Interpretable Interaction: Users can check LLM outputs, challenge/override decisions, enabling human-in-the-loop;
3. Complex Time-Series Processing: Supports long time series, irregular sampling, and multi-variable joint modeling;
4. Robust LLM Modeling: LLM decisions are embedded in a state machine, with numerical checks (e.g., prior predictive checks) at each step;
5. Fast MCMC: Leverages Corenflos 2025 research results, GPU-associated Kalman filtering achieves O(log T) inference, completed in minutes;
6. AI Assistant Compatibility: Supports Codex and Claude-Code during the interaction phase.

Application Scenarios: Personal Decision-Making Examples

• Health Management: Analyze the effectiveness of three sleep supplements;
• Learning Efficiency: Compare the impact of early-morning vs. late-night study on test scores;
• Athletic Performance: Evaluate the contribution of interval running vs. long-distance jogging to VO2max;
• Mental Health: Explore the causal relationship between meditation app usage and anxiety scores.

Limitations: Faced Challenges

1. Data Quality Dependence: Results are highly dependent on data quality and completeness;
2. Causal Identification Limitations: Some problems cannot identify effects from observational data;
3. Prior Sensitivity: Bayesian results are influenced by prior choices;
4. Technical Threshold: Understanding results requires certain statistical knowledge.

Summary: Causal Inference from Academia to Personal Life

N-of-1 Causal Lab brings rigorous causal inference from academia to personal life, allowing everyone to use data-driven approaches to answer their own questions. Its open-source nature allows any individual/developer to use and improve it without relying on giants. It is suitable for readers interested in personalized medicine, self-quantification, causal inference, and LLM applications to research and contribute.