🤖 machine learning

FeatEHR-LLM: Leveraging Large Language Models for Feature Engineering in Electronic Health Records

FeatEHR-LLM is a privacy-preserving framework that utilizes Large Language Models and tool-augmented generation to automatically produce clinically meaningful, executable feature-extraction code from irregularly sampled and sparse Electronic Health Record (EHR) time series.

Original authors: Hojjat Karami, David Atienza, Jean-Philippe Thiran, Anisoara Ionescu

Published 2026-04-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hojjat Karami, David Atienza, Jean-Philippe Thiran, Anisoara Ionescu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master chef trying to create the perfect recipe for a complex, multi-course meal. However, there’s a catch: your ingredients arrive at your kitchen at completely random times. Sometimes the salt arrives in five minutes; sometimes the steak doesn't show up for three hours. Even worse, some ingredients are missing entirely, and you don't know if they were forgotten or if they just weren't needed for that specific dish.

In the world of medicine, Electronic Health Records (EHRs) are like those chaotic ingredients. A patient’s data (heart rate, blood sugar, oxygen levels) doesn't come in a neat, rhythmic beat. It’s "irregularly sampled"—doctors only take measurements when something happens. This makes it incredibly hard for standard computers to "cook" a reliable prediction about whether a patient is getting better or worse.

This paper introduces FeatEHR-LLM, a new way to organize this kitchen chaos using the "brainpower" of Large Language Models (like ChatGPT).

The Problem: The "Messy Kitchen" of Medical Data

Most AI models are like rigid robots. They expect data to be perfectly organized in rows and columns, like a neatly laid out spreadsheet. When they encounter the "messy kitchen" of real hospital data—where a heart rate might be recorded every minute, but a blood test only happens once a day—the robots get confused. They either try to "guess" the missing data (which can be dangerous) or they simply fail to see the patterns.

The Solution: The "Expert Sous-Chef" (FeatEHR-LLM)

Instead of forcing a rigid robot to work, the researchers decided to hire an Expert Sous-Chef (the LLM).

Here is how the Sous-Chef works:

Privacy First (The "Recipe-Only" Rule): To protect patient privacy, the Sous-Chef never actually sees the "raw meat" (the actual sensitive patient names or private records). Instead, they are only shown the "Menu" (the Data Schema). They see: "We have a variable called 'Heart Rate' measured in beats per minute." This way, the AI learns how to cook without ever knowing who the diner is.
Writing the Instructions (Code Generation): Instead of doing the math themselves, the Sous-Chef writes customized instruction manuals (Python code). They say: "If the heart rate is jumping up and down wildly every ten minutes, write down a 'Stability Score' of 2."
Univariate & Multivariate Thinking (The "Flavor Profile"):
- Univariate (Single Ingredient): The Sous-Chef looks at one thing at a time. "How much is the sugar level fluctuating?"
- Multivariate (The Whole Dish): This is where the magic happens. The Sous-Chef understands how ingredients interact. They might ask: "How is the relationship between blood pressure and oxygen levels changing?" This is like a chef knowing that if the sauce is too salty, you need more acid to balance it.
The Validation Loop (The "Taste Test"): The Sous-Chef writes the code, runs it, and if the code crashes or makes no sense, they fix it. They keep refining their "recipes" until they work perfectly.

Why does this matter?

By using an LLM to turn messy, irregular time-series data into clean, meaningful "features" (like "Organ Stability Score" or "Respiratory Stress Level"), the researchers found they could predict patient outcomes much more accurately.

In their tests, this method beat almost all the old "robotic" ways of doing things, sometimes improving accuracy by as much as 6%. In a hospital setting, a 6% improvement in predicting a crisis can literally be the difference between life and death.

Summary in a Nutshell

FeatEHR-LLM takes a "smart" AI, shows it the structure of medical data (but not the private details), and asks it to write clever, custom math formulas that can handle the messy, unpredictable timing of real-world hospital monitoring. It turns a chaotic pile of timestamps into a clear, organized dashboard for doctors.

Technical Summary: FeatEHR-LLM

FeatEHR-LLM: Leveraging Large Language Models for Feature Engineering in Electronic Health Records

1. Problem Statement

Feature engineering for Electronic Health Records (EHR) is uniquely difficult due to the nature of clinical time-series data. Unlike standard sensor data, EHR data is characterized by:

Irregular Sampling: Measurements are taken at non-uniform intervals, often driven by clinical events rather than fixed schedules.
Structural Sparsity: Missingness is often "informative" (e.g., a test is not ordered because the patient is stable) rather than purely statistical.
Complexity: Clinical signals involve complex physiological interactions across multiple variables (e.g., the relationship between pH, $PaCO_2$ , and respiratory rate).

Existing automated methods (like tsfresh or OpenFE) either lack clinical domain awareness—leading to a "curse of dimensionality" with meaningless features—or assume clean, regularly sampled data, making them unsuitable for real-world ICU settings.

2. Methodology

The authors propose FeatEHR-LLM, a framework that uses the reasoning capabilities of Large Language Models (LLMs) to generate executable, clinically motivated Python code for feature extraction.

Key Architectural Pillars:

Privacy-Preserving Metadata-Level Operation: To protect patient privacy, the LLM never sees raw patient records. It only receives the dataset schema (variable names, units, summary statistics) and the clinical task description.
Tool-Augmented Generation: The LLM is equipped with a specialized library of programmatic tools (e.g., is_variable_increasing, is_variable_unstable, count_missing_intervals). This allows the LLM to write code that explicitly handles irregular temporal patterns and structural sparsity.
Two-Stage Feature Generation:
1. Univariate Engineering: The LLM proposes functions for individual variables (e.g., calculating the volatility of urine output) based on the variable's metadata.
2. Multivariate Engineering: The LLM first generates high-level clinical questions (e.g., "How compromised is the microvascular perfusion?") and identifies the necessary variables. It then writes code to compute scores (0–10) or boolean values based on those multi-variable interactions.
Validation-in-the-Loop: The framework includes a robust pipeline for syntax verification and runtime error handling. It uses an iterative strategy to refine features based on their utility.

3. Key Contributions

A Novel Framework for Irregular EHRs: Bridges the gap between LLM reasoning and the technical complexities of irregular, sparse clinical time series.
Clinical-Semantic Feature Synthesis: Moves beyond mathematical transformations to "semantic" transformations that reflect medical logic (e.g., acid-base balance or metabolic stress).
Privacy-First Design: Demonstrates that high-performance feature engineering can be achieved without exposing sensitive raw data to the LLM.
Hybrid Engineering: Combines univariate statistical trends with multivariate organ-system/disease-level reasoning.

4. Results

The framework was evaluated on eight clinical prediction tasks across four major ICU datasets (PhysioNet 2012, PhysioNet 2019, MIMIC-III, and eICU-CRD).

Superior Performance: FeatEHR-LLM achieved the highest mean AUROC on 7 out of 8 tasks, outperforming strong baselines like tsfresh, OpenFE, and other LLM-based methods (CAAFE, FeatLLM).
Significant Gains: Improvements of up to 6 percentage points in AUROC were observed over baseline statistical features.
Robustness:
- Data Efficiency: The framework showed consistent gains even in low-resource settings (using only 10% of the data).
- Generalizability: Performance gains were consistent across different backbone LLMs (Gemini, GPT-4o-mini) and various downstream classifiers (LightGBM, XGBoost, CatBoost).

5. Significance

FeatEHR-LLM represents a significant step toward interpretable AI in healthcare. By generating explicit Python code rather than opaque latent embeddings, the features produced are human-readable and auditable by clinicians. This addresses a major barrier to the clinical deployment of machine learning: the "black box" problem. Furthermore, by automating the most labor-intensive part of the ML pipeline—feature engineering—it allows researchers to rapidly prototype complex clinical models that are both high-performing and clinically grounded.