💻 computer science

Individualized Machine Learning–Grounded Large Language Models for Within-Person Simulation of Physical Activity Psychological Mechanisms and Behavior in Adults with Prehypertension: A Methodological Evaluation

This methodological evaluation demonstrates that embedding individualized machine learning-derived behavioral rules into large language model prompts enhances the simulation of within-person psychological and physical activity patterns in adults with prehypertension, though the hybrid approach did not surpass the predictive accuracy of dedicated participant-specific machine learning models for structured exercise outcomes.

Original authors: Haoming Ma, Zihao Liu, Runyuan Pei, Zhining Yang, Yuyang Zhang, Meihua Piao

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haoming Ma, Zihao Liu, Runyuan Pei, Zhining Yang, Yuyang Zhang, Meihua Piao

Original paper licensed under CC BY 4.0 (https://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

Keeping the heart healthy often comes down to a simple, daily choice: moving the body. For adults whose blood pressure is already creeping up but has not yet crossed the line into full-blown hypertension, regular exercise is one of the most powerful tools to stop the progression. Yet, the decision to exercise is rarely a fixed habit. It is a fluid, daily negotiation between a person's mood, their energy, the weather, and the pressure of a busy schedule. One day, a person might feel motivated and run; the next, the same person might feel tired and skip it. Understanding these daily shifts is crucial for doctors and researchers who want to help people stay active, but capturing the unique rhythm of a single individual's life is incredibly difficult.

To solve this, scientists are exploring a new kind of digital assistant. Imagine a computer program that can learn the specific patterns of one person's life and then act as a mirror, simulating how that person might think and act on any given day. This is the goal of a recent study involving adults with prehypertension. The researchers wanted to see if they could teach a large language model—a type of artificial intelligence known for its ability to understand and generate human language—to predict daily exercise behavior. The twist was that they did not want the AI to guess based on general knowledge. Instead, they wanted to ground the AI in the hard data of that specific person's past, using a hybrid approach that combines the pattern-finding power of traditional machine learning with the conversational flexibility of modern AI.

The study began with forty-six adults living in Beijing. Over the course of twelve weeks, these participants reported their daily lives in detail. They tracked their psychological states, such as their motivation, confidence, and feelings of social support, alongside their actual exercise habits and external factors like the weather or whether it was a weekday or weekend. This created a rich, day-by-day record of how each person's internal world influenced their physical actions. The researchers then took this data and split it into two parts. First, they used a standard machine learning tool to analyze the history of each individual and extract specific, personal rules. For example, the tool might learn that for one specific person, a drop in confidence on a rainy Tuesday almost always leads to skipping a workout, while for another, a strong sense of social support overrides a bad mood.

These personal rules were then fed into a large language model. The researchers designed a test where the AI had to act as a simulator. Given a snapshot of a person's previous day and the current day's context, the AI was asked to predict two things: how the person's psychological state would change and whether they would exercise. The team tested different ways of giving instructions to the AI. They tried giving it just the raw data, they tried giving it the data plus the personal rules, and they tried showing it examples of how to make the prediction before asking it to do the work. They also tested different versions of the AI to see which one was best at this task.

The results offered a clear picture of what works and what does not in this new field. The study found that the AI performed best when it was given the full context of the person's life and when it was shown examples of how to make the prediction before being asked to do it. Crucially, the AI's predictions became much more accurate and consistent when the researchers included the specific, data-derived rules about that individual. When the AI was told, "Here is how this specific person usually reacts to stress," it was far better at guessing their next move than when it was left to rely on general knowledge. The researchers also discovered that describing these rules in plain, natural language worked better than using complex mathematical formulas. The AI understood the story of the person's habits better when the rules were told as a narrative rather than a code.

However, the study also drew a firm line in the sand regarding what this technology can currently achieve. While the AI improved significantly when grounded in personal data, it still could not outperform the traditional machine learning model when it came to the simple, structured task of predicting whether someone would exercise. The traditional model, which is designed specifically for this type of number-crunching, remained the most accurate predictor. This suggests that the large language model is not a replacement for the specialized tools that doctors might use to calculate risk. Instead, its value lies in something else: it can take the complex, hidden patterns found in a person's data and translate them into a readable, human-like simulation.

The researchers concluded that this hybrid approach is a promising step forward for creating "digital twins"—virtual versions of people that can be used to test different health strategies without risking real-world harm. By combining the precision of traditional data analysis with the flexible, narrative power of language models, scientists may eventually be able to build systems that help individuals understand their own behavior. These systems could one day offer personalized advice that feels less like a generic medical instruction and more like a conversation with a partner who truly knows the person's daily struggles and triumphs. For now, the work remains a methodological evaluation, a proof of concept that shows how to make artificial intelligence listen more closely to the unique story of a single human life.

Technical Summary: Individualized Machine Learning–Grounded Large Language Models for Within-Person Simulation

Problem Statement
Regular physical activity is critical for preventing the progression from prehypertension to hypertension, yet daily exercise behavior is highly dynamic, fluctuating with psychological states and contextual factors. Traditional behavioral models often emphasize between-person differences, failing to capture the within-person variation and time-varying contexts that drive daily decisions. While Large Language Models (LLMs) offer potential for person-specific behavioral simulation, unconstrained LLM outputs often lack grounding in observed individual patterns, leading to generic reasoning that may not reflect a specific user's empirical behavioral regularities. The core challenge is to determine whether LLMs can be effectively "grounded" using individualized data to simulate daily psychological changes and exercise behavior without simply replacing established machine learning (ML) prediction models.

Methodology
The study employed a methodological evaluation using intensive longitudinal data from 46 adults with prehypertension collected over a 12-week observational period. The framework integrated conventional ML with LLMs in a hybrid design:

Data Structure: Daily records included 12 psychological variables (e.g., intention, self-efficacy, habit strength, outcome expectancies), prior-day exercise behavior, and same-day contextual factors (weather, weekday/weekend, holidays).
Individualized ML Rule Extraction: For each participant, a personalized Random Forest (RF) model was trained on temporally ordered historical data (80% training, 20% held-out test) to predict moderate-to-high-intensity exercise. From these models, two types of knowledge were extracted:
- Feature Importance: Identifying key drivers for that specific individual.
- Decision Rules: Translated into human-readable, natural-language statements describing conditional patterns (e.g., how specific psychological states correlate with exercise for that person).
LLM Simulation Framework: The LLM was tasked with two sequential prediction tasks:
- Task 1: Predicting the directional change (increase, decrease, or no change) of the 12 psychological variables from day $t-1$ to day $t$ .
- Task 2: Predicting whether the participant engaged in moderate-to-high-intensity exercise on day $t$ .
Experimental Variables: The study compared various configurations:
- LLMs: GPT-5, Claude-haiku-4.5, Gemini-2.5-pro, and DeepSeek-V3.
- Prompting Strategies: Zero-shot, Few-shot, and No-Chain-of-Thought (direct prediction).
- Prompt Components: Full prompts vs. reduced prompts (ablation studies).
- Rule Representation: Natural-language descriptions of ML rules vs. formula-based representations.
- Feature Subsets: Using all 12 psychological variables vs. subsets (top 5 or 10).
- Thresholds: Sensitivity analysis on the definition of "change" (using a 10-point threshold on 0–100 scales).

Key Results

Performance Hierarchy: The participant-specific Random Forest benchmark significantly outperformed all LLM conditions for structured exercise prediction (RF Accuracy: 0.878 vs. best LLM Accuracy: 0.650). This confirms that LLMs are not superior to conventional ML for pure structured prediction tasks in this domain.
Impact of Grounding: Embedding individualized ML-derived rules into LLM prompts improved simulation fidelity compared to prompts containing only profile and context data.
Optimal Configurations:
- Prompting Strategy: Few-shot prompting generally yielded the strongest performance across models.
- Rule Representation: Natural-language descriptions of ML rules outperformed formula-based or direct mathematical representations.
- Prompt Completeness: Full prompts (including all components) outperformed reduced prompts.
- Feature Selection: Using all 12 psychological variables yielded better performance than reduced subsets.
Model Specifics:
- Exercise Prediction: Claude-haiku-4.5 achieved the highest accuracy (0.650) among LLMs.
- Psychological Change Prediction: DeepSeek-V3 achieved the highest accuracy (0.626) among LLMs.
Threshold Sensitivity: Higher thresholds for defining psychological change were associated with higher prediction accuracy, indicating that outcome definitions significantly impact simulation metrics.

Key Contributions
The primary contribution of this work is the development and evaluation of a hybrid ML-LLM framework that uses conventional machine learning to extract person-specific empirical regularities and then grounds LLM prompts with these rules. The study demonstrates that:

LLMs can be effectively constrained by individualized data to simulate within-person behavioral dynamics, reducing generic reasoning.
The form of the grounding matters: natural-language rule descriptions are more effective for LLMs than mathematical formulas.
This approach creates an interpretable simulation layer that integrates participant history, psychological states, and context, distinct from the statistical regularity detection of pure ML models.

Significance and Claims
The authors position this work as a methodological evaluation rather than a demonstration of LLM superiority over conventional ML. The significance lies in:

Grounding LLMs in Health: Addressing the risk of hallucination and generic outputs in healthcare by anchoring LLMs in participant-specific evidence derived from longitudinal data.
Digital Twin and JITAI Potential: Providing a computational intermediate layer for future "digital twin" architectures and Just-in-Time Adaptive Interventions (JITAIs). In this vision, ML models extract the "rules," and LLMs translate these into interpretable constraints for simulating future states or configuring personalized behavior strategies.
Interpretability: Offering a simulation process that is more human-readable and context-aware than a "black box" statistical model, potentially aiding in the design of personalized health interventions.

The paper explicitly states that the framework did not outperform participant-specific ML for structured prediction and that its value is in the translation of empirical regularities into an interpretable simulation layer, not in raw predictive accuracy. Future work is needed to prospectively validate whether these simulations improve real-world health outcomes or intervention timing.

Technical Summary: Individualized Machine Learning–Grounded Large Language Models for Within-Person Simulation

More like this