Re-thinking the Nature of Planning for Safe and Personalized Treatment Management Planning using Large Language Models

Authors: Ayoub et al.
Conference: IEEE International Conference on Bioinformatics and Biomedicine (BIBM 2025)
Keywords: Clinical AI, Large Language Models, Treatment Planning, Personalized Medicine, Reasoning, Evaluation Dataset

---

🧠 Overview

This repository accompanies our IEEE BIBM 2025 paper
"Re-thinking the Nature of Planning for Safe and Personalized Treatment Management Planning using Large Language Models".

While Large Language Models (LLMs) have made substantial progress in diagnostic reasoning, their capacity to deliver safe, accurate, and personalized treatment management plans remains limited.

Our work introduces a new planning paradigm that redefines how LLMs conceptualize treatment—not as a fixed goal optimization task, but as the systemic modulation of a patient’s illness state toward a healthier state.

---

🖼️ Conceptual Framework

Below is an illustration of our proposed method:

Figure: Overview of the proposed illness planning process: Dataset construction process (a) and proposed planning method (b) used in this study.

---

🚀 Core Contributions

🧩 1. Novel Evaluation Dataset

We release an evaluation dataset of 1,015 real-world patient cases, each containing:

• Complex, realistic illness narratives
• Expert-authored field-specific treatment management plans

This dataset supports rigorous benchmarking of LLMs’ ability to generate clinically safe, context-aware, and personalized treatment plans.

🧭 2. Novel Planning Method

We propose a new formulation of treatment planning as a field modulation process, introducing three core components:

• Attractive Tendencies: Latent directional vectors representing desirable shifts toward healthier states.
• Field Mapping: Identification of modifiable life and health domains to define personalized illness management goals.
• Field Sculpting: Generation of actionable, safe, and individualized interventions to guide patients toward resilient health trajectories.

📊 3. Strong Empirical Performance

Our approach achieves substantial performance gains over baseline reasoning methods (e.g., CoT, Reflexion):

• +12 BLEU-4
• +11 METEOR
• Maintains high clinical relevance and computational efficiency

---

📘 Citation

If you use this work, dataset, or methodology, please cite our IEEE BIBM 2025 paper:

@inproceedings{Ayoub2025BIBM,
  title={Re-thinking the Nature of Planning for Safe and Personalized Treatment Management Planning using Large Language Models},
  author={Ayoub et al.},
  booktitle={IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
  year={2025}
}