• Sorted by Date • Classified by Publication Type • Classified by Topic • Grouped by Student (current) • Grouped by Former Students •
Yonatan Vernik, David Izhaki, Alexander Tuisov, Chana Weitman, Alexander Shleyfman, and Gal A. Kaminka. Personalized Medication Planning via Direct Domain Modeling and LLM-Generated Heuristics. In Proceedings of the ICAPS Workshop on Language Models for Planning (LM4Plan), 2026.
(unavailable)
Personalized medication planning involves selecting medications and determining a dosing schedule to achieve medical goals tailored to an individual patient. Previous work has shown that automated planners using domain-independent heuristics can generate personalized treatment plans when the domain and planning problems are modeled in PDDL+. However, this approach was practically limited to scenarios involving no more than seven medications---an unrealistic constraint in clinical settings. In this paper, we explore the use of automatically generated heuristics combined with general search as a method for scaling medication planning to levels that enable closer collaboration with clinicians. We model the domain using the Explicit Successor Generator (ESG) approach, in which the initial state, successor generation procedure, and goal test are implemented in an off-the-shelf programming language (Rust). A commercially available reasoning LLM is then used to generate a domain-specific heuristic that guides a fixed search algorithm (GBFS). The results show substantial improvements in both coverage and planning time. The approach scales to scenarios involving at least 28 medications, bringing automated medication planning significantly closer to practical clinical use.
@inproceedings{lm4planws,
title = {Personalized Medication Planning via Direct Domain Modeling and LLM-Generated Heuristics},
author = {Yonatan Vernik and David Izhaki and Alexander Tuisov and Chana Weitman and Alexander Shleyfman and Gal A. Kaminka},
booktitle = {Proceedings of the {ICAPS} Workshop on Language Models for Planning ({LM4Plan})},
year = {2026},
abstract = {Personalized medication planning involves selecting medications and determining a dosing schedule to achieve medical goals tailored to an individual patient.
Previous work has shown that automated planners using domain-independent heuristics can generate personalized treatment plans when the domain and planning problems are modeled in PDDL+. However, this approach was practically limited to scenarios involving no more than seven medications---an unrealistic constraint in clinical settings.
In this paper, we explore the use of automatically generated heuristics combined with general search as a method for scaling medication planning to levels that enable closer collaboration with clinicians. We model the domain using the Explicit Successor Generator (ESG) approach, in which the initial state, successor generation procedure, and goal test are implemented in an off-the-shelf programming language (Rust). A commercially available reasoning LLM is then used to generate a domain-specific heuristic that guides a fixed search algorithm (GBFS).
The results show substantial improvements in both coverage and planning time. The approach scales to scenarios involving at least 28 medications, bringing automated medication planning significantly closer to practical clinical use.
},
}
Generated by bib2html.pl (written by Patrick Riley ) on Wed Jun 03, 2026 16:27:46