Open Foundation Models in Healthcare: Challenges, Paradoxes, and Opportunities with GenAI Driven Personalized Prescription

Mahdi Alkaeed; Sofiat Abioye; Adnan Qayyum; Yosra Magdi Mekki; Ilhem Berrou; Mohamed Abdallah; Ala Al-Fuqaha; Muhammad Bilal

cs.CLFeb 4, 2025

Open Foundation Models in Healthcare: Challenges, Paradoxes, and Opportunities with GenAI Driven Personalized Prescription

Authors: Mahdi Alkaeed, Sofiat Abioye, Adnan Qayyum, Yosra Magdi Mekki, Ilhem Berrou, Mohamed Abdallah, Ala Al-Fuqaha, Muhammad Bilal,

Organizations: Department of Computer Science and Engineering, College of Engineering, Qatar University, Doha, Qatar · Birmingham City University, Birmingham, United Kingdom · College of Science and Engineering, Hamad Bin Khalifa University (HBKU), Doha, Qatar · College of Medicine, Qatar University, Doha, Qatar ·

Abstract

In response to the success of proprietary Large Language Models (LLMs), there is a growing interest in developing open, non-proprietary LLMs and AI foundation models (AIFMs) for transparent use in academic, scientific, and non-commercial applications. Open models have substantial potential to improve healthcare applications. In this paper, we provide a comprehensive survey and a taxonomy of state-of-the-art open LLMs and AIFMs for clinical tasks. In addition, we empirically evaluate open-weight LLMs on a personalized-prescription task and compare them with a proprietary model with and without Retrieval-Augmented Generation (RAG). Across 100 synthetic patient profiles, the open-weight models matched or exceeded the proprietary baseline on overall suitability accuracy without retrieval augmentation (0.810--0.850 versus 0.780), although the confidence intervals overlap; with retrieval augmentation, every model improved, and the ranking closed to within 0.04. Gains from RAG were field-dependent rather than uniform. To highlight the clinical relevance of LLM-driven personalized prescriptions, we also conduct a subjective evaluation in which two raters, blinded to model and condition, scored responses independently of each other. Their ratings did not follow the accuracy gains: retrieval raised overall suitability accuracy for every model while lowering the two-rater means for two of the four on all five dimensions. Finally, we outline the ethical implications and potential risks associated with the misuse of advanced LLMs and AIFMs, highlighting the importance of responsible and transparent deployment in healthcare contexts. \textit{Data and code used in the paper will be released upon publication.

Figures & tables

Ref.	Type	Domain	Primary Focus	OS	OS FM Tax.	Emp. CS	OS vs. CS	Human
He et al. [ 14 ]	Methodological review	Healthcare	Provides a comprehensive review of healthcare FMs, including multi-modal FMs.	✗	$\approx$	✗	✗	✗
Khan et al. [ 15 ]	Methodological review	Medicine	Provides a taxonomy of FMs in medicine and discusses their challenges	✗	$\approx$	✗	✗	✗
Mahesh et al. [ 16 ]	Review	Healthcare	Explores the adoption of FMs in healthcare along with challenges and ethical considerations.	✗	✗	✗	✗	✗
Li et al. [ 17 ]	Review	Biomedical Healthcare	Surveys FMs with federated learning for biomedical applications.	$\approx$	$\approx$	✗	✗	✗
Zhang et al. [ 18 ]	Survey	Computational healthcare (CH)	Data-centric FMs in CH applications and associated challenges	✗	$\approx$	✗	✗	✗
Bilal et al. [ 19 ]	Survey	Pathology	Overviews applications of FMs in computational pathology	✗	✗	✗	✗	✗

TABLE I: Comparison of related works on foundation models in healthcare with our paper. Legend: OS

\rightarrow

Open Source; FMs

\rightarrow

Foundation Models; Tax.

\rightarrow

Taxonomy; Emp. CS

\rightarrow

Empirical Case Study; OS vs. CS

\rightarrow

Open-Source vs. Closed-Source benchmarking; Human

\rightarrow

human or expert evaluation of model outputs reported; ✓

\rightarrow

covered;

\approx

\rightarrow

partially covered; ✗

\rightarrow

not covered.

Fig. 1: Levels of Open-Source Access for LLMs and AIFMs. The figure shows the spectrum from Fully Closed Source to Fully Open Source, detailing key aspects and descriptions for each level. (Adapted from [ 28 ] ).

Fig. 2: Applications of LLMs in Healthcare. This figure shows six key areas: Clinical Documentation, Disease Prediction and Diagnosis, Drug Discovery, Virtual Health Assistants, NLP in EHRs, and Telemedicine Support.

Model (year)	Released params	Ref.	W	license	C	D	OSAID
Biomedical text encoders and generators
BioBERT (2020)	110M, 340M	[ 69 ]	✓	none	$\approx$	✗	✗
BioMegatron (2020)	345M	[ 113 ]	✓	bsp	✓	✗	✗
BioELECTRA (2021)	110M	[ 114 ]	✓	none	✗	✗	✗
KeBioLM (2021)	n/s	[ 115 ]	✓	none	$\approx$	✗	✗
BioBART (2022)	140M, 406M	[ 116 ]	✓	osi	✓	✗	✗

TABLE II: Open AIFMs for healthcare applications. W : weights ( ✓ public, ◗ gated, ✗ not released). license : weight licensing ( osi OSI-approved, ccl community/use-restricted, nc non-commercial, nc-nd non-commercial/no-derivatives, bsp bespoke terms, none none declared, n/a not applicable). C : training code; D : training data;

\approx

partial or access-controlled release. OSAID : compliance with the Open Source AI Definition v1.0 [ 37 ] and “n/s” denotes not stated.

Dataset	Modality	Task type	Size	Access	License
Open access
Observed Antibody Space [ 154 ]	Antibody sequences	Protein modeling	1.5B unpaired sequences from 80 studies and paired sequences from 5 studies	Open	CC-BY-4.0
Big Fantastic Database (BFD) [ 155 , 156 ]	Protein sequences	Protein folding, MSA	Clustered from 2.5B sequences	Open	CC-BY-4.0
RNAcentral [ 157 ]	ncRNA sequences	Structure prediction	$>$ 45M sequences, $>$ 30M secondary structures (Release 26)	Open	CC0
ZINC20 [ 158 ]	Chemical compounds	Molecule generation	1.4B compounds, 310 catalogs, 150 companies	Open	Free-use terms
ZINC-22 [ 159 ]	Chemical compounds	Molecule generation	$>$ 37B compounds, $>$ 4.5B ready-to-dock 3D compounds	Open	Free-use terms

TABLE III: Publicly documented healthcare datasets used for LLM and AIFM development and evaluation, grouped by access tier. Open : anonymous download. Registered : free account or click-through agreement. Credentialed : identity verification, human-subjects training, and a signed data use agreement. Controlled : approval by a data access committee.

License class	Models	Share
OSI-approved (Apache-2.0, MIT)	16	35.6%
Bespoke terms of use	8	17.8%
Community license (use-restricted)	6	13.3%
Non-commercial	4	8.9%
Non-commercial, no derivatives	4	8.9%
None declared	6	13.3%

TABLE IV: Release openness of the 45 healthcare and healthcare-adjacent foundation models in Table II .

Fig. 3: Illustration of the proposed methodology for evaluating open-weight and proprietary LLMs using two settings, i.e., with and without Retrieval-Augmented Generation (RAG).

Context
You are an experienced and helpful prescribing assistant named Charlie. Charlie can support prescribers to carry out relevant checks when prescribing medication based on the patient’s medical profile and medical knowledge. You should tell if the medication is suitable for the user based on the following patient profile and provide the result in JSON format:
$<$ patient profile $>$
Instructions
For each check, use “Suitable”, “Risky”, or “N/A”, and provide a reason and evidence field. Cite the source filename and page when retrieved context is supplied. Return an overall result, a score from 0 (clearly unsuitable) to 100 (clearly suitable), and a reason.
Output Format (JSON)
{ "Age": {"result": <result>, "reason": <reason>}, "Dose": {"result": <result>, "reason": <reason>}, "Comorbidities": {"result": <result>, "reason": <reason>}, "Contraindications": {"result": <result>, "reason": <reason>}, "Pregnancy": {"result": <result>, "reason": <reason>}, "Lactation": {"result": <result>, "reason": <reason>}, "Warnings": {"result": <result>, "reason": <reason>}, "Genetics": {"result": <result>, "reason": <reason>}, "Overall Suitability": {"result": <result>, "score": <score>, "reason": <reason>} }

TABLE V: Medication Suitability Evaluation Prompt Template

Model	Field	No RAG				RAG
Model	Field	Acc.	P	R	F1	Acc.	P	R	F1
GPT-5.6 Luna	Age	0.880	0.872	0.813	0.836	0.810	0.586	0.479	0.517
	Comorbidities	0.940	0.961	0.897	0.922	0.860	0.611	0.533	0.562
	Contraindications	0.750	0.516	0.501	0.508	0.770	0.597	0.511	0.549
	Dose	0.640	0.736	0.709	0.638	0.820	0.600	0.553	0.572
	Genetics	0.950	0.500	0.475	0.487	0.980	0.500	0.490	0.495

TABLE VI: Field-level and overall-suitability performance on 100 synthetic profiles using the first of two runs. P, R, and F1 are macro-averaged. Mean field accuracy is the unweighted mean of the eight field accuracies.

Fig. 4: Two-reviewer mean scores (scale 0–5) for each of the 12 reviewed profiles. Each subplot is one rating dimension; the final panel averages the five dimensions.

Model	Cond.	MSA	DID	PSDA	PSS	GA
GPT-5.6 Luna	RAG	$4.83\pm 0.39$	$4.71\pm 0.45$	$4.54\pm 0.58$	$4.58\pm 0.56$	$4.54\pm 0.58$
	No RAG	$4.67\pm 0.49$	$4.58\pm 0.47$	$4.25\pm 0.62$	$4.46\pm 0.50$	$4.46\pm 0.50$
Llama 4 Scout	RAG	$4.75\pm 0.45$	$4.58\pm 0.47$	$4.29\pm 0.58$	$4.54\pm 0.54$	$4.50\pm 0.52$
	No RAG	$4.92\pm 0.29$	$4.79\pm 0.33$	$4.54\pm 0.50$	$4.67\pm 0.49$	$4.62\pm 0.43$
Mistral Small 3.2 24B Instruct	RAG	$4.92\pm 0.29$	$4.79\pm 0.40$	$4.58\pm 0.56$	$4.75\pm 0.45$	$4.75\pm 0.45$
	No RAG	$5.00\pm 0.00$	$4.79\pm 0.26$	$4.46\pm 0.45$	$4.62\pm 0.48$	$4.58\pm 0.47$

TABLE VII: Mean

\pm

SD of the two-reviewer mean score across 12 patient profiles (scale 0–5; observed scores 3–5). RAG = with Retrieval-Augmented Generation; No RAG = without.

Medical LLMs

Medical large language models are AI systems adapted to clinical and biomedical data for health reasoning, information extraction, and communication.

See all 396 papers

BIT.UA-AAUBS at ArchEHR-QA 2026: Evaluating Open-Source and Proprietary LLMs via Prompting in Low-Resource QA
May 5, 2026Richard A. A. Jonker, Alexander Christiansen, Alexandros Maniatis +4LLM Grounding Healthcare
Aalborg University Business School Fibigerstræde 2, Aalborg East 9220, Denmark · IEETA, DETI, LASI, University of Aveiro Campus Universitário de Santiago, Aveiro 3810-193, Portugal
Role-Specialized Mixture-of-Agents with Open-Weight LLMs for Clinical Prediction
Aug 23, 2026Jun Hou, Yi Fang, Xuan WangMixture-of-Agents Clinical Prediction
Virginia Tech
Prompt Sensitivity and Answer Consistency of Small Open-Source Language Models for Clinical Question Answering in Low-Resource Healthcare
Mar 1, 2026Shravani HariprasadClinical Language Model Evaluation Prompt Sensitivity
Independent Researcher

May 5, 2026J/K move · Enter open · S save

BIT.UA-AAUBS at ArchEHR-QA 2026: Evaluating Open-Source and Proprietary LLMs via Prompting in Low-Resource QA

Richard A. A. Jonker, Alexander Christiansen, Alexandros Maniatis, Rúben Garrido+3

Aalborg University Business School Fibigerstræde 2, Aalborg East 9220, Denmark · IEETA, DETI, LASI, University of Aveiro Campus Universitário de Santiago, Aveiro 3810-193, Portugal

LLM Grounding Healthcare LLM Prompting Clinical QA Clinical Language Model Evaluation

Open paper AskPDF arXiv

This paper presents the joint participation of the BIT.UA and AAUBS groups in the ArchEHR-QA 2026 shared task, which focuses on clinical question answering and evidence grounding in a low-resource setting. Due to the absence of training data and the strict data privacy constraints inherent to the healthcare domain (e.g. GDPR), we investigate the capabilities of Large Language Models (LLMs) without weight updates. We evaluate several state-of-the-art proprietary models and locally deployable open-source alternatives using various prompt engineering strategies, including task decomposition, Chain-of-Thought, and in-context learning. Furthermore, we explore majority voting and LLM-as-a-judge ensembling techniques to maximize predictive robustness. Our results demonstrate that while proprietary models exhibit strong resilience to prompt variations, domain-adapted open-source models (such as MedGemma 3 27B) achieve highly competitive performance when paired with the right prompt. Overall, our prompt-based approach proved highly effective, securing 1st place in Subtask 4 (evidence citation alignment) and 3rd place in Subtask 3 (patient-friendly answer generation). All code, results, and prompts are available on our GitHub repository: https://github.com/bioinformatics-ua/ArchEHR-QA-2026.

Abstract

Medical LLMs

Explore similar work

BIT.UA-AAUBS at ArchEHR-QA 2026: Evaluating Open-Source and Proprietary LLMs via Prompting in Low-Resource QA

Role-Specialized Mixture-of-Agents with Open-Weight LLMs for Clinical Prediction

Prompt Sensitivity and Answer Consistency of Small Open-Source Language Models for Clinical Question Answering in Low-Resource Healthcare