A proposed artificial intelligence framework for optimizing medication treatment would give AI an advisory role while leaving responsibility for medication decisions with clinicians. The work is a conceptual design, not a report of a tested clinical system. It sets out how support could be organized, but does not establish that the framework improves care in practice.
The stated aim was to develop a relevance-driven, clinician-supervised hybrid AI framework for pharmacotherapy optimization. Hybrid AI, as defined in the paper, combines retrieval-augmented generation, a way to bring relevant evidence into the system for use, with deterministic safety rules, or fixed checks, and large-language-model reasoning. This combination is part of the proposed design, not a performance finding.
The framework was developed through an iterative, design-science-informed process. An interdisciplinary research group used repeated discussions to shape the design, rather than following a fully predefined design-science process.
Seven ideas behind the blueprint
The resulting blueprint rests on seven design principles. They call for clinical tasks to be broken into parts, information to be prioritized by relevance, and hybrid reasoning to operate under clinician oversight. They also require the system to include patient goals, identify its evidence sources clearly, support longitudinal medication optimization through a governed closed loop, and treat evaluation as part of the design itself.
The human role is explicit. AI is assigned an advisory, non-autonomous role, while professional responsibility for medication decisions remains with the clinician. The design keeps the final medication decision within clinical practice rather than assigning it to the system.
A system built to show its work
At the architectural level, the principles were translated into a framework integrating structured clinical data, patient preferences, longitudinal patient information and evidence retrieval within clinician-governed decision support. Patient preferences are treated as an input to optimization alongside clinical records and information collected over time.
The proposal also includes a governed closed loop. It would document clinician actions, patient feedback and observed outcomes, supporting transparency and accountability while allowing the system to be refined iteratively. Longitudinal optimization is therefore described as an ongoing process rather than a single recommendation.
The evidence gap
But the framework remains untested. The paper reports no validation, implementation outcomes or performance metrics for the proposed framework or model. That leaves no empirical basis in this paper for deciding how well the design would work in practice.
The process itself also has limits. The iterative discussions were not predefined, formally documented or independently validated, which limits reproducibility. Patients were not directly involved in the framework's development, review or appraisal. This makes the article a record of a proposed design process, not an independently repeatable evaluation of the framework.
The seven principles and architecture are proposals, and the paper does not report evidence that the proposed AI components perform reliably in practice. Questions about implementation and performance remain open until the framework is tested and its outcomes measured.
The paper's status and disclosures
The report says institutional requirements did not call for formal ethics approval for this conceptual framework development. Open-access funding was provided by Paracelsus Medical University, and all listed authors declared no conflict of interest. The dataset is available on request from the corresponding author. The front matter reports receipt on 25 March 2026 and acceptance on 22 July 2026.
Paper data and sources
Original title: Development of a hybrid artificial intelligence framework for pharmacotherapy optimization.
Authors: Olaf Rose, Stephanie Clemens, Andreas Leiherer et al.
Journal/Repository: International journal of clinical pharmacy
Status: Peer-reviewed
First online: 2026-08-21
DOI: 10.1007/s11096-026-02205-0
Original paper · Full text