Design principles

The framework development process resulted in seven design principles for AI-supported pharmacotherapy optimization. These principles were derived from prior feasibility work, published literature on clinical decision support and AI in healthcare, recurring challenges in pharmacotherapy optimization, and iterative interdisciplinary discussions during framework development.

1.

Decomposition of clinical activities.

Pharmacotherapy optimization consists of distinct clinical activities with different risk profiles, information requirements, and frequencies [12, 15, 26]. Decomposition may facilitate meaningful digital support and avoid indiscriminate automation of complex clinical decision-making [26]. For example, a comprehensive medication review can be decomposed into distinct tasks such as acquiring data, performing the patient assessment, setting therapeutic goals, assessing medication safety or evaluation of guideline adherence. Such decomposition enables targeted digital support while preserving clinician oversight of the overall decision-making process.

2.

Relevance-based prioritization.

CDSS should prioritize information according to clinical relevance, patient vulnerability, and context-specific actionability. Within this framework, AI structures clinical relevance by identifying issues that most urgently require professional judgement rather than replacing decision-making. For example, a clinically significant drug-related problem associated with serious harm would be prioritized over minor informational alerts, allowing decision support to focus attention on issues most relevant to patient outcomes and therapeutic decision-making [20].

3.

Hybrid reasoning under human oversight.

Safe pharmacotherapy decision support may benefit from hybrid reasoning that combines deterministic safety rules, established risk models, retrieval-augmented evidence, and contextual AI-based synthesis [20]. AI augments but does not replace clinical reasoning, and human oversight remains essential.

4.

Explicit integration of patient goals and treatment burden.

Clinical relevance depends on patient priorities, including treatment goals, tolerability thresholds, symptom burden, and adherence barriers. Explicit integration of these factors is necessary to avoid technically correct but clinically misaligned recommendations.

5.

Transparency and retained professional authority.

AI-generated recommendations must remain advisory, with clinicians retaining full authority to accept, modify, or reject them. Transparent presentation of underlying evidence supports verification, accountability, and informed decision-making [20].

6.

Longitudinal pharmacotherapy optimization within a governed closed loop.

Pharmacotherapy optimization is dynamic and requires systems that capture outcomes, overrides, and safety events to enable iterative refinement. Continuous learning must occur within defined governance structures to ensure auditability, compliance, and trust.

7.

Evaluation as a design requirement.

Evaluation should be embedded in system design and initially focus on usability, workflow integration, transparency, and reliability of recommendations. Only after stable use is demonstrated should evaluation extend to clinical outcomes and patient-level effects.

Conceptual framework

This framework is intended to inform governance, design, and evaluation of AI-supported pharmacotherapy rather than to prescribe a specific technical implementation. The conceptual framework described below operationalizes the design principles outlined above by translating them into an integrated architecture for AI-supported pharmacotherapy optimization. It positions hybrid AI as an augmentation layer in pharmacotherapy, supporting clinicians in managing therapeutic complexity, reducing cognitive and administrative burden, and enabling focus on patient interaction. To this end, the framework proposes the integration of structured clinical data, guideline-based evidence, and patient-specific preferences to support medication-related clinical decision-making.

Information foundation and multimodal inputs

Pharmacotherapy optimization requires a longitudinal and integrative clinical perspective. Relevant information extends beyond the current medication list and may, depending on availability, include diagnoses, laboratory values, renal and hepatic function, vital signs, pharmacogenomics, comorbidities, and prior treatment courses. Increasing availability of EHRs enables access to many of these data points, while patient preferences and symptom trajectories provide complementary insight into lived treatment experience. Clinical guidelines and safety recommendations frequently convey therapeutic logic through flow charts, tables, and visual algorithms. Consequently, the proposed framework is based on a multimodal information foundation that integrates structured clinical data, unstructured clinical narratives via natural language processing, patient-reported information, and guideline artefacts via computer vision-based representation learning. By aligning heterogeneous information sources within a unified representation, the system aims to support context-aware reasoning while minimizing additional documentation burden for clinicians.

Dual engine “hybrid” reasoning: combination with knowledge base

Safe and clinically meaningful recommendations require explicit grounding in established therapeutic evidence. RAG enables LLM to reference up—to-date guideline content, product information, safety communications, and curated evidence repositories at the point of care [27, 28]. Within the proposed framework, hybrid reasoning combines contextual AI-based synthesis with deterministic safety rules and established risk models. Together, these components are intended to support transparent and evidence-anchored recommendations (Fig. 2). Uncertainty signaling is integral to the system design. When data are incomplete or evidence is limited, the system should explicitly communicate uncertainty and prompt clinical verification. Hybrid reasoning is intended to augment, not replace, professional judgment. To support reliable and traceable evidence retrieval, source documents such as clinical guidelines require systematic preparation. Large documents are converted into machine-readable formats, stripped of non-informative elements, and segmented into semantically coherent units (chunking). These units are indexed individually to enable precise contextual retrieval.

Fig. 2

Example of a user interface with dual engine reasoning, providing the source of the recommendation. CCS: chronic coronary syndrome, PCSK9: proprotein convertase subtilisin/kexin type 9, PRN: pro re nata/as needed

Integration of patient goals, patient preferences and treatment burden

Modern pharmacotherapy requires alignment with patient goals, values, tolerability and daily capabilities. Adherence is influenced by symptom severity, regimen complexity, health literacy and expectations. Therefore, the framework integrates structured documentation of patient goals and burden thresholds, such as preferred treatment intensity, acceptable side effect levels, priorities related to cognition, mobility or sleep and willingness to engage in monitoring. Incorporating these parameters ensures personalized alignment rather than algorithm-centric optimization. For CDSS, this important information on complaints, preferences and treatment goals needs to be translated into relevant and contextualized prompts. A digital scale is required to prepare these inputs for the AI system. However, patient-reported inputs may be incomplete or imprecise, and data collected by wearables or other devices are often unspecific and not routinely integrated into medication reviews. An application on a mobile device, for example an avatar-based digital assistant, could elicit medication-related preferences, collect longitudinal data, and complement information obtained during patient-clinician encounters. This digital health assistant may also provide guidance when predefined red-flag patterns are detected.

Human oversight, autonomy and accountability

Professional responsibility for medication decisions remains with the clinician. AI within the proposed framework serves an advisory function and does not operate autonomously. Pharmacists, physicians and other health professionals retain full authority to accept, modify, or reject system-generated recommendations based on clinical judgment, patient context, and interprofessional considerations. To support accountable oversight, recommendations are accompanied by transparent reasoning, linked evidence sources, and concise rationale summaries. Explicit uncertainty signaling informs clinicians when a system-generated recommendation should not be applied directly without further review. In such cases, the signal prompts clinicians to reassess the underlying data, consider missing patient information, or seek additional evidence before making a final decision. These mechanisms are intended to preserve professional agency while enhancing efficiency and situational awareness. By embedding decision authority, explanation, and documentation within existing workflows, the framework aligns digital support with ethical obligations, regulatory expectations, and established standards of professional practice [19]. Moreover, the framework enables clinicians to provide feedback on system recommendations, which can be used to evaluate the model’s performance, facilitate learning, and continuously validate and refine the AI, ensuring its ongoing improvement.

The clinician governed closed loop AI model

Pharmacotherapy optimization is inherently dynamic. Patient conditions evolve, therapies change, and new information emerges over time. The proposed framework therefore suggests to adopt a closed-loop architecture that supports longitudinal monitoring and iterative refinement of recommendations. As illustrated in Fig. 3, patient-reported information on symptoms, goals, well-being, medication-related problems, and adherence could be captured longitudinally through digital interfaces.

Fig. 3

Example of a digital patient feedback panel, which can be augmented with a talking avatar

These inputs complement data obtained during clinical encounters and enable early identification of emerging issues that may otherwise remain unrecognized between visits. Beyond a standard patient-facing interface, an avatar-based assistant may facilitate a more structured and continuous capture of patient-reported information. It can add information for the categorization of chief complaints and treatment-related concerns, translate them into clinically relevant representations, and enhance patient understanding of therapy and medication use. By providing tailored guidance and timely feedback, it may contribute to improved adherence. In addition, the system can generate alerts and prompt referrals to clinicians when predefined red-flag patterns are detected. Information retrieved through the interface is fed back into the AI model, thereby closing the loop for pharmacotherapy optimization.

Multimodal clinical inputs, including structured electronic health record data, laboratory results, medication plans, patient preference measures, and guideline artefacts, would be continuously integrated into the system. Evidence layers based on RAG access to clinical guidelines, summaries of product characteristics, and safety communications inform a hybrid reasoning process that combines contextual AI-based synthesis with classical risk models and uncertainty scoring. A conceptual operationalization of the proposed hybrid AI architecture, including the closed-loop mechanism, is shown in Fig. 4.

Fig. 4

Patient-centered hybrid AI model designed as a clinician-governed closed loop. ADR: adverse drug reaction, AI: artificial intelligence, LLM: large language model, ML: machine learning, RAG: retrieval augmented generation

System-generated therapy recommendations remain subject to human oversight. Digital support can facilitate adherence reminders and track symptoms. Patients may actively signal concerns or alerts to the care team through the same interface. Clinician actions, patient feedback, and observed outcomes are documented to support transparency, accountability, and iterative system refinement. Within defined governance structures, this feedback enables continuous learning on priorities, while ensuring that adaptation remains evidence-grounded, auditable, and aligned with professional standards rather than uncontrolled self-optimization.

Safety, audit and continuous learning

Robust safety mechanisms are a foundational requirement for digital decision support in pharmacotherapy optimization. Within the proposed framework, safety guardrails include verification of doses, contraindications, and clinically relevant drug-drug interactions, as well as mechanisms to limit hallucinations and unsupported inferences in AI-assisted reasoning, thereby providing the required level of reliability and helping to uphold ethical standards in clinical decision making [19]. All clinician interactions with system-generated recommendations, including overrides, modifications, errors, and adverse events, should be systematically documented. This audit trail may enable traceability, accountability, and structured safety monitoring across real-world use. Feedback loops can support iterative improvement based on observed performance while preserving transparency and professional responsibility. Continuous learning can be constrained by governance structures. System adaptation ideally remains evidence-grounded and does not rely on unconstrained data generation or autonomous self-optimization. Deployment within privacy-preserving, jurisdiction-compliant infrastructures and centrally validated, version-controlled updates should ensure regulatory alignment, auditability, and prevention of uncontrolled model drift. By combining safety verification, auditability, and supervised AI, the framework aims to increase transparency, strengthen clinician trust, and support safe longitudinal use in routine clinical practice.