Rx Strategist: Modular Drug Verification

• Rx Strategist is a prescription verification system that combines LLM agents, knowledge graphs, and structured drug data to assess indication fit, dosage appropriateness, and drug interactions.
• The system employs a multi-stage pipeline—including OCR, ICD matching, and dosage retrieval—to deliver performance comparable to experienced clinical pharmacists with reduced computational demand.
• Beyond verification, Rx Strategist embodies a broader strategic framework for adaptive medication recommendation, causal treatment-effect estimation, and evidence generation, despite limitations in data coverage and localization.

Searching arXiv for papers directly about “Rx Strategist” and closely related uses of the term. “Rx Strategist” denotes several distinct but conceptually related uses in recent technical literature. In the most direct sense, it names a prescription-verification system that combines LLM agents, knowledge graphs, and structured drug data to evaluate indication fit, dosage fit, and potential drug interactions in prescription review (Van et al., 2024). In adjacent literatures, the term also functions as a strategic abstraction for fine-grained medication recommendation, causal treatment-effect estimation, dose optimization, and evidence generation, where “Rx” refers variously to prescriptions, treatment policies, or randomized designs (Chen et al., 14 May 2026, Wang et al., 27 May 2026, Uehara, 21 Jan 2026, Wang et al., 6 May 2025). Across these usages, the unifying theme is the replacement of monolithic decision rules with modular, staged procedures that combine domain knowledge, explicit constraints, and adaptive decision criteria.

1. Prescription verification system

The paper “Rx Strategist: Prescription Verification using LLM Agents System” defines Rx Strategist as a prescription-verification system that combines multiple LLM-based agents with structured drug knowledge and knowledge graphs to approach the performance of an experienced clinical pharmacist, while being more reliable than a single monolithic LLM (Van et al., 2024). It is designed for environments like Vietnamese hospitals, where the rate of incorrect doses can approach $40\%$ and pharmacist coverage is low (Van et al., 2024).

Its core task is to decide, for every active ingredient in a prescription, whether the indication is appropriate for the patient’s diagnosis, whether the prescribed dose is appropriate for the patient’s age and other characteristics, and to surface potential drug–drug interactions (Van et al., 2024). The system formalizes two internal states: fit of indication, $FI$, and fit of dosage, $FD$ (Van et al., 2024).

The architecture is a multi-stage pipeline comprising an information extractor, an ICD Finder, an ICD Matcher, a Dosage Retriever, and a Checker (Van et al., 2024). Upstream OCR converts prescription images to text, after which GPT-4o-mini reformats the content into a structured representation containing patient information, diagnosis, and medication details (Van et al., 2024). The ICD Finder maps active ingredients to therapeutic indications and then to ICD-10 codes, using fuzzy matching against a database of 1,780 active ingredients and LLM fallback when a drug is absent from the database (Van et al., 2024). The ICD Matcher compares the ICD-10 categories associated with each drug’s usages against the patient diagnosis categories and labels the ingredient APPROPRIATE or INAPPROPRIATE for indication (Van et al., 2024).

Dosage verification is handled through a dosage knowledge graph derived from Drugs.com AHFS DI monographs and Long Chau Pharmacy data, transformed from cleaned HTML and Markdown into structured JSON and then into graph nodes and relations such as TREATS, INITIAL_DOSAGE, and SPECIFIC_DOSAGE (Van et al., 2024). Retrieval uses the active ingredient, age group, diagnosis, and patient age to return dosage options from the graph; if the diagnosis text does not exactly match a graph disease, an LLM is used to match it semantically (Van et al., 2024). In the reported experimental setup, the system uses the “initial dosage” for each disease as the baseline recommended dose because parameters such as weight, renal function, and hepatic function are not consistently available in the prescription dataset (Van et al., 2024).

The Checker combines $FI$, $FD$, and interaction context to generate final verdicts and explanations per medication (Van et al., 2024). The evaluation prompt explicitly instructs the model to consider age, kidney failure, liver disease, pregnancy, and allergies when present (Van et al., 2024). This staged decomposition is presented as a remedy for hallucinations, shallow pattern matching, and the context overload associated with one-shot prompting (Van et al., 2024).

2. Knowledge grounding and retrieval structure

A defining property of Rx Strategist is its reliance on structured external knowledge rather than sole dependence on internal LLM recall. The active-ingredient dataset is assembled from Drugs.com and Long Chau Pharmacy and includes indications, administration, dosage, and adverse effects (Van et al., 2024). Dosage content is normalized by removing extraneous characters, standardizing hyphens, and converting micrograms to milligrams via regular expressions (Van et al., 2024).

The dosage knowledge graph encodes drugs, diseases, and dosage entities, with age group and administration route attached to edges (Van et al., 2024). Algorithm 1 in the paper constructs the graph by iterating over active elements, age groups, and diseases, then using a LLM to generate dosage nodes and relations (Van et al., 2024). Algorithm 2 retrieves dosage recommendations by searching the graph for diseases treated by an active ingredient in the relevant age group, mapping patient age to graph age ranges, and returning dosage options when available (Van et al., 2024).

A separate interaction knowledge graph is built from Drugs.com interaction data for 27 common active components (Van et al., 2024). LLaMA 3–8B extracts triplets from textual interaction descriptions, and the triplets are embedded using the gte-Qwen2-1.5B-instruct embedding model (Van et al., 2024). Retrieval operates by cosine similarity between a query embedding and precomputed triplet embeddings, after which the retrieved triplets are summarized and inserted into the evaluation prompt as reference material (Van et al., 2024).

This design places symbolic lookup, fuzzy matching, ICD normalization, graph traversal, and semantic summarization in separate modules rather than treating verification as a single generative step (Van et al., 2024). A plausible implication is that the term “Rx Strategist” designates not only a single model but a workflow in which retrieval, coding, comparison, and explanation are distributed across specialized operators.

3. Empirical performance and pharmacist comparison

Evaluation is conducted on 20 real-world prescriptions collected from Vietnamese hospitals, with personally identifiable information removed (Van et al., 2024). Three clinical pharmacists with 1, 3, and 5 years of experience independently label each active ingredient for indication appropriateness and, if appropriate, dosage appropriateness; a council of three highly experienced hospital pharmacists provides the gold standard (Van et al., 2024).

The reported metrics are Accuracy, Precision, Recall, and $F_{0.5}$, where $F_{0.5}$ is used because false positives—incorrectly labeling inappropriate medications as appropriate—are considered more harmful than false negatives (Van et al., 2024). The paper gives the standard $F_{\beta}$ form and specifies the $\beta=0.5$ case (Van et al., 2024).

The main results are as follows.

System	Accuracy	Precision	Recall
5-year pharmacist	75.93	81.01	85.33
GPT-4o-mini	70.37	73.12	90.67
Rx Strategist	75.93	82.67	82.67

Rx Strategist achieves Accuracy 75.93%, Precision 82.67%, Recall 82.67%, and $F_{0.5}$ 82.67 (Van et al., 2024). The 5-year pharmacist achieves Accuracy 75.93%, Precision 81.01%, Recall 85.33%, and $FI$0 81.84 (Van et al., 2024). The paper interprets this as performance comparable to, or slightly better than, a highly experienced clinical pharmacist in this benchmark setting (Van et al., 2024).

The system also outperforms several open- and closed-source LLM baselines, including LLaMA 3.1–8B, LLaMA 3.1–70B, Qwen2–72B, LLaMA 3.1–405B, Claude 3.5 Sonnet, and GPT-4o-mini, on the same task (Van et al., 2024). Integrating the interaction graph with small LLaMA models improves indication verification accuracy by $FI$1, bringing an 8B model to performance comparable to much larger models (Van et al., 2024).

In runtime terms, Rx Strategist generates far fewer tokens than the large-model baselines, with 1,223 total generated tokens versus roughly 13,000–15,000 for several comparison models, while achieving similar or better accuracy (Van et al., 2024). This supports the paper’s claim that staged reasoning and structured retrieval reduce memory demands and improve correctness (Van et al., 2024).

4. Limitations and boundary conditions

The paper identifies several limitations. Database and knowledge-graph coverage is incomplete; when active ingredients are not found in the 1,780-drug database, ICD Finder falls back to LLM-generated ICD-10 codes, which the authors note are less reliable and require additional validation (Van et al., 2024). Dosage modeling is limited by missing patient-specific variables such as weight, renal function, liver function, and pregnancy status, so initial monograph doses are treated as the reference baseline (Van et al., 2024). Interaction coverage is restricted to 27 common components, and interaction severity is not represented numerically (Van et al., 2024). The resources and evaluation setting are strongly localized to Vietnamese-language prescriptions and Vietnamese drug information sources, leaving generalization to other languages and formularies untested (Van et al., 2024).

These constraints distinguish the system from a full prescribing engine. The paper positions Rx Strategist as decision support rather than autonomous prescribing, emphasizing explanation-rich outputs and human oversight (Van et al., 2024). This is consistent with later medication-recommendation work that treats safety verification and traceability as separate stages rather than latent properties of a single generator (Wang et al., 27 May 2026).

5. Relation to prescription-level evaluation and medication recommendation

A second usage of “Rx Strategist” appears as a conceptual label for systems that support nuanced, patient-specific prescribing. The benchmark “RxEval” defines inpatient medication recommendation at prescription level rather than admission level, framing the task as repeated per-timepoint selection of exact medication–dose–route triples from detailed patient profiles and time-ordered clinical trajectories (Chen et al., 14 May 2026). RxEval comprises 1,547 questions spanning 584 patients, 18 diagnostic categories, and 969 unique medications (Chen et al., 14 May 2026). Evaluation of 16 LLMs yields F1 ranging from 45.18 to 77.10, and the best Exact Match is only 46.10%, indicating that prescription-level reasoning remains difficult even for frontier models (Chen et al., 14 May 2026).

RxEval is especially relevant because it operationalizes failure modes that a prescription-focused Rx Strategist must avoid: oversight errors, in which models ignore explicit chart facts such as allergies or current glucose values, and reasoning errors, in which they fail to derive higher-order clinical conclusions such as renal impairment or withdrawal-related sympathetic surge (Chen et al., 14 May 2026). A probing experiment reported in the paper shows that when the allergy list is repeated at the end of the prompt, 85.4% of allergy-violating decisions become correct, suggesting that failure often reflects inconsistent use of available information rather than absence of medical knowledge (Chen et al., 14 May 2026).

SafeRx-Agent approaches this problem by moving from broad ATC-L3 medication categories to fine-grained ATC-L4 medication recommendation, grounding recommendations in MEDI indications, TWOSIDES drug–drug interactions, and openFDA contraindications (Wang et al., 27 May 2026). It uses a multi-agent sequence—Route, Summarize, Generate, Critique, FindFlags, Verify—to produce medication sets while controlling drug interactions, contraindications, and medication set size (Wang et al., 27 May 2026). On MIMIC-III and MIMIC-IV, the paper reports that SafeRx-Agent improves fine-grained medication prediction accuracy while controlling drug interactions, contraindications, and medication set size (Wang et al., 27 May 2026). A plausible implication is that the original Rx Strategist verification system and SafeRx-Agent occupy adjacent points on a continuum: the former verifies prescriptions already written, whereas the latter constructs and prunes candidate regimens before final selection.

6. “Rx” as strategy beyond prescriptions

Outside medication verification, the label “Rx” also appears in methodological literatures where it abbreviates a strategic randomized or robust procedure rather than a prescription system. “Robust X-Learner” introduces the RX-Learner for heterogeneous treatment-effect estimation under extreme imbalance and heavy-tailed outcomes (Uehara, 21 Jan 2026). The method replaces MSE-based regressions in the X-Learner with a redescending $FI$2-divergence objective equivalent to the Welsch loss under Gaussian assumptions, combined with a Proxy Hessian grounded in Majorization–Minimization principles (Uehara, 21 Jan 2026). On a semi-synthetic Criteo Uplift dataset, RX-Learner reduces the Precision in Estimation of Heterogeneous Effect metric by 98.6% compared to the standard X-Learner (Uehara, 21 Jan 2026). This usage broadens “Rx Strategist” into the domain of treatment-policy learning, where robustness to outliers and imbalance is central.

In imaging, “Randomized RX for target detection” applies random Fourier features to approximate the Gaussian kernel in kernel RX, keeping nonlinear accuracy while reducing computational cost (Nar et al., 2020). This is unrelated to prescribing, but it illustrates a separate lineage in which “RX” denotes the Reed–Xiaoli detector rather than medication. The coexistence of these uses makes the term polysemous; in biomedical and health-AI contexts, disambiguation depends on whether the surrounding discussion concerns prescriptions, treatment effects, or signal detection.

A further strategic use appears in “ROSE: Randomized Optimal Selection Design for Dose Optimization,” which addresses selection of an optimal biological dose under Project Optimus (Wang et al., 6 May 2025). The design compares two dose arms with a simple decision boundary on the difference in response rates and is calibrated by the probability of correct selection (Wang et al., 6 May 2025). Although not named “Rx Strategist,” it belongs to the same family of decision frameworks in which randomized evidence and explicit operating characteristics are used to optimize treatment strategy rather than merely estimate efficacy.

7. Evidence strategy, adaptive orchestration, and future directions

The broader ecosystem surrounding Rx Strategist also includes frameworks for evidence generation and adaptive control. “A Causal Roadmap for Generating High-Quality Real-World Evidence” defines an explicit, itemized, iterative process with seven steps for translating clinical or regulatory questions into causal estimands, data requirements, identification assumptions, statistical estimands, estimators, sensitivity analyses, and simulation-based design comparisons (Dang et al., 2023). This suggests that any future Rx Strategist deployed in clinical or regulatory settings would need not only accurate verification or recommendation performance but also a causal evidence strategy for demonstrating validity across populations and workflows.

In retrieval-augmented systems, “Adaptive Defense Orchestration for RAG: A Sentinel-Strategist Architecture against Multi-Vector Attacks” introduces a Strategist that selectively deploys defenses based on a Sentinel’s risk assessment, recovering retrieval utility relative to a fully static defense stack while mitigating membership inference and poisoning (Pallerla et al., 22 Apr 2026). Although the domain is RAG security, the architecture is conceptually relevant: a Strategist functions as a control-plane policy that decides which safeguards to apply given query context and estimated risk (Pallerla et al., 22 Apr 2026). A plausible implication is that future prescription-oriented Rx Strategist systems may adopt similar orchestration logic, selectively invoking allergy checks, dosage verification, formulary filters, or knowledge-grounding modules according to the structure and uncertainty of a case.

Taken together, the literature presents Rx Strategist as both a concrete prescription-verification system and a broader design pattern. In its narrow sense, it is a multi-agent LLM pipeline for prescription verification with knowledge-grounded checks for indication, dosage, and interaction safety (Van et al., 2024). In its broader sense, it denotes a strategy-centered approach to therapeutic decision support: modular rather than monolithic, explicit rather than implicit in its constraints, and grounded in structured evidence, adaptive evaluation, and safety verification (Chen et al., 14 May 2026, Wang et al., 27 May 2026, Dang et al., 2023).