Implementation of LLM-supported therapeutic drug monitoring reporting in a laboratory information system

Therapeutic drug monitoring (TDM) reports require integration of drug concentrations, dosing schedules, sampling times, pharmacokinetics, metabolites, co-medication, drug interactions, and patient-specific context. In a previous proof-of-concept, standalone LLM-generated TDM reports were feasible but showed limitations regarding steady-state assessment, trough-level interpretation, conflicting reference information, and incomplete clinical context. We therefore implemented a follow-up workflow inside medicalvalues Laboratory Intelligence, the laboratory information system (LIS) of medicalvalues.

The system receives patient data including detailed medication information, drug and metabolite concentrations, therapeutic range and dosage level information, and optional additional clinical data. Report generation is performed by custom TDM LLM agents integrated into the LIS. The agents use several explicit knowledge sources: drug interaction information present in UMLS through relations between medical concepts; additional drug interaction relations and medical scientific sources curated by medicalvalues from public PubMed, Rote Liste, Fachinformation, and package leaflets as well as TDM-specific guidelines. Test cases were based on real patient constellations and were intentionally designed to include incomplete or flawed input. This allowed assessment of whether the system generates clinically useful draft reports while explicitly marking uncertainty and avoiding unsupported conclusions.

The implementation separates deterministic preprocessing from generative report drafting. Pharmacokinetic calculations, source retrieval, interaction extraction, and patient-bundle parsing are performed before text generation. Generated reports remain subject to expert review before release. Governance measures include traceable source display, validation of clinically relevant failure modes, and risk-oriented handling of hallucination, missing data, and contradictory sources.

The project demonstrates that LLM-supported TDM reporting can be embedded into an LIS workflow when structured patient data, curated pharmacological knowledge, RAG-based source retrieval, and human oversight are combined. Key lessons are that TDM agents require explicit data provenance, deterministic handling of pharmacokinetic parameters, transparent uncertainty reporting, and user interfaces that allow laboratory experts to inspect, correct, and approve generated report text.