KI in der Pharmakotherapie

AI in Pharmacotherapy: Promise and Peril

AI spans drug development, bedside management, and personalized dosing, but overreliance risks skill decay. After three months of AI-assisted endoscopy, adenoma detection falls from 28.5% to 22.5% when support stops. Treat AI as a tool; preserve clinical reasoning.

LLMs for Therapy: Context and Caution

LLMs assist therapy design if given full clinical context. The gabapentin case shows renal function monitoring and dose adjustment at GFR 30 ml. Curated, locally hosted models improve reliability. Best uses include decision support and education; always validate, acknowledge hallucinations.

Personalized, Explainable, Clinician-Guided Dosing

Personalized dosing integrates classical pharmacological models and TDM with EHR-driven neural models to predict concentrations. Curate AI identifies a Goldilocks Zone, improving tolerability, especially in geriatric/palliative settings. Keep clinician and patient in loop. Mitigate black-box risk, applicability domain, data quality, and bias via explainable AI.

The continuing education session “KI in der Pharmakotherapie,” organized by Klinik Barmelweid and delivered by PD Dr. med. et phil. Felix Hammann (Clinical Pharmacology, Inselspital Bern), surveys rapidly evolving AI applications in drug development, bedside management, and personalized dosing, while emphasizing a critical, self-determined use of AI as a tool rather than an authority. He presents evidence of skill decay with AI assistance, citing a Polish endoscopy study in which adenoma detection rates drop from 28.5% to 22.5% three months after removing AI support. A gabapentin case illustrates opportunities and pitfalls of LLMs in therapy design: only when renal function (eGFR 30 mL/min) is explicitly provided does the model prioritize dose adjustment, and a locally curated, Switzerland-hosted LLM yields more cautious, mainstream-aligned recommendations. Educational use cases include role-play with LLMs to generate structured, adaptive learning plans for new clinical roles. For personalized dosing, he contrasts classical a priori/a posteriori PK and TDM (including a pilot web tool for antibiotics) with emerging EHR-integrated neural networks and highlights CurateAI to navigate a “Goldilocks zone” for combination therapies, improving tolerability and adaptability in complex settings such as geriatrics and palliative care. He discusses networked infusion pumps linked to EHRs that enhance medication safety and enable dynamic dosing, but also notes applicability-domain limits, data-quality vulnerabilities, and black-box concerns; explainable AI methods (e.g., saliency, feature importance, SHAP) help build trust and expose biases. The session underscores risks of algorithmic bias (e.g., cost-based risk models that disadvantage Black patients) and the heterogeneous Swiss context (languages, regional practice, demographic and socioeconomic diversity) as challenges for validation. In the discussion, early Dutch studies in long-term care are noted where wearable-informed AI supports delirium treatment dosing, and overall the recommendation is to deploy AI where validated—particularly in education—while preserving clinical reasoning, context, and clinician oversight.