FDA finalises guidance on AI/ML-based Software as a Medical Device (SaMD) for drug dosing clinical decision support

FDA's final guidance AI/ML-Based Software as a Medical Device (SaMD): Clinical Decision Support for Drug Dosing establishes the regulatory framework distinguishing device CDS from non-device CDS. Non-device CDS provides general pharmacology information (e.g., a reference tool showing published PK parameters) — clinicians are expected to independently review and apply clinical judgment. Device CDS makes specific, patient-individualised dosing recommendations that a busy clinician may implement with minimal independent verification. AI/ML-based Bayesian dosing software (e.g., dose individualisation tools for vancomycin, aminoglycosides, tacrolimus, and busulfan) fall into the device CDS category and require FDA clearance or De Novo classification. Guidance is available at FDA SaMD guidance page ↗. See our FDA AI ANDA review news.

Bayesian pharmacokinetic software uses population PK priors (published model parameters: CL, Vd, inter-patient variability) combined with individual observed drug concentrations (TDM samples) to compute maximum a posteriori (MAP) estimates of a patient's individual PK parameters — then predicts the dose needed to achieve a target AUC or concentration. This is called Model-Informed Precision Dosing (MIPD). The 2020 ASHP/IDSA/SIDP vancomycin consensus guidelines recommended AUC-guided monitoring using Bayesian software as the preferred approach, supported by studies showing reduced nephrotoxicity while maintaining efficacy. Tools such as DoseMe, InsightRX, and BestDose are in clinical use across major US health systems. The FDA guidance requires these tools to demonstrate algorithmic transparency, prospective clinical validation, and real-world performance monitoring. See our TDM guide.

The guidance specifies four requirements for AI/ML dosing SaMD: (1) Algorithm transparency: the PK model source (population study citations, patient demographics), software version, and prediction uncertainty must be documented; (2) Clinical validation: prospective or retrospective validation in a patient population representative of the intended clinical use — including adequate representation of special populations (renal impairment, obesity, paediatrics); (3) Real-world performance monitoring (RWPM): post-deployment tracking of prediction accuracy, clinical outcomes (target attainment, adverse events), and demographic performance gaps; (4) Change control: any update to the underlying model or algorithm that could materially change dosing recommendations requires FDA notification. AI/ML-based software that learns from new patient data post-deployment must follow the FDA Predetermined Change Control Plan framework ↗.

For clinical pharmacists using Bayesian TDM software, the guidance means that tools used for vancomycin AUC-guided dosing, tacrolimus post-transplant monitoring, aminoglycoside extended-interval dosing, and busulfan conditioning regimens will increasingly require FDA clearance. Institutions should verify that their software has appropriate 510(k) clearance or De Novo authorisation. Uncleared software may continue for research use but not as the basis for clinical dosing decisions. The guidance aligns with MIPD guideline development from the Population Approach Group Europe (PAGE) ↗ and the International Society of Pharmacometrics ↗. See our nonlinear PK glossary and TDM guide.