FDA's 2024 LDT rule was vacated March 31, 2025 and formally rescinded September 19, 2025 but FDA device oversight of AI-enabled software functions remains fully intact. The classification question still matters.
If the function diagnoses, predicts, stratifies, or recommends treatment for disease, FDA device analysis is triggered first.
Transparent support for a healthcare provider (HCP) may fit non-device CDS. Opaque or directive outputs push the function toward device/SaMD treatment.
Calling something an LDT does not answer whether the software behaves like IVD-SaMD. The function and workflow do.
The more the software analyzes medical signals or images, produces patient-specific diagnostic output, and cannot be independently checked by the clinician, the more it looks like regulated IVD-SaMD.
Intended to diagnose, predict, classify, risk-stratify, or guide treatment for disease or state of health?
Processes pathology images, assay signals, waveforms, genomic features, or other IVD-derived data?
Produces a diagnosis, probability, classification, or recommendation for a specific patient or specimen?
Can the healthcare provider (HCP) meaningfully review the basis for the output, rather than simply accepting a black-box result?
Q1–Q3 Yes and Q4 No: output analyzes IVD-derived data, produces a patient-specific result, and cannot be independently verified. Treat as device-like pending resolution of FDA's legal reach over LDTs.
The 2024 LDT rule is gone, but FDA still classifies AI/ML software based on what it does: its intended use, whether it analyzes medical signals or images, and whether the clinician can independently review the reasoning behind its output. The January 2026 FDA CDS guidance update allows enforcement discretion for software that gives a single clinically appropriate recommendation — but the ability to independently review the output remains the key factor. What the software does determines how it is classified. Where it is deployed does not.
Use the examples below to help determine where your tool likely fits
Rule-based triage aid that displays ordinary clinical information and lets the pathologist independently evaluate the basis. Transparent rationale, no patient-specific diagnostic call.
AI risk score layered into a lab workflow — transparent rationale, required human confirmation before reporting. Classification depends on how directive the output is in practice.
Black-box molecular or pathology classifier that converts specimen data into a diagnostic call or treatment recommendation without independent verifiability.