Written and maintained by CASRAI Editorial Board
Last updated
An FDA De Novo medical device classification request—governed by Section 513(f)(2) of the Federal Food, Drug, and Cosmetic Act (FD&C Act) (21 U.S.C. § 360c(f)(2)) and codified in 21 CFR Part 860 Subpart D (§§ 860.200–860.260)—is a risk-based premarket regulatory pathway that enables manufacturers of novel, low-to-moderate-risk medical devices lacking a legally marketed predicate device to obtain Class I or Class II marketing authorization without submitting a Class III Premarket Approval (PMA) application. Under Section 513(f)(1) of the FD&C Act, any new post-1976 device that is not “substantially equivalent” to an existing Class I or Class II predicate is automatically classified by statute into Class III. The FDA De Novo classification pathway provides a direct mechanism to reclassify such novel devices—particularly first-of-a-kind Software as a Medical Device (SaMD), clinical AI/ML algorithms, and digital therapeutics—when General Controls alone (Class I) or General and Special Controls (Class II) provide a reasonable assurance of safety and effectiveness.
When Should a Novel SaMD or Clinical AI Product Use the FDA De Novo Classification Request Pathway?
When determining the appropriate regulatory submission strategy across 510(k) vs. De Novo and FDA clearance vs. approval pathways, a medical software or hardware developer selects the De Novo pathway when two statutory conditions are simultaneously met:
- Absence of a Valid Legally Marketed Predicate Device: Search of the FDA 510(k) database and the FDA De Novo database confirms that no legally marketed Class I or Class II device shares the same intended use, or that the novel device introduces fundamentally different technological characteristics (for example, an autonomous deep-learning fundus diagnostic algorithm or a prescription closed-loop digital therapeutic) that raise new questions of safety and effectiveness not addressed by any existing predicate.
- Low-to-Moderate Risk Profile Amenable to General and Special Controls: Probable clinical risks of false-positive outputs, false-negative missed diagnoses, software malfunction, or user misinterpretation can be mitigated through identifiable engineering standards, clinical performance thresholds, and labeling controls without requiring Class III PMA oversight.
Under current FD&C Act Section 513(f)(2)(A)(ii) (as amended by FDASIA in 2012), sponsors may submit a Direct De Novo request immediately—often following a Pre-Submission (Q-Sub) meeting with the FDA Digital Health Center of Excellence and the relevant CDRH Office of Health Technology—without first filing a doomed 510(k) and waiting for a Not Substantially Equivalent (NSE) determination.
Mandatory 21 CFR § 860.220 Submission Checklist and eSTAR Requirements
Effective October 1, 2025, CDRH requires all De Novo classification requests (unless specifically exempted) to be prepared and submitted electronically using the interactive PDF eSTAR (electronic Submission Template And Resource) via the CDRH Portal. Once received, the submission undergoes a 15-calendar-day Acceptance Review (21 CFR § 860.230) followed by a substantive review with a 150-FDA-review-day MDUFA performance goal. Every complete De Novo request under 21 CFR § 860.220 must contain the following required sections:
| 21 CFR § 860.220 Section | Mandatory Regulatory Content | SaMD & Clinical AI Engineering Evidence |
|---|---|---|
| 1. Classification Summary & Predicate Search (§ 860.220(a)(6)) | Recommended class (Class I or Class II), explanation of why the device is novel, and a comprehensive search summary of existing 21 CFR classifications, 510(k)s, PMAs, and De Novos showing no valid predicate exists | Detailed comparison explaining why existing CADe/CADx or clinical decision support product codes do not cover the new intended use or autonomous workflow |
| 2. Proposed Special Controls (§ 860.220(a)(8)) | Mandatory for all Class II De Novo requests: an explicit, itemized list of proposed Special Controls showing how each identified probable risk to health is mitigated | Crosswalk table mapping each clinical hazard (e.g., subgroup bias, sensor dropout, AI model drift) to specific validation, labeling, and post-market controls |
| 3. Probable Benefits vs. Probable Risks (§ 860.220(a)(7)) | Quantitative benefit-risk assessment demonstrating that probable clinical benefits outweigh probable risks when Special Controls are applied | Sensitivity/specificity gains, time-to-triage reduction, or clinical endpoint improvement weighed against false-positive/false-negative clinical consequences |
| 4. Software, Cybersecurity & Non-Clinical Data (§ 860.220(a)(9)–(10)) | Complete technical bench/analytical testing, IEC 62304 software lifecycle documentation, and FD&C Act Section 524B cybersecurity package | Basic or Enhanced Software Documentation, SBOM, adversarial threat modeling, standalone analytical performance on locked multi-site test cohorts, and optional PCCP |
| 5. Clinical & Human Factors Validation (§ 860.220(a)(11)) | Valid scientific evidence (21 CFR § 860.7) from prospective or multi-reader multi-case (MRMC) clinical studies, plus IEC 62366-1 human factors validation | Pivotal clinical study report (compliant with 21 CFR Parts 50, 56, and 812 IDE rules) stratified by patient demographics, scanner vendor, and clinical site |
What Are Special Controls and Benefit-Risk Determinations in an FDA De Novo Submission for Medical Software?
For Class II De Novo requests, drafting the Risks to Health and Proposed Special Controls matrix is the single most consequential element of the submission. Unlike a 510(k) that argues equivalence to a predicate, a Class II De Novo sponsor is effectively co-authoring a new federal regulation for an entire category of medical devices.
In granted SaMD and AI/ML De Novo authorizations (such as autonomous diabetic retinopathy detection under 21 CFR § 886.1100 or digital therapeutic software regulations), FDA Special Controls routinely mandate five enforceable pillars:
- 1. Algorithm Training and Locked Test Dataset Independence: Documented proof that the pivotal validation cohort was completely sequestered from training and hyperparameter tuning datasets and represents the intended U.S. clinical population across age, sex, race/ethnicity, and disease severity.
- 2. Standalone and Assisted Clinical Performance Thresholds: Pre-specified minimum sensitivity, specificity, positive predictive value, or primary clinical efficacy endpoints demonstrated in a pivotal clinical study.
- 3. Input Data Quality and Failure-Mode Handling: Automated software checks that detect poor-quality, out-of-distribution, or corrupted input signals (e.g., ungradable retinal images or noisy ECG leads) and refrain from outputting an unreliable diagnostic result.
- 4. Human Factors and Clinician/Patient Labeling Disclosures: Mandatory “Model Card” style labeling in the Instructions for Use (IFU) disclosing the algorithm’s architecture, training cohort demographics, known subgroup limitations, and exact user training requirements.
- 5. Change Management and Real-World Performance Plans: Post-market change control governance—increasingly paired with a Predetermined Change Control Plan (PCCP) under Section 515C of the FD&C Act—to prevent unvalidated model drift.
How Does a Granted De Novo Classification Establish a New Class I or Class II Predicate Device?
When CDRH grants a De Novo classification request under 21 CFR § 860.260, three immediate regulatory consequences occur:
- Marketing Authorization Order: The sponsor receives a written De Novo Grant Order authorizing immediate commercial distribution of the device in the United States.
- Creation of a New 21 CFR Classification Regulation and 3-Letter Product Code: The FDA publishes a final order in the Federal Register establishing a brand-new section in Title 21 of the Code of Federal Regulations (e.g., 21 CFR Parts 862–892) along with a new three-letter CDRH Product Code and the codified Special Controls.
- Establishment of the First 510(k) Predicate for Follow-On Competitors: Under Section 513(f)(2)(B)(i) of the FD&C Act, the newly granted De Novo device immediately becomes a legally marketed predicate device. Any subsequent manufacturer developing a device of the same generic type does not file another De Novo; instead, follow-on competitors submit a standard 510(k) Premarket Notification demonstrating substantial equivalence to the De Novo pioneer and compliance with the newly published Special Controls.
Frequently Asked Questions
When should a novel SaMD or clinical AI product use the FDA De Novo classification request pathway?
A novel SaMD or clinical AI product should use the FDA De Novo pathway under Section 513(f)(2) of the FD&C Act when it has no legally marketed Class I or Class II predicate device (due to a new intended use or novel algorithmic technology) but presents a low-to-moderate risk profile that can be safely controlled through General and Special Controls rather than a Class III Premarket Approval (PMA).
What are Special Controls and benefit-risk determinations in an FDA De Novo submission for medical software?
Special Controls are device-type-specific regulatory requirements—such as independent multi-site clinical validation thresholds, software verification under IEC 62304, input-quality guardrails, cybersecurity testing, and transparency labeling—that work alongside General Controls to ensure that the device’s probable clinical benefits outweigh its probable risks.
How does a granted De Novo classification establish a new Class I or Class II predicate device?
When the FDA grants a De Novo request, it issues a marketing authorization order, codifies a new device classification regulation in 21 CFR with a new product code and Special Controls, and designates the granted device as the foundational predicate device that future competitors can cite in standard 510(k) submissions.
