Device Software Function: FDA’s 2025 AI Draft Definition
- bengeelim
- 5 days ago
- 3 min read
TL;DR — What’s a Device Software Function (DSF)?
A device software function (DSF) is stand-alone software that meets the FDA’s definition of a medical device.
In the 2025 AI draft guidance, FDA states:
“For purposes of this guidance, a device software function is a software function that meets the definition of a device under section 201(h) of the FD&C Act.” (lines 148–151, Jan 2025 draft)
If your code diagnoses, treats, predicts, or guides clinical decisions—and it runs independently—it almost certainly qualifies as a DSF and falls under FDA oversight.
Why This Term Matters for AI Teams?
SaMD has been the headline for years, but the AI draft guidance shifts focus to the function-level unit FDA regulates.
Knowing whether each function in your app is a DSF:
determines if you need a 510(k), De Novo, or PMA,
dictates whether you must submit a Predetermined Change Control Plan (PCCP), and
shapes your evidence package (validation, cybersecurity, data lineage).
Shortcut: If any code block influences a diagnosis or therapy decision, treat it as a DSF first—justify exclusions later.
The Legal Definition Behind The Term
FD&C Act § 201(h)
A device is “any product… intended for use in diagnosis… cure, mitigation, treatment, or prevention of disease… which does not achieve its primary intended purposes through chemical action or being metabolized.”
Draft guidance (2025) clarification
FDA simply applies that device definition to software functions: intent-driven, not platform-driven. Your hosting environment (phone, cloud, on-prem) is irrelevant; the intended use is everything.
Function vs Feature vs Device Software
Term | Plain-English meaning | Reg status |
Software feature | Capability inside an app (“zoom”, “export PDF”) | Usually unregulated |
Software function | Discrete intended use (“flag likely tumour”) | May be regulated |
Device software | Software that as a whole meets device definition | Regulated |
Device software function (DSF) | The individual function within that software that triggers FDA oversight | Regulated & focal point of 2025 AI draft |
Example: A mobile app has five features; only one analyses ECG to detect AFib. That one feature is the DSF.
Real-world AI DSF examples
Still unsure? Let’s look at a few real-world examples.
Imaging Classifier
An AI tool that flags potential tumors in MRI scans.
✅ Intended for diagnosis
✅ Used by clinicians
✅ Device software function
Insulin-Dosing Algorithm
Software that calculates recommended insulin doses using patient data.
✅ Directly affects treatment
✅ Regulated as a Class II or III DSF
Clinical Decision Support Flag
An app that highlights abnormal EKG patterns.
☑️ May qualify as DSF if it interprets or guides care—not just displays data
3-Question Litmus Test: Is Your Code a DSF?
Before worrying about submissions, start here.
What is the intended use?
If it diagnoses, treats, monitors, or predicts health conditions—watch out. FDA cares about intent, not disclaimers.
Who is the end user?
If it informs decisions by clinicians or patients, you’re probably in DSF territory.
Is it standalone or embedded?
Standalone software is a prime DSF candidate. If it’s embedded in a device’s firmware, it might fall under that device’s classification instead.
Why This Matters for Submissions
So your software qualifies as a DSF—now what?
PCCP Requirements
If your algorithm changes over time (think retraining, updating thresholds), the FDA now expects a Predetermined Change Control Plan. That means documenting:
What you’ll change in the future
How you’ll validate those changes
How risks will be mitigated
Other Supporting Documents
You’ll also need:
DSF description
Intended use statement
Model architecture diagram & training data lineage
Input/output data types
Performance claims and validation methods
Cybersecurity & SBOM per 2024 guidance
Skipping this will lead to delays—or worse, rejection.
FAQs
Q: Is firmware considered a Device Software Function?
Not typically. Firmware is usually considered part of the hardware device itself. DSFs are usually standalone or app-based.
Q: Do wellness algorithms count as DSFs?
Only if they make medical claims. If it just tracks steps or sleep, you’re probably safe. But once it says “reduce anxiety” or “improve glucose,” you may cross the line.
Q: How does PCCP tie into DSFs?
If your DSF includes AI/ML that evolves over time, the FDA expects a PCCP showing how you’ll manage those changes safely.
