Industries / Health / Medical devices
Software in devices
This is the code that runs inside devices, or that is a device by itself, and the rules for writing, changing and protecting it.
7
Milestones achieved, from 1987 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1987
Software overdoses in radiotherapy
The Therac-25, a radiation therapy machine made by Atomic Energy of Canada Limited (AECL), gave massive overdoses in six known accidents between June 1985 and January 1987, with deaths and serious injuries. Unlike AECL's earlier machines, it relied on software for safety checks that hardware interlocks, physical switches that block an unsafe state, had done before, and it reused code from those machines. The machine was recalled in 1987.
LessonSoftware should not be a machine's only safeguard, and code carried over from an older machine can hide faults that its hardware used to catch.
Credit
the investigation by Nancy Leveson and Clark Turner, published in IEEE Computer in July 1993
2006
A life cycle for code
The International Electrotechnical Commission (IEC) published IEC 62304 on 9 May 2006. It sets the life cycle requirements for medical device software: the processes, activities and tasks for developing and maintaining it, whether the software is a device by itself or part of one.
Credit
IEC
2013
Software alone is a device
In 2013 the International Medical Device Regulators Forum (IMDRF), a voluntary group of device regulators from around the world, formed a working group on software as a medical device (SaMD), chaired by the FDA. It agreed that SaMD is software "intended to be used for one or more medical purposes" that performs them "without being part of a hardware medical device", a definition the FDA quotes.
Credit
IMDRF SaMD Working Group, chaired by the FDA
2017
A pacemaker patch as recall
On 29 August 2017 the FDA issued an alert about Abbott's voluntary recall, for a firmware update, of pacemakers first sold by St. Jude Medical, and urged the 465,000 US patients who had them to talk to their doctors. Without the update, someone with commercially available equipment could reach a pacemaker and change its programming, which could drain its battery or pace the heart wrongly. The update was given in person and took about three minutes. Josh Corman of the Atlantic Council said the FDA had issued similar notices before, but "this is the first to affect implanted devices".
LessonAn implant's software may need security fixes long after it is placed, and each fix can mean a visit to a clinic.
Credit
Abbott; FDA
2023
Open-source dosing app cleared
On 23 January 2023 the FDA cleared Tidepool Loop, an app that adjusts insulin doses automatically, as an interoperable automated glycemic controller (510(k) number K203689). It grew out of DIY Loop, open-source software that Nate Racklyeft created and released and that volunteers built on. Tidepool, a nonprofit, calls it "the first ever community led innovation in the diabetes space to be FDA cleared", and at clearance said the app was not yet available for use.
Credit
Tidepool; Nate Racklyeft and the DIY Loop community
2023
Cyber rules written into law
From 29 March 2023, section 524B of the US Federal Food, Drug, and Cosmetic Act applies to any "cyber device", one with the maker's software that can connect to the internet. Whoever seeks FDA authorization for one must submit a plan to find and fix vulnerabilities after sale, provide patches on a regular cycle and fix critical flaws as soon as possible, and give the FDA a software bill of materials (SBOM), a list of the commercial, open-source and off-the-shelf components in its software.
Credit
US Congress, in the Food and Drug Omnibus Reform Act of 2022
2024
Updates cleared in advance
A predetermined change control plan (PCCP) describes the changes a maker plans to make to a device's software, such as retraining an AI model, and how it will develop, test and roll them out. Once the FDA has reviewed the plan with the device, changes that follow it need no new marketing submission. Congress added PCCPs to the law in December 2022, and the FDA issued final guidance for AI-enabled device software on 4 December 2024. In February 2023 the FDA had classified Caption Health's ultrasound software that estimates how well the heart pumps as a new device type "with predetermined change control plan". The FDA's list of AI-enabled devices it has authorized, updated 4 September 2026, has 1,614 entries by a count of its table rows made for this file.
Credit
FDA; Caption Health
Watch for
A chat model in dosing
UpDoc says its product is the first FDA-cleared clinical AI platform built on patient-facing large language models (LLMs, the kind of AI behind chat assistants). The FDA cleared a device named UpDoc on 23 December 2025 as a drug dose calculator (510(k) number K253281), for insulin management in adults with type 2 diabetes. Innolitics, a regulatory consulting firm that read the public record, found that it names a "Conversation Service (UpDoc Agent)", computes insulin instructions from limits that clinicians set, names no AI model, and does not settle whether the June 2026 announcement refers to this clearance.
Achieved whenan FDA record, or a study by people outside the company, confirms that a language model is part of the cleared device and reports how it performed with patients.
Claimed by
UpDoc, June 2026
Dosing code ported, then compared
Trio is an open-source app that runs the oref dosing algorithm, written in JavaScript. Researchers led by Thomas Screven and Samuel T. King rewrote that algorithm in Swift with known bugs fixed, ran both versions side by side on 736,480 recorded runs from eight Trio users, and report that the two disagreed in under 1.3% of calls for each part of the algorithm. In simulation both kept glucose in range 84.9% of the time.
Achieved whenthe work passes peer review, or another group reruns the comparison on its own data and gets matching results.
Claimed by
Screven and colleagues, in a preprint, June 2026
Open
01
SBOMs for open dosing apps
For each release of five open-source diabetes projects (AndroidAPS, Loop, Trio, xDrip+ and Nightscout), a software bill of materials, the list of third-party components in a program, made with an open tool and checked against public databases of known vulnerabilities. Findings go privately to each project's maintainers under its security policy, and the ledger is published once fixes are out. The work reads code only; it never touches anyone's phone, pump or server.
The briefCredit · unclaimed02
Prove oref's dose limits
A machine-checked proof that the core dosing function of OpenAPS, determine-basal in the oref0 code, can never set a temporary basal rate above its maxSafeBasal cap or push insulin on board above the user's max_iob setting, for any input. The function is ported to a proof language and the port is tied to the JavaScript by testing both on the same inputs. Nothing is run on a person.
The briefCredit · unclaimed
Open 01 · the brief
SBOMs for open dosing apps
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
For each release of five open-source diabetes projects (AndroidAPS, Loop, Trio, xDrip+ and Nightscout), a software bill of materials, the list of third-party components in a program, made with an open tool and checked against public databases of known vulnerabilities. Findings go privately to each project's maintainers under its security policy, and the ledger is published once fixes are out. The work reads code only; it never touches anyone's phone, pump or server.
Why it is still open
Section 524B makes SBOMs part of a request for FDA authorization, so it does not reach open-source projects that do not seek it. A search found no published SBOM or dependency audit of these five projects. The nearest work looked elsewhere: a June 2026 preprint by Screven and colleagues looked for bugs in the dosing math, and a 2025 Queen's University study checked AndroidAPS's own code for private data leaking into its logs, not its third-party components.
Built with what exists
Syft or the CycloneDX tools to make SBOMs; Google's OSV-Scanner or Grype to check them against the OSV and GitHub Advisory databases; an AI model to sort which flagged components the app actually calls (see the Note "Cyber defenders first").
First step
Run Syft and OSV-Scanner on the current AndroidAPS release and read the project's security policy.
Done looks like
SBOMs for every tagged release of the five projects over six months. Each known-vulnerable component is listed with whether the app's code reaches it, and a second person checks every finding. Each maintainer's acknowledgment and fix date is recorded, and the median days from report to fix is published.
Open 02 · the brief
Prove oref's dose limits
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A machine-checked proof that the core dosing function of OpenAPS, determine-basal in the oref0 code, can never set a temporary basal rate above its maxSafeBasal cap or push insulin on board above the user's max_iob setting, for any input. The function is ported to a proof language and the port is tied to the JavaScript by testing both on the same inputs. Nothing is run on a person.
Why it is still open
The authors of GlucOS, at the University of California, Davis, write that this function is 1,192 lines of JavaScript with 63 inputs and settings and 90 branches. Instead of proving it, they proved 25 lines of Swift that limit whatever dose any algorithm proposes, and found 9 bugs in their own code doing so. Screven and colleagues (2026) compared a Swift port of Trio's oref by running both versions, not by proof. Bara' Nazzal, Manar Alalfi and James Cordy checked by static analysis, reading the code without running it, that AndroidAPS validates its dosing values against safety limits, and found one value "not explicitly validated" at the plugin level; their 2024 survey of the field found "a gap in the research when it comes to using formal verification on real implementations". A search found no machine-checked proof of the oref0 or Loop dosing code.
research.cs.queensu.caresearch.cs.queensu.cabob.cs.ucdavis.edugithub.com
Built with what exists
Dafny or Lean, languages in which a computer checks every step of a proof; AI proof agents (see the Note "Proven in four weeks"); the oref0 code on GitHub; random input testing; the open-source artificial pancreas testbed of Zhou, Kouzel, Ren and Alemzadeh (2022) for simulated runs.
First step
Port the part of determine-basal that caps the basal rate at maxSafeBasal to Dafny and prove that one property.
Done looks like
The proof checks in Dafny or Lean, with every assumption written down. The port and the JavaScript give the same output on 1 million random inputs. Any input on which the JavaScript breaks a limit is reported privately to the maintainers. A second group reruns the checker and gets the same result.