3. :rotating_light: _Go check your saline. B. Braun recalled three lots of 0.9% Sodium Chloride Injection nationwide for iron oxide and plastic particulate — and it's the bag you dilute everything in
FDA · company announcement posted Sep 3, 2026 · September 3, 2026Pharmacy
- What was recalled, exactly. B. Braun Medical Inc. is voluntarily recalling three lots of 0.9% Sodium Chloride Injection USP, 100 mL, in a 150 mL PAB® container, to the hospital / healthcare-facility / user level. Catalog S8004-5264. NDC 0264-1800-32. Lots J6B435, J6B444, J6B457. Expiration 30APR2027 — so this is stock that is nowhere near expiring and will be sitting on shelves. Distributed nationwide to hospitals and healthcare facilities from February 27 through August 3, 2026.
- The contaminant list got longer after the fact. FDA's page carries the note "On September 3, 2026 firm updated their press release to include additional contaminants." The reason for the recall is now particulate matter identified as iron oxide, inorganic, cellulose or polystyrene particulate matter. It started narrower.
- FDA's risk statement, verbatim, because it's blunt. Administration "has a reasonable probability of causing life-threatening adverse health consequences including pulmonary emboli … occlusions of other blood vessels (which can lead to tissue death and possible organ damage), and/or phlebitis." Systemically, infused foreign particles can cause "systemic activation of the immune system, organ dysfunction, and hemolysis." To date there are no reports of serious injury, death or other adverse events.
- Why this specific product matters more than a normal recall. FDA's own description: this product "is designed for use as a diluent and delivery system for intermittent intravenous administration of compatible drug additives." It is not just hydration — it is the bag your compatible drugs go into. Returns and replacements are being arranged by B. Braun through distributors; questions go to <mailto:recalls@bbraunusa.com>, Mon–Fri 8am–5pm EST.
Why it matters for an independent practice: every IV suite, every med spa, every regen practice running infusions or diluting anything keeps 100 mL saline on the shelf, and almost nobody has a person whose job is reading FDA recall notices. Three concrete moves. One, today: someone physically reads the lot numbers on your saline stock against J6B435 / J6B444 / J6B457 — and while they're standing there, re-check the epinephrine lots from Friday's wrap (25128L1C0, 25280L1C0, 26108L1C0, NDC 0517-3030-01), because that's the same cabinet. Two, the systems point, which is the real one: two recalls in three weeks means the manual process failed twice. A standing FDA recall check is the single cheapest automation a practice can run — a scheduled pull of the recall feed filtered to the NDCs you actually stock, into a channel someone reads. That is a 30-minute build, and it is worth more than most of the AI tools that get pitched. Three, P2: the practice that can tell a patient "we checked, our lot isn't affected" within an hour of a recall is running a different operation than the one that finds out from a distributor email in October.
4. :car: _The car stopped, the driver assumed it saw the cyclist, and it hadn't. MIT built a readout that catches the AI being right for the wrong reason.
The Rundown AI · Jennifer Mossalgue · September 2, 2026Buildable AI
- What it is. MIT and autonomous-driving company Motional built CW-Net, a system that makes part of a self-driving car's planning process visible while it drives — surfacing named concepts like "close to cyclist" and "approaching stopped vehicle" next to the trajectory it has chosen. Published in Nature on September 2; MIT announced it the same day; deployed in a Motional robotaxi on a private track generating explanations in real time. Crucially, the concept layer feeds the stage that scores trajectories — so the explanation isn't a narration bolted on afterward, it has a role in the decision.
- The test case is the whole item. A safety driver watched the car stop and believed it had stopped because it accounted for a nearby cyclist. The concept display said otherwise: cyclist activation was low. Investigation found the perception system had detected the cyclist, but the experimental planner was not configured to consume cyclist inputs at all — and a backup braking system was overriding unsafe trajectories. The car did the right thing. The reason was broken. Without the readout, the correct-looking stop would have increased confidence in a planner with a serious blind spot.
- A second one, same shape. A driver blamed a traffic cone for unexpected braking. The stopped-vehicle concept pointed elsewhere — and removing the cone did not eliminate the stop. The explanation turned a wrong guess into a testable hypothesis.
- What was and wasn't shown, stated as the authors state it. Two studies: an online study with 9 Motional experts and 30 non-experts, and a situational-awareness study analyzing 99 participants after attention checks. Reported improvements in predicting vehicle behavior, including during surprising events. "Whether the system reduces crashes remains unestablished." The supplementary validation shows errors in concept classification — a readable label doesn't prove what's happening. As of Sep 6 the authors' own project page calls it a proof of concept, tested on limited scenarios and one planner type, and notes drivers needed time to interpret the probabilities.
Why it matters for an independent practice: zero medicine in this, and it is the most directly transferable item in today's digest. One — the failure mode has a name now, and every practice using clinical AI has it: right answer, wrong reasoning. Your scribe produces a clean note. Your intake bot routes the patient correctly. Your imaging tool flags the right knee. Each correct output raises your confidence in a system that may be ignoring an input entirely — and you will only find out on the case where the backup human isn't paying attention. The cyclist was detected and then discarded, which is exactly how a scribe drops the one sentence about prior failed injections. Two — the practical version: when you evaluate any clinical AI, stop scoring it on whether the output is right and start asking what it used to get there. "Show me which inputs the model actually consumed" is a vendor question almost nobody asks and very few can answer, and it's the same question as the audit-trail item from Sep 3 and the vendor-disclosure item from Sunday. Three — the honest limit: MIT is explicit that the explanation itself can be wrong. A confident dashboard about a confident model is two layers of unearned reassurance. The value isn't certainty, it's a reason to look closer — which is what a physician's skepticism is supposed to be for.
