Hi!

Welcome to AIMedily.

This week, I’ll be helping test the Variable Stiffness Orthosis (VSO) at the Neurobionics Lab at the University of Michigan.

One of the things I enjoy most about being close to this work is seeing how thoughtful engineering can help improve what is possible in rehabilitation medicine.

You can read more about this orthosis here.

Let’s dive in.

Researchers tested whether AMIE, a conversational diagnostic AI system, could reason through uploaded medical images and documents during simulated telehealth visits.

Study design: Randomized, blinded, exploratory study

AI system: AMIE, using Gemini 2.0 Flash with a state-aware dialogue framework

Scenarios: 105 simulated telehealth cases using multimodal clinical scenarios

Data included:

• dermatology photographs
• electrocardiograms (ECGs)
• clinical documents

Comparison: Board-certified primary care physicians

Evaluators: 18 specialist physicians

Outcomes included:

• diagnostic accuracy
• history-taking
• management reasoning
• communication quality
• empathy
• multimodal reasoning

• AMIE had higher diagnostic accuracy than primary care physicians (P < 0.001).

• Specialists rated AMIE higher on 29 of 32 evaluation areas.

• AMIE performed better on 7 of 9 multimodal reasoning measures.

• Patient actors rated AMIE higher for explaining findings from uploaded medical artifacts.

• AMIE appeared more robust when image quality was lower.

• This was a simulated study, not a real-world clinical trial.

• This study moves beyond text-only medical chatbots.

• Multimodal AI may be more useful when it can reason through what patients already share: photos, test results, and documents.

• Strong simulation results do not mean the tool is ready for clinical use.

• Real-world validation is still needed.

🔗 Saab K, Park C, Strother T, et al. Advancing conversational diagnostic AI with multimodal reasoning. Nature Medicine. 2026;32:1726-1736. doi:10.1038/s41591-026-04371-0

Researchers tested whether an AI risk model could help high-risk patients receive faster mammography results and same-day follow-up when needed.

Study design: Prospective controlled study

Setting: Urban safety-net hospital

AI model: Mirai 1-year breast cancer risk score

Patients: 4,145 women undergoing screening mammography

Workflow: Patients above a top 10th-percentile risk threshold were flagged as high risk.

On enrollment days, they were offered immediate screening interpretation and same-day diagnostic evaluation when indicated.

Outcomes included:

• feasibility
• time to screening result
• time to diagnostic evaluation
• time to biopsy
• cancer detection rate

• 525 patients were flagged as high risk.

• 100 women consented to expedited care.

• 94% received immediate screening interpretation.

• 26 women received same-day diagnostic evaluation.

• Median time to screening result decreased from 191.9 minutes to 13.0 minutes.

• Median time to diagnostic evaluation decreased from 852.8 hours to 1.27 hours.

• Median time to biopsy decreased from 59.0 days to 20.1 days.

• Among patients with screen-detected cancers, mean time decreased by 99.1% for screening results, 99.1% for diagnostic evaluation, and 87.2% for biopsy.

• Cancer detection rate was 60.0 per 1,000 in the expedited high-risk group and 24.4 per 1,000 in high-risk controls.

• The AI model helped prioritize patients who may benefit from faster follow-up.

• The model did not replace radiologists. It was used for triage only.

• The safety-net setting matters because delays after abnormal mammography can worsen disparities.

• Larger studies are needed to test scalability, false positives, cost, and long-term outcomes.

🔗 Chung M, Davis E, Greenwood H, et al. Prospective deployment of AI-based risk stratification to enable expedited mammography workflow in a safety-net setting. npj Digital Medicine. 2026. doi:10.1038/s41746-026-02743-x.

• A clinical decision-support tool built around images and differential diagnosis.

• You can use it to compare visual findings, review condition-specific images, and think through possible diagnoses.

• A platform for creating medical and scientific visuals for teaching, research, patient education, and presentations.

• You can use it to create clearer figures, diagrams, and visual explanations without starting from scratch (you can try it free).

• An AI notepad that turns meeting audio and rough notes into organized summaries.

• You can use it to summarize discussions, identify action items, and search across past notes.

This is what’s been happening in AI in medicine this week:

Papers worth your attention this week:

Which medical discovery started after a scientist noticed that a covered tube made a nearby screen glow?

A) Ultrasound
B) X-rays
C) MRI
D) Pacemaker

Now, time for last week’s TriviaRX answer: ✅ A) Visible cyanosis or arterial blood gases.

Before pulse oximetry, clinicians often relied on cyanosis or arterial blood gas testing to detect low oxygen levels.

That’s it for today.

As always, thank you for taking the time to read.

If this issue helped you understand where medical AI is heading, feel free to send it to a colleague 📩 AIMedily (thanks for sharing!)

See you next week.

Itzel Fer, MD PM&R

Follow me on LinkedIn | Substack | X | Instagram

Forwarded this email? Sign up here

P.S. Enjoying AIMedily? 👉 Write a review here (it takes less than a minute).