Healthcare AI Navigator

The latest in healthcare AI. Insights, breaking news, journal findings & real-world applications for clinicians and healthcare leaders. #HealthcareAI

[Healthcare AI Navigator @HealthAINav]

An interesting one next week: the European Society of Cardiology is hosting its Digital & AI Summit in Switzerland.

One discussion I'm particularly interested in is whether hospitals should build their own AI solutions or purchase existing ones. As more tools enter clinical practice, that decision is becoming increasingly important.

[Healthcare AI Navigator @HealthAINav]

A smaller, hospital-trained AI model outperformed much larger general-purpose models on several real-world prediction tasks.

NYU Langone's Lang1 was trained using clinical records from more than 10 million patients and tested on tasks including 30-day readmissions, in-hospital mortality, length of stay and insurance denials.

The researchers report an average AUROC of 85.2% across five tasks for their fine-tuned 1-billion-parameter model. The largest gains over general-purpose systems were for administrative predictions, not every clinical outcome.

This is an interesting shift away from asking AI medical exam questions toward predicting what hospitals actually need to plan for.

But better prediction scores do not automatically mean fewer readmissions, safer discharges or lower costs. Those benefits still need to be tested in clinical workflows.

The future of healthcare AI may depend as much on specialized data and implementation as model size.

Source: NYU Langone, November 2026; Lang1 preprint, March 2026.

[Healthcare AI Navigator @HealthAINav]

AIMed26 starts today in Orlando, bringing clinicians and AI developers together to discuss how AI is actually being used in healthcare.

I'm especially interested in the practical implementation discussions. There's no shortage of promising AI tools, but getting them safely integrated into clinical practice is another challenge entirely.

[Healthcare AI Navigator @HealthAINav]

https://t.co/gBdWETxFbn

[Healthcare AI Navigator @HealthAINav]

At the American College of Rheumatology’s annual conference in Orlando this weekend, a speaker asked a question worth bringing back to every hospital:

Are we measuring the success of healthcare AI the wrong way?

Speaking at ACR Convergence 2026 on November 7, Brian Anderson, CEO of the Coalition for Health AI, challenged how we evaluate tools like AI scribes.

Rather than focusing on how many extra patients clinicians can see, he argued we should ask whether AI helps them listen more closely, document more accurately, and keep patients’ voices central to their care.

He also called for specialty societies and frontline clinicians to help shape AI standards, with greater transparency about model limitations and accountability for AI agents in clinical workflows.

Saving time matters, but it shouldn't be our only measure of success.

As AI becomes part of everyday care, are we measuring what actually matters to patients and clinicians?

Source: ACR Convergence Today, November 8, 2026.

[Healthcare AI Navigator @HealthAINav]

Could a single question in the emergency department help identify patients at risk of PTSD?

Researchers at NYU Langone tested PREVIA, an AI system that analyzes patients' descriptions of the traumatic event that brought them to the ED.

In a study of 145 patients across two trauma centers, it reported 95.4% sensitivity for identifying PTSD one month later. But precision was only 46.8%, meaning fewer than half of its positive predictions were correct.

In a separate head-to-head assessment of 30 patient narratives, including nine PTSD cases, PREVIA identified all nine, compared with eight identified by an experienced ED clinician.

It's a promising approach to identifying people who might otherwise miss follow-up support. But this was a small study within one health system, not evidence that using the tool improves outcomes.

If clinicians get an early warning, what follow-up pathway needs to be in place for it to be useful?

Source: Nature Communications, November 2026.

[Healthcare AI Navigator @HealthAINav]

Could a patient's voice warn clinicians that their heart failure is getting worse, weeks before they need care?

In the DETECT-HF study, 563 patients with chronic heart failure recorded six short sentences on a smartphone each day. Cordio Medical's HearO AI analyzed subtle voice changes.

It detected 52 of 76 first heart failure events (68.4% sensitivity), on average 24.1 days before clinical presentation. Daily recording adherence was 85%.

Promising, but there's a catch: the study narrowly missed its prespecified sensitivity target, and clinicians did not act on the AI alerts as part of the trial.

So we know the technology can identify an early warning signal in many patients. We don't yet know whether responding to those alerts reduces admissions or improves outcomes.

The practicality is appealing: no implant or special hardware. HearO is not yet cleared for sale in the US. The next step is proving what happens after the alert.

Source: DETECT-HF, JACC: Heart Failure, November 2026.

[Healthcare AI Navigator @HealthAINav]

@HealthcareAIGuy @pewresearch Interesting that people are even less comfortable with AI making insurance claim decisions than medical diagnoses. I wonder if that reflects concerns about the technology itself, or a broader lack of trust in how insurers might use it to make decisions about their care.

[Healthcare AI Navigator @HealthAINav]

@npjDigitalMed @KyleVerrier @AchilleNazaret I can see some interesting applications for dietitians here, particularly for patients who struggle with detailed food diaries. But I'd be curious whether more accurate predictions actually translate into behavior change.

[Healthcare AI Navigator @HealthAINav]

I wonder if we're at risk of recreating the same alert fatigue problems we've seen with clinical decision support systems. If AI agents are constantly blocking legitimate actions or requiring unnecessary approvals, clinicians may eventually stop trusting the safeguards altogether.

[Healthcare AI Navigator @HealthAINav]

@AegisOpsLLC @FDA The gains for junior pathologists are interesting, but I'd also like to see what happens to specificity. Are we catching more melanomas without significantly increasing false positives and unnecessary follow-up? That balance will be important in practice.

[Healthcare AI Navigator @HealthAINav]

@AMIAinformatics It makes me think multilingual AI shouldn't be treated as a single capability that can simply be switched on. Each language really needs its own validation before these tools reach patients.

[Healthcare AI Navigator @HealthAINav]

@angcharlesli Achieving 98.4% with less than half the parameters is pretty impressive. Makes you wonder how much computing power we really need for some of these specialized biological AI tasks.

[Healthcare AI Navigator @HealthAINav]

@rediminds I think the finding about standalone detectors is particularly interesting. A tool can be very good at detecting the conditions it was designed for, but that doesn't necessarily make it safe for deciding which patients can wait. Two very different jobs.

[Healthcare AI Navigator @HealthAINav]

@pash22 @JacobKnorr10 I actually think one of the biggest opportunities here is freeing up researchers to spend less time extracting data and more time analyzing it. I'd be interested to know how much human review would still be needed to maintain accuracy at scale.

[Healthcare AI Navigator @HealthAINav]

This is a pretty incredible use of AI for medical research.

Being able to pull together pathology images and clinical information from hundreds of papers in minutes could save researchers an enormous amount of time.

I'd be interested to see how accurately it extracts and validates the data, but the potential here is huge.

[James Zou @james_y_zou]

Paperclip Datasets is very useful for creating multi-modal datasets!

For example, in minutes it curated a dataset of post-immunotherapy pathology images from hundreds of papers. https://t.co/IXBUigJcjM

[Healthcare AI Navigator @HealthAINav]

@primeai9 I wonder how many healthcare organizations are actually reviewing their AI audit logs rather than just collecting them. Logging activity is one thing, but having someone responsible for identifying and investigating potential issues is another.

[Healthcare AI Navigator @HealthAINav]

@David_Ouyang Interesting that around half of all preliminary reports still needed substantial changes, regardless of who or what drafted them. Makes me wonder how much of this reflects genuine errors versus differences in how cardiologists interpret the same images.

[Healthcare AI Navigator @HealthAINav]

@zakkohane The 43 vs 138 days is pretty striking. I'd be interested to know how much of that improvement came from the AI identifying high-risk patients versus simply getting those patients into rheumatology faster. Could the referral process itself be improved without AI?

[Healthcare AI Navigator @HealthAINav]

OpenEvidence just secured access to another major source of clinical evidence.

The American College of Physicians (ACP) has announced a licensing agreement that will bring its clinical guidelines and content from Annals of Internal Medicine into OpenEvidence's AI platform.

That's a significant addition. ACP is one of the largest physician organizations in the US, and its guidance is widely used in clinical practice.

For clinicians, the appeal is obvious: faster access to trusted recommendations during consultations, without having to search through multiple resources.

But there's an important distinction between giving AI access to high-quality evidence and proving that its answers are accurate, complete and appropriately applied to individual patients.

We're seeing clinical AI platforms compete not just on their models, but on the medical knowledge they can access.

Will the quality of their evidence libraries become just as important as the AI itself?

Source: American College of Physicians, November 2026.

#healthcareai #digitalheatlh