200: Artificial Intelligence in Healthcare: From Diagnosis to Rehabilitation

Artificial Intelligence in Healthcare: From Diagnosis to Rehabilitation. Witek K, Nowocien M, Gerlach J, et al. Cureus 2026 Jan 25;18(1):e102286.

Episode Summary: In this journal club deep dive on the Digital Pathology Podcast, we completely bypass the venture capital hype and science fiction narratives to look strictly at the hard clinical evidence surrounding artificial intelligence in medicine. We examine a monumental 2026 narrative review synthesizing a full decade's worth of data across the entire healthcare continuum—from diagnosis to rehabilitation. We explore the proven clinical benefits, the structural limitations, and the highly unpredictable reality of integrating these advanced algorithms into live clinical workflows.

In This Episode, We Cover:

• The Diagnostic Powerhouse: Why AI truly shines in visually intensive specialties like radiology, ophthalmology, dermatology, and digital pathology. We also unpack the crucial bottleneck: why algorithms that achieve board-certified performance in "open book" retrospective lab settings often struggle when faced with the messy, artifact-heavy reality of a live clinic.

• Laboratory Medicine & LIS Optimization: How AI is functioning as a massive force multiplier behind the scenes. We discuss AI-driven lab test checkers that run continuous delta checks, acting as an algorithmic safeguard against inevitable human cognitive traps like anchoring bias during high-stress, 12-hour shifts.

• Physical Rehabilitation & Robotics: AI stepping out of the computer monitor and interacting directly with the physical world. We explore robotic hand exoskeletons that process real-time electromyiography data to adapt to stroke patients millisecond by millisecond, and the use of large language models (LLMs) to design personalized therapy programs. We also discuss why massive multi-center prospective validation is required before these become the standard of care.

• Conversational Agents (Chatbots): The delicate deployment of chatbots to bridge gaps in patient education and hold the line with immediate interventions for vulnerable individuals stuck on mental health waitlists. We emphasize why these agents must remain strictly as clinical adjuncts and triage tools, not replacements for empathetic human caregivers.

• The Four Pillars of Friction: The massive structural hurdles preventing immediate global deployment: generalizability and algorithmic bias, the "black box" of algorithmic transparency, infrastructure limitations, and the scramble by organizations like the FDA and EU to establish proper regulatory oversight.

Key Takeaway: The ultimate takeaway from a decade of data is that AI is a supportive clinical decision support technology, emphatically not a replacement for human healthcare professionals. The future of healthcare is the convergence of human and artificial intelligence; by letting algorithms absorb the heavy lifting of routine data verification, we may finally create the necessary breathing room to make clinical medicine profoundly human again

Get the "Digital Pathology 101" FREE E-book and join us!

Transcript

Welcome trailblazers. You are listening to the digital pathology podcast.

We are uh incredibly glad you could join us today.

Yeah. For today's journal club style deep dive. We know you are out there on the front lines of healthcare right now

navigating some really massive very rapid technological shifts in your departments.

Exactly. So today we're completely bypassing the usual hype. We're ignoring the venture capital buzzwords and you know the science fiction narratives surrounding artificial intelligence and medicine,

right? because we need to put everything under the microscope and look strictly at the hard clinical evidence.

It's the only way to do it.

It really is. The landscape of medical tech is shifting so quickly that it's well, it's become genuinely difficult to separate the theoretical promises from actual boots on the ground clinical utility.

Right.

So, our mission today is to critically curate the current applications of AI across the entire healthcare continuum for you

for the trailblazers listening.

Exactly. We are going to examine the proving clinical benefits, the very real structural limitations, and the practical day-to-day challenges of actually implementing these tools today.

To do that, we're focusing our deep dive on a really comprehensive narrative review. It was just published in January 2026 in the journal Curious,

right? Titled artificial intelligence in healthcare from diagnosis to rehabilitation.

And we are looking at a monumental effort here. It's a massive cross-disciplinary team led by Carolina Wake

a huge team. Martin Oosian, Joanna Gerlac, Natalyia Gusk,

Barbara Blevitz, Lucas Sewick, Carolina Lichuala,

Olivia Sepiora, Yakub, Andra Zijovich, and Monica Cleopella. They took on this completely herculean task of synthesizing a full decade's worth of data.

Yeah. From major biomedical databases, PubMed, Scopus, Web of Science, Ambass,

it really is a panoramic view of where AI stands in modern medicine right now.

By qualitatively synthesizing evidence across everything from diagnostic imaging to laboratory diagnostics,

even physical rehabilitation technologies and conversational agents.

Yeah. Whitek and her team have basically provided a road map of current capabilities

and our goal is to extract the most critical insights from that road map so you can understand exactly how these developments are going to impact your workflows

and obviously your patient outcomes.

Right.

Okay. Let's unpack this. We should probably start with what is arguably AI's most mature and famous medical application.

The diagnostic powerhouse.

Yeah. Imaging and pathology.

The diet fights team gathered clearly validates why this is the primary association most professionals have with medical AI.

They evaluated systems utilizing convolutional neural networks and deep learning architectures. Right.

Yes. And the evidence shows these systems are achieving diagnostic performance that is fully comparable to human healthare professionals

in highly specific well-defined tasks.

Exactly. The specialties where AI truly shines right now are inherently visual. They rely on massive data sets of pixel level information. So the review highlights significant measurable triumphs in radiology. Mimography for breast cancer detection.

Opthalmology is a big one specifically looking at the identification of diabetic retinopathy across multithnic populations.

That's amazing. And dermatology is another major focus area particularly in the automated classification of skin cancer lesions.

And naturally digital pathology is right at the forefront of this imaging revolution.

Of course these algorithms are extraordinarily adept at pattern recognition.

Yeah. Picking up on subvisual features within tissue architecture that the human eye might honestly just gloss over.

I mean, it makes sense when you think about the computational power involved. You have these deep learning systems acting as a tireless second set of eyes

trained on millions of annotated whole slide images or radioraphs.

But I do have to push back a little here.

Go for it.

If the data shows these models are matching board certified specialists in say identifying a melanoma or flagging a micro metastasis on a lymph node slide, Why isn't every single pathology department and radiology reading room running autonomously right now?

That is the million-dollar question.

Like what is the actual bottleneck?

What's fascinating here is the crucial caveat the authors emphasize regarding how we evaluate that performance.

Okay.

Yes, the high performance results are statistically impressive, but you have to look closely at the environment in which they were achieved.

Wow.

Wid and her colleagues point out that these triumphs are predominantly happening under retrospective valid. ation or highly controlled study conditions,

which is a massive distinction for anyone deploying these tools. We really need to clarify what that actually looks like in practice.

Exactly.

Retrospective validation essentially means the AI is taking an open book test,

right? It's being fed perfectly curated, digitized, pristine data from past cases.

The slides are perfectly stained. The lighting in the dermatology photos is optimal. The LIS data is neatly categorized.

It is a totally sterile environment. That is a fundament ally different ecosystem from the chaotic unpredictable reality of a live clinical setting.

Totally. In the real world, a digital pathology slide might have a fold in the tissue section

or the H& staining might be weak because the regent was old.

There might literally be a fingerprint on the glass.

Exactly. In dermatology, the patient might have a tattoo right near the lesion or the lighting in the exam room is just terrible.

And when these pristine algorithms encounter realworld artifacts,

their performance metric often degrade rapidly.

A controlled laboratory benchmark is just not a bustling Tuesday morning clinic.

Not at all.

So, while the diagnostic potential is immense, we cannot just assume an algorithm that aced a retrospective test is going to flawlessly transition to live patient care.

It requires rigorous prospective validation,

putting the AI into the messy daily workflow and seeing if it still holds up over time,

which the review notes is still severely lacking in any clinical domains. The translation from the controlled lab to the live clinic is where the friction happens.

It's not just about having a smart algorithm.

No, it's about having an algorithm robust enough to handle the infinite variability of human biology

and clinical workflows which perfectly bridges us to the next major area the review tackles. Right?

We are moving from the highly visual world of diagnostic imaging to the slightly less glamorous but absolutely vital world of laboratory medicine. For all the trailblazers managing lab workflows or relying on LIS data, it turns out AI is quietly becoming a massive force multiplier behind the scenes.

It isn't just about reading a slide anymore. It's about managing the entire data pipeline.

In laboratory medicine, we're seeing AI based tools move way beyond pattern recognition. We're talking systemic workflow optimization.

The review details how these systems are assisting in complex multivariable result interpretation

and providing really essential clinical decision support. Think about the immense volume of data flowing through a modern clinical lab on any given day.

It is staggering.

It really is. AI is being deployed to manage that volume efficiently and safely functioning as this intelligent layer over the laboratory information system.

The review brings in some very compelling concrete data here too. They specifically cite a 2024 prospective cohort study by Sumillis and colleagues.

Right. They looked at AIdriven lab test checkers,

automated systems designed specifically to improve diagn agnostic accuracy and safety. They run continuous delta checks and cross reference patient history against incoming lab values.

They act as automated safety nets, catching subtle inconsistencies or complex physiological red flags in a metabolic panel before those results ever reach the treating position.

And that safety net is critical because the review makes a fascinating connection to the reality of cognitive limitations in medical decisions.

Yeah. In a high stress, high volume laboratory setting, human error isn't a possibility, it's an inevitability.

Fatigue sets in at the end of a 12-hour shift.

And we are all susceptible to cognitive biases, particularly when interpreting complex lab panels under extreme time pressure.

Let's contextualize that for a second. Take anchoring bias for example.

Good example.

In a lab setting, anchoring bias is when a tech or a clinician sees one critically abnormal number, say a highly elevated potassium level,

and their brain just hyperfocuses on that single critical value.

Ex. Exactly. Because they are anchored to the potassium, they might completely miss a subtle creeping trend in the patient's creatinine levels.

A trend that signals an entirely different impending crisis.

Right?

A human brain just has a limited bandwidth for parallel processing. An AI lab test checker does not anchor.

Well, it doesn't get tired at 3:00 a.m. either.

No. It evaluates all 40 parameters of a comprehensive metabolic panel simultaneously, weighing them against the patient's historical baselines and known pharmacological interactions without prejudice.

So what does this all mean for you the trailblazers listening to this right now?

If you are managing these workflows, it means AI is stepping in to fundamentally reduce your cognitive load.

It is an algorithmic safeguard against the very human biases that can cloud medical decisions when you're overworked and overwhelmed.

It's not taking over the interpretation of a complex leukemia panel. It's just doing the heavy lifting of routine data verification

so you have the clearest, most accurate data set possible. before you apply your clinical judgment.

It allows the laboratory professional to actually operate at the top of their license,

freeing you up to focus on the truly complex cases,

right? The anomalous ones that require deep human expertise and nuanced contextual judgment

rather than getting bogged down in routine high volume verification that a machine can handle instantly.

Exactly.

Well, here's where it gets really interesting. We have spent the first half of this deep dive talking about analyzing gigapixel images and optimizing LIS data which are very screenbased digital applications

very softwareheavy

but W's review completely shifts the paradigm by diving into physical rehabilitation

yes

AI is actually stepping out of the computer monitor and interacting directly with the physical world and the patient's body

this was one of the most dynamic areas of the synthesized research moving from diagnostic software to AI enabled physical systems

the review highlights advanced robotics like hand exoskeletons used for poststroke rehabilitation.

And these are not just mechanical braces. They use embedded AI to process surface electromyiography or kinematic data in real time.

They adapt to the patients specific fluctuating motor deficits, dynamically adjusting the assistive torque they provide during a single therapy session.

The mechanics of that are just incredible. A robotic exoskeleton that learns how your hand is recovering from a neurological event and modulates its support millisecond by millisecond.

The review also brought up motion analysis platforms. Specifically for upper extremity rehab in patients with spinal cord injuries.

Using advanced computer vision, these systems can precisely track joint angles and movement trajectories without requiring the patient to wear clumsy markers,

providing real time AIdriven BOF feedback to optimize neuromuscular re-education.

And the integration doesn't stop at hardware either. The review sites fascinating developments where large language models LLM are being integrated into the rehab space.

Yeah, there is observational data showing LLM being used to design highly personalized adaptive rehab programs for chronic conditions

like knee osteoarthritis.

By continuously processing a patients evolving clinical history, daily reported pain levels and mobility goals, the AI can generate and adjust a tailored exercise regimen on the fly.

It sounds like having a highly educated, infinitely patient physical therapy assistant integrated directly into your clinical equipment.

It does, but I have to throw the flag again here.

Okay.

Are we really at the point where we are trusting an LLM to write a physical therapy script?

Mhm.

And trusting a robot to physically apply torque to a recovering stroke patient.

It's a valid concern.

What is the reality check here?

Well, the authors of the review are very explicit about throwing a red flag on this exact issue themselves.

Oh, good.

While the functional recovery enabled by these AI systems shows incredible promise in pilot studies, the evidence base currently is highly heterogeneous.

We're looking at a fragmented landscape of different proprietary devices, different software algorithms, and inconsistent measurement metrics.

Meaning, we're comparing apples to oranges across these studies.

Exactly.

More importantly, there is a glaring lack of massive multi-enter perspective validation.

Before these systems become the standard of care, we need rigorous randomized control trials to definitively prove they are safe,

are consistently superior to traditional human-led physical therapy over a long-term recovery. So if you are working in the rehab or orthopedic space, the technology is undeniably arriving, but it needs to be rigorously tested in the wild before it replaces traditional modalities.

Definitely.

Now, speaking of AI interacting directly with patients, let's move to a topic that has absolutely exploded in both clinical circles and the public consciousness.

Chat bots.

AI chat bots and conversational agents. The review asks a heavy question. Can an algorithm safely handle patient communication, triage, and even mental health intervent? It is a highly sensitive clinical area and the review approaches the data carefully.

What the synthesized research shows is that conversational agents demonstrate immense potential particularly in bridging gaps in patient education and providing low-level mental health interventions.

For example, Wik's team highlights research on empowering breast cancer clients through specialized AI chatbots.

These tools are deployed to transform knowledge and attitudes by answering routine yet vital questions about treatment side effects or medication schedules

providing consistent scientifically accurate information at any hour of the day which is huge.

It is and the mental health applications really jumped out of the data too.

The review references a 2025 trial focusing on depression interventions using AI chat bots that actually incorporate dynamic social cues into their text generation.

And even more compelling is a study on using a chatbot for a single session intervention for people stuck on weight lists for eating. disorder treatment.

That weightless scenario is a brilliant use case for this technology.

It really is.

Think about the reality of the mental health landscape right now. When human resources are stretched to the breaking point and a patient reaches out for help only to be told they are on a six-month wait list,

that is an incredibly dangerous window of vulnerability.

An AI chatbot can step into that immediate gap and provide evidence-based coping strategies like cognitive behavioral therapy exercises right in that critical moment.

It isn't curing them. But it is holding the line.

Exactly.

If we connect this to the bigger picture, the review makes a very firm crucial distinction regarding deployment.

Okay.

These conversational agents are proving highly effective for reducing baseline patient anxiety, for managing routine communication workflows, and for initial triage.

Right.

However, the review is emphatic that they must be viewed strictly as clinical adjuncts.

They are supportive tools.

They are not under any circumstances ready to act as primary ary caregivers or to replace the nuanced empathetic connection and complex risk assessment capabilities of a human physician or mental health professional.

They are the warm-up act, the triage desk, not the main event.

Perfectly said.

Which brings us to the grand synthesis of this entire review. We've looked at the data across digital pathology, laboratory LIS optimization, physical robotics, and conversational agents.

Pulling all these diverse domains together, White's team asks the ultimate question.

What is the friction? Then what are the structural hurdles holding us back from total seamless integration into every hospital in the world?

The authors identify four persistent major hurdles that the medical community must overcome. We can view them as the four pillars of friction.

Okay, let's go through them.

The first is generalizability, which ties directly into data representativeness.

An AI model is only as good as the data it was trained on.

The review points specifically to studies revealing severe performance perities in dermatology AI when evaluated on diverse clinical image sets.

Right? If an algorithm is trained predominantly on images of lighter skin tones, its diagnostic accuracy drops precipitously when applied to darker skin tones.

It is a fundamental data integrity issue that prevents the model from generalizing safely across a diverse patient population.

That is a massive structural flaw. You cannot deploy a diagnostic tool globally if it only works accurately for a specific genetic background.

Exactly. And that leads into The second and third pillars, algorithmic transparency and the necessary infrastructure.

We are dealing with the blackbox problem.

Clinicians need to understand how a deep learning model arrived at its conclusion, especially when it contradicts their own clinical judgment.

And furthermore, hospitals need the massive data infrastructure to securely process these gigapixel whole slide images and continuous robotic telemetry

without compromising patient privacy. Yes. Finally, the fourth pill is regulatory oversight, which is currently scrambling to keep pace with the technology.

How do you regulate a software tool that continuously learns and changes its own parameters?

It's tough. The review mentions frameworks like the FDA's artificial intelligence and machine learning software as a medical device action plan,

as well as the sweeping 2024 European Union artificial intelligence act.

Regulators are working aggressively to lay down harmonized international rules for deployment, but the algorithms are often evolving faster than the legislation can be drafted. It's a delicate balancing act promoting rapid clinical innovation while ensuring absolute uncompromised patient safety.

Which leads us to the final verdict of this massive narrative review.

The core conclusion from Whitek and her team across all these domains is incredibly clear.

AI is a supportive clinical decision support technology. It is emphatically not a replacement for human healthare professionals.

That is the ultimate takeaway from a decade of data. The future Healthcare is not artificial intelligence replacing human intelligence.

It is the convergence of the two.

Moving forward, the research must pivot away from celebrating retrospective controlled lab triumphs and focus entirely on perspective real world effectiveness.

We have to prioritize responsible validated integration into routine clinical practice. Ensuring these tools actually serve the workflows of the practitioners and enhance the outcomes for the patient.

The convergence of human and artificial intelligence.

Yeah.

Trailblazers. Thank you so much for joining us for this Journal Club deep dive on the digital pathology podcast.

We have covered a massive amount of ground today

from the first convolutional neural networks scanning a digital pathology slide to the algorithms running continuous delta checks in your lab workflows

all the way to the real-time kinematic adjustments of robotic rehab and the triage support of conversational agents.

The scope of this curious review proves that AI is touching every single corner of the clinical continuum.

This raises an important question for to consider as we wrap up today.

Okay.

If AI successfully absorbs the heavy lifting of diagnostic sorting, lab optimization, and preliminary patient education, how will that empty space change your daily workflow?

That is the real question.

Will the time saved make the clinical environment feel more automated and distant? Or will it finally give you the breathing room to make healthcare profoundly human again?

That is a phenomenal thought to leave on. As these tools move from the lab to your clinic, how we fill that newly created space is going to define the future of medicine.

Absolutely.

Keep pondering that. Keep critically questioning the data you see and keep blazing trails in your respective fields. We will catch you on the next deep dive.