Watch the full webinar:AI Fundamentals for Wound Care Professionals on YouTube
Speakers:
- Dr. Lucian Vlad
- Wound care clinician presenting from the practitioner’s perspective
- Dr. Jordon Gilmore, PhD
- Associate Professor of Bioengineering, Clemson University
Moderator: Dr. Mohamed El Masry, Assistant Professor of Surgery, University of Pittsburgh; Chair, WHS Education Committee
Artificial intelligence (AI) in wound care is moving quickly from research labs into clinical practice, raising important questions for clinicians about safety, effectiveness, and real-world utility.
A Quick History That Matters
The term “artificial intelligence” was coined in 1956 at the Dartmouth conference, almost 80 years ago. For most of that time, AI progress was stuck. Scientists in the 1980s predicted human-level machine intelligence within decades, and it didn’t happen. The field went through multiple “AI winters” where a new program would generate hype, investment would drop, and the project would go quiet.
What changed in the last decade is raw computing power. Moore’s Law (transistors on a chip doubling every two years) produced hardware that can handle what the algorithms actually need. Specialized chips designed for AI workloads, plus cloud computing, made the tools we’re now seeing possible. The democratization of this specialized hardware is what pushed AI into clinical conversations over the last five years.
For context on scale: the global AI market is growing rapidly, and healthcare AI accounts for about 15 percent of it, the largest and fastest-growing segment. Investments hit $15 billion in 2023. North America leads in adoption, Asia-Pacific is accelerating thanks to government investment in countries with massive healthcare systems (China and India, especially), and Europe is moving more slowly due to stricter data privacy rules under the GDPR.
The IBM Watson Cautionary Tale
When IBM Watson beat human champions on Jeopardy in 2011, there was enormous optimism that the same technology could solve cancer. IBM partnered with MD Anderson, fed Watson Health a mountain of oncology data, and expected the system to recommend perfect treatments.
Unfortunately, it didn’t work. The system produced unsafe and incorrect cancer treatment recommendations in some cases. The collaboration ended after about seven years and roughly $4 billion in acquisition and R&D costs.
Why did it fail?
- The training data leaned heavily on case studies rather than randomized controlled trials
- The system couldn’t integrate properly with hospital EMRs
- There was insufficient clinical validation before scale-up
- Clinicians weren’t kept in the loop during development and deployment
The lessons from that failure are now the guardrails for every clinical AI deployment: keep clinicians in the loop, integrate with the EMR, ensure diversity in training data, start small, and validate clinically before scaling.
What AI Actually Is (And What It Isn’t)
AI is essentially the attempt to use computers to mimic human intelligence. Dr. Vlad described it as “monkey see, monkey do”, or essentially, the system imitates patterns that humans produce.
There are three theoretical categories:
- Narrow AI. Every AI system you currently interact with. Designed and trained for specific tasks. Image recognition for skin lesions. Readmission risk prediction. Language models like ChatGPT. All narrow AI.
- General AI (AGI). A hypothetical system with human-like intelligence across any domain. Doesn’t exist yet.
- Super Intelligence. Also hypothetical. The Skynet category. More philosophy than engineering at this point.
We’re entirely in the Narrow AI phase, according to Dr. Vlad.
How Machines Actually Learn
Traditional programming works by a human writing explicit rules. If the wound is X size, classify it as Y severity. This gets impossible fast because real clinical situations have thousands of variables and edge cases.
Machine learning flips the approach. Instead of writing rules, you show the system huge amounts of data and let it find the patterns. It’s how babies learn, through repetition, pattern recognition, and no explicit instructions.
There are three main approaches:
- Supervised learning. You give the system labeled data. “This is a venous ulcer. This is stage 2.” The system learns to match inputs to labels.
- Unsupervised learning. You give the system raw data with no labels and let it find patterns on its own.
- Reinforcement learning. A hybrid where the system gets feedback as it goes, correcting its predictions over time.
Dr. Gilmore shared a great side story on how this works in practice. Those CAPTCHA and reCAPTCHA systems where you type in distorted text to prove you’re human? That was originally created by graduate researchers who developed the early CAPTCHA security systems to distinguish human users from fake “bot” users. The creators quickly realized they could use this system to solve another challenge – reading and digitizing old texts, but they needed a massive amount of labeled data to train their models (supervised learning). Every time you solve a CAPTCHA or a reCAPTCHA puzzle, you’re actually labeling training data for AI models. You’ve probably already contributed to AI training without knowing it.
The Black Box Problem
Most of the concern around clinical AI comes down to this: with deep learning models, data goes in, predictions come out, and the layers in between are opaque even to the people who built the system. This is the black box problem, and it matters for a few reasons:
- Explainability. If an AI tool recommends a treatment, clinicians reasonably want to know why. Black box models can’t always tell you.
- Trust. Without explainability, clinicians hesitate to act on AI recommendations, especially when the stakes are high.
- Automation bias. When a system keeps producing outputs that sound reasonable, users stop questioning them. Errors compound.
- Hallucinations. AI models can confidently make up information that isn’t real. Fake news, essentially, but dressed up as clinical insight.
A lot of current AI research is focused on making the black box more transparent through techniques like SHAP values, which help show which variables most influenced a specific prediction. Dr. Gilmore showed a real example of this in the webinar, using a model predicting whether a wound would heal in four weeks, with each variable’s contribution to the prediction made visible. Tools like that are how explainability starts to move from theory to practice.
Where AI Is Already Working in Medicine
This is where the conversation gets grounded. AI has already hit solid clinical validation in several fields:
- Radiology. FDA-validated software reading CT scans and X-rays at levels comparable to human radiologists
- Ophthalmology. Google’s AI can identify retinopathy from fundus photos, which matters in regions with limited specialist access
- Cardiology. The smartwatches tracking heart rhythms are doing narrow AI work
- Pathology. Significant validation for digital slide analysis
Wound care is earlier on the adoption curve, but real applications are emerging.
AI in Wound Care Today
The presenters walked through several categories of applications currently being developed or validated.
Large Language Models
ChatGPT, Claude, Gemini, and domain-specific models like AvoMD are being used for:
- Literature reviews
- Patient education material development
- Summarizing and structuring clinical notes
- Interpreting clinical guidelines
Dr. Gilmore showed an example flyer he generated in seconds by prompting ChatGPT with “create a flyer for self-care diabetic foot ulcers with key components of care and safety” in shades of green. The model filled in reasonable content on its own. Not perfect, but a solid starting point that a clinician can review and refine.
Dr. Vlad mentioned Google NotebookLM as a free collaborative tool where you can control exactly which sources the AI draws from. His group is using it to analyze wound care standards across multiple guidelines and flag areas where standards differ.
Image-Based Wound Diagnosis
Image recognition is a natural fit for wound care because clinicians are already making decisions based on visual inspection. Models are being trained to:
- Measure wounds automatically
- Classify tissue types in the wound bed
- Detect signs of infection or osteomyelitis
The hard part is data. Most imaging AI progress in radiology happened because large, open, free image datasets existed for researchers to train on. Wound care doesn’t have that yet. Building those datasets with proper diversity across skin tones, wound types, and clinical contexts is one of the most important things the field needs next.
Predictive Analytics
Models are being developed to forecast:
- Healing time
- Infection risk
- Treatment optimization
- Outcome prediction
The key consideration here is bias. If your training data overrepresents certain patient populations, your model’s predictions will be biased in ways that harm underrepresented patients. Dr. Gilmore gave an example: if dark skin is overrepresented in amputation outcomes in your training data, the model may incorrectly associate dark skin with amputation risk.
This is why thoughtful data curation matters as much as algorithmic sophistication.
Remote Patient Monitoring
This is an area already showing meaningful clinical results. Remote monitoring platforms allow patients to:
- Send wound images directly from home
- Track compliance with treatment protocols (including NPWT device use)
- Get just-in-time feedback
One study cited showed that digital tracking was associated with healing in 15 days versus 35 days without digital tracking, along with fewer nursing visits required. That’s a significant clinical signal.
Other remote monitoring innovations include smart mats that measure plantar pressure and temperature, socks and shoe inserts with embedded sensors, and smart offloading boots that ping a watch reminding the patient to put the boot back on.
Virtual Reality and AI Agents
Virtual reality is being explored for pain management and patient education. Conversational AI agents (think of how Siri and Alexa work) are being tested for patient coaching and adherence support outside the clinic.
What To Watch For If You’re Considering AI Tools
Both presenters made clear that this is a first-in-series webinar and much of the evaluation framework will come in later sessions. For clinicians evaluating AI tools, the presenters emphasized five practical questions to ask before adoption:
Start with the data questions. AI engineers will tell you the quality and diversity of training data matters more than the algorithm. If a vendor can’t explain what their model was trained on, that’s a red flag.
Demand validation specific to your patient population. A model validated on one demographic may not perform well on another. Ask about validation across skin tones, ages, and comorbidities that match your patients.
Insist on explainability. Any tool that can’t explain why it reached a conclusion makes it harder to catch errors and harder for clinicians to trust.
Connect AI output to something clinically familiar. Dr. Gilmore’s strong recommendation was that any AI decision should trace back to a clinically validated measure that clinicians already know and trust. Images are a natural anchor because clinicians already reason from them.
Understand the regulatory and privacy path. Where does patient data go when you use the tool? How long is it retained? Is the tool FDA-cleared for the use you’re putting it to?
This article is a recap of a WHS Education Committee webinar presented October 8, 2025, featuring Dr. Lucian Vlad and Dr. Jordon Gilmore, moderated by Dr. Mohamed El Masry. It is intended as an introduction to AI concepts in wound care and does not constitute clinical or technology implementation guidance.