AI to Improve Healthcare Delivery | International Perspectives

Session Focus

Barrow works at the Cambridge Innovation Center with international startups and scaleups entering the US market, helping with strategy, introductions and coaching on how to present to a US audience. All four panel companies are European and have worked with her previously.

Her framing noted that several of them were doing AI before it was called AI. The discussion covered what each company’s technology does, how it produces value for hospital customers, and what changes when a European company enters the US market.

What the Companies Do

Frederik Bender described Piur Imaging, a medical device company from Vienna providing tomographic 3D ultrasound. The product is an add-on for third-party ultrasound systems that brings the modality from 2D into 3D, then runs AI over volumetric datasets to help doctors and sonographers standardize ultrasound, work more efficiently and produce more reproducible outcomes.

Declan Kelly MD, an emergency physician by background, founded Eolas Medical, named for the Irish word for knowledge. The problem came from his own practice: treating a patient having a major heart attack, or one whose cancer treatment has failed, clinicians struggle to access knowledge held within their own health system, because the information is scattered and often sitting on SharePoint. The company indexes that content with AI so a clinician can ask a question and receive an answer within seconds. It is used by roughly 90 percent of UK hospitals, many hospitals across Ireland and about 40 US centers.

Harald Schnidar PhD, a genetic engineer who worked in clinical skin drug research, founded Scarletred in Vienna with a subsidiary at the Cambridge Innovation Center. The problem originated in a study on subjective skin assessment, and he illustrated it from the stage: his skin looks different under the bright panel lighting than it would under other conditions. The company’s smartphone application, fully certified as a medical device, auto-calibrates images taken by a clinician or by a patient at home, with AI accounting for lighting, distance, angle and other parameters that normally prevent quantitative extraction. It can assess any of roughly 3,000 skin diseases. Customers are biopharmaceutical companies using it for safety and efficacy assessment in clinical trials, with expansion underway into remote care.

Jukka Tapaninen MSc leads Aiforia, whose name stands for artificial intelligence for image analysis. The company is listed on NASDAQ Helsinki, with entities in France and the US and a Boston office. It supports pathologists in diagnosing tissue samples, offering both a development platform for building models and a portfolio of regulatory-approved models covering breast, lung, prostate, colon and gastric, with 12 CE IVD marked solutions. Five years of investment went into building the offering and reaching production customers, the largest US customer being Mayo Clinic for three years, with the Paris hospital region the largest in Europe alongside Milan, several Spanish regions and the Nordics.

Where the Value Comes From

Tapaninen located the value in a workforce problem. There are roughly 100,000 pathologists globally, and a US study found 40 percent are over 60 years old. The shortage is already present and will worsen.

His system works on digitized tissue images, producing heat maps, quantification and a prepared report. The pathologist’s job becomes verifying results, after which the case moves automatically to the next step, which delivers both efficiency and accuracy alongside more detailed data supporting treatment decisions. His figure: pathologists can do two to three times more work with better accuracy.

Bender explained why ultrasound is unusually suited to this. Unlike CT or MR, where images are captured into a volumetric dataset and reviewed retrospectively, ultrasound is real-time, meaning the person holding the probe is diagnosing while the patient is present. That makes it highly operator-dependent and slow.

In thyroid, his focus example, many US sites allocate 30-minute slots per patient. Conventional practice involves placing the probe on the neck, rotating it, generating 2D screenshots, measuring, generating more screenshots and measuring again. The alternative is a quick sweep over each side of the thyroid taking 10 to 20 seconds, captured as a volumetric dataset.

AI does two jobs. During acquisition it estimates the probe’s trajectory to reconstruct the volume. The more valuable part comes afterward, when it works through the dataset identifying thyroid and nodule borders, placing measurements, adding annotations and preparing the report. The clinician corrects what needs correcting, because no AI is perfect, and submits.

Studies showed roughly a 60 percent reduction in time from scan to report, with fewer repeat examinations as a secondary benefit of reduced operator dependency.

Kelly offered the clearest framework for adoption. Any AI solution has to fall into one of three buckets for a health system to want it: improving quality and safety of care, improving efficiency of care, or reducing the cost of care delivery.

His company addresses the first two. Indexing health system content and streamlining it to the point of care, accessible by mobile, desktop or directly in the electronic record, significantly increases adherence to policies, procedures and guidelines, with downstream clinical consequences including effects on prescribing behavior. On efficiency, finding information has been shown to be ten times faster than existing systems.

His stated next objective is proving that this increases capacity to see more patients, rather than producing only deflationary economics.

Building the Infrastructure Before the AI

Schnidar described a deliberate sequencing decision that distinguishes his company. Scarletred did not call itself an AI company before AI became fashionable, despite having AI systems in place, because the first problem it solved was subjectivity and standardization. AI came second.

The infrastructure built first is now the company’s key asset for scalability. The underlying issue is that images captured on different hardware cannot be compared: sensors differ, resolution differs. A physician taking a skin image is therefore not automatically generating clinical data.

The standardization layer, which relies on a small skin patch applied before capture to enable auto-calibration, allows standardized images from any smartphone and makes documentation uniform.

Once an image is uploaded, AI performs the assessment, validated through roughly 100 studies worldwide, many conducted in the US. The automated measures include photometric changes such as erythema and pigmentation, texture change, and planimetry covering two and three dimensional change in a wound.

Key Takeaways

1. Three buckets determine adoption: quality and safety, efficiency, or cost of delivery. Anything outside them struggles to get bought.

2. Pathology’s workforce problem is arithmetic. Roughly 100,000 pathologists globally and 40 percent of US pathologists over 60.

3. Ultrasound’s operator dependency is the opportunity. A 10 to 20 second sweep replaced manual 2D capture, cutting scan-to-report time by about 60 percent.

4. Health system knowledge is scattered and hard to reach. Indexing it improved guideline adherence and made retrieval ten times faster.

5. Standardization precedes intelligence. Scarletred built calibration infrastructure first, which became the scalability asset.

6. Efficiency gains still need a capacity argument. Kelly named proving increased patient throughput as the outstanding question.