Integration of artificial intelligence into modern healthcare systems is progressing at an increasingly rapid pace. Nevertheless, its large-scale implementation in real-world clinical practice still faces certain barriers.
Today, the main challenge is no longer merely developing high-precision algorithms — but rather how effectively the technology simplifies the physician’s daily workflow, saves time, and improves the quality of face-to-face communication with patients.
Ludwig Christian Hinske, Head of the Institute for Digital Medicine at Augsburg University Hospital and Professor of Data Management and Clinical Decision Support, spoke with us about the practical, real-world clinical value of artificial intelligence and the existing barriers to its integration.
Today, AI is discussed everywhere, yet its large-scale implementation in real clinical practice still lags behind. In your experience, what is the primary reason for this?
When we started working in this field, we thought the main task was building databases, training algorithms, and achieving high accuracy metrics. Over the years, however, we realized that this is only the first and easiest step. The real challenge begins when we have to translate this technology directly into the clinic, into the doctor’s daily workflow.
There is a huge gap between the thousands of successful algorithms published in the scientific literature and their actual integration into clinical practice. Fewer than 20% of physicians actually use AI tools. The main issue is that products are often built purely for algorithmic precision, while forgetting the clinical workflow. If an AI system doesn’t save the clinician time or is clumsy to use, they simply won’t adopt it.
Back in the 1970s — roughly 50 years ago — one of the first AI systems, MYCIN, was created. In diagnosing infectious diseases, its accuracy even outperformed human specialists. Despite that, it never caught on in clinical practice because it wasn’t integrated with other hospital systems, and doctors had to enter data manually. In other words, the architecture and algorithms existed half a century ago, but the primary bottleneck to this day remains their practical integration into clinical processes.
You also emphasized that AI is not merely a diagnostic tool and that it should support clinicians in other ways. What does this mean in practice?
Medicine isn’t just about algorithms, medicine is a human relationship. When a patient is sick, they don’t want a computer deciding their treatment — they need a doctor who will listen, show empathy, and offer hope.
Today, doctors — especially in intensive care units — spend 80 to 90% of their time at the computer filling out documentation, and only 10% with the patient. That should be the main focus of artificial intelligence: taking on the administrative burden, optimizing documentation, and preventing errors, thereby freeing up time for the physician to engage in face-to-face communication with the patient.
Alongside clinical adoption, one of the most pressing issues is ethics and bias. How does this problem manifest in medical data?
This is a massive challenge. Artificial intelligence learns from the data we feed it. Even if we deliberately withhold information regarding a patient’s sex, skin color, or ethnic background, AI still manages to infer these traits through indirect biomarkers — such as blood pressure curves or chest X-rays. As a result, AI covertly absorbs whatever cultural biases we might hold, which is why rigorous verification is critically important. If a model isn’t validated across diverse populations, it risks making erroneous or discriminatory decisions.
You cited an example where expensive AI systems turned out to be ineffective in practice. Why does a technology work perfectly from an algorithmic standpoint, yet fail to improve the final patient outcome?
This is what we call the “Evidence Gap.” There are AI devices on the market with exceptional technical parameters, yet clinical studies show that their implementation hasn’t actually reduced patient mortality or complications. For instance, a predictive algorithm for hypotension (low blood pressure) successfully reduced episodes of low blood pressure, but this didn’t meaningfully impact the overall clinical outcome for the patient. Meanwhile, manufacturers doubled the price of the equipment and sensors. Essentially, the hospital was paying twice as much for technology that yielded virtually no Clinical Value.
Therefore, before purchasing an AI product, we must ask: does this technology genuinely improve the patient’s health?
Implementing AI is not purely a technological endeavor. What main challenges do you foresee for the healthcare system in terms of scaling and funding?
Today in medicine, we are shifting from episodic care to continuous care (Continuous Care). Thanks to modern wearable devices and AI sensors, it is already possible to monitor a patient’s condition not just during clinical visits, but 24/7 from home, preventing complications proactively before they require hospitalization. However, this means that a doctor or system is no longer monitoring 100 or 1,000 patients, but entire regions — millions of people simultaneously under continuous surveillance. This raises a fundamental question: how do we pay for this, and how do we fund such a system?
Existing healthcare and insurance models are tailored to the traditional approach — paying per discrete clinical visit, rather than for continuous monitoring. When observing patients at this scale continuously, technological, organizational, and financial models must be completely re-evaluated. Furthermore, AI models require ongoing validation and updating over time as patient demographics and clinical workflows shift.
Implementing AI isn’t a one-off event; it’s a socio-technical process where technology, clinician, and patient form a single ecosystem. This requires training medical staff, engaging them on how and why they want to use technology, and fostering the right organizational culture.
How should the true clinical value of artificial intelligence be evaluated?
Artificial intelligence holds enormous potential, but we shouldn’t trust it blindly. For AI to deliver real Clinical Value, several factors must align simultaneously. It resembles mathematical multiplication, where the general formula looks like this: Clinical Value = Usability x Actionability x Outcome. Just like in any multiplication, if any single component in this formula is zero, the total Clinical Value becomes zero as well. If a product is technically brilliant but cumbersome for the clinician (Usability = 0), or fails to provide actionable clinical advice (Actionability = 0), or doesn’t alter the patient’s health outcome (Outcome = 0) — the overall Clinical Value is zero.
We need to shift from merely creating “pretty algorithms” to building an AI ecosystem that saves doctors time, helps patients, and adds tangible value to the healthcare system.
