Summary of the Panel Discussion “AI in Healthcare”: Theory vs. Practice – Part 1

After the successful TeamUp x Health event, we want to remind you of the key knowledge and information that emerged from the panel discussion on “AI in Healthcare.” The panel featured Mladen Fernežir, Marko Poljak, Dr. Tomislav Šmuc, and Tomislav Strgar. The panelists provided valuable insights into the application of artificial intelligence in healthcare and the latest technological trends shaping the future of medicine. The panel discussion was divided into 3 segments, and from each, we gained new and interesting insights. In this article, we will focus on the first segment: The Application of AI in Healthcare – Theory vs. Practice.

AI u Zdravstvu
Panelisti (s lijeva prema desno): Bernard Ivezić (moderator), Mladen Fernežir, Marko Poljak, dr.sc. Tomislav Šmuc i Tomislav Strgar

Our moderator, Bernard Ivezić, initiated the discussion by presenting information on the importance of artificial intelligence in healthcare, emphasizing that its current application is still limited. At the beginning of the panel discussion, two graphs were displayed to illustrate the perception of artificial intelligence and its actual utilization in the industry.

Izvor: https://www.businessinsider.com/ai-medical-diagnosis-report

 

From the first graph, we can observe that the perception of American healthcare providers regarding the utility of artificial intelligence is very high. A significant 65% of respondents stated that they consider AI to be crucial for performing their job, while only 14% reported currently using artificial intelligence in their work. Although the data is from a 2018 graph, it remains relevant and vividly portrays the current state of using AI-based tools.

 

Barriers hindering the better utilization of AI in healthcare

Complementary innovations are crucial for the successful adoption of artificial intelligence and other information technologies in healthcare. For instance, the successful implementation of electronic medical records required innovations in integrating software systems and changes in the working methods of doctors and pharmacists, which is not easy to accomplish.

 

There are several major challenges in adopting artificial intelligence in healthcare, including:

 

  • Algorithmic Limitations: AI is based on neural networks, but the problem is that we sometimes cannot understand how AI makes certain decisions. This reduces the confidence of healthcare professionals in AI, as they will be responsible for the decisions AI makes. Developing more transparent AI systems is crucial for their broader use.

 

  • Limited Data Access: High-quality data is essential for the proper functioning of AI. However, collecting medical data is difficult and often imperfect. Connecting data from different hospitals is also a challenge. Improving data collection and sharing is necessary to make AI more beneficial.

 

  • Regulatory Barriers: Strict privacy regulations make the collection and sharing of health data challenging. Additionally, approving new medical technologies takes time and careful analysis. Innovations in AI regulations are needed to facilitate easier AI adoption.

 

  • Fear of Job Loss: Introducing AI can threaten the jobs of some healthcare professionals, which may reduce the willingness to adopt the technology. Educating staff about artificial intelligence and its applications can promote adoption and alleviate fears of job loss.

 

From the second graph, we can infer that artificial intelligence in healthcare and pharmaceuticals is most widely used in quality control (60% of cases), followed by patient/user care (44% of cases), and diagnostics and monitoring (42% of cases). This data is not surprising, as these are scenarios where it is possible to collect the largest amount of data to train machine learning models to facilitate work.

.

Marko Poljak, CEO of Newton Technologies, shared experiences from his early days with AI technologies and touched on the product Newton Dictate, which can be described as “Alexa for doctors,” enabling voice filling of forms or allowing healthcare staff to “write” with their voice.

 

The technology enabling the existence of such a product is called Natural Language Processing (NLP). NLP is a field of artificial intelligence that deals with computers and human language. The goal of NLP is to understand, interpret, and generate human language using machine learning techniques, natural language processing, and linguistics.

 

Key tasks of NLP include tokenization, lexical and morphological analysis, syntactic and semantic analysis, speech recognition, natural language generation, language translation, emotion detection, and question-answering systems.

 

NLP has a wide range of applications in areas such as internet search, sentiment analysis on social media, virtual assistants, customer support automation, and many others.

 

By using NLP technologies, healthcare staff can be relieved of administrative burdens, ultimately resulting in faster and more efficient work, leading to the treatment of a larger number of patients.

 

More about NLP can be found here.

AI u Zdravstvu
Mladen Fernežir, Co-founder Velebit AI (lijevo) i Marko Poljak, CEO Newton Technologies (desno)

laden Fernežir, Co-founder of Velebit AI (left) and Marko Poljak, CEO of Newton Technologies (right)

 

Mladen Fernežir, Co-founder of Velebit AI, emphasized many opportunities for AI in healthcare and the potential for improving existing products.

 

For example, artificial intelligence can already be used for the early diagnosis of diseases in fetuses, helping detect chromosomal abnormalities and predicting which diseases parents and the child will encounter in the future.

 

Machine learning also enables the analysis of a large amount of scientific research and medical studies to better understand the mechanisms of genetic diseases and identify new therapeutic methods. This approach supports the rapid and efficient development of new drugs and therapies targeted at specific genetic mutations. More about artificial intelligence and genetics can be read here.

 

Additionally, artificial intelligence can already be used to analyze patient report data. It is possible to more easily detect patients who have the potential for diseases such as diabetes and cardiovascular diseases. Therefore, medical staff can perform better disease prevention and enable earlier treatment to preserve the patient’s health.

 

Insufficient (Quality) Data – A Challenge Hindering Faster Adoption of Artificial Intelligence in Healthcare

 

Dr. Tomislav Šmuc from the Ruđer Bošković Institute emphasized the need for quality data for the successful training of AI models. The problem of insufficient quality data, on which artificial intelligence could be trained, is one of the current major challenges preventing better application of artificial intelligence in healthcare.

 

There are several key reasons why insufficient quality data pose a challenge for training AI models, including:

 

  • Errors and Noise: Data may contain errors, ambiguities, noise, or uncertain values. If the data is inaccurate or unclear, AI models can learn incorrect patterns and make inaccurate predictions.

 

  • Missing Data: If there is not enough data, the AI model will lack sufficient information to create adequate generalized models. Missing data can lead to overfitting or the model’s inability to make correct decisions.

 

  • Unbalanced Data Sets: If the data set used to train the AI model is unbalanced, meaning it contains too many examples from one class compared to another, the model may be biased and have difficulty recognizing less represented classes.

 

  • Poorly Labeled Data: Accurately labeled data is crucial for training AI models. If the data is poorly labeled or contains errors in annotations, the model will not learn correct patterns and may make inaccurate decisions.

 

  • Lack of Diversity: If the data set is not diverse enough, the AI model may struggle to generalize to new and unknown instances. Data diversity is crucial for reliable predictions in real-world conditions.

 

  • Insufficient Data Quantity: AI models, especially deep models requiring large amounts of data, can be limited by insufficient available data. Insufficient data quantity can lead to overfitting, where the model “memorizes” the data instead of learning general patterns.


Addressing the “insufficient quality data” issue requires careful preparation of the data set, proactive identification and correction of errors, and additional strategies such as data augmentation techniques to ensure data diversity and quality. Thoughtful selection and preparation of data are key steps in building reliable, precise, and efficient AI models.

dr.sc. Tomislav Šmuc, Institut Ruđer Bošković

The conclusion of the first part of the panel discussion “AI in Healthcare – Theory vs. Practice” is that artificial intelligence in healthcare is in its infancy, with ample room for improvement and the adoption of new technologies.

We also invite you to share this article with people you think might be interested and to continue following us, as we will continue to publish summaries and key learnings from the other two topics in the “AI in Healthcare” panel discussion.