果冻传媒

Dave Hamersma already had considerable experience in data science when he started the program at the EAISI Academy. For his master鈥檚 thesis in Health Sciences at the University of Twente, he of breast cancer treatments in the Netherlands and Norway using federated learning (a technique that enables computations on data stored decentrally, without copying the data itself). Daniel Kapitan (DK) spoke with him about his motivation to deepen his expertise and how he further developed his Mastering Project at Sensire.

DK: Great to speak with you again! To start, can you tell us how you ended up working in the field of data science and AI in healthcare?

"Looking back on my career so far, it wasn鈥檛 obvious that I would become a data scientist. After completing my vocational education in International Business Studies, I went on to study physiotherapy. Even then, I found the research-oriented courses particularly interesting. During my studies, I encountered many fascinating and exceptional case studies. However, working as a physiotherapist turned out to be too repetitive and didn鈥檛 quite match my interests. That鈥檚 when I decided to pursue a degree in Health Sciences. There, I was introduced to statistical learning and learned the basics of machine learning using R. The idea of contributing to healthcare through data science appealed to me, and that鈥檚 the direction I chose to follow."

DK: What made you decide to enroll in our program?

"Not long after completing my degree, I started working at Sensire, a large healthcare organization active in nursing homes, elderly care, and home care. When I joined, we were mainly focused on setting up the basics: extracting data from source systems, generating business reports, and so on. Later, we decided to launch a comprehensive digitalization program. With the shift to Google Cloud, it became much easier to engage in machine learning, which motivated me to deepen my knowledge."

DK: Coincidentally, I was involved at Sensire at that time as an advisor for that cloud migration.

"That definitely made it easier for me to make the choice! What made the program particularly attractive was that I could join the Applications Track directly without starting from scratch. My goal was to gain more hands-on experience and build confidence in independently executing end-to-end projects. Although I was already familiar with many of the practice assignments, the extensive feedback I received helped me dive much deeper into the subject matter. We also went far beyond just writing a script to train a model. I remember a question posed by Pieter Overdevest (one of the lecturers) during a workshop: 鈥極nce you鈥檝e trained a model, what are you going to do with it?鈥� That question pushed me to dive deeper into various aspects of data and software engineering to actually implement and manage AI systems in practice."

DK: Every day, there鈥檚 a new story about an LLM breakthrough. Can you tell us more about your project and what you learned from it?

"I find it fascinating how quickly developments in this field are moving, and that鈥檚 what makes my work exciting. In my research, I first focused on crafting effective prompts to avoid bizarre hallucinations in the summaries. For example, during one experiment, a summary suddenly stated that a client 鈥榯alks a lot about God and is very religious鈥欌�攅ven though that wasn鈥檛 mentioned anywhere in the original text! Another issue is that LLMs easily confuse left and right. If the notes say someone just had surgery on their left knee, the summary might suddenly say it was the right knee instead. Fine-tuning these models is essential for practical use. In this case, I also spent a lot of time collaborating with nursing colleagues to develop a solid test dataset, enabling us to measure the quality of the generated summaries. That鈥檚 still a challenge, though鈥攈ow do you objectively assess the quality of a summary? That鈥檚 something I plan to research further."

DK: How did you eventually bring the summarization tool into production?

"After completing my studies, I developed several versions of the tool and explored how to integrate it as simply as possible into a web browser. In November 2024, we officially launched SensAI. We eventually chose Google Gemini 2.0 as our foundational model because it delivered the best results and integrated well with our data platform. "

DK: Congratulations! We all know how challenging it is to actually bring a MACHINE-LEARNING project into production. Do you have any advice for future participants?

"This program has mainly taught me a new way of thinking: where you can apply data science and where it truly adds value. Of course, it was exciting to ride the wave of Generative AI innovation, but the knowledge and skills I gained go far beyond that."