ºÚÁϸ£ÀûÍø

To address this, PhD researcher studied how artificial intelligence can learn treatment strategies in simulated patient environments, with a focus on keeping these strategies transparent and interpretable.

He defended his PhD thesis at the Department of Mathematics and Computer Science on Thursday, May 7.

Learning treatment strategies in simulated patients

Because real patients cannot be used for experimentation, Liventsev worked with patient simulation environments that mimic how a patient’s condition changes over time in response to different treatments. These simulators are built using clinical data, medical knowledge, or a combination of both.

Within these virtual environments, he developed methods that allow artificial intelligence to explore and compare many possible treatment strategies in a safe way. Instead of only predicting what might happen, the system actively tests different strategies in simulation and gradually improves them based on the results.

A key feature of this approach is that the learned treatment strategies are written as readable computer programs. This makes them easier for clinicians to understand, since they can inspect the exact steps the system follows when making decisions.

From program synthesis to interpretable decision support

Liventsev developed a framework in which AI systems learn treatment strategies by automatically creating and running computer programs in simulated patient environments. The system then evaluates how well these strategies perform and gradually improves them through repeated testing and refinement.

This approach combines program synthesis, a method in which computers generate their own programs to solve a task, with reinforcement learning, in which systems learn by trial and error based on feedback from outcomes. The result is that the AI produces treatment strategies in the form of readable and executable programs, rather than hidden patterns inside a complex model that are difficult to interpret.

Advances in automatic programming and AI systems

To enable this approach, Liventsev developed several contributions in automatic programming and machine learning.

He introduced a new programming language designed specifically for teaching AI systems how to write programs. He also developed a neurogenetic programming framework, which combines ideas from neural networks and evolutionary processes to generate code in simple programming languages such as BF++.

In addition, he contributed a tree-based variational autoencoder, a machine learning model that learns structured representations of computer programs. This helps the system better understand the structure of code rather than treating it as a flat sequence of text.

Finally, he developed SEIDR, which stands for Synthesize, Execute, Debug, and Rank. This system allows large language models to generate programs, test them, identify and fix errors, and iteratively improve them without human intervention.

The research shows that combining different types of AI methods, such as neural networks, symbolic reasoning, and evolutionary algorithms, leads to better performance than using any single method alone. It also shows that including human expert knowledge in the learning process can further improve results.

Simulating emergency and intensive care scenarios

Liventsev developed several simulation environments to test and evaluate these methods in healthcare settings.

These include Auto-ALS, an interactive training environment for emergency care, a framework for simulating ultrasound images, and a benchmark dataset for intensive care scenarios based on real clinical data. Together, these environments allow researchers to safely test how AI systems behave in situations that resemble real clinical practice.

Experiments show that the proposed approach can discover treatment strategies that perform better than traditional methods while still remaining interpretable to clinicians. However, the most complex emergency-care scenarios remain challenging. This highlights important directions for future research, particularly in improving reliability in high-stakes medical situations.