The LarvixON AI system is a cross-platform application analysing larval movement to help identify substances in the patient's body. The project won third prize in the CreatiWITy 2026competition, in the team project category. It was developed by graduate students of Applied Computer Science in cooperation with the Medical University of Wrocław.

The project "LarvixON AI – serum toxicity diagnostics based on larval behaviour analysis" was carried out by graduates of Applied Computer Science: Patryk Łuszczek, Krzysztof Kulka, Mikołaj Kubś and Martyna Łopianiak. It was supervised by Dr. Natalia Piórkowska.

As can be learnt from the project's presentation, its aim was to create a system that provides data supporting doctors in quickly (under 20 minutes) identifying substances such as illegal drugs or xenobiotics that may be present in a patient's body.

Inspiration from clinical practice

The project was inspired by real-life medical situations, in which quick diagnosis is obstructed by the patient.

“There are cases in which the patient does not want to or is unable to say what they have taken. At the same time, different substances require different treatment methods, and their improper selection can pose a risk,” explains Krzysztof Kulka.

The authors of the solution recount that the situation which inspired them to develop this project took place in one of the hospitals in Wroclaw. A boy with drug overdose symptoms didn't want to say what he had taken.

The authors emphasize that tests detecting substances can be time-consuming.

“Standard laboratory methods may require any time from several minutes to several hours. In the case of our solution, the analysis may take approx. 10 minutes,” emphasizes Patryk Łuszczek.

From larval behaviour to substance identification

The system functions based on the use of moth larvae. They respond differently to different chemical substances, and their behaviour can be analysed using machine learning methods.

As the authors explain, the test employs the patient's serum, which is administered to the larvae. Their behaviour in the laboratory, in the so-called Petri dish, is recorded, and the system analyses the video material and tries to determine what substance could have caused a given reaction.

“Knowing how larvae behave under the influence of specific substances, we can observe their movement in a Petri dish and infer what substance was given to them, and – indirectly – what was in the patient's serum,” says Mikołaj Kubś.

Experiments with larvae

The project was developed in cooperation with the Medical University of Wrocław, which provided research data and conducted behavioural analysis of the larvae. The authors mainly relied on caffeine and alcohol at various concentrations.

“Studies have shown that larvae behave in a characteristic way depending on the administered substance. For example, they move faster after caffeine and lose sensitivity to light after alcohol,” explains Krzysztof Kulka.

The authors point out that publicly available datasets were insufficient for the project.

“We were looking for ready-made datasets, but we didn't find any that included recordings of larvae labelled by the administered substance. Therefore, a new dataset was developed in collaboration with the Medical University,” explains Patryk Łuszczek.

How does the system work?

LarvixON AI is an application that takes a video recording of larvae as input data, and then returns probabilities indicating what substance could have been administered.

The system utilizes a hybrid machine learning model. A CNN network, i.e. the first of the models, is responsible for extracting the characteristic features from the image, and LSTM, which is the second machine-learning model, analyses changes over time and uses this information for classification purposes.

“First, we convert the video into a set of features describing the movement, and then the model analyses their variability and tries to match them to a specific substance,” explain the authors.

One of the biggest challenges was the quality and consistency of the video material.

“A big part of our work involved preparing data. The recordings were of varying quality, they included noise or elements not related to the test,” says Krzysztof Kulka.

Further steps

The model was trained on the basis of resources from the Wroclaw Centre for Networking and Supercomputing.

The project is currently in the prototype phase. Its implementation in clinical practice would require further research, additional infrastructure, and compliance with formal requirements.