Freezing of Gait (FOG) is a common motor symptom in Parkinson’s disease, where a person may suddenly have difficulty starting or continuing to walk. These episodes can be short and unpredictable, so they may not always happen during a clinical visit. This makes FOG difficult to detect and monitor. It can also affect daily activities and increase the risk of falls.

In this study, we investigated whether wearable sensor data and machine learning could be used to automatically detect FOG episodes. We used a publicly available dataset containing 3D accelerometer data collected from the lower back of people with Parkinson’s disease in home-based settings. The dataset includes different types of FOG events, such as start hesitation and freezing during walking and turning. In our analysis, we mainly focused on freezing during turning.

We tested several machine learning methods and evaluated the models using both subject-dependent and subject-independent settings. In the subject-dependent evaluation, the Decision Tree model achieved an F1-score and recall of 99%. In the more realistic subject-independent evaluation, where the model was tested on people whose data were not used during training, LightGBM achieved an F1-score of 80.9% and a recall of 93.4%.

These results show the potential of combining wearable sensors and machine learning for automatic FOG detection. In the future, similar approaches could support continuous monitoring of gait symptoms in daily life and provide more objective information about changes in Parkinson’s motor symptoms.

This work was supported by the AI4HOPE project and was conducted by Ayse Kosal Bulbul, Muhammad Irfan, Maria K. Jaakkola, Riku Klén, and Abdulhamit Subasi. The work was presented at ECMS 2026, University of Agder (UiA), Grimstad, Norway, 23–26 June 2026, by Maria K. Jaakkola.