Utilização de técnicas de AutoML para a detecção do congelamento da marcha em pacientes com doença de Parkinson

Abstract

This study aimed to identify and classify episodes of FOG in different phases of walking using AutoML (Automated Machine Learning) techniques. Accelerometer data from patients with DP were utilized, and various classification techniques were explored using AutoGluon, an AutoML library. Four models were trained with the same data, differing only in the size of temporal windows for each training, which were of sizes 1, 2, 3, and 6. The results indicated that the chosen approach did not exhibit good performance in classifying FOG events. However, the Mean Average Precision (mAP) metric of the best model surpassed the result of the 5th place in the Kaggle competition. Future work includes evaluating the contribution of larger time windows and advanced feature engineering. The study contributes to advancements in the detection and treatment of Freezing of Gait in patients with DP.