Machine Learning-Based Aggression Detection in Children with ADHD Using Sensor-Based Physical Activity Monitoring

Abstract

Aggression in children is highly prevalent and can have devastating consequences, yet there is currently no objective method to track its frequency in daily life. This study aims to investigate the use of wearable-sensor-derived physical activity data and machine learning to objectively identify physical-aggressive incidents in children. Participants (n = 39) aged 7 to 16 years, with and without ADHD, wore a waist-worn activity monitor (ActiGraph, GT3X+) for up to one week, three times over 12 months, while demographic, anthropometric, and clinical data were collected. Machine learning techniques, specifically random forest, were used to analyze patterns that identify physical-aggressive incident with 1-min time resolution. A total of 119 aggression episodes, lasting 7.3 ± 13.1 min for a total of 872 1-min epochs including 132 physical aggression epochs, were collected. The model achieved high precision (80.2%), accuracy (82.0%), recall (85.0%), F1 score (82.4%), and area under the curve (89.3%) to distinguish physical aggression epochs. The sensor-derived feature of vector magnitude (faster triaxial acceleration) was the second contributing feature in the model, and significantly distinguished aggression and non-aggression epochs. If validated in larger samples, this model could provide a practical and efficient solution for remotely detecting and managing aggressive incidents in children.

Keywords: aggression; machine learning; pediatrics; remote patient monitoring; wearables.

Figure 1

A wearable sensor (GT3X+, ActiGraph Corp., Pensacola, FL, USA) and its placement for physical activity assessments.

Figure 2

(Top): An example of the vector magnitude (vm) signal, which is one of the sensor-derived physical activity variables, for both aggression and non-aggression epochs. Each rectangle represents an epoch, consisting of 1-min worth of physical activity data. The red-filled boxes indicate that the epoch occurred during a reported aggression episode, while the green ones indicate the absence of aggression. The large orange rectangles represent 30-min worth of data occurring before and after an aggressive incident. These were excluded from the analyses to minimize contamination risk. (Bottom): This diagram illustrates the different types of variables that are associated with each epoch for inclusion in the model. These variables include demographic, clinical, sensor-derived physical activity, and observed day type variables.

Figure 3

This flowchart illustrates the data preprocess and feature selection using a machine learning model, as well as the evaluation of the resulting model. The random forest classifier model was employed, with 100 bootstraps utilized to assess model performance. The evaluation metrics included mean ± SD values for model accuracy, recall, precision, F1 score, and area under the curve (AUC).

Figure 4

(A) The duration of reported aggression (verbal and physical) as a function of time of day. (B) The duration of reported aggression (verbal and physical) as a function of weekdays and weekend days. The majority of physical aggression epochs occurred during the afternoon and evening, and on weekend days.

Figure 5

Figure (A) shows the ranking of 20 features based on their significance in distinguishing physical aggression from non-aggression epochs, as determined by the random forest classifier algorithm. Meanwhile, Figure (B) demonstrates the model’s effectiveness in distinguishing between the two groups, as measured by AUC, F1 score, accuracy, recall, and precision. The figure utilizes several abbreviations, such as CBCL (child behavior checklist), MPH (methylphenidate), CGI-S (clinical global impression-severity), AUC (area under curve).