Detection of mobile device use in epilepsy video telemetry
Hiillos, Matias (2024)
2024
Teknis-luonnontieteellinen DI-ohjelma - Master's Programme in Science and Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-10-08
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202410029064
https://urn.fi/URN:NBN:fi:tuni-202410029064
Tiivistelmä
The purpose of this thesis was to integrate mobile device detection into a video-based seizure detection system and to explore the possible improvements in the performance of the system. Classical machine learning is used in the epilepsy seizure detection pipeline, with the most important extracted signal being an optical flow-based oscillation detection. Mobile devices are commonly used by patients in a clinical recording setting, and the use of these devices contributes noise to the oscillation signal. Object detection is a widely researched topic in machine learning, with various open-source models already available for detecting mobile devices in images. By applying a filter that utilizes these machine learning tools to detect mobile device areas, these device regions can be excluded from processing. The goal of this was to decrease the signal strength from these non-physiological components, such as mobile phone usage.
A single daytime recording was chosen for 16 patients. From these recordings, events emitted from a pre-trained clonic seizure detection based on decision-tree output of temporally-sampled video-tracking features under model rating of 0.09 were filtered out from a previous analysis to reduce the amount of data to be analyzed. This was done to focus only on events that would be considered false positive detections under normal clinical use. Out of 387 filtered events, 241 were visually confirmed to have a mobile device in the view. Mobile devices were correctly detected for 177 events out of these events, with a sensitivity of 73.4% and specificity of 65.1%. Detected events containing a mobile device were re-processed by subtracting the detected area from the scene.
The results were analyzed separately for true positive and false positive detections. For the total number of 177 true positive events, 106 events saw a reduction in the clonic model rating with a median change of -0.034 and a median relative change of -22.6%. 42 events experienced a small increase in the model rating, with a median change of 0.008 and a median relative change of 5.7%. For 29 events there were no changes in the clonic model rating.
For the total of 59 events with a falsely detected mobile device, the effects of subtraction in the resulting rating were negligible, with a median change of -0.015 and a median relative change of -11.6% for 23 events with a reduced rating. 11 events had a similar increase in the ratings, with a median change of 0.007 and a median relative change of 4.8%, and 25 events resulted in no changes.
The derived oscillation signals were also separately analyzed to understand how the signals change when excluding an area from being processed, and how they impact the clonic model rating results. For true positive detections, oscillation changes varied for each frequency bin. For false positive detections, changes were more consistent for lower frequencies (0.5–3 Hz), where most oscillation occurs. The changes in oscillation became more dramatic for both cases at higher frequencies due to the low amount of oscillating pixels.
A single daytime recording was chosen for 16 patients. From these recordings, events emitted from a pre-trained clonic seizure detection based on decision-tree output of temporally-sampled video-tracking features under model rating of 0.09 were filtered out from a previous analysis to reduce the amount of data to be analyzed. This was done to focus only on events that would be considered false positive detections under normal clinical use. Out of 387 filtered events, 241 were visually confirmed to have a mobile device in the view. Mobile devices were correctly detected for 177 events out of these events, with a sensitivity of 73.4% and specificity of 65.1%. Detected events containing a mobile device were re-processed by subtracting the detected area from the scene.
The results were analyzed separately for true positive and false positive detections. For the total number of 177 true positive events, 106 events saw a reduction in the clonic model rating with a median change of -0.034 and a median relative change of -22.6%. 42 events experienced a small increase in the model rating, with a median change of 0.008 and a median relative change of 5.7%. For 29 events there were no changes in the clonic model rating.
For the total of 59 events with a falsely detected mobile device, the effects of subtraction in the resulting rating were negligible, with a median change of -0.015 and a median relative change of -11.6% for 23 events with a reduced rating. 11 events had a similar increase in the ratings, with a median change of 0.007 and a median relative change of 4.8%, and 25 events resulted in no changes.
The derived oscillation signals were also separately analyzed to understand how the signals change when excluding an area from being processed, and how they impact the clonic model rating results. For true positive detections, oscillation changes varied for each frequency bin. For false positive detections, changes were more consistent for lower frequencies (0.5–3 Hz), where most oscillation occurs. The changes in oscillation became more dramatic for both cases at higher frequencies due to the low amount of oscillating pixels.