Ladybird Performance in Detecting Pathological Epileptic Events and Resistance to False Positives

📄 Description

This repository showcases a component of Ladybird that is focused on the efficient automatic detection of Fast Ripples (FR) in epilepsy.

More specifically, this notebook presents the performance of Halyzia in the detection of Fast Ripples (FR) at different signal to noise ratios: 0dB, 5 dB, 10 dB and 15 dB. Step 1 of Halyzia consists of detecting pathological shapes on EEG scalograms transformed into images with a convolutional neural network (CNN). Halyzia's Step 2 filters false positives using a homemade algorithm consisting of multiple clustering and probabilistic steps.

⚒️ Installation

Prerequisites

• Python 3.11
• Python libraries

  pip install -r requirements.txt

📝 Usage

Detailed analysis here can be accessed and performed from main.ipynb.

Input data

The data used for this work are the gold standard of the literature [1], used by several authors to evaluate their Fast Ripples, Ripples or Intercritical Epileptic Spikes detector. These simulated EEG data are composed of 40,320 events that can appear alone or in combination, as show in the table below.

Results

Event type	0 dB	5dB	10dB	15dB	TOTAL
FR alone	n=1140	n=1140	n=1140	n=1140	n=5,760
FR + Ripple	n=1140	n=1140	n=1140	n=1140	n=5,760
FR + Spike	n=1140	n=1140	n=1140	n=1140	n=5,760
FR + Spike + Ripple	n=1140	n=1140	n=1140	n=1140	n=5,760
Spike + Ripple	n=1140	n=1140	n=1140	n=1140	n=5,760
Spike alone	n=1140	n=1140	n=1140	n=1140	n=5,760
Ripple alone	n=1140	n=1140	n=1140	n=1140	n=5,760
TOTAL	n=10,080	n=10,080	n=10,080	n=10,080	n=40,320

Detector evaluation methods

• Events that contain an FR and are detected by the algorithm are considered true positives (TP).
• Events that do not contain an FR and are detected by the algorithm are considered false positives (FP).
• Events that contain an FR and are not detected by the algorithm are considered false negatives (FN).
• Events that do not contain FR and are not detected by the algorithm are considered true negatives (TN).

Metrics

• Sensitivity (sens) was calculated as follows:
  [ \text{sens} = \frac{\text{TP}}{\text{TP} + \text{FN}} ]

• Precision (prec) was calculated as follows:
  [ \text{prec} = \frac{\text{TP}}{\text{TP} + \text{FP}} ]

• F measure was calculated as follows:
  [ F_1 = 2 \times \frac{\text{prec} \times \text{sens}}{\text{prec} + \text{sens}} ]

Summary figure

We compared the performance of Halyzia with other detectors evaluated on the same dataset (see [1]).

📚 References

[1] Roehri, N., Pizzo, F., Bartolomei, F., Wendling, F., & Bénar, C. G. (2017a). What are the assets and weaknesses of HFO detectors? A benchmark framework based on realistic simulations. PLoS ONE, 12(4). https://doi.org/10.1371/journal.pone.0174702

About Ladybird

Ladybird, developed during L. Gardy's doctoral research under the guidance of E. Barbeau (neuroscientist, CNRS) and C. Hurter (engineer, ENAC), is a sophisticated software designed for the automatic detection of fast-ripples (FRs) in epilepsy. Originally developed by L. Gardy, with significant input from various researchers and medical professionals the early trials in the neurology department at Toulouse Hospital, along with support from various academic and economic entities, led to widespread enthusiasm and substantial funding, allowing for further development and eventual patenting (Brevet: FR3128111). Ladybird was later rebranded as Halyzia©, a name change necessitated for trademark reasons, and is now being further developed and commercialized by the French startup Avrio MedTech.

👤 Author

• LinkedIn: Ludovic Gardy
• Doctoral thesis: PDF