Escribir un comentario

Summary: 

The Purpose of the paper was to exhibit how the usage of machine learning (ML) techniques can help to predict epileptic seizures. The paper was well-written, giving a comprehensive review on ML. We addressed some Pros, Cons, and Suggestions for improvement. 

PROS:

Abstract

• Studies unique point is to study machine learning techniques for ECG prediction of epileptic seizure activity

Introduction

• A clear explanation of what epilepsy is and the need for better seizure prediction given its harmfulness

• Clear explanation on the relationship between epilepsy and the cardiovascular system, specifically the acute and chronic physiological changes caused by seizures

• Proves awareness of the current landscape regarding epileptic research by citing multiple sources, specifically those showing a clear link between seizures and the cardiovascular system

• Clearly identifies the limitations of EEG-based approaches (e.g., variability, artifacts, impracticality for long-term monitoring) and motivates the shift toward ECG-based methods

• The introduction ends with a detailed outline of the paper’s structure

Section 2: Epileptic Seizure and Cardiovascular System

• Thoroughly connects seizure prediction capabilities with improvements in patient safety and quality of life

• Effectively explains the limitations of EEG use such as artifacts and impracticality for long term use and explains why these limitations do not exist for ECG

• Cites other studies connecting tachycardia and bradycardia patterns to seizures which enhances the credibility of ECG based prediction

• Connects seizures to ANS responses which increases the credibility of HR and HRV for prediction

Section 3: Epileptic Seizure Prediction 

• There is a good amount of diversity in the sample size of the study which makes the seizure prediction findings more applicable to the general population

• Good use of diagrams to model the relevant data metrics

• A strong representation in time frequency

  • DWT, and CWT makes capturing localized and frequency specific changes a key in ECG for specific dynamic anomalies

• Advantages of STFT

  • STFT spectrograms have the best predictive performance as frequency domain specificity can be discriminative for seizure detection

• Reconstruction based modeling

  • Features align well with anomaly detection transformer-based models by reconstruction, reinforcing consistency

• Patient specific customization 

  • Feature extraction allows for adapting to patient specific ECG dynamics 

• Threshold fixing with statistics

  • Threshold selection based on mean and standard deviation of errors introduces sensitivity, adaptable to feature quality.

Discussion + Conclusion

• Clear articulation of the benefits of ECG (non-invasive, more broadly available) over EEG in predicting seizures

• Good emphasis on personalized models and novel biomarkers (e.g., T-wave heterogeneity

• Extensive discussion of two fundamental approaches for seizure prediction: anomaly detection and classification and limitations of each framework (e.g., false positives for anomaly detection)

• Identifies areas of promise (e.g., XAI, automated thresholds, advanced ML e.g., GNNs)

• Presents a compelling case for clinical translation and improvements in the quality of life of patients

—————————————————————————————————————————————————

CONS

Abstract

• Doesn’t clarify the approaches “classification” and “Anomaly Detection” at this point of the paper

Introduction

• The section could be strengthened by a more in-depth analysis regarding machine learning analysis and how this system is created (make the section slightly more technical)

• Seems repetitive in certain sections (e.g., mentions the need for seizure prediction and detection multiple times, we already know this)

• Does not define the specific research gap that exists regarding machine learning approaches for ECG-based seizure prediction

• Uses terms such as HRV and RRI to define what is measured using an ECG and terms such as ANS and CAN to determine systems that epilepsy affects, but does not go into detail regarding what and how HRV and RRI are measured, and what aspects of these systems are disrupted

• The paper could be strengthened by focusing more on a clinical context, as these are common medical conditions. Does not connect to the real-world implications of healthcare as much as it could

• Briefly mentions what the research objectives are in various sections of the introduction. It could benefit by clearly defining objectives at the end in bullet points, sort of like a hypothesis or thesis statement

Section 2: Epileptic Seizure and Cardiovascular System

• Lack of data regarding the HR and HRV changes mentioned

• Methods for measuring HRV not mentioned

• The time intervals for seizure stages are vague and sometimes listed as unspecified

• While EEG limitations are mentioned regarding long term data collection, the predictive ability of EEGs in relation to that of ECGs are not mentioned

Section 3: Epileptic Seizure Prediction

• Pre ictal intervals vary between patients which makes the findings not fully applicable to everyone and there is a need for personalized modeling to predict seizures accurately

• No measure for what is defined as a “high quality data set”

• Little discussion on how HRV can be used outside a lab setting in real patients

• Using too many transforms may be redundant, highly correlated features could inflate data and dimensionality to different models

• The review stops at the selection of the best performing feature type. It doesn't specify if they use feature specific selection to further their optimization of their feature set

• Time frequency is often very opaque as the model could very well not perform up to the certain standard. Understanding which feature components drive predictions still is difficult

Discussion + Conclusion

• The challenges of dataset standardization and validating ambulatory testing were overlooked

• There needs to be clearer pathways to recommend models to implement, relating to real-world contexts

• The computational/clinical feasibility of some advanced ML (e.g. GNNs, self-supervised learning) were not mentioned

• The place of inter-patient variability and real-world noise (e.g. movement artifacts) needed more attention and assumptions about the inevitable lack of standardization

• The paper did not have a clear EEG compared to ECG sensitivity/specificity comparison and therefore did not set up how trade-offs can be framed

___________________________________________________________________

SUGGESTIONS FOR IMPROVEMENT

• Make a table comparing ECG/EEG pros/cons

• Propose hybrid approaches to balance sensitivity/specificity

• Include case studies of ML applied to specific biomarkers

• Apply clustering and ranking to possibly reduce duplication and enhance frequency

• Introduce specific feature selection protocols

Competing interests

The authors declare that they have no competing interests.