LIME for TimeSeries is an open-source project dedicated to advancing the interpretability of machine learning models focused on time series data. Utilizing the Local Interpretable Model-agnostic Explanations (LIME) technique, this project aims to demystify the decision-making processes of complex models. By integrating perturbation-based explanations, it provides insights into model predictions, enhancing transparency and trust in AI applications across various domains. Ideal for researchers, data scientists, and anyone invested in explainable AI (XAI), this repository offers tools, documentation, and examples to facilitate the understanding and application of LIME in time series analysis. By employing LIME, we aim to uncover which segments of an electrocardiogram (ECG) signal most influence the model's classification decisions, enhancing the interpretability of time-series models in healthcare.
Time-series classification, particularly in the context of ECG signal analysis, plays a crucial role in diagnosing cardiovascular diseases. While deep learning models offer promising results, their "black-box" nature hinders clinical adoption due to the lack of interpretability. This project leverages LIME, a technique for explaining predictions of any classifier in an interpretable and faithful manner, by perturbing the input signal and observing the changes in predictions.
The dataset utilized in this project is derived from the PhysioNet/Computing in Cardiology Challenge 2020, focusing on the classification of 12-lead ECGs to identify various cardiac abnormalities. Compiled from multiple sources, including the CPSC Database, INCART Database, PTB and PTB-XL Database, the Georgia 12-lead ECG Challenge Database, and an undisclosed database, it encompasses a diverse collection of ECG recordings from different demographics and geographic locations. The data, hosting ECG recordings across four classes (Normal Sinus Rhythm, Atrial Fibrillation, Ventricular Tachycardia or Ventricular Fibrillation, and Myocardial Infarction (Heart Attack)) representing different individuals, aims to advance automatic detection and classification of cardiac abnormalities, contributing significantly to the early and accurate diagnosis of cardiac conditions. For more detailed information: Classification of 12-lead ECGs: The PhysioNet/Computing in Cardiology Challenge 2020 by Perez Alday et al., 2020.
- Data Preprocessing: Techniques for transforming raw ECG signals into a format suitable for model training.
- Model Training: A Convolutional Neural Network (CNN) approach for ECG signal classification.
- LIME Explanations: Implementation of LIME to identify influential signal segments contributing to each classification decision.
- Visualization Tools: Utilities for visualizing ECG signals, their perturbations, and the influence of different segments on the model's predictions.
Clone the repository to your local machine:
git clone https://github.com/mdhabibi/LIME-for-Time-Series.git
cd LIME-for-Time-Series
-
Data Preparation: Start by preparing the ECG dataset. The
data_preprocess.pymodule provides functions for loading and preprocessing the data. -
Model Training: Use the
model_training.pymodule to train a CNN on the prepared ECG data. The module outlines the model architecture and training procedure. -
Applying LIME: The
lime_explanation.pymodule contains the implementation of LIME for time-series data. It includes functions for generating perturbations, applying perturbations to the signal, and fitting an interpretable model to the perturbed data. -
Visualization: The
visualization.pymodule offers visualization utilities to plot the original and perturbed ECG signals, class distribution, and the impact of each segment on the model's predictions.
A detailed example of using these modules can be found in the main.ipynb within the notebooks directory.
Explaining time-series predictions, especially in the medical domain, is crucial for trust and interpretability. Here, we detail the recipe for applying Local Interpretable Model-agnostic Explanations (LIME) to time-series data, focusing on ECG signals.
Before applying LIME, ensure you have a trained model ready. The model should be capable of classifying ECG signals into various categories (e.g., normal, arrhythmia types).
- Preprocess your ECG dataset.
- Train a Convolutional Neural Network (CNN) model, or any suitable classifier, on the processed ECG data.
- Select an instance from the ECG signals and divide the signal into segments. (Here, we have used a fixed-width segmentation.)
Perturbations simulate small changes in the data to observe how the model's predictions vary. For ECG signals, perturbations can involve modifying segments of the signal to reflect potential variations in the heart's electrical activity.
- Generate perturbed versions of a signal by modifying some segments (inactive segments) with perturbation function (here, a mean value of the signal voltage in each specified segment) while leaving others unchanged (active segments).
With perturbations ready, LIME can be applied to explain individual predictions:
- For each perturbed signal, predict the class probabilities using the trained model.
- Calculate the similarity between each perturbed signal and the original signal. This typically involves measuring the cosine distance.
- Fit a simple interpretable model (e.g., linear regression) to the perturbations, using the similarity scores as weights. This model aims to approximate the complex model's behavior around the original signal.
The coefficients of the interpretable model indicate the importance of each segment in influencing the model’s prediction:
- Identify the segments with the highest coefficients as the most influential for the model's prediction.
- Visualize the original signal, highlighting these influential segments (yellow segments) to provide insight into which parts of the signal were most significant in reaching the classification decision.
Contributions to improve the project are welcome. Please follow the standard fork-pull request workflow on GitHub to submit your improvements to the dev branch.
This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details or visit GNU General Public License v3.0.
This project is inspired by the original paper on LIME: "Why Should I Trust You?" Explaining the Predictions of Any Classifier by Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin.