University:
Boston University
Course:
SAR HP 151 | Introduction to Health Professions
Academic year:

2023
Views:

163

Pages:

1
Author:

levedzgetmo7f

Automated Classiﬁcation of Electrocardiogram Data Using Machine Learning Introduction Pipeline Methods Training 80% Validation 20% Background: 1. Preprocess Training Data: Utilizing PeacLab’s supervised machine 2 275,200 ECG exams ECG Exam 68,800 exams ● tsfresh feature extraction 1 learning (ML) framework that: Training Dataset from time-series data ● Uses multivariate time series data from Total: 344,000 exams ● Chi-square feature selection large-scale computing systems to diagnose Data Preprocessing Pipeline performance anomalies Feature 2. Train Model: One vs. Rest Classiﬁer with LGBM Feature Selection ● Extracts statistical features and then 3 Extraction ECG Exam ● Fits a model for each class separately (250 feats.) Training Test Dataset selects important features to reduce data ○ Classes: 1dAVb, RBBB, LBBB, SB, AF, ST features dimensionality ○ Each exam can have multiple anomalies ● Trains tree-based ML models Test data: 827 ECG exams Data Preprocessing Pipeline 3. Evaluate Model: Problem: Feature ● Perform the same feature extraction, scaling, Feature ● Investigating medical data manually can be Selection Model Training Extraction and selection used to process training dataset (250 feats.) time-consuming and lead to delays in ● Compare the predicted labels and true labels for each diagnosis Testing anomaly type separately Model ● ML models can automate the investigation ● Report precision, recall, and F1-score for each anomaly process while maintaining high accuracy 1dAVb RBBB LBBB SB AF ST ● Drawbacks of deep learning (DL) models: Pred 0 1 1 0 1 0 ○ Require large quantities of labeled data Performance Truth 0 0 1 0 0 1 Evaluation ○ Costly to train and maintain in a resource limited environment Goal: Results ● Determine the applicability of Peaclab’s ECG Anomaly Precision Recall F1-Score Overall: Our model exhibited poorer performance supervised ML framework to a public 4 compared to the DL-baseline model LGBM Baseline LGBM Baseline LGBM Baseline medical dataset of electrocardiogram (ECG) 1dAVb 0.48 0.87 0.50 0.93 0.49 0.90 Potential Explanation: Only 15% of the ECG data exams RBBB 0.87 0.90 0.78 1.00 0.82 hh 0.94 used to train the DL-baseline model was public LBBB 1.00 1.00 0.88 1.00 0.93 1.00 ● Determine if any of 6 ECG anomalies Better Performance by LGBM SB 0.99 0.83 0.72 0.94 0.80 0.88 (1dAVb, RBBB, LBBB, SB, AF, ST) are present Similar Performance (about <10% difference) AF 0.49 1.00 0.49 0.77 0.49 0.87 based on a patient’s ECG exam data Better Performance by DL-Baseline4 ST 0.48 0.95 0.50 0.97 Discussion & Future Work 0.49 0.96 Discussion: ● Existing model would not be able to be implemented in real-world situations as the model only correctly identiﬁes 49% to 88% of anomalies ● Potentially fatal misdiagnosis ● Proof of concept → if improved, similar models could be utilized in the future ○ Tree-based classiﬁers can be faster and less resource intensive compared to DL-models Future Work: ● Train the model with more data ● Experiment with other feature selection techniques + ML models ● Experiment with hyperparameter tuning