Rare-Class Collapse in ECG-Based Ventricular Tachycardia and Fibrillation Detection: A Systematic Benchmark of Class-Imbalance Mitigation from Reweighting to Cascade Classification

Technical Summary: Rare-Class Collapse in ECG-Based Ventricular Tachycardia and Fibrillation Detection

Problem Statement
The automated detection of lethal arrhythmias, specifically Ventricular Tachycardia (VT) and Ventricular Fibrillation (VF), is critically hindered by severe class imbalance in clinical ECG datasets. In these settings, lethal events often constitute fewer than 22% of segments. This imbalance leads to a failure mode termed "rare-class collapse," where standard classifiers achieve high overall accuracy by predominantly predicting the majority class (normal rhythm) while missing a significant proportion of life-threatening events. Conventional evaluation metrics, such as accuracy and Area Under the Receiver Operating Characteristic Curve (AUROC), often obscure this clinical failure because they are dominated by the large number of negative samples. The study posits that without specific evaluation at a fixed clinical operating point and the use of metrics focused on the minority class, the safety of these detection systems cannot be reliably assessed.

Methodology
The study conducted a systematic benchmark of six distinct imbalance-handling approaches using a pooled dataset of 15,614 ECG segments derived from three PhysioNet databases: VTaC, VFDB, and CUDB. The final cohort exhibited a normal-to-lethal ratio of approximately 3.6:1.

• Data Preprocessing: Signals were sampled at 250 Hz, band-pass filtered (0.5–40 Hz), normalized, and segmented into non-overlapping 6-second windows (1,500 samples).
• Feature Extraction: A set of 22 handcrafted features was computed per segment, encompassing time-domain statistics (e.g., mean absolute amplitude, zero-crossing rate) and frequency-domain properties (e.g., spectral entropy, dominant frequency, band power ratios).
• Evaluation Framework: To ensure clinical relevance, all models were evaluated at a fixed operating point of 95% specificity (5% false-alarm rate). Performance was measured using Recall, Area Under the Precision-Recall Curve (AUPRC), and the Missed-Lethal-Event Rate (MLER). Decision curve analysis was employed to estimate net clinical benefit.
• Models Evaluated:
  1. Naive Baseline: L2-regularized logistic regression without class weighting.
  2. Inverse-Frequency Reweighting: Logistic regression with weights inversely proportional to class frequency.
  3. LDAM Loss: A Multi-Layer Perceptron (MLP) trained with Label-Distribution-Aware Margin loss to enforce larger decision margins for the minority class.
  4. Cost-Sensitive Training: An MLP trained with cross-entropy loss where missing a lethal event is penalized 10x more than a false alarm.
  5. Two-Stage Cascade: A screening Random Forest (high recall, minority weight 20) followed by a balanced Random Forest to filter false alarms.
  6. Anomaly Detection: An Isolation Forest trained exclusively on normal segments to detect VT/VF as anomalies.

Key Results

• Failure of Naive Approaches: The naive logistic regression model achieved 84.1% accuracy and an AUROC of 0.822 but suffered from rare-class collapse, detecting only 45.1% of lethal events (MLER = 0.549). It missed 564 out of 1,027 lethal events in the test set.
• Limited Efficacy of Simple Reweighting: Inverse-frequency reweighting yielded negligible improvement, increasing recall only to 45.5% and missing 560 events.
• Superiority of Structural and Objective Modifications: Methods that altered the learning objective or model architecture significantly outperformed simple reweighting.
  • LDAM improved recall to 58.9%.
  • Cost-Sensitive Training achieved a recall of 61.8% and the highest AUROC (0.927).
  • Two-Stage Cascade achieved the best overall performance with a 65.2% recall and an AUPRC of 0.821. This approach reduced the number of missed lethal events by 37% (from 564 to 357) compared to the naive baseline while maintaining a similar false-alarm count.
• Source-Specific Performance Disparity: A critical finding was the divergence in performance between arrhythmia types. The cascade model achieved near-complete detection for VF (recall up to 0.975 on the VFDB dataset) but performed poorly on organized VT (recall of only 0.183 on the VTaC dataset).
• Feature Space Limitations: Analysis of feature discrimination (Cohen's d) revealed that the most discriminative feature (spectral entropy) had an effect size of only 0.52. The study attributes the poor VT detection to the spectral similarity between organized VT and rapid sinus rhythm, suggesting that the handcrafted feature set imposes a ceiling on performance regardless of the classification algorithm.

Significance and Claims
The paper claims that the choice of evaluation metrics fundamentally dictates the visibility of rare-class failure. By shifting focus from accuracy to recall at a fixed specificity and the missed-lethal-event rate, the study demonstrates that standard reweighting techniques are insufficient for lethal arrhythmia detection. The primary contribution is the identification of a feature-space limitation rather than a model-design limitation for VT detection; the spectral overlap between VT and rapid sinus rhythm prevents handcrafted features from achieving high recall.

Consequently, the authors argue that for clinical deployment, detection systems must be evaluated based on the number of missed events rather than overall correctness. They conclude that while two-stage cascades offer the best current performance among the tested methods, further improvement for VT detection likely requires moving beyond handcrafted features to models that learn directly from raw ECG waveforms (e.g., deep learning approaches). The study serves as a methodological proof of concept, highlighting that current feature-based systems have a practical ceiling in distinguishing organized VT from normal rhythms.