Application of machine learning to predict timing of ventricular tachycardia onset

Ventricular tachycardia (VT) poses a significant threat in acute care settings, where early detection is crucial. Traditional alarm systems are reactive and plagued by false alarms, offering limited utility for early intervention. Previous models have attempted to classify alarms as true or false, but few have explored the feasibility of predicting VT onset in advance.

This study aims to explore whether it’s possible to predict true VT events several seconds in advance using only routinely available pre-alarm physiological signals.

We designed a deep learning model based on a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) architecture. The model was trained using 30-second segments of waveform data, including two-lead ECG, plethysmography (PLETH), and arterial blood pressure (ABP), ending 10 seconds before a documented VT event. Training ran for 39 epochs and was automatically stopped early when the model’s performance plateaued on the validation set. We evaluated the model using ROC curves and confusion matrices across two thresholds: one default and one optimized for higher sensitivity.

The model achieved an AUC of 0.744, showing moderate predictive ability. At the default threshold (0.50), recall was 72%, with 175 true positives, 69 false negatives, and 168 false positives. When tuned to a threshold of 0.34, recall improved to 80%, capturing more true VT cases but also led to more false alarms (212). This sensitivity-specificity trade-off gives clinicians flexibility in prioritizing early detection versus alarm burden.

These results offer a promising proof-of-concept that true VT events can be anticipated in advance using only pre-alarm ECG and waveform data. Our CNN-LSTM model shows that these signals hold valuable predictive information. With further development, such models could one day serve as intelligent early-warning tools, giving clinicians precious seconds to intervene before a VT episode unfolds.Figure 1:

ROC