Medtronic Internship 2024
U-Net Deep Learning Model for Arrhythmia Segmentation

Introduction and Background

During the summer of 2024, I interned with Medtronic’s AI Research Team within the Cardiac Rhythm Management (CRM) Research & Technology (R&T) sector. My work focused on training a U-Net deep learning model to segment episodes of a specific arrhythmia (an irregular heartbeat) under investigation by the team.

The arrhythmia in question serves as a critical risk indicator for more severe cardiac conditions, which can lead to sudden cardiac arrest and death. Accurate detection and segmentation of this arrhythmia within a patient’s electrocardiogram (ECG)—a data-driven representation of the heart’s electrical activity—are essential for effective patient monitoring and care.

Figure 1: Example of an electrocardiogram (ECG) of the heart (American Heart Association).

Furthermore, with Medtronic’s LINQ™ Implantable Cardiac Monitor (ICM)—a small device designed to measure and collect patient ECG data—the team aimed to integrate the detection model into the CareLink™ system. This would enable the segmentation of arrhythmia episodes alongside existing AI algorithms.

The LINQ™ ICM is a compact, minimally invasive device implanted just beneath the skin of the chest. Positioned near the heart, it continuously monitors the patient’s electrocardiogram (ECG), providing long-term cardiac data. This small device is particularly valuable for detecting irregular heart rhythms and transmitting the data wirelessly to the CareLink™ system for real-time analysis and physician review.

The ECG data used for the segmentation project was collected from the LINQ™ ICMs of thousands of patients.

Figure 2: LINQ™ ICM (Medtronic).

Figure 3: Upper chest implantation location of the LINQ™ ICM (Medtronic).

Data Collection and Preprocessing

The dataset for this project consisted of ECG recordings collected from LINQ™ devices across multiple patients. Each recording was annotated by a clinical annotation team, identifying the presence and type of arrhythmias. This provided a rich source of labeled data for training and evaluation.

In the preprocessing phase, I filtered the dataset to exclude arrhythmia types that were not relevant to the segmentation task, ensuring the training data was highly focused. Additional preprocessing steps included standard techniques commonly applied in deep learning workflows for time-series data, such as normalization, signal filtering, and segmenting the ECG recordings into fixed-length windows for model input. These steps helped enhance data quality and model performance.

Why Segmentation and U-Net?

In the context of arrhythmia detection, segmentation plays a crucial role by enabling the precise identification of regions within an ECG signal that correspond to specific arrhythmias. Unlike simple classification, segmentation provides detailed information about when and where these events occur, which is critical for clinical diagnosis and treatment planning.

The U-Net architecture was chosen for this task due to its proven effectiveness in medical image and signal segmentation. Its encoder-decoder structure captures both high-level contextual information and fine-grained details, making it well-suited for the complex patterns in ECG data. Additionally, U-Net’s skip connections preserve spatial resolution, ensuring accurate segmentation of arrhythmias even in challenging regions of the signal.

This combination of segmentation and U-Net provides a robust framework for detecting arrhythmia episodes with high precision, meeting the demands of clinical applications.

My Role

Identified and resolved annotation anomalies in patient ECG data using Python scripts.
Filtered and preprocessed ECG data, generating approximately 3,500 binary masks for model training.
Visualized ECG-mask combinations to ensure data quality and alignment.
Trained a U-Net deep learning model in PyTorch Lightning, achieving a competitive mIoU score of 70%.
Established the groundwork for integrating the model into Medtronic’s LINQ™ ICM device to enhance arrhythmia detection.
Presented project results and documentation to Medtronic’s CRM R&T Department and AI Research Team, showcasing its advantages over standard rate-based detection algorithms.

These efforts demonstrated the application of deep learning to improve the precision and reliability of arrhythmia detection in real-world medical applications.

Key Outcomes

Visuals - Model Performance

Acknowledgements

During the summer of 2024, I interned with Medtronic's AI Research Team within the Cardiac Rhythm Management (CRM) Research & Technology (R&T) sector. My work focused on training a U-Net deep learning model to segment episodes of a specific arrhythmia (an irregular heartbeat) under investigation by the team.