Case Studies
Remote cardiac monitoring platform
Key Results
- Real-time arrhythmia detection
- Distributed processing for wearable ECG data
- IEC 62304 engineering documentation
About the project
PerformaCode developed an ML-driven ECG monitoring platform for a European medtech startup developing wearable cardiac devices for remote patient monitoring.
The platform supported early product development by combining publicly available ECG datasets with recordings collected from pilot wearable devices. This allowed the arrhythmia detection pipeline and distributed processing architecture to be evaluated before larger clinical deployments.
The system processed ECG data from Holter monitors and other wearable cardiac monitoring devices. It was designed to detect heartbeat abnormalities automatically from streamed ECG signals and notify responsible users when stable signs of urgent cardiac conditions appeared.
The platform combined ECG signal preprocessing, QRS/QRST complex detection, analytical arrhythmia detection, and neural-network-based recognition of urgent cardiac conditions. A three-layer neural network was trained on PhysioBank ECG datasets, while analytical methods handled rhythm abnormalities such as bradycardia, tachycardia, and sinus arrhythmia.
The backend used distributed open-source components for data ingestion, buffering, compute distribution, and long-term storage. Apache Ignite provided the distributed processing layer, Kafka and MQTT handled transmission buffers, and Hadoop-based storage supported offline analysis of recorded ECG data.
PerformaCode also prepared the engineering evidence around the functional core, including requirements, code inspection, architecture and design documentation, software testing, bug reporting, and IEC 62304 / ISO 13485 process artifacts.
2-5
engineers
24
months
FP
delivery model
Client challenges
Remote cardiac monitoring is a streaming problem as much as a machine learning problem. ECG signals arrive continuously from distributed wearable devices, contain measurement noise and motion artifacts, and require processing fast enough to recognize clinically significant rhythm abnormalities without interrupting the data flow.
The platform had to combine several processing stages into a single pipeline. Raw ECG signals first required filtering, normalization, decimation, and QRS/QRST complex detection before analytical algorithms and a neural network could classify arrhythmias. Each stage depended on the accuracy and timing of the previous one, making signal quality and processing latency critical to the overall system.
At the same time, the architecture needed to scale beyond a single server. The system was designed to receive continuous data streams from large numbers of monitoring devices, distribute processing across heterogeneous compute nodes, and store recordings for long-term analysis while maintaining interoperability through HL7 standards.
As the platform matured, the functional implementation also had to be supported by engineering documentation and verification artifacts aligned with IEC 62304 and ISO 13485. Existing source code and system design had to be analyzed, documented, inspected, and tested to establish the engineering evidence expected for medical device software.
Tasks performed
- Designed a distributed processing architecture for real-time ingestion and analysis of ECG data from wearable cardiac monitoring devices.
- Developed ECG signal preprocessing algorithms for normalization, filtering, noise reduction, and frequency decimation.
- Implemented QRS and QRST complex detection using a customized Pan-Tompkins algorithm for ECG feature extraction.
- Developed analytical arrhythmia detection for rhythm abnormalities including bradycardia, tachycardia, sinus arrhythmia, and heterotopic arrhythmias.
- Built a three-layer neural network trained on PhysioBank ECG datasets to recognize urgent cardiac conditions from processed ECG signals.
- Integrated Apache Ignite as the distributed compute layer for horizontally scalable ECG processing.
- Integrated Kafka and MQTT as transmission buffers for continuous data streams from monitoring devices.
- Implemented Hadoop/Hive-based storage and analysis for long-term ECG recordings and offline statistical processing.
- Implemented HL7 v2/v3 interoperability for healthcare data exchange between monitoring devices and health information system components.
- Integrated OpenEMR as the healthcare information system component within the monitoring platform.
- Developed web services for remote monitoring, alerting, and historical ECG data access.
- Implemented event correlation workflows to compare detected ECG abnormalities with external context data, including weather conditions.
- Built BI/reporting workflows using Qlik for analysis of stored monitoring data.
- Implemented Java and C++ platform components across data processing, integration, and monitoring workflows.
- Created software requirements for the existing functional core.
- Produced architecture and design documentation aligned with IEC 62304 expectations.
- Performed source code inspection for the platform’s functional core.
- Executed software testing and defect reporting under ISO 13485 and IEC 62304 quality processes.
- Prepared verification evidence for the documented functional core.
Project results
5–8 arrhythmia classes
The recognition pipeline identified up to eight rhythm abnormality classes by combining analytical ECG rules with neural-network classification on processed ECG features.
50+ devices monitored
The platform supported 50+ wearable ECG devices in pilot-scale monitoring by routing device streams through a common ingestion and processing pipeline.
10,000+ ECG records
Arrhythmia recognition was developed and evaluated against 10,000+ public ECG records, supplemented with ECG data collected from pilot wearable devices.
3-node deployment
The system ran on a three-node distributed deployment by using Apache Ignite for compute distribution and Kafka/MQTT for stream buffering.
<10 s event-to-alert
Stable arrhythmia detections reached the notification workflow in under 10 seconds by keeping ECG preprocessing, classification, and alert routing inside the streaming pipeline.
500+ HL7 records/min
Healthcare data exchange handled 500+ HL7 records per minute through standardized interfaces between monitoring devices and health information system components.
24/7 ECG streaming
Continuous ECG streams were buffered, processed, and stored without stopping device communication by separating ingestion, processing, and long-term storage.
Historical ECG archive
Recorded ECG data remained available for offline statistical analysis by storing long-term signal data in Hadoop/Hive-based infrastructure.
IEC 62304 evidence
The functional core was supported with requirements, architecture, design, testing, verification, and code inspection artifacts aligned with IEC 62304 and ISO 13485 processes.
Value we bring
Open-source platforms for regulated medical software
Open-source infrastructure can be used in regulated medical systems, but only when the surrounding engineering discipline is strong enough. We work with open-source components while defining architecture boundaries, integration paths, testing scope, operational behavior, and documentation needed for medical software development.
Streaming architectures for physiological signals
Physiological data is continuous, noisy, and time-sensitive. We design processing architectures that separate ingestion, buffering, signal processing, event detection, storage, and reporting so monitoring systems can handle live device data without treating it like ordinary batch analytics.
Growing prototypes into medical products
Working software is not the same as product-ready medical software. We take functional cores that already process data or support workflows and add the missing requirements, architecture records, design documentation, code inspection, testing, defect reports, and verification evidence needed for regulated development.
Technologies
- Java
- C++
- Apache Ignite
- Apache Kafka
- MQTT
- Apache Hadoop
- Apache Hive
- HL7 v2/v3
- OpenEMR
- Neuroph
- Qlik
Other Case Studies
ECG Artifact Recognition for Portable Defibrillation
Algorithm development and validation for a dual-mode portable defi...
