The problem
1) Innovation Dooel produces an online heart monitoring software ViewECG with CE mark as a medical device implementing self-diagnostic AI-based services for long-term real-time continuous monitoring.
2) Training DL-based algorithms to detect arrhythmia requires extremely large number of processing hours on sophisticated GPUs.
3) A monitoring center prototype sends alerts in case of dangerous arrhythmia, which is a Big data streaming problem requiring large processing power.
4) Find the optimal HPC platform for real-time monitoring of 10K patients.
Approach
1) Develop an improved ML/DL algorithm with extensive ECG benchmark databases trained on thousands of GPU cores.
2) Realize an experiment to simulate 10K virtual patients to determine the optimal HPC platform.
2a) Generate 10K virtual patients that stream ECG.
2b) Three different HPC environments:
2b.1) Classical approach with HPC CPU cores.
2b.2) xAFCL orchestration workflow engine.
2b.3) Serverless architecture.
Expected results
Innovation Dooel
1) Design DL-algorithm in several weeks, not solvable by other means
2) Reduce time to design DL-based algorithm from years to weeks (50x)
3) Improved DL-based algorithm (accuracy 80% increased to 90%)
4) Reduced costs compared to physical methods by 30%
5) Reduced costs for large-scale demonstration by 50%
6) Increase revenue (3x annually)
HPC Expert and Competence Center
1) Publish project results, and disseminate use of HPC to industry partners
2) Update learning material, new courses, seminars, summer schools, hackathons, etc.
Key performance indicators
1) Design DL-algorithm in several weeks, not solvable by other means:
2) Reduce time to design DL-based algorithm from years to weeks (50x)
3) Improved DL-based algorithm (accuracy 80% increased to 90%)
4) Reduced costs compared to physical methods by 30%
5) Reduced costs for large-scale demonstration by 50%
6) Increase revenue (3x annually)
Impact
1) A heart monitoring center for remote long-term continuous monitoring of patients in their home environment outside of a hospital, based on wearable ECG sensors will shorten hospitalization time and prevent serious heart damages, providing better healthcare.
2) Large-scale demonstration of processing 10K simultaneous ECG data streams with real-time responses implementing efficient methods and approaches on HPC resources capable of processing several TFLOPS for computationally-intensive signal processing algorithms to classify arrhythmia.
3) Improve the existing DL solution with advanced HPC services, as they cannot be computed with conventional computing resources.