TL;DR
This paper introduces a machine learning approach using a sequentially-reduced Feedforward Neural Network (FNN) to efficiently extract crucial cardiovascular health indicators (intrinsic frequencies, IFs) from arterial pressure waveforms. This ML model bypasses the computationally intensive optimization methods, enabling faster and more accessible analysis for real-time clinical applications.
What Problem Does This Paper Address?
Analyzing cardiovascular waveforms is vital for understanding health and disease. The "intrinsic frequency" (IF) method is a promising technique that extracts physiologically relevant information from a single arterial pressure waveform. However, the current standard method for calculating IFs relies on a computationally complex L2 optimization solver. This complexity makes real-time deployment of the IF method challenging, hindering its widespread clinical adoption. The authors aim to overcome this bottleneck by developing a machine learning-based approach that can determine IF parameters efficiently and accurately.
Methods
The core of this work is a sequentially-reduced Feedforward Neural Network (FNN). Unlike traditional FNNs, this architecture is designed with a progressively decreasing number of neurons in each hidden layer. This "sequentially-reduced" structure is key to mapping the high-dimensional input (arterial waveforms) to a lower-dimensional output (IF parameters) efficiently. The FNN is trained to directly predict the IF parameters from normalized carotid waveforms and the dicrotic notch time. This approach bypasses the need to solve the non-convex L2 minimization problem inherent in the conventional IF method. The methodology also includes detailed procedures for data pre-processing (normalization and resampling) and model training.
Datasets
The authors utilized a heterogeneous database for training, validation, and testing:
- Clinical Databases:
- Framingham Heart Study (FHS): Carotid waveforms from N=6697 participants measured via applanation tonometry.
- HMRI iPhone Heart Study: Carotid waveforms from N=2312 participants measured using three distinct devices: traditional tonometry, optical tonometry (Vivio), and an iPhone camera.
- Synthetic Database: Generated waveforms with exact IF values to ensure adherence to the mathematical formulation of IF and to enrich the training data, particularly in underrepresented physiological ranges.
A portion of the clinical data was set aside for a blind clinical test (external validation) using 3009 additional FHS waveforms.
Key Insights and Ideas
- ML as a Computational Shortcut: The FNN acts as an efficient surrogate for the complex L2 optimization, significantly reducing computational burden without sacrificing accuracy.
- Device and Sampling Rate Independence: A crucial finding is that the FNN model's performance is largely unaffected by the measurement device or the signal's sampling rate, making it highly adaptable to real-world clinical data.
- Importance of Data Pre-processing: Normalization and resampling are critical steps to ensure that waveforms from different devices and with varying lengths can be uniformly processed by the FNN.
- Sequentially-Reduced Architecture: The specific FNN design, with its decreasing neuron count across layers, is tailored to effectively capture the essential features of the waveform for IF prediction.
Key Results
- High Agreement with Standard Method: The FNN-based IF parameters showed excellent agreement with those computed by the standard L2 optimization method, with correlation coefficients (R) as high as 0.98 for the first IF (ω₁) and 0.93 for the second IF (ω₂).
- Low Errors: In blind clinical tests, the RMSE for ω₁ was 0.6 (relative error 0.71%) and for ω₂ was 1.49 (relative error 2.4%). Other IF parameters also showed low errors.
- Computational Efficiency: The FNN model performs online predictions in near real-time (3.94e-5 seconds per waveform on a laptop), a significant improvement over the computationally expensive L2 optimization.
- Robustness: The model's accuracy was found to be largely independent of the training data size beyond 80% of the total dataset.
Limitations and Future Work
- Clinical Value Assessment: The authors state their future goal is to directly assess the clinical value of the predicted IF parameters in terms of diagnostic and prognostic capabilities for cardiovascular diseases.
- Extension to Multiple Decoupling Points: While the current model handles one decoupling point (typical for a single cardiac cycle), the authors suggest the methodology could be extended to systems with multiple decoupling times for more complex physiological scenarios. They hypothesize that an ML approach would still be significantly faster than brute-force minimization for extracting multiple IFs.
Summary updated