Multiple instance learning framework can facilitate explainability in murmur detection.

OBJECTIVE: Cardiovascular diseases (CVDs) account for a high fatality rate worldwide. Heart murmurs can be detected from phonocardiograms (PCGs) and may indicate CVDs. Still, they are often overlooked as their detection and correct clinical interpretation require expert skills. In this work, we aim to predict the presence of murmurs and clinical outcomes from multiple PCG recordings employing an explainable multitask model. APPROACH: Our approach consists of a two-stage multitask model. In the first stage, we predict the murmur presence in single PCGs using a multiple instance learning (MIL) framework. MIL also allows us to derive sample-wise classifications (i.e. murmur locations) while only needing one annotation per recording ("weak label") during training. In the second stage, we fuse explainable hand-crafted features with features from a pooling-based artificial neural network (PANN) derived from the MIL framework. Finally, we predict the presence of murmurs and the clinical outcome for a single patient based on multiple recordings using a simple feed-forward neural network. MAIN RESULTS: We show qualitatively and quantitatively that the MIL approach yields useful features and can be used to detect murmurs on multiple time instances and may thus guide a practitioner through PCGs. We analyze the second stage of the model in terms of murmur classification and clinical outcome. We achieved a weighted accuracy of 0.714 and an outcome cost of 13612 when using the PANN model and demographic features on the CirCor dataset (hidden test set of the George B. Moody PhysioNet challenge 2022, team "Heart2Beat", rank 12 / 40). SIGNIFICANCE: To the best of our knowledge, we are the first to demonstrate the usefulness of MIL in PCG classification. Also, we showcase how the explainability of the model can be analyzed quantitatively, thus avoiding confirmation bias inherent to many post-hoc methods. Finally, our overall results demonstrate the merit of employing MIL combined with handcrafted features for the generation of explainable features as well as for a competitive classification performance.

An extensive quantitative analysis of the effects of errors in beat-to-beat intervals on all commonly used HRV parameters.

Heart rate variability (HRV) analysis is often used to estimate human health and fitness status. More specifically, a range of parameters that express the variability in beat-to-beat intervals are calculated from electrocardiogram beat detections. Since beat detection may yield erroneous interval data, these errors travel through the processing chain and may result in misleading parameter values that can lead to incorrect conclusions. In this study, we utilized Monte Carlo simulation on real data, Kolmogorov-Smirnov tests and Bland-Altman analysis to carry out extensive analysis of the noise sensitivity of different HRV parameters. The used noise models consider Gaussian and student-t distributed noise. As a result we observed that commonly used HRV parameters (e.g. pNN50 and LF/HF ratio) are especially sensitive to noise and that all parameters show biases to some extent. We conclude that researchers should be careful when reporting different HRV parameters, consider the distributions in addition to mean values, and consider reference data if applicable. The analysis of HRV parameter sensitivity to noise and resulting biases presented in this work generalizes over a wide population and can serve as a reference and thus provide a basis for the decision about which HRV parameters to choose under similar conditions.

Automatic Generation of Labeled Data for Video-Based Human Pose Analysis via NLP applied to YouTube Subtitles.

With recent advancements in computer vision as well as machine learning (ML), video-based at-home exercise evaluation systems have become a popular topic of current research. However, performance depends heavily on the amount of available training data. Since labeled datasets specific to exercising are rare, we propose a method that makes use of the abundance of fitness videos available online. Specifically, we utilize the advantage that videos often not only show the exercises, but also provide language as an additional source of information. With push-ups as an example, we show that through the analysis of subtitle data using natural language processing (NLP), it is possible to create a labeled (irrelevant, relevant correct, relevant incorrect) dataset containing relevant information for pose analysis. In particular, we show that irrelevant clips (n = 332) have significantly different joint visibility values compared to relevant clips (n = 298). Inspecting cluster centroids also show different poses for the different classes.

Estimating Thoracic Movement with High-Sampling Rate THz Technology.

We use a high-sampling rate terahertz (THz) homodyne spectroscopy system to estimate thoracic movement from healthy subjects performing breathing at different frequencies. The THz system provides both the amplitude and phase of the THz wave. From the raw phase information, a motion signal is estimated. An electrocardiogram (ECG) signal is recorded with a polar chest strap to obtain ECG-derived respiration information. While the ECG showed sub-optimal performance for the purpose and only provided usable information for some subjects, the signal derived from the THz system showed good agreement with the measurement protocol. Over all the subjects, a root mean square estimation error of 1.40 BPM is obtained.

Design of Hardware Accelerators for Optimized and Quantized Neural Networks to Detect Atrial Fibrillation in Patch ECG Device with RISC-V.

Atrial Fibrillation (AF) is one of the most common heart arrhythmias. It is known to cause up to 15% of all strokes. In current times, modern detection systems for arrhythmias, such as single-use patch electrocardiogram (ECG) devices, have to be energy efficient, small, and affordable. In this work, specialized hardware accelerators were developed. First, an artificial neural network (NN) for the detection of AF was optimized. Special attention was paid to the minimum requirements for the inference on a RISC-V-based microcontroller. Hence, a 32-bit floating-point-based NN was analyzed. To reduce the silicon area needed, the NN was quantized to an 8-bit fixed-point datatype (Q7). Based on this datatype, specialized accelerators were developed. Those accelerators included single-instruction multiple-data (SIMD) hardware as well as accelerators for activation functions such as sigmoid and hyperbolic tangents. To accelerate activation functions that require the e-function as part of their computation (e.g., softmax), an e-function accelerator was implemented in the hardware. To compensate for the losses of quantization, the network was expanded and optimized for run-time and memory requirements. The resulting NN has a 7.5% lower run-time in clock cycles (cc) without the accelerators and 2.2 percentage points (pp) lower accuracy compared to a floating-point-based net, while requiring 65% less memory. With the specialized accelerators, the inference run-time was lowered by 87.2% while the F1-Score decreased by 6.1 pp. Implementing the Q7 accelerators instead of the floating-point unit (FPU), the silicon area needed for the microcontroller in 180 nm-technology is below 1 mm2.

Exploring novel algorithms for atrial fibrillation detection by driving graduate level education in medical machine learning.

During the lockdown of universities and the COVID-Pandemic most students were restricted to their homes. Novel and instigating teaching methods were required to improve the learning experience and so recent implementations of the annual PhysioNet/Computing in Cardiology (CinC) Challenges posed as a reference. For over 20 years, the challenges have proven repeatedly to be of immense educational value, besides leading to technological advances for specific problems. In this paper, we report results from the class 'Artificial Intelligence in Medicine Challenge', which was implemented as an online project seminar at Technical University Darmstadt, Germany, and which was heavily inspired by the PhysioNet/CinC Challenge 2017 'AF Classification from a Short Single Lead ECG Recording'. Atrial fibrillation is a common cardiac disease and often remains undetected. Therefore, we selected the two most promising models of the course and give an insight into the Transformer-based DualNet architecture as well as into the CNN-LSTM-based model and finally a detailed analysis for both. In particular, we show the model performance results of our internal scoring process for all submitted models and the near state-of-the-art model performance for the two named models on the official 2017 challenge test set. Several teams were able to achieve F1scores above/close to 90% on a hidden test-set of Holter recordings. We highlight themes commonly observed among participants, and report the results from the self-assessed student evaluation. Finally, the self-assessment of the students reported a notable increase in machine learning knowledge.