The power of CNN & DL in arrhythmia prediction
Detection & Validation of 21 categorizations of arrhythmias
Accuracy score of 98% and a Precision score of 95% & growing
Affirmed by MD American board-certified panel of Physicians
Trained over a million data sets that are manually annotated by the American CCT/CRAT techs.
Built in fusion with Computer vision and Artificial neural network to detect arrhythmias
Works across a diversified set of digital ECG data sets; such as Holters, MCT, Events & 12 lead
Runs on secure HIPAA cloud infrastructure, enabling easy integration and customisation for IDTF’s & Wearables
The conception of machine learning became engrossing in almost every sector as the healthcare industry stumbles its application as well. We are in devoir of an automated edifice that would administer the detection of assorted arrhythmias from the captured ECGs at the same time be mindful of the signal quality and analogous artifacts.
The approach we have adopted here is the implementation of Deep Learning for arrhythmia detection through a fresh approach utilizing discriminative visualization to enhance transparency and the interpretability of the Deep Neutral Network in an effort to imitate “Human eye interpretation”.
Training and Validation samples(n)
We use two-dimensional Deep Convolutional Neural Network comprising of 14 layers to divvy up cardiac rhythms into 21 categorizations, established using around 315488 single-lead 10 second ECG strips from almost 120063 individual patients. We evaluated the model accuracy using a test set of 5069 ECGs from 4780 patients to achieve desirable test accuracy. It was commentated by five American board certified CCT panel followed by review from a senior CCT and effectuating Grad-Cam analysis on 235 images of the test data set to arrive at a mean Intersection over Union score.
Different devices in the external validation set
The data source comprises of a dataset collected from FDA approved round the clock vitals monitoring devices such as Lifesignals, Bittium Faros, Apple Watch and 12 vanguard ECG machines inclusive of Philips, BPL, GE, etc of 4780 patients to have a profusion of sources.
Intending to tackle the intricacies concomitant with the data processing from disparate sources and their multiple facets of information gain, speed, sampling rate, hardware specific physical and digital min/max calibration values and the data scale, the process had to undergo data preprocessing/preparation phase of the CRISP-DM to create the input image data fed into the model.
1. Signal regularization: The ECG data from discrepant scales pertinent to their respective hardware are standardized to mV signal
2. Befitting lead selection:Felicitous data channel is designated just in case channel information is proven to be enigmatic or inaccessible
3. ECG strip production and dynamic cropping: Transformation of raw data for One-Dimensional to Two-Dimensional raw data transformation for classification.
Signal Processing – Noise Removal and Baseline Correction
The high gullibility of the ECG signals to aspects like power interferences, motion artifacts, muscle movements or hardware-level issues. Such interferences will be acute in the event of cardiac monitoring wearable tech. Especially noise elimination and baseline correction, a zero-phase Butterworth bandpass filter of 0.5 – 40 Hz of order 4 is actionable. Thus the high-frequency noises and low-frequency baseline is eliminated thus preserving the morphological features. The blue signal corresponds to the raw sample inclusive of the noise and baseline drift and the red signal corresponds to the processed signal.
The standard lead name conventions cannot be used here owing to the fact that continuous data from sources like ECG patches might have ambiguity or wouldn’t follow standard lead positions unlike conventional 12 lead ECG machine or 12 lead holter devices. When the lead information is found to be incongruous, an appropriate limb lead detection algorithm is took advantage of to minimize the false positive values and to date the model never detected precordial leads or its facsimile signals during model training.
The idiosyncratic feature of the conventional V1, V2, V3 precordial leads have a unique morphology of deep S waves, small R peaks and high T wave amplitude from other lead patterns which greatly helps in contrasting these lead patterns. The detection algorithm first filters the signal and identifies the QRS peaks in the signal. Then it extracts the QRS beats to calculate the R to S amplitude ratio for individual beats to arrive at the complete nature of the signal to select the pertinent leads for strip generation.
The dismembering of the lead data into a strip of 10 second time frame and further its plotted as the standard gain(10mm/mV) against speed(25mm/sec). This approach affirms that no amplitude is omitted or temporal details of the signal precluded thereby aping a cardiologist’s assessment. Once the ECG images are apprehended they are subjected to information preprocessing in order to get rid of the background using the Hue, Saturation, Value Thresholding to pivot around signal morphology.
Pre-processing of ECG images
Natural images have many statistical properties that are invariant to translation. Convolutional Neural Networks take this property into account by sharing parameters across multiple image locations. It has boundless practical applications for image dataset and replenishes extensive performance in medical image diagnosis and analysis, this method was best suited for our deep learning model. Foremost the ECG image dataset was subjected to preprocessing to eliminate the grid background using the Hue, Saturation, Value thresholding. Now the Convolutional Neural Network which has been employed for the deep learning model has been enriched with the EfficientNet, which is a newly designed Convolutional Neural Network that delivers higher accuracy and computational efficiency. It uses the B3 scaled architecture as a keystone upon which a series of custom layers are built. We use dropouts between average pooling and fully connected layers to regularize the model to prevent overfitting. The final layer includes a softmax activation function.
The search for good hyperparameters can be cast as an optimization problem. 21 different arrhythmia can be classified by the model. Since it is a clinically aligned model it is very important for the classification categories to be tenuous. The model has improved generalization and more optimized owing to the fact that it has unerring class labeling thereby furthering continuous leaning of the model. The model is evaluated using Area Under the Curve (AUC) score to measure the tradeoff between true positive rate and false positive rate and the F1 score that provides the balance between precision and recall.
We're helping some of the most respected names in healthcare deliver measurably better outcomes. Let us show you what personally Human & AI integrated solution can do for your organization. While filling the form, please fill in the information more specifically that you are looking for.
Thank you for your query! We will get back to you shortly!!