Dear Milad Eyvazi Hesar,

Yes, mathematical modelling & scientific computing is a way of visualization cardiac activities and its current activities over the human argons. Now, a days robotic heart surgery is also even possible and active use of machine learning algorithms have been utilizing in this area. See for example beating heart surgery in Article Implementation of a beating heart system for training in off...

Many statistical/mathematical modelling frameworks have been implemented for all such heart related diseases monstering. You can see a lot such work simply by googling it.

Muhammad Ali thanks for sharing this info. I agree with what you mentioned. It supports my sayings. The question is all about wearables and the challenges to using this data and ML or AI.

If you have a lot of data, then the answer is definitely Yes, it's useful!

If you have a small data set (like a few hundred or a few thousand patients), then you have to be careful with which AI/ML model you choose. From my experience with heart disease data, the most challenging part is selecting the right ML algorithm.

For example, I found association rules outperform some ML algorithms for a few hundred patients, surprising! Although feature selection plays an important role, the accuracy doesn't vary too much due to the addition/removal of features.

Dear Milad Eyvazi Hesar Data science is likely to lead to major changes in cardiovascular imaging. Problems with timing, efficiency, and missed diagnoses occur at all stages of the imaging chain. The application of artificial intelligence (AI) is dependent on robust data; the application of appropriate computational approaches and tools; and validation of its clinical application to image segmentation, automated measurements, and eventually, automated diagnosis. AI may reduce cost and improve value at the stages of image acquisition, interpretation, and decision-making. Moreover, the precision now possible with cardiovascular imaging, combined with “big data” from the electronic health record and pathology, is likely to better characterize the disease and personalize therapy. This review summarizes recent promising applications of AI in cardiology and cardiac imaging, which potentially add value to patient care.

Also, these papers will be useful:

Article Artificial Intelligence in Cardiovascular Imaging

https://www.frontiersin.org/articles/10.3389/fcvm.2020.00025/full

https://www.mdpi.com/2076-3417/12/1/519/htm

Dear Sikder Tahsin Al-Amin , thanks for your hint. I also find the selection of the model based on the amount of available, quality data very important. I am wondering if you can explain more about your experience with different particular models and different amounts of data or introduce references with different models for different case studies particularly on time series collected from wearables.

Dear Masoum Mohammadi Gharagoz , I appreciate your answer, and the review articles that you shared are interesting to me. However, I am only interested in AI/ML in time series, or in other words, the recordings collected from electronic wearables such as ECG, PPG, etc., and their challenges.

Dear Milad Eyvazi Hesar , I mostly worked on data collected from hospitals. So, many data were missing and the number of data points was not that high. I had a good model with Logistic Regression, Decision Trees, and SVM. Association rules worked better for a few hundred data points. I tried Deep NN, but it didn't work, mostly due to lack of data I think.

Keeping in mind the answer of Kurt Benke , I would rather say that ML models are the only way of reliably using these large amounts of data. You do not want someone to sit and watch these data flow manually during operation (obviously), and coming up with rules to automatically take decisions is difficult (if not infeasible) given the large amounts of variables and varieties of cases that might arise.

However, large amounts of data alone is not always a road to success. It is important that you have large amounts of correctly labeled data, which should ideally somewhat balanced. Especially in your cases, most of your data will be from one class (healthy) and only few data points for cardiac problems will be seen in your data. Training a classifier on this biased dataset will end up learning the majority case only. This will end up with a classifier with a very high accuracy, but a very bad sensitivity/specificity.

Deep learning techniques will suffice for big data generated from wearable devices capturing cardiac abnormalities.

Well, first of all, AI/ ML is beneficial in the case in case of cardiac disease as well as health monitoring. Machine learning models are bound by their mathematical functions somehow, so I suggest you use Deep learning for this. All we need with deep learning is that explanation. It's about the black-box nature of deep learning while serving a good result. You can follow many methods in the case of image-based or text-based descriptions of data. More researchers are working on this domain right now.

AI/ML are promising fields for automatic classification of cardiac diseases. However, it is expected that real-world datasets in the fields of cardiac disease treatment are collected, pre-processed for training and testing purposes. Also, because of the sensitivity and privacy issues in such data, privacy enhancing techniques should be put in place by the researchers building such datasets, prior to releasing for AI/ML researches.

Dear Kurt Benke , I really appreciate your comment here. looks like there must be concrete regulations to make consistent recordings for any further medical analysis. I believe it is the task of FDA, TGA, EMA or MHRA.

In your opinion, how much accuracy is enough to convince medical society that the output of AI/ML models are pretty reliable even in medical diagnoses?

Dear Manuel Günther , I pretty much agree with you on bold words. My curiosity though- do you think creating augmented, meaningful data which are already confirmed as particular disease, is a good idea to resolve the issue of biased dataset and low sensitivity/specificity?

Dear Sefinat Muhammad-Thani , thanks for your comment. What do you think about the lack of labeled data, balanced dataset, or consistency between different devices?

Dear Arkaprabha Basu , thanks. you think if we mathematically explain what is happening inside a DL model, it is not a black box model anymore. Can we just say we ignore what is happening inside and we focus on the performance of the output of final trained model?

Dear Akinyemi Moruff Oyelakin , thank you for sharing your thoughts on addressing privacy. Isn’t anonymizing enough for it?

Yeah, Milad Eyvazi Hesar, you can, unless explaining the mathematics behind this, you can go for a concept called explainable AI, which is a method of generating explanations or highlighting the ROI on the query from the dl model.

Balancing your data could also be done with easier methods such as oversampling and undersampling, if data augmentation turns out to be difficult. Also, modifications of the loss functions might be a good idea, such as in: Conference Paper MOON: A Mixed Objective Optimization Network for the Recogni...

Arkaprabha Basu I would definitely disagree with your reasoning. Particularly because deep learning models are a black box, we should be more careful in their application. At least we need to clearly understand what they are supposed to do.

I also disagree with Milad Eyvazi Hesar. Although we know mathematically how the deep learning models achieve their goals, and how to train them, we have absolutely no knowledge of what they learn. Hence, they are (and I assume that they will always be) a black box.

I will partly agree with Sir Manuel Günther. Yes, class imbalance handling techniques like SMOTE(JAIR), M2m(CVPR), GAMO(ICCV) will surely give you a better result in case you have an imbalance dataset to train. You can also use class-based weighting in that case, but if you want to go with better applicability of deep learning where we are not only getting the result but also we can say where the deep learning focuses on making its prediction, you can have a look at:

1) CAM: Class activation Maps

a better one than CAM with a mathematical explanation why we don't need retraining

2) Grad-CAM: Gradient-weighted Class Activation Mapping

and a new one that explains we should take care of the positive maps

3) Grad-CAM++: Generalized Gradient-Based Visual Explanations for Deep Convolutional Networks

These are potential explainable AI methods that can explain why the model has decided on the deep learning models. Suppose you are a doctor and a patient has a chest x-ray with them. Your brain is trained with the various patterns and lesion points so you can tell what the abnormalities are, but WHY WOULD THE PATIENT TRUST YOU? That's what the explainable AI is trying to achieve. To explain the result of deep learning, I think the explainable AI is very much necessary in the case of medical imaging as the patient is unaware of why the chest x-ray has TB.

Manuel Günther & Arkaprabha Basu : I see many interesting suggestions for tackling the issue of the imbalanced datasets. I also got the point of Manuel about the meaning of blackbox, which was clearly pointed out- you are right. In addition to Explainable AI that Arkaprabha mentioned, I see some attention-based techniques are used to resolve this issue as well as probabilistic graphical models. Basically, I think it seems like the steps (layers) a ML model takes to give the output are not necessarily similar to the logical steps that the human mind takes to answer the question.

Back to the main question, do you think using models like Transformers, LSTM, etc., in particular in time series, will make the blackbox whiter?

In the case of using AI models for wearables and their recordings, it seems that we may give hints to medical doctors, but not a diagnosis unless we know what logical steps the ML model takes in each step.

Milad Eyvazi Hesar , Manuel Günther & Arkaprabha Basu

Class imbalanced data should be treated with precautions, we proved that many SMOTE variants and alike methods provide fake examples to increase the performance of the system on paper, but in reality, it is a different story because most of these synthesized examples are majority examples and forced to be minority due to their similarities, please find our new study at:

https://ieeexplore.ieee.org/document/9761871