Automating fetal heart monitor using machine learning
- 1. Automating Fetus Health Monitor Using Machine Learning Md. Tamjid Rayhan Department of Electrical and Electronic Engineering University of Chittagong Session : 2012 - 2013 1
- 2. Outline of this seminar • How I started my journey in Machine Learning and what plan do I suggest for the newbies? • How did I approach a Biomedical Engineering problem using machine learning? • Question and Answer session. 2
- 3. Machine learning: Beginner to Professional • Which tutorials, programming languages, books and blogs suits the beginners best? • Do I need GPU’s , TPU’s, Cloud computing resources? If so, can I get them for free? • Where do I get my data for free? • How do I know whether I have become a Pro in Machine Learning? 3
- 4. Where to Start? • C /C++ / JAVA / MATLAB / R/ Python? • There’s a github repo that can help you decide • https://github.com/josephmisiti/awesome-machine-learning 4
- 5. Advantages of python • Easy to learn • Easy to find materials • Easy to write code in python • Enormous community who are very supportive • There is a library / framework of python for almost any ML task • Python can also be used for Web and GUI development 5
- 6. How to start learning python? https://python.maateen.me/ 6
- 7. Why not official documentation? https://docs.python.org/3/ 7
- 8. What about video tutorials? https://www.py4e.com/ 8
- 9. Which IDE should you use? • Here’s a link to all available options. Find one that suits you best: • https://wiki.python.org/moin/PythonEditors • I use iPython/Jupyter notebooks for data analysis and model development • I currently use VS code for my python development • I started coding in the python IDLE that comes built in with python installation. 9
- 10. Data structure and algorithms Grokking Algorithms Introduction to algorithms 10
- 11. Machine learning is all about MATH!!! • This is the truth , whether you are happy about it or sad about it! • Do I need to be a math genius to start learning Machine Learning. Well the answer is no! • The prerequisite maths of learning machine learning is very simple: Matrix multiplication, vector dot product, differentiation, mean, variance, histogram. • You need to love math though! Because each and everything you learn about will involve math! 11
- 12. Get your hands dirty asap! With Kaggle! 12
- 13. Essential Kaggle Micro-courses! 1. Python 5. Deep Learning 2. Pandas 6. Feature engineering 3. Intro to Machine Learning 7. Data visualization 4. Intermediate Machine Learning 8. Intro to SQL 13
- 14. Outcome of Kaggle Micro courses Python Library Introduction Machine learning introduction Pandas Your first ML Project Scikit-learn Your first ML competition Tensorflow Your first DL project Matplotlib and Seaborn Data collection using SQL 14
- 15. The signature Machine Learning Course https://www.coursera.org/learn/machine-learning 15
- 16. Outcome of “Machine Learning” course Statistics Machine Learning Algorithms Linear and Logistic Regression Neural Networks implementation Concept of cost function and gradient descent SVM implementation Concept of bias and Variance K-Means clustering implementation Performance Metrics of an ML algorithm Recommender system implementation 16
- 17. Do I need GPU or TPU? • No, instead use the free online virtual runtimes provided by Kaggle and Google Colab • In Google Colab you get a runtime with 12GB RAM and you have access to GPU and TPU. Also you can mount your google drive as permanent storage. • In Kaggle you can upload your own dataset , do analysis on it using their free runtime. 17
- 18. Where to get the data from? • Kaggle competitions and datasets • UCI Machine learning repository • Physionet.org • Google Dataset search • Web scraping • Retrieving data from relational databases using SQL • Collecting and Building your own dataset • Collecting data from existing products and services 18
- 19. What are the best books to learn Machine learning? Hands on Machine Learning using scikit-learn and tensorflow Deep Learning – Ian Goodfellow https://www.deeplearningbook.org/ 19
- 20. Dive into deep learning! https://www.coursera.org/specializations/deep-learning 20
- 21. Miscellenius Resources • Google crash course on machine learning • https://developers.google.com/machine-learning/crash-course/ • This blog is for computer vision enthusiasts • https://www.pyimagesearch.com/ • Subscribe the newsletter to deeplearning.ai • https://www.deeplearning.ai/thebatch/ • Having difficulty setting up your GPU? • https://course.fast.ai/#using-a-gpu 21
- 22. End to End learning vs. multiple stage learning 22
- 23. Small data vs Big data https://www.industryweek.com/technology-and-iiot/digital-tools/article/21122846/making- ai-work-with-small-data • Unlike consumer Internet companies, which have data from billions of users to train powerful AI models, collecting massive training sets in manufacturing is often not feasible. • In a recent MAPI survey, 58% of research respondents reported that the most significant barrier to deployment of AI solutions pertained to a lack of data resources. • Synthetic data generation, Transfer learning, Self-supervised learning, few- shot learning, One-shot learning, anomaly detection, Human-in-the-loop 23
- 24. Solving a problem using Machine Learning • Why did I choose to work on automating Fetus Health Monitor? • Why Machine Learning is needed to solve this problem? • What were my results? Did I answer my research questions? • Is my work ready to get implemented in a hospital yet? If no, why not? 24
- 25. Problem Stillbirth Global problem Application of technology is inadequete SDG of attaining zero stillbirth by 2030 28 out of 1000 births are stillbirths in Bangladesh 98% of these death in Developing countries 2.1 million stillbirth in 2015 25
- 26. Solution Automate Effective decision in emergency Save time Enable Tele- Monitoring 26
- 27. Cardiotocography features Feature Definition Ideal Value Baseline Heart rate at rest 110 – 160 beats per minute Variability Fluctuations from the baseline > 6 beats per minute Acceleration Abrupt increase from baseline of 15 beats/min that lasts for 15 seconds Must be present in a healthy fetus Deceleration Decrease from baseline of 15 beats/min that lasts for 15 seconds Decelerations are non- reassuring, Should not be present in healthy fetus 27
- 28. Cardiotocography data of a normal fetus 28
- 29. Cardiotocography data of a pathologic fetus 29
- 30. Similar Problems • Detection of heart disease from ECG data • Predicting epileptic seizure from analyzing EEG data • Detecting Pneumonia from Chest X-ray Images 30
- 31. Objectives • Research question 1: Is it possible to predict fetus health automatically from cardiotocography data? • Research question 2: If it is possible to predict fetus health automatically from the Cardiotocography data, then how accurate is that model in predicting fetus health? • Research question 3: If it is possible to predict fetus health automatically from the Cardiotocography data, then which features of cardiotocography data are most important in predicting fetus health? 31
- 32. Data preprocessing Exploratory data analysis Handling categorical variables Missing value interpolation Outlier detection and removal Feature selection Train Test spliting 32
- 33. Model Development Choose algorithm Algorithm applicability by drawing learning curve Tune Hyperparameters from validation curves Train model Calculate performance metrics, draw confusion matrix and ROC Compare the performance of all built models 33
- 34. Learning Curve for SVM Learning curve for SVM with Linear kernel Learning curve for SVM with Gaussian kernel 34
- 35. Comparisn between built models Model Description Sensitivity (Pathologic) Sensitivity (Suspected) Precision (Normal) Accuracy Logistic regression with selected features Good(0.946) bad(0.767) Excellent(0.976) Good(0.812) Logistic regression with all features Excellent (0.973) Good(0.86) Excellent(0.988) Average (0.804) Random forest with selected features Good (0.892) Good(0.837) Excellent( .974) Excellent( 0.953) Random forest with all features Good(0.892) Good(0.884) Excellent( .98) Excellent( 0.953) SVM with selected features Good(0.919) Good(0.837) Excellent(0.992) Good(0.841) SVM with all features Good(0.946) Good(0.814) Excellent(0.984) Good(0.833) 35
- 36. Answers to research questions • Question 1: Is it possible to predict fetus health automatically from Cardiotocography data? • Answer: Yes, It is possible. • Question 2: Then how accurate is that model in predicting fetus health? • Answer: Based on the performance of models I selected two models as most implementable • One of them Random forest with selected features has high overall accuracy of 95.3% • Another one Logistic Regression with all features has compromised overall accuracy of 80.4% to obtain excellent sensitivity in pathologic of 97% 36
- 37. Answers to research questions • Question 3: Which features of cardiotocography data are most important in predicting fetus health? • Answer: According to feature importance of random forest model the feature importance of the five most important features are as below: Feature Importance Percentage of time with abnormal short term variability (ASTV) 0.15 Percentage of time with abnormal long term variability (ALTV) 0.13 Histogram mean 0.1 Mean value of short term variability (MSTV) 0.08 Acceleration (AC) 0.07 37
- 38. How do you know you have become a Professional in Machine Learning? • Be a competitions, notebooks, datasets, discussions master in Kaggle. 4x grandmaster Abhishek Thakur explains: • https://www.youtube.com/watch?v=z15TKkAPNUM • Publish the results of your research in an impactful journal. • Get the Machine learning engineer job you wanted • Get accepted in your favourite PhD program • Build a good freelancing portfolio in this field 38
- 39. Yann Lecun’s Advice • So my advice is, if you want to get into this, make yourself useful. • So make a contribution to an open source project, for example. • Or make an implementation of some standard algorithm that you can't find the code of • online, but you'd like to make it available to other people. • So take a paper that you think is important, • and then re-implement the algorithm, and then put it open source package, • or contribute to one of those open source packages. • And if the stuff you write is interesting and useful, you'll get noticed. • Maybe you'll get a nice job at a company you really wanted a job at, • or maybe you'll get accepted in your favorite PhD program or things like this. 39
- 40. A business example! Digital Marketing! • A company has passenger ships from Tokyo to Hokkaido • The company also have a dataset of Internet usage data for it’s potential customers (Potential Tourists from Tokyo to Hokkaido) • Can this company deliver custom advertisements based on the users behavior and characteristics, so that the user is most likely to buy a ticket to Hokkaido from this company? 40
- 41. K-Means clustering to group similar customers 41
- 42. What are the things this group of customers are interested in? 42
- 43. My team’s solution for this cluster! • There will be a fortune telling corner on the ship! • There will be Kabuki shows on the journey! 43
- 44. Our Advertisement for this cluster! Omikuji Kabuki Art performance 44
- 45. 45
- 46. Thank you! 46