An honest reaction to Andrew Ng’s AI for medicine specialization

Original article was published on Deep Learning on Medium


An honest reaction to Andrew Ng’s AI for medicine specialization

Sometime ago, the world’s most affable and recognizable AI leader, Andrew Ng launched a specialization called AI for medicine through his MOOC institution, deeplearning.ai. I have always been a big fan of Andrew Ng, and it was he who had introduced me to the world of machine learning through his grainy Youtube videos of Stanford lectures back in 2012.

I was very excited that finally, Andrew Ng has finally turned his attention to the critical shortage of AI experts in the medical field . Truth be told, AI in the medical world has not seen as much progress as other domains like personalized advertisements, recommendations, autonomous driving etc. There are lot of complex issues like data privacy, small sample sizes etc. which I would prefer to discuss in depth in another post.

Naturally, I was hooked to the news of the launch and tuned in to the online keynote event that featured key opinion leaders like Dr. Eric Topol, a renown cardiologist and the author of the bestseller Deep Medicine. At the same event, I was introduced to Pranav Rajpurkar, the person who was going to be the main instructor.

If you do not know Pranav, he is the author of the famous CheXNet paper, which to the best of my knowledge, introduced deep convolutional neural nets (CNNs) to medical imaging by diagnosing pneumonia from chest X-rays. I was very impressed that a person of such caliber would be leading the instruction.

Recently, I found the time to go through the lectures on Coursera and in this post, I provide a concise review of this specialization. I only provide my opinion on the overall content, but not on the coursework as I only audited the courses. I would also refrain from providing a rating as I do not believe in them.

What are my qualifications?

I am a fourth year PhD researcher in the field of clinical data science in respiratory medicine. My research involves application of statistics, traditional machine learning and deep learning to domains like biomarker discovery, diagnostics, medical data quality assurance and personalized intervention. I have been very fortunate in publishing at some of the top medical journals and conferences in my field. I also collaborate with a startup called Artiq to translate my research into clinical practice.

Overview of AI for medicine

This specialization consists of three courses in order, which are as follows:

  1. AI for medical diagnosis: Focuses on the application of CNNs for disease detection and segmentation on imaging data.
  2. AI for medical prognosis: Introduces models for disease risk prediction (prognosis) and models for survival analysis (to quantify probability of survival in the future).
  3. AI for medical treatment: Introduces causal inference to estimate treatment effect of drug in randomized clinical trials (RCT), methods to extract labels from clinical reports and applications of NLP, and various model interpretation techniques.

All the lectures are in the form of short videos of around 2–4 minutes.

What I liked

The specialization does an excellent job at familiarizing the students with different types of medical data and analytical techniques. It introduces common issues like handling imbalanced and missing data, the importance of internal vs external validation and different ways of quantifying model performance in a clinical setting. It also acquaints the students with state-of-the-art deep learning techniques to analyse medical imaging data.

I was very impressed with Pranav and his lectures were extremely thorough. By employing intuitive figures and examples in his presentations, he makes even the most nuanced topics like calculating C-statistics for individualized treatment effect easy to follow. Pranav easily maintains the high standards of instructing, that have been set by Andrew Ng.

What I did not like

I had a hard time grasping the overall structure of the specialization. It starts with advanced topics like deep learning in the first course, but focuses on common analytical methods like risk prediction, survival, data imputation, model interpretation etc. later in the specialization. Routine clinical data analysis involves structured or tabular data collected from the laboratory like blood samples, lung function etc. It requires classical statistical approaches or conventional machine learning. However, the current course structure gives an impression that deep learning on imaging data, which is relatively expensive to obtain, is the most important analytical technique in a clinical setting.

Secondly, it is not immediately clear whom this course is intended for. It requires a fairly high foundational knowledge of machine learning and deep learning, which immediately scares off a medical resident or a biologist from signing up. If this course is intended as a boot-camp for engineers to find jobs in the healthcare domain, the requirements for such jobs almost always include proven experience with real biological data. I am not sure if an introductory course to medical data analysis and projects on toy datasets cuts the line.

Final thoughts

With people like Pranav Rajpurkar and Andrew Ng at the helm, the instruction quality of this course is stellar. However, it still needs some changes in its content to fully realize its potential.

From my own interactions with colleagues in the medical world and in animal research, there is a big audience out there who is hungry to learn about novel data analysis methods like deep learning. However, they fail to find the right resources that covers the entire spectrum of data analysis, starting with classical statistics to deep learning.

With that in mind, I take the liberty of proposing an modified structure for this specialization:

  1. Basics of AI for medicine: Covers foundations of statistical inference and supervised learning on structured data.
  2. AI for medical prognosis and treatment: Covers risk prediction models, survival models and causal inference
  3. Deep learning in medicine: Covers CNN for imaging data, NLP for extracting data from clinical reports and recurrent networks for real-time monitoring data

Whatever the outcome maybe, I congratulate Pranav and Andrew Ng for their incredible effort in democratizing the knowledge of AI across the world.