AI Trends Interview: David Ledbetter, Children’s Hospital LA: Extracting from Data to Guide Patient Care


David Ledbetter is a Senior Data Scientist at Children’s Hospital Los Angeles Pediatric Intensive Care Unit (CHLA PICU). He is an expert on decision theory, with particular emphasis on accurately predicting performance on new data in real-world scenarios. Recently he has been analyzing electronic medical records from 15 years of data collected from the CHLA PICU and developing Recurrent Neural Networks to personalize treatment recommendations to optimize patient outcomes.

David chairs AI World’s AI in Healthcare Summit on Monday, December 11. He spoke with AI Trends Editor John Desmond recently to update readers on his work.

Q. How would you summarize how AI is being incorporated into your work?

Our primary goal is to take all data from the pediatric ICU, one of the most instrumented units in the hospital and with the highest fidelity medical data available, and use AI to extract information from that data to help patients coming through the doors.

We use deep learning techniques such as Recurrent Neural Networks. We take time series data and understand the trends and correlations taking place, so when a question arises about treatment A or B, we can make an informed suggestion for each scenario.

Q. How long have you been doing this?

It’s been two-and-a-half years for me personally, and our group, the Virtual Pediatric Intensive Care Unit, has been funded by the Laura P. and Leland K. Whittier Foundation since 2012. That was when the field was becoming aware of how critically important data is to taking advantage of more advanced techniques.

Q. Are these data science practices widespread in healthcare now?

We are seeing extreme interest in it right now, for very good reasons. Healthcare moves methodically, because we have regulations and need to ensure HIPPA compliances to make sure we walk before we can run. When you get an answer wrong in this field, the repercussions can be dire. So we need clinical trials, followed by more clinical trials and next FDA approvals. We are seeing a lot of interest in the field right now, but not widespread adoption yet of these more advanced techniques.

Q. Can we get an update on the overall Recurrent Neural Network project?

Because in the ICU, we have some of the highest fidelity data in the hospital, we are on the leading edge of people doing research in the field. One of our initial projects was to take all the data — physiological readings from patients, lab results, drugs administered to the patients — to understand the interactions to generate a severity of illness. In this way, we can tell the doctors who is the most sick, and who they need to take a second look at. Also, it provides us with an early warning system, in order to provide extra scrutiny.

Q. How’s it working out?

We have many checks and balances. In terms of a retrospective analysis, we are able to generate state-of-the-art mortality predictions. Given an observation window for a patient, we are better able to determine who is going to survive. Now that we have this information, we are starting to generate plans for clinical trials to capture and quantify our ability to provide guidance to the clinical teams.

We always follow two paths: one is pure research, when we work on new and exciting algorithms, and the other is the clinical path, where we work closely with the clinical teams on deployment to the unit.

Q. What are the organizational challenges you face?

By and large, within the field itself, it’s probably Information Technology. Because of how much protection is around your personal health records, for good reason, it makes it difficult. It’s not a negative; it’s just the way it is. We always try to add value.

Q. How do you handle data security?

That is the domain of institutional IT. We have an entire group dedicated to the data engineering side, security, integrity, and the machinations of pulling data from all the databases, organizing it and getting it ready for the data scientists to extract value.

Q. Can you find the help you need from a workforce point of view?

That is one of the things I’ve been most pleasantly surprised about. We recently added two new members to the team, after we advertised for capable data scientists. The response we got was phenomenal. For example, new graduates of the University of San Francisco, University of Southern California, or programs like were really well-matched. Their curriculum is really matching the needs of industry right now. I was ver impressed with the quality of the candidates. The fact we were only able to hire two was a great disappointment.

Q. What would you say are the top trends in the data science profession right now?

Keeping in mind what happens to the patient on the path of treatment. There is a lot of great work using Convolutional Neural Networks applied to image data: in radiology, mammography, detecting circulating tumor cells. And also for medical records and health records to detect patterns using Recurrent Neural Nets to provide the ability to integrate the time series.

Q, What suggestions would you have for students or professionals interested in learning more about data science and AI?

I would point to, run by Jeremy Howard and Rachel Thomas. They provide online courses free of charge.

To train up someone new in the role as a data scientist, I myself, might have a different take on the strengths needed to get someone to full capacity to become a contributing member of the team. I am looking for math, computer science, physics, chemistry and biology. Those strong fundamentals are so critical to every aspect of what we do. The specifics of coding in Python or constructing deep learning models with Keras, learning [Python] pandas to munge the data — those are easier to train. But the fundamentals are the foundation on which everything rests.

Q. What was your education background?

I had a degree in math. I worked for a company that did a lot of digital signal processing, using fundamentals of detection theory. Then I did machine learning work, then deep learning work, and lot of different detection analysis – on radar, sonar, optical data. Once you are thinking about it abstractly, pulling signals out of data, the transition to the medical field is not that extreme. When we get readings for patients in the ICU, we look for signals showing why they are sick and how they are going to get better. So many of the skills are the same, and being at Children’s Hospital Los Angeles, we get extremely high fidelity data.

We have a fellowship program in the pediatric ICU, a post-doctorate program for doctors, where 50% of their time is dedicated to research. We have a close collaboration with the doctors, to leverage their medical expertise and fold it into our data scientist expertise. We are both together trying to look at the same problem to come up together with the best overall solution.

Q. Do you have anything to add or elaborate on?

One thing to emphasize is the importance of domain. Many times, people in AI moving into deep learning think they will only rely on data; that’s the AI mindset. My experience is how critically important it is to have access to doctors to help us interpret the data. We couldn’t do that as well without the level of experience they have, and I think that is true of many fields.

We have people in droves doing work on self-driving cars, for instance. The best way to make them better is to start where they are. If you ignore the stakeholders, like the truck drivers who know a thing or two about driving a truck, you do so at your own peril. We tend to focus on what we think is going to happen most frequently, but it’s the edge cases that will get you. In real life applications, those gotchas are potentially peoples’ lives.

This content is original to AI Trends.