Earlier this week, I attended the first Penn conference on big data in population health sciences. This was my first conference where I got to attend all the talks, and it was gratifying. The organizers did a wonderful job on selecting a great breadth of topics to cover, from electronic databases and biobanks to digital and mobile health. I learned so much! Some of these topics, according to my PI, were right up my alley, Maybe that’s why I really enjoyed these talks? and I will go a little deeper in some of them in later posts. But for now, I thought I’d share some of my key takeaways from this conference.
Big data can be useful.
I remember “big data” as a buzzword when I was an undergrad. Dan Ariely tweeted,
Big data is like teenage sex: everyone talks about it, nobody really knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it.
That was back in 2012. Things have changed a bit since then, and I believe everyone in the room understands that this buzzword is simply a shorthand for the complex data or databases we all are working with. In addition to a good chunk of heart/brain imaging, whole exome/genome sequencing, gene expression and Electronic Medical Records (EMR) data, it was neat to see analyses of various unfamiliar types of data, including wearable device, social network and geographical. One striking result came from Mitesh Patel’s work on leveraging the EMR to identify patients eligible for cardiac rehabilitation and alert their physicians: patient attendance goes from 5% up to 40% simply by changing the referral process from opt-in to opt-out. Opt-out increases cardiac rehab referral rates As another example, Casey Green showed that models trained on big data compendia can reveal patterns associated with rare diseases in smaller datasets. Even of different types, e.g., microarray vs RNASeq
Useful objectives are important.
Just like in every evolutionary algorithm design (or any other optimization technique for that matter), defining a good objective is crucial for the discovery of an optimal solution. Many of the presented works go beyond a binary diagnosis classification problem, cleverly reframe the questions and implement very innovative ideas to solve them. Taki Shinohara transferred batch effect correction tools in gene expression analysis to harmonize multi-site imaging data. Harmonization of multi-site DTI data Lyle Ungar linked social media with EMR to reveal individual’s clinically relevant information from their Facebook posts. Predicting medical conditions from social media posts Facebook language predicts depression in medical records Daniel Rueckert applied reverse classification accuracy for automated quality control of image segmentation. Automated quality control in image segmentation Unsupervised machine learning was mentioned at different points to reduce bias and enables disease subphenotyping. The importance of refining objectives in predictive analytics were further emphasized by Amol Navathe and Ziad Obermeyer via the selective labels problem.
We still have a ways to go.
While methods like David Madigan’s P value calibration, Jinbo Chen’s anchor variable framework and Ian Barnett’s incorporation of random effects in neural nets have been proposed to address several technical challenges, I think a lot more are needed. Big data are inherently messy and require a lot of domain knowledge to be cleaned, transformed and analyzed. For wearables with fine time resolution in multiple facets, I would not even know where to start looking. Further, these challenges are not limited to techniques. Rosalind Picard reminded us to take caution in interpreting individual-level data, especially before translating it to clinically meaningful actions. Marylyn Ritchie and Daniel Rueckert noted the heterogeneity and disparity in both imaging and genetic studies, and this issue will remain until more studies of diverse ancestries are carried out. Li Shen reiterated that no drug has yet cured Alzheimer’s disease or even slowed its progression, and the same go for many other diseases. And what about privacy? Do our works get us any closer to the ultimate goal of improving health? Is our objective function right?
Conclusion
A big thank you to all of the speakers for the delightful talks and enlightening conversations! And once again, kudos to the organizers for a successful conference!
P.S. During a break, I spotted a spotted lantern fly on the right arm of an audience member sitting right in front of me. Now, if you haven’t heard, these flies are harmful invasive species that should be killed. So, like a true West Philadelphian, I asked the man:
- Sir, may I swat your arm?
- I’m sorry, what?
- There’s a bug on your right arm. May I swat it?
- Huh? Oh! Uhhh… okay.
I smacked his sleeve with my rolled up papers, and the fly fell on the floor. I apologized and tried to explain my bizarre approach just to learn that he was the last speaker of the day, Nilanjan Chatterjee 🤦♀.