Team: I’m in Sweden right now, they’ve got some exciting research going on that involves linking various national registries (including prescription, hospitalization, and cancer) with a new dataset that pulls out radiology images of tumor sites, that can then be used for predictive modeling via deep learning and other algorithms. The team at Karolinska Institute have already demonstrated successful ETL for most of the registers, but as a community, we don’t yet have a common solution for storing the imaging files and whatever associated records to link to them. Has anyone in the community worked on this problem, whether it be for oncology or for other areas? @Rijnbeek, does the work you’ve led in EKG imaging have some applicability here?
Thanks for sharing Patrick. I am very interested in the algorithmic pipeline they use for this including the deep learning part. As you know we have started with increasing our deep learning experience at EMC with a new postdoc and collaboration with NYU.
Regarding storing other type of data in the CDM, I think there is indeed an interesting area to work on. I know that at EMC a lot of work is done on predictive image analysis and combining this with other clinical information is been explored intensively in some ongoing initiatives (see http://www.bigr.nl/website/). There have been some exploration on Transmart etc, but there is no reason not to look at OMOP CDM (actually there are very good reasons to do look at it )
Regarding the ECG analysis work we are doing there might indeed also be some opportunities. We are doing predictive analytics using ECG parameters in combination with clinical predictors. These ECGs are recorded periodically from all 15,000 persons in a large cohort that is being followed in Rotterdam (started in 1991, see http://www.epib.nl/research/ergo.htm). Currently we use the automatically extracted ECG parameters (using a in house build computer program) for prediction. ERGO is not in the OMOP CDM but if it would be and we would have a way to store the derived variables or the ECG signals themselves this would open some big opportunities. Think for example about running a predictive study that incorporates ECG derived measurement together with others that have ECG data would be an unique opportunity. This would be awesome!
I also think that @rwpark and his group would love to have a way to add the ECG information for there predictive work. I know that they are storing ECG signals in a continuous manner from the ICU and also have paper scans from the Hospital. You could store the images (less preferred I think in the ECG case), the signals (would allow pattern recognition tools to run on top of OMOP CDM), or only te derived data.
@cfstrand, our posts crossed, but this should be useful discussion for you.
@Rijnbeek We are only at the planning stage right now. We work from the radiology perspective focusing on the images, but we will also plan to add relevant data from the registers. At the moment there is no analytical pipeline, we are considering how to do this without sharing the actual images (which might be an ethical/legal issue). There are ways to “bring the model to the data” instead of the other way round, but when it comes to deep learning that requires that you host a lot of storage and processing power for others to use. Any thoughts on this? It would be nice with a “CERN” for deep learning work (how about “ERCAPAI”, European Research Centre for Applied Artificial Intelligence") - with massive central processing powers for rent connected to a restricted access raw data storage facility for each data contributor. Then, any researcher with a good idea and model could have that model tested using the processing power, and only get the output data fed back from the system (not the actual images).
Also, what I really wonder if there are any OMOP CDM standards for how to deal with the more or less standardized DICOM tags.
As @Rijnbeek mentioned, we are storing all the continuous waveform signals from the 30 patient monitoring devices in ICU, which constitutes ECG lead 2, SpO2, Respiration, arterial pulse wave, central venous pressure and etCO2 up to 250 Hz sampling rate. By next year, we will collect the signals from 80 patient monitoring devices and ventilators in ICU. Basically the waveform signals are stored in CSV files with associated metadata of it. We are considering how to integrate the waveform signal data together with CDM.
For the radiology image, we are facing similar situations. We are now receiving research collaboration requests using CDM clinical data, radiology image and deep learning algorithms for prediction of disease risk or diagnosis. So we are also in the bottom line how to integrate the waveform signal data and/or radiology image into CDM.
@cfstrand We are applying the deep learning algorithm to the waveform signal and clinical data. For the computing power for deep learning while not sharing the waveform signal or radiology image, we prepared the same deep learning machine and software setting, which the research partner has. Thus if the research partner send their deep leaning training model to us, we can run the model against to our data and can send the resulting topology and parameters of the model.
Do you want to make a proposal to the CDM Working Group of how to model that? Is EKG and Imaging the same, or different?
Would love to be involved in this discussion. We at Google have also been working on deep learning on various types of linked data including EHR / telemetry / images. We are right now doing some standardization in an ad-hoc manner, but would be fantastic if OMOP has a standard on it. I’ve heard from lots of AMCs that they have similar issues that impede their progress!
+1 on this. At Georgia Tech we have folks interested in working with EEG data as well as the other modalities mentioned in this thread already. We’d strongly prefer to take a community approach.
We are Medical Convergence Research Center of Wonkwang University Hospital in Korea(prof. Kwon-Ha Yoon).
We has been working on building medical image big data, and proposing a new table for radiology based on OMOP CDM.
It is a table related to medical imaging. We named this as Radiology_CDM and documented the definition of the attributes of each table.
The attributes of Radiology_CDM are defined to be obtained from Tag information of DICOM which is medical image standard.
We prepared for Radiology _CDM ver.1.0 supported by Ajou university hospital research team(prof. Rae Woong Park)
We would like to make a working group and discuss it together.
I think integration of radiology data into CDM would be one of the the most important works in CDM. We’ll join this working group, definitely!
@mkwong, i remember you were interested and already working on mapping ECG data?
We would like to make a working group and discuss it together.
Suggested tables looks great. An optional reference to a radiology report in the NOTE table would be nice. Maybe add RADIOLOGY_NOTE_ID to RADIOLOGY_OCCURRENCE.
Also, yes, I know about FACT_RELATIONSHIP.
It’s very good point, @mgurley!
I think the note table should be revised to capture the information where it comes from
(current note table doesn’t store the pathology or radiology reports), as discussed here.
I made a suggestion and posted it to the github.
Please continue the EKG conversation in this thread.
Two things, @SiHyung_No: Could you be so kind and put this proposal into the OMOP CDM Github? We also need use cases. What kind of questions are we trying to answer, and how do all these fields support them.
One of use cases would be developing machine learning algorithm analyzing X-ray, CT or MRI using radiology extension model.
I started to take the Neurohacking course in Coursera on every Saturday.
I plan to convert the sample data of this course into radiology CDM and apply the analysis on CDM, which I learn in the course.
can you please clarify:
are we talking about CDM modification allowing us to put there images and reports
are we talking about the
I’m not sure that I understand correctly,
But what @SiHyung_No suggested is that storing path for the images and its informations, not image itself.
Image reports can be stored in the Note table after slight modification.
I’m sorry for the incomplete message posting, but you answered exactly what I wanted to know:)