Apply HN: TemporalHealth Search and Aggreg. engine on millions of health records

11 points by aub3bhat ↗ HN
(To quote Linus Torvalds “Talk is cheap show me the code”) A working demo on 6 Million hospital visits: http://www.computationalhealthcare.com

Government agencies collect medical records on millions of patients for various purposes such as reimbursement, medical research, healthcare quality/delivery assessment. Even though these datasets are available to researchers and analysts, they are severely underutilized. Further due to security & privacy concerns, modern analytics tools cannot be directly used. We have developed a privacy preserving analytics platform. The platform uses aggregates statistics pre-computed in an exhaustive manner to enable exploratory analysis. By allowing researchers and physicians to search on these aggregate statistics using medical concepts (diagnosis, procedure & drugs), we can significantly reduce security & privacy concerns, while greatly enhancing benefits of these datasets. Given the enormous privacy concerns due to amount of information present in such datasets, we are deeply committed to transparency. We will Open Source core parts of platform for researchers. While we would prefer such system strong and useful enough for public use. In short term we aim for providing access to all practicing physicians and medical students in United States.

Consider a physician who discharged a patient with Sarcoidosis, within a week the patient went to an ED complaining severe headaches. The physician would be curious if any other patients have had suffered headaches following sarcoidosis and if she should admit him. Due to availability of 140 Million visits and 40 Million patients our system can provide useful guidance to the physician in spite Sarcoidosis being a rare disease.

We have the data. We have infrastructure. There is no real reason why it should not exist. Hence we have built it. We have a fully functional platform for research use. We have been working on this for 4 years and its my PhD thesis. Also applied to YC S16.

11 comments

[ 3.3 ms ] story [ 37.3 ms ] thread
Cool idea. I looked at your demo, but all I could get from it was a report of visit counts by diagnosis/symptom. Can you explain how the example question you offered (Sarcoidosis, headache, readmit) could be answered?

Is your goal to build a business or a non-profit?

Who pays for it? There are lots of potential users of this data, but some already have access to this data; do you open up the data to users who wouldn't have been able to access it before?

Hi, the Texas data which is publicly available to download, only has individual visits. While the HCUP data (california state datasets) has information for ED, Ambulatory Surgery and Inpatient visits linked to a unique patient identifier spanning over multiple years. Using this data, you can set up complex query types such readmission following a hospitalization with a DRG. We do not offer these results online. But you can see an example (fake data) of readmissions following sub-endocardial infarction (a heart attack) in which the patient underwent intubation in the next visit. http://www.computationalhealthcare.com/N2/HCUP/Entry/D41071_...

>> Is your goal to build a business or a non-profit?

We plan on building a business. We plan to operate a publicly available service for researchers and physicians that also acts as marketing vehicle. In addition to the central platform and we will allow individual enterprise users to run it on their premise with their own dataset.

>> Who pays for it?

Hospital, Insurers and Government agencies already have their own data in same format as we do. We plan to create a simple hosted drop in solution. By creating a single platform for all three types of customers and a public platform we can divide the development costs. Further we plan to offer add-ons specific to requirements of each customer type.

We do not "own" the raw data. The "raw" data is merely licensed to use for specific use cases by the government. We plan on using the data provided by the government to generate insights in form of aggregate statistics. These insights are then in turn provided to physicians and researchers.

Variations of this have been tried and didn't go well including one serious attempt by me. I agree this should exist but it is very much an uphill battle.

Email rpedela @ datalanche .com if you would like to discuss further.

Good luck!

Thanks I have actually observed several startups attempt this over years in one form or the other.

My hope is that in case we fail I still leave behind a good theory/framework, academic papers and hopefully some quality open source code from my research that allows the next guy to pick up. There are some good developments over last year as far as access to data is concerned. We also have some good medical discoveries using this platform that we expect to publish over the next year. I will follow up on email tomorrow.

This is great. We and few others including rpedela who commented below did something similar working with over 50M+ patient events from clinical data from our partners however found it incredibly hard to sell. Hope you find some success in this area. It is always interesting to see another healthcare entrepreneur. Feel free to ping me at rahurkar@gmail.com if you want to talk.
Thanks, I will email you. One particular feature of our platform is that the bulk of data collection & normalization work is already done by government agencies such as AHRQ HCUP and various state agencies. These databases have several users already. We are literally standing on shoulder of giants. Further while these agencies have been very good at developing databases. They suffer from lack of good tools to query them efficiently.

E.g. Here is the tool used by Texas, http://healthdata.dshs.texas.gov/

This is the one used by AHRQ HCUP http://hcupnet.ahrq.gov

New York state simply strips large number of fields and makes all individual visits information available via https://health.data.ny.gov/Health/Hospital-Inpatient-Dischar...

There is a strong need to create a single easy to use platform that allows patients/physicians/researchers to query them while optimizing for usability, privacy and security. We think that Computational Healthcare does that and is a step in the right direction.

Have any of the users of these databases and/or agencies committed to using/buying your tool? Have any signed a letter of intent?
I spent a reasonable amount of time pitching this space. I think we need more brilliant people working on hard problems like this.

That being said, I think you might be a bit off the mark and I'm going to share my story and the evidence before me. (Feel free to reach out to me, I'm happy to help anyone trying to tackle problems in this space).

A blog post I wrote several years ago about this kind of thing: www.engineersf.com/2015/07/04/do-we-need-a-human-data-projecthdp/

Our Stab At It My cofounder and I went around pitching over 100 MDs and disease researchers telling them we had the full medical records of tens of thousands of patients(hint: We didn't).

What They Told Us What they told us is that the data isn't all that useful for research purposes because it's inaccurate to some extent, but as well the ability to compare and juxtapose patient cases is just really tough. Ron Shigeta @rshigeta on twitter quickly told us the queries they would have to run would be several pages long even for research purposes.

How The Data Could Be Valuable to MDs and Researchers The REAL pervasive utility in a large number of records is to find the Gene Regulation Pathways, and as humanity we do that through clinical trials. The MDs and Researchers wanted to use it to recruit patients for trials. Fast forward to today and we've focused all our efforts on patient recruitment and screening.

There's another company that's also a YC company that went down a different route and focuses on diagnoses. It might be worth checking them out: humandx.org. (Stands for Human Diagnosis Project)

As well cancer is often highly mutated....ummm... http://blogs.sciencemag.org/pipeline/archives/2008/09/08/the...

The example you've outlined might be a bad one for the simple reality that cancer is so complicated. Cancer is a cell to cell battle.

Evidence 1: http://blogs.sciencemag.org/pipeline/archives/2011/04/07/mor... Evidence 2: http://blogs.sciencemag.org/pipeline/archives/2011/04/05/so_...

The Quote for me that really catches my attention in the pipeline blog: "Recent work from Bert Vogelstein’s group at Johns Hopkins (with a host of collaborators) and from the CGA itself now show that there are an average of 63 mutations in pancreatic cancer cells, and 47 in glioblastomas, two of the nastiest tumors around. The first impulse might be to think “Great! Plenty of drug targets to go around!” But hold on. For one thing, even though these mutations are surely not all equal, the fact that there are so many makes you wonder about whether attacking any one of them alone can make much of a difference. And different patients can have varying suites of those mutations, so it’s difficult to imagine that going after just one or two of those targets will be enough to treat a majority of cases. This work follows up on earlier studies in other tumor lines, all of which seem to point in the same direction: patients who are currently classed as having the same type of cancer really don’t. This won’t come as a surprise to most oncologists, who have seen for themselves the widely varying responses to current therapies. The challenge is to figure out what these various changes mean, and how to classify patients to give them the best therapy. It’s not going to be easy. Just doing the math on the possible interactions of several dozen mutations with a list of p...

Thank you for your comment. The HCUP data we use is being collected for more than 15 years. There have been at least 2000 peer reviewed papers published using the HCUP data in journals such as New England Journal of Medicine, JAMA , Annals, etc. My co-founder has deep experience with both medicine (as an anesthesiologist), Outcomes research and healthcare system. We have already published several papers using Computational Healthcare.

The example quoted is not too far off from a real scientific study that we are preaparing to publish soon. Also Sarcoidosis quoted in the example is an auto-immune disease and not a cancer.

The chemo / stem cell example is from a published article which investigated timing of chemotherapy prior to transplant.