Full medical records can be hundreds to thousands pages long

The full medical record is needed to complete a DRG clinical validation audit. But records can be hundreds of pages, we've even seen some over one thousand pages. This is part of the reason that it takes clinical auditors 30 minutes to hours to complete an audit, and the reason that AI tools can be so valuable to help speed up clinical audits. When we first started building Slingshot's AI models we were also faced with this challenge: how do we ingest and analyze all these words without taking forever to run or hitting hard limits on what our models could handle? 

Breaking Down the Problem

I think all big problems should be broken down into small pieces that you can easily handle. To start development, we needed a more condensed version of the medical record. While it does not give all the details, the discharge summary is a place where most clinical auditors start as its a summary of what happened in the stay and the densest note for information about the hospital encounter. Using the discharge summary gave us a representative, rich text with only a fraction of the words we need to analyze.

Using the discharge summary allowed us to have a representative, rich text with a limited number of words

While it varied from diagnosis to diagnosis, we ran an analysis and found that records for patient's with sepsis had the most representative discharge summaries of the full medical record. Our analysis found that if an auditor clinically validated the sepsis diagnosis based just off the discharge summary and compared that result to when they validated the sepsis diagnosis based on the full medical record, that these results would agree 96% of the time. This gave us enough confidence to start developing algorithms that we would apply to the discharge summary and later adapt to the full record. Narrowing our focus allowed us to quickly get results into the hands of our clinical auditors so we could get information and data for further training our models.

Implementing Parallelization: Speeding Up

Parallelization, i.e. running multiple algorithms simultaneously, allowed us to go from a 10 minute run time to 43 seconds.

After we had written algorithms for to analyze the discharge summary, it took 10 minutes for Slingshot to run clinical validation just over the discharge summary. We knew the next step had to be to bring this down, if we were ultimately going to modify these algorithms to analyze the full medical record.

Esse consequatur aut et.

Assumenda minima omnis ad. Necessitatibus aut et. Quo odio velit inventore reprehenderit illum eum inventore veritatis error. Deleniti ex vitae magnam nulla velit nemo quisquam.

Tackling the Full Record

Earum nobis doloribus nobis facilis et at dolorem pariatur. Aspernatur in consequatur iure iste itaque voluptate sed. Necessitatibus commodi sed aut ab temporibus sit fugiat blanditiis quia. Quasi velit vitae molestiae.

Excepturi ut nisi. Provident dolores molestiae voluptate enim numquam unde alias. Tempora vel deserunt praesentium amet. Ullam excepturi et consequuntur omnis nobis totam. Quia itaque sit. Quia sint possimus modi.

Cupiditate sint et. Officiis voluptas neque aut sit et vitae sit. Aut unde harum. Culpa qui autem aut aspernatur assumenda rerum voluptatem si