My data science career has focused to date on the R&D side of healthcare — mostly cancer research projects. Typically, by the time a dataset arrives on my computer, patients have been reduced to database IDs and clinical information has been distilled down to just a few key variables (e.g. Survival: 1=yes, 0=no). While the medical data I receive is relatively simple, the challenge of my work lies in analyzing the large volume of omics datameasured from the patients’ tissue samples.
I’m usually tasked with finding patterns in omics data with respect to clinical variables like survival time and drug response. To achieve this, I write code to extract+transform+model the data, then write code to perform statistics and machine learning algorithms. After the code completes its work, I communicate the results of the analysis to oncologists and biologists, who use that information to find potential drug targets, patient segments, diagnostic biomarkers, and make biological discoveries. Here’s an example of an end result from that process.
In short, I help translate data patterns into the >em class="markup--em markup--p-em"> of human researchers, who use their expertise, intelligence and curiosity to drive discovery.
At least that’s how I used to operate, until Tag.bio automated the process with software so oncologists and biologists could do it all themselves — faster than sending me an email asking for help.
Healthcare outcomes, medical records and billing data
Starting at the end of 2017, my startup has been working with the top hospital in California for the pilot application of our platform in healthcare. There’s a rich, complex and intimidating world of data stored in patient medical and billing records —a wealth of information I never had access to in my omics data world. It’s really exiting.
All this high-value data is worth nothing without a strategy for using it. And from what I’ve learned so far, the state-of-the-art in healthcare analytics is a bit…depressing.
Would you like KPIs with that?
The standard approach for hospitals is to create dashboards showing known key performance indicators for generalized patient cohorts — KPIs such as length-of-stay, readmission, cost-of-care, and other outcomes. Patient cohorts are specified in the dashboard via a few general options such as date range, diagnosis, payor or admission type.
Coming from omics research, this seems like an egregious waste of high-value data — simply scratching the surface of what could be discovered. Where are the specialized software tools to let physicians explore the space of questions they have about highly-specific patient cohorts, care paths, outcomes and costs-of-care?
What are the unknown KPIs in healthcare data?
Get in line
There are a number of reasons why healthcare analytics to date has been so simplistic — as I have learned — and the biggest reasons seem to be organizational, not technical. As I often say to my co-founder — technically, anything is possible.
There’s a concrete wall between doctors/administrators and their IT/Analytics department. Physicians with time-critical data questions have to queue up for service from IT, and service time is measured in weeks (!). Alternatively, the hospital may pay for and wait for a consulting firm to do the analysis. Healthcare analytics consulting is a huge industry.
Consider the following real-world question from a physician:
“How do costs+revenues+outcomes compare for transfer patients with a specific diagnosis, compared to other patients with that same diagnosis?”
The standard process for answering this question is:
- See if the KPI dashboard can answer that question. No, it can’t.
- Send an email to Decision Support to set up a meeting to discuss a data pull, wait a few weeks to get the data, then work with a data analyst to decide how to slice the data properly and answer the question with statistics. w>em class="markup--em markup--li-em">ait…
Let me Google that for you
Think about it this way — what if you had to send an email and schedule a meeting with IT just to have them do a Google search for you? That’s the status quo in healthcare data analytics.
Tag.bio, in contrast
Asking that question above in our software platform takes a physician only a few seconds to do themselves — the statistics and descriptive language returned provide clear, reportable answers.
The impact on our industry-leading hospital partner has already been significant — data questions about highly-specific patient cohorts are now answered in the very same meeting when they arise.
— easy-to-use software, flexible and powerful enough to support complex queries and instantly return back a list of answers that physicians can understand and act upon.
A note about security and compliance
It’s astounding how easy it is today to develop a HIPAA-compliant cloud-based enterprise software platform. The ability to sign a BAA with Amazonfacilitates everything. It’s more difficult for a startup to document HIPAA policies and procedures (e.g. >em class="markup--em markup--p-em">“Who does code reviews when there’s only one coder?”), than it is to build a secure, scalable AWS architecture.
The blind rush to AI
On the distant other end of the data utilization spectrum, many healthcare organizations — albeit fewer than projected — are scrambling to adopt artificial intelligence solutions. AI in healthcare has a lot of promise, to be sure — especially when combined with rich, accurate omics, wearable-sensor, and imaging data.
From my omics career perspective on the space of healthcare data utilization, there’s an incredible opportunity being ignored — the human intelligence of the doctors, nurses, pharmacists and administrators that each hospital has on the front lines of patient care. They have questions, they have curiosity, they understand what the data fields in patient billing and medical records represent, and they each have decades of experience in the industry. And for the next few decades at least, doctors and nurses will be responsible for making critical decisions from individual patient care to high-level hospital policies, care paths, and payor agreements.
But they haven’t been able to access or query data like they need to — like they should be able to — at least, not until now.
“Tag.bio has helped us create an army of physician data scientists” — UCSF physician
I’m happy to discuss this more, if you have any questions, ideas, or feedback. Thank you for reading, and thanks to my team at Tag.bio and our collaborators at UCSF.
This article originally appeared in Towards Data Science