Subsalt

Quality wants 30-day readmissions after TEAM episodes. Marketing wants to know if last quarter’s campaign moved new-patient volume. Neither question needs a patient’s name, but both teams make data requests that will sit for months before they get answers because of HIPAA privacy laws. Subsalt lets both teams ask and get answers instantly, without touching PHI.

Healthcare employees reach for shadow AI when there’s no approved option -- Subsalt gives your whole health system an approved option in less than half the time it takes to respond to a data request today. Yep, we've created the solution to committee purgatory. General-purpose AI chat tools connect to a de-identified twin of your patient data, and PHI stays inside your security boundary. So the people who run your health system can ask questions that need patient data without joining the data analyst queue or submitting a BI build request.

Finding the balance between taking care of patient data and moving quickly is really important. While ethical and privacy reviews are being done, cures and interventions are delayed. Worse, because there is so much anticipated waiting around, many projects never even start. Not everyone has the luxury of time. We have got to find more ways to move fast without breaking things. That is what we're doing at Subsalt. https://lnkd.in/dChYzjWn

A team I worked with was building an ML model on a combination of de-identified datasets. But they'd lost before they even started. They started with 600 columns. Before anyone could figure out which ones actually mattered for the predictions, 40% of them were dropped in a second de-identification pass. Combine two de-identified datasets and you almost always trigger more redactions. The part that struck me was that they didn't mourn (for lack of better word) the loss of those features. Sure, if you asked them if they'd rather have those columns back and... of course they'd say yes. But there was nothing they could do about it so they just accepted that 40% of the analytical value in that dataset was vaporized. It's the classic: ask a fish "how's the water?" and it'll say "what's water?". So common and ubiquitous and expected that its nearly imperceptible. What is actually sitting in that 40%? A signal nobody will ever find…a better prediction? For a cure? A new intervention? We'll never know! The model came out fine, fwiw. Building a good model isn't and won't ever be the hard part. It's the never-to-be-seen upside that should haunt your dreams (we all dream about redaction-induced data loss, right?). Thankfully, this is changing in no small part due to advances in AI. There are now ways to keep the privacy compliance and keep the value - that's what we're building at Subsalt. So. Keep the columns. All six hundred of them. Sit with the long tail a while - they have the best stories. Order the granular codes. Order the dates, too. Let the data tell you where it's been. Because the good stuff - the marrow, the AUC, F1 score - lives in the unredacted version of everything. Has anyone at your org ever estimated what the cost of what de-identification deletes? Opportunity cost is too infrequently considered imo. #HealthcareAI #ResponsibleAI #HealthData #MachineLearning #DataStrategy

Health systems are being asked to move AI from pilots to enterprise impact using data environments, approval processes, and governance models designed for a different era. Join leaders from CHAI, Northwestern Medicine, and Subsalt for a practical discussion on what must change to make sensitive health data safely usable for AI, analytics, research, and innovation. September 9 | 2:00 PM ET | Virtual Live discussion + Q&A

AI creates an urgent new question for health systems: who can use sensitive data, for what purpose, and under what controls? Existing permission models were built for people, applications, and conventional reporting. Rapidly expanding AI, analytics, research, and data-sharing use cases make it nearly impossible for human-paced governance committees to keep up, keeping innovation on a lengthy timeline. There's a better way. Join leaders from CHAI, Northwestern Medicine, and Subsalt for a live discussion on rethinking health-data access and permissions for enterprise AI—without compromising privacy, trust, or control.