Clinical AI & EHR Data Platform
A US healthcare product needed to put decision support in front of clinicians, but the data underneath was spread across fragmented EHR and clinical databases that did not agree with each other. I rebuilt the data foundation and then the AI layer that sits on it, because in healthcare one is worthless without the other.
What was actually broken
Clinical records lived in mismatched schemas across multiple systems, so any feature that needed a complete patient picture had to stitch the data together at read time. That made the product slow, brittle, and impossible to reason about clinically. You cannot recommend a medicine safely on top of data you do not trust.
How it fits together
Source normalization
Refactored fragmented EHR and EMR databases into normalized, consistent schemas so a patient is one coherent record rather than fragments across systems.
ELT on ADF and Databricks
Azure Data Factory orchestrates extraction and load, with Databricks handling the transformation and modeling into curated clinical tables on ADLS.
Decision-support layer
A module that surfaces medicine recommendations and a clear, chronological patient-history view built on top of the curated data.
The parts that did the work
- A normalized clinical data model that gives every downstream feature one consistent source of patient truth.
- A decision-support module that turns history into medicine recommendations clinicians can actually read.
- Patient-history visualization that makes a fragmented record legible at a glance.
The hard parts, and how I got past them
Clinical source data was inconsistent and privacy sensitive, with the same concept modeled differently in each system.
A normalization pass mapped every source into a shared schema with validation, so the AI layer reasons over clean, consistent records instead of raw fragments.
Recommendations have to be explainable for the people relying on them.
I kept the recommendation surface tied to the visible patient history, so a clinician can see the record the suggestion is grounded in rather than a black box.
What changed for the business
What I took awayHealthcare AI lives or dies on the data underneath it. Most of the value here came from the unglamorous work of normalizing fragmented records first, which is exactly the pattern I keep seeing: the model is only as good as the platform it stands on.
More work
Let's build something that ships.
Hiring for a senior Data/AI role, or need a data platform that actually holds up in production? Let's talk.
or email me directly at muhammaduzairkhan329@gmail.com