OncoHealth manages oncology benefits, and everything valuable it knew was buried in unstructured clinical documents. We built the machine learning that reads those records at scale, then the live surface that put the answers in front of the teams who needed them.
the close rate on new sales, using the live intelligence tool
taken out of the sales turnaround
from the tool going live to those results
THE PROBLEM
The answers existed. Nothing could reach them.
Disease state, genetic markers, metastasis, the treatments tried and how each patient responded: all of it was already recorded, and none of it was reachable. Clinical documents are written to be filed, not queried.
So analysts read records by hand. Executives waited days for answers that already existed somewhere in the archive, and sales conversations stalled on a promise to follow up once somebody had gone and looked.
THE ENGINE
Machine learning that reads the record at scale
We built production pipelines that read clinical documents and extract structured patient intelligence: classification of the document itself, then extraction of disease state, genetic markers, metastasis, treatment, response and progression.
Precision was the constraint, not throughput. Every extracted finding stays linked to the document and passage it came from, so a claim can be checked rather than trusted, and the evaluation tooling around the models made failure analysis and clinician feedback part of the build instead of an afterthought.
THE SURFACE
Evidence in the room, not after it
An engine nobody can see does not change a business. Working with the sales team, we built a real-time monitoring surface over patient and programme data for the executives who had been waiting days on it.
The answers now arrive while the conversation is still happening. Deal teams walk in with live evidence instead of a follow-up promise, and the extraction work behind it finally shows up where the company earns.
THE RESULT
Two days out, and double the close rate
Two days came out of the sales turnaround, and the rate of deal closure doubled inside the first six months on the tool.
The same extraction engine underneath went on to serve clinical and operational teams, not just the commercial one.
Also delivered at OncoHealth
Delivered well beyond the project above.
- Clinical NLP pipelines classifying documents and extracting disease state, genetic markers, metastasis, treatment, response and progression
- Transformer and clinical-domain models for classification, named-entity extraction and longitudinal analysis
- Distributed data and enterprise search across Databricks, Spark, Elastic and Azure
- A clinical copilot for searching patient histories and inspecting the evidence behind each finding
- Early GPT-2 and GPT-3 capabilities for clinical summarisation and conversational assistance
- Evaluation and review tooling to compare model versions, analyse failures and fold in clinician feedback
- A full-stack application for oncology prior-authorisation decision evaluation
- Standards for traceability, privacy, security, observability and human oversight
- The production clinical AI architecture, presented at the Databricks Data + AI Summit
Discuss your next system.
Start with a free, one-hour engineering review. Leave with a practical next step.