The Challenge
The organisation’s prescribing guidance product relied on local configuration of large libraries of clinical messages. That created three practical problems:
-
Inconsistent profile performance across regions
Some regions had strong adoption and clear ROI, while others struggled to translate message coverage into meaningful clinician behaviour change.
-
Alert fatigue at the point of care
When clinicians are presented with too many messages in a single workflow, acceptance drops. The team wanted a way to reduce “noise” while still preserving safety and best-practice guidance.
-
Hidden co-trigger patterns
Messages can be similar, overlap, or daisy-chain. That can lead to multiple alerts for effectively the same clinical moment, which accelerates fatigue and reduces trust.
The Solution
We designed a pragmatic, staged solution on Azure Databricks focused on measurable clinical product outcomes.
1) A Databricks-backed recommendation engine for profile optimisation
We defined the profile optimisation problem as a recommendation task: “Which messages should be enabled or disabled for a given region, based on what works elsewhere and what that region is trying to optimise for (cost, safety, long-term conditions)?”
The approach intentionally starts simple and scales in sophistication:
- Rule-based baselines for transparency and quick wins
- Collaborative filtering and matrix-based approaches for data-driven recommendations
- A feedback loop to avoid repetitive recommendations and manage “content fatigue” over time
2) Co-trigger analysis using graph techniques
To explain why certain messages appear together (and where they cause unnecessary duplication), we designed a method to remodel message relationships as a property graph. This helps identify hotspots and high-risk co-trigger pathways, and it creates a clearer justification layer for recommendations.
3) Alert-fatigue reduction via acceptance classification
We defined a binary classification problem to estimate the likelihood of message acceptance, with the intent of suppressing low-probability messages in scenarios where too many alerts would otherwise surface. Where additional user-level signals are available, this approach becomes even more powerful over time.
4) Built for production on Azure Databricks
Alongside the modelling approach, we produced a delivery backlog and build plan that included standing up the core platform components needed for scaled ML, including a Databricks workspace, supporting data lake components, and a structured DevOps approach for repeatable delivery.
Tools & Technologies
This solution was designed for a modern Databricks stack:
- Azure Databricks for scalable model development and scoring
- A structured ML discovery and requirements approach using our Machine Learning Canvas
- A delivery backlog with the core platform items needed to move from discovery to MVP and beyond
The Results
The engagement delivered a clear, production-oriented blueprint for improving prescribing guidance outcomes:
- A practical recommendation approach to standardise and improve profile performance across regions.
- A defined method to reduce alert fatigue through message suppression and prioritisation based on predicted acceptance.
- A graph-based approach to explain and mitigate co-trigger and duplication effects that erode clinician trust.
- A delivery backlog that explicitly included provisioning and using Databricks as part of the implementation plan.
This work shows how “AI in healthcare” can be genuinely practical: not a shiny chatbot, but a measurable improvement to clinical workflows. When the right guidance appears at the right moment, clinicians spend less time swatting pop-ups and more time making decisions that matter.
