Redesigning Patient Support Workflows Around AI Segmentation and Adherence Forecasting
Patient support triage redesigned around two production models - segmentation to group patients by likely barriers, and adherence forecasting to flag drop-off risk - so care teams intervene earlier and route effort where it matters most.
The Challenge
Patient support teams could not tell in advance which specialty-medicine patients were most at risk of never starting therapy or dropping off
Adherence signals sat in fragmented clinical, dispensing, and program-enrolment systems with no common patient view
No forecasting layer existed - risk was discovered after a patient had already lapsed
PSP care-team capacity was spread evenly across cohorts, not weighted to highest-risk patients
Deliver in partnership with the Scale AI acceleration program
Approach
Redesigned the patient-triage workflow around two models feeding the same care-team queue
Built a segmentation model to cluster patients by likely adherence barriers (clinical, behavioural, access-related)
Built an adherence-forecasting model to score drop-off risk for each patient across the therapy timeline
Integrated clinical and program data into a unified feature store designed to generalize across demographics and specialty programs
Routed high-risk patients to earlier and higher-touch outreach; routed lower-risk cohorts to lighter self-serve pathways
Embedded model outputs into care-team tooling so scoring drives action, not a separate dashboard
What Was Delivered
Segmentation model clustering patients by clinical and behavioural attributes into actionable cohorts
Adherence-forecasting model producing per-patient drop-off risk scores across the therapy timeline
Feature engineering pipeline on Microsoft Fabric spanning ingestion, cohort assignment, and risk scoring
Power BI reporting layer surfacing cohort composition and drop-off risk to PSP leadership and care teams
Deployment pattern designed to extend across specialty programs without rebuilding the data layer
Business Impact
Bayshore’s AI initiative has introduced a new level of precision into its patient support programs.
By enabling data-driven triage and outreach, the organization is better positioned to intervene early with at-risk patients.
Segmentation and forecasting models running in production on Microsoft Fabric, extensible to additional Bayshore programs without rebuild.
The scalable model infrastructure has also improved the efficiency of PSP operations by focusing care team efforts where they are most needed.
These early outcomes establish a foundation for broader analytics adoption across Bayshore’s national care network.
Microsoft Fabric - Power BI - Segmentation model - Adherence forecasting model
Through AI, we're not just predicting patient needs; we're actively shaping a future where every individual receives the right care at the right time, enhancing the quality of life for all Canadians.
Frequently asked questions
- How does workflow redesign apply to patient support programs?
- Patient support programs run on care-team capacity - a finite resource spread across thousands of patients at different stages of therapy. Without a way to predict who will lapse, that capacity gets distributed evenly instead of where it changes outcomes. Redesigning the workflow means the models decide who to prioritize and when; the care team decides how to reach that patient and what to do next. AI changes the routing. People still own the judgment.
- What changes for care teams when patients are risk-scored before drop-off?
- The intervention shifts from reactive to pre-emptive. Care coordinators stop discovering a lapse from a missed refill or a return call - they see a rising risk score days or weeks earlier and act on it. That reallocates the same team's hours toward the patients most likely to drop off, which is why the model outputs sit inside care-team tooling, not in a separate analytics dashboard. A lift in early-intervention rate and reduction in time to therapy are the operational signals that the shift is holding.
- Why two models instead of one?
- Segmentation and adherence forecasting answer different questions. Segmentation asks what kind of patient this is and which barriers are most likely to matter for them. Forecasting asks whether this specific patient is about to drop off the therapy timeline. Running them together lets Bayshore tailor the outreach approach and prioritize which patient to reach first as two separate decisions - a single blended model would collapse them and make both harder to act on.
Related industry
AI for Senior Care and HealthcareNext step
Ready to prove it in your workflows?
Book an AI Jumpstart. Identify the workflow. Establish the baseline. Prove the value in 5-7 weeks.