We built an ML forecasting tool for a hospital pharmacy network that uses patient volumes and procedure schedules to stock costly drugs with a short shelf life. Waste fell 18%, and the fill rate rose 22%.
Client: Frontier RadiologyIndustry: HealthcareService: AI & Automation
Overview
Client
Frontier Radiology
Multi-specialty hospital pharmacy network
Engagement snapshot
The challenge
Stock levels were set by older averaging models that ignored seasonal changes in patient numbers and local disease outbreaks. Fixed min/max reorder points meant high-cost biologics were overstocked again and again, and $12M of drugs expired and were thrown away each year.
Excess stock was only spotted once drugs were within 30 days of expiry, and demand data reached analysts as CSV exports that sometimes still carried patient identifiers.
Fixed ordering left large volumes of specialty drugs to expire, eating directly into the health system's margin.
Manual ordering added a 72-hour delay between a spike in patient numbers and the replenishment order.
The older demand analysis risked HIPAA violations by exposing patient-level treatment patterns to third-party vendors.
The solution
We built a secure data flow that pulls patient encounter trends and procedure schedules straight from the EHRs, using HL7 health-record standards.
A probabilistic (Bayesian) ML engine ranks each drug's stock risk by shelf life, how fast it is used locally and how unstable regional supply is.
Automated agents raise purchase orders within pre-set financial limits and send any exception to a pharmacy director.
Key features
A Bayesian model forecasts demand for costly, short-life drugs from patient volumes and procedure schedules.
Business impact
18% less drug waste
Ordering agents raise purchase orders within set limits, and surplus stock moves between sites before expiry.
Business impact
22% higher fill rate
Custom scoring weighs how quickly a drug expires against how critical it is clinically, so life-saving drugs do not run out.
Business impact
99.2% fill rate
A de-identification layer means the forecasting models see only anonymised clinical trends, never patient identities.
Business impact
100% audit readiness
The platform spots surplus at oncology centres and redirects it to community clinics before it expires.
Business impact
18% less waste
Before and after
Before
Min/max levels
Fixed triggers ignored upcoming procedures and expiry dates.
After
Based on real demand
AI triggers orders from local usage speed and clinical schedules.
Before
Manual exports
CSV exports for analysis often still contained patient identifiers.
After
De-identified stream
Data is de-identified as it is collected, which keeps analysis within HIPAA rules.
Before
Spotted too late
Excess stock was found only once drugs were within 30 days of expiry.
After
Moved before it expires
ML flags drugs at risk of expiring early and moves stock between facilities.
Results
18% Waste Reduction
Matching stock to procedure volumes broke the old cycle of over-ordering and then throwing out expired drugs.
Before: High DecayAfter: 18% Lower
22% Fill Rate Lift
Putting critical procedures first in the reorder queue improved fill rates for life-saving medicines.
Before: FragileAfter: Resilient
100% HIPAA Audit Trace
Every procurement decision is logged and linked back to the procedure-demand forecast behind it.
Before: VagueAfter: Verified
Testimonial
“With Coretus we now forecast demand for costly, short-life drugs from patient volumes and procedure schedules. Waste is down 18%, fill rate is up 22%, and surplus moves between sites before it expires. All patient data stays de-identified.”
Dr. Antonio RossiDirector of Pharmacy, Frontier RadiologyKey results
Trust and control
Every workflow is built to HIPAA requirements, runs on cloud infrastructure covered by a BAA and strips out personal data it does not need.
HIPAAThe forecasting agents give a reason code for each replenishment order, based on shelf life and volume.
TraceableThe cloud-based workflow keeps ordering reliable when demand spikes.
Scales with demandThe client owns 100% of the IP in the inventory models and replenishment agent code.
100% ownedFaster delivery
These tested components saved about 10 weeks, which went into the forecasting models and moving stock between sites.
Gave the project a tested, HIPAA-aligned pipeline for pulling de-identified data from the hospitals' EHRs.
Tracked stock levels and expiry dates for every drug at every facility.
Followed regional supply changes, which fed the forecasting engine's view of supply risk.
Enforced the financial limits on automatic orders and routed exceptions to pharmacy directors.
Its ordering relied on fixed min/max levels and older averaging models that ignored seasonal patient numbers and local outbreaks. High-cost biologics were routinely overstocked and expired, wasting $12M a year. Manual ordering also added a 72-hour delay between a spike in patients and the replenishment order.
It streams de-identified encounter trends and procedure schedules from the EHRs. A Bayesian forecasting model ranks each drug's risk by shelf life, local usage speed and regional supply volatility. Agents raise purchase orders within set financial limits, send exceptions to pharmacy directors and move surplus between facilities before it expires.
A de-identification layer removes patient identities before data reaches the forecasting models, so they only see anonymised clinical trends. This replaced CSV exports that sometimes carried identifiers. Every procurement decision is logged with a reason code and linked to its forecast, which gives a complete HIPAA audit trail.
If fixed reorder levels leave your pharmacy overstocked on expensive drugs, we can help you forecast demand from patient volumes without exposing patient data.