AI for Hospital Operations: Demand, Stock, Staff and Fraud

When hospital directors and health ministries talk about AI, the discussion usually turns to diagnosis. Yet much of a health system's budget, and many of its avoidable failures, sit in operations: medicines that run out in one district while expiring in another, wards short of nurses during a predictable winter surge, and payments for services that were never delivered. These are forecasting and anomaly-detection problems, and they suit machine learning well. They also carry less clinical risk than diagnostic AI, which makes them a sensible early investment for a national programme.

1. Forecasting medicine and consumable demand

Most public health systems still forecast medicine needs from last year's consumption plus an adjustment. That method misses seasonality, facility-level variation and the distortion caused by past stock-outs: a facility that ran out records low consumption, and is then sent less. Machine learning models can combine consumption history, stock levels, service volumes, seasonality and disease trends to produce facility-level forecasts.

Côte d'Ivoire: facility-level forecasts from routine data

One well-documented example comes from Côte d'Ivoire. Pendulum (formerly macro-eyes), working with the Ministry of Health, USAID and the Gates Foundation, trained models on data from the national DHIS2 health information system and the eSIGL and mSupply logistics systems. According to the project's write-up for the DHIS2 community, forecasts for 11 family-planning commodities across more than 2,500 health facilities were 38% more accurate than traditional methods, and tests on 36 malaria commodities reduced forecasting error by an average of 37%. Forecasts were delivered each month to support facility ordering.

The lesson for other countries is that the improvement came from data already in national systems. The prerequisite was integration, not new data collection.

What a good forecasting service looks like

• Horizon matched to the procurement cycle. A 30–90 day horizon fits most replenishment and tendering decisions.
• Forecasts with uncertainty ranges. Buyers need to know how far demand could vary, not only the central estimate.
• Correction for stock-outs. Recorded consumption during stock-outs understates real demand and must be adjusted.
• Human review. Pharmacists and procurement officers approve orders. The model recommends.

2. Monitoring stock across institutions

A forecast helps only if you know what is on the shelf. Many systems lack reliable, near-real-time visibility of inventory across hospitals, and that visibility is the foundation for any AI on top.

NHS Scan4Safety: visibility first

England's Scan4Safety programme put GS1 barcode scanning of patients, products and places into six demonstrator NHS trusts over two years (2016/17 to 2017/18). Every item used in a procedure, including implants, was recorded against the patient and location. The evaluation, reported in July 2020, found recurrent inventory savings of nearly £5 million and one-off inventory reductions of £9 million, and estimated 140,000 hours of clinical time released, largely because staff spent less time managing stock. Leeds Teaching Hospitals cut the time to trace recalled products from more than eight days to under 35 minutes.

Scan4Safety is not an AI project. It shows the data foundation operational AI needs. With item-level, location-aware stock data, models can do things that are otherwise impossible:

• Flag facilities heading for a stock-out days before it happens.
• Spot items nearing expiry in one hospital that another hospital needs, and suggest transfers.
• Detect unusual consumption that may point to waste, diversion or recording errors.

3. Forecasting patient demand and planning the workforce

Workforce is the largest cost in most health systems and the hardest constraint to relieve. The World Health Organization projects a global shortfall of about 11 million health workers by 2030, mostly in low- and lower-middle-income countries. High-income systems face the same pressure. NHS England's Long Term Workforce Plan, published on 30 June 2023, estimated that without action the NHS would face a shortfall of between 260,000 and 360,000 staff by 2036/37.

Short-term: predicting arrivals to staff shifts

At the operational end, Assistance Publique–Hôpitaux de Paris (AP-HP) piloted a machine learning model at four of its hospitals. It used ten years of admission records together with external data to predict emergency arrivals and admissions by day and hour, up to 15 days ahead. The stated aim was simple: give managers enough notice to bring in extra staff when high demand was expected.

Long-term: modelling supply and demand of professionals

Strategic workforce planning asks a different question: how many doctors, nurses, radiographers and pharmacists of each specialty will the system need in five or ten years, and how many will it have? Useful models combine:

• demographic projections and disease-burden trends on the demand side;
• age profiles, retirement patterns, emigration and training pipelines on the supply side;
• regional distribution, since national averages can hide severe rural shortages.

These models do not need deep learning. They need clean HR data linked to activity data, and scenario tools that let planners test the effect of more training places, retention measures or new care models.

4. Detecting fraud, waste and anomalies in health spending

Health spending attracts error, waste and fraud: billing for services not delivered, upcoding, duplicate claims, inflated procurement prices and phantom staff. Rule-based audits catch known patterns. Machine learning adds anomaly detection, flagging providers, claims or transactions that differ from their peers in ways rules did not anticipate.

United States: predictive analytics before payment

The best-documented national example is the US Centers for Medicare & Medicaid Services (CMS) Fraud Prevention System. Required by the Small Business Jobs Act of 2010, it has run predictive analytics against Medicare fee-for-service claims nationwide since 30 June 2011. CMS reported that in its second year of operation the system identified or prevented more than $210 million in improper payments, double the first year, and led to action against 938 providers and suppliers. Its distinguishing feature is timing: claims are analysed before payment, not recovered afterwards.

Design principles for anomaly detection

1. Score, do not accuse. A model output is a lead for investigators, never a finding of fraud.
2. Compare like with like. Peer groups must account for specialty, case mix and region, or legitimate high-volume providers will be flagged.
3. Close the loop. Investigation outcomes should feed back into the model so it learns which alerts matter.
4. Protect due process. Providers need a route to explain or contest findings, and models should be explainable enough to support that.

Common foundations

All four use cases share requirements that are worth planning for together:

Requirement	Why it matters
Unified product and facility master data	Forecasts and stock comparisons fail if the same item has different codes in different hospitals.
Near-real-time data feeds	A stock alert based on last month's data is history, not monitoring.
Secure, in-country processing	Claims, HR and prescribing data are personal data and often health data.
Clear decision ownership	Every model needs a named team that acts on its output.

Healthcare management AI in Moldova

Healthcare management is one of the five AI domains of Moldova's National Healthcare Digitalization Program (2025–2030). It covers forecasting purchases of medicines and consumables 30–90 days ahead, monitoring stock across institutions, forecasting workforce shortages, and detecting anomalies and potential fraud in accounting data, for a health system serving around 2.36 million people. Because these models use administrative, prescribing and personnel records, they follow the programme's rule that citizens' data is processed only inside Moldova.

Import Medical provides sovereign, on-premise, high-performance AI computing capacity in Moldova for these workloads, alongside our work in medical equipment import, which keeps us close to the realities of healthcare procurement and supply chains. Explore our healthcare management solution, and see how operational forecasting connects to disease prediction and our AI infrastructure. Related reading: predictive analytics for population health and Estonia's national eHealth system. To discuss a pilot for your institution, get in touch.

hospital operations demand forecasting supply chain workforce planning fraud detection

Sources

1. Leveraging DHIS2 data to create AI-driven health facility forecasts in Côte d'Ivoire — DHIS2 Community of Practice
2. Barcode scanning improves safety and efficiency — Healthcare Financial Management Association (HFMA)
3. Paris Hospitals Use Big Data To Predict Admissions — Healthcare IT Today
4. NHS Long Term Workforce Plan — NHS England
5. Health workforce — World Health Organization
6. CMS Fraud Prevention System identified or prevented $210 million in improper Medicare payments in its 2nd year — Centers for Medicare & Medicaid Services