Blog · Procurement and supply chain · Healthcare
Predictive analytics in healthcare supply chain needs validation against actual medical supply usage and inventory exposure. Compare forecast errors by item and facility, then examine shortages and expiration without hiding local exceptions.
Predictive analytics in healthcare supply chain is useful when forecasts are tested against actual supply usage, availability, and expiration exposure at each facility. Save the prediction before the period begins, match it to the later operational records, and measure the miss. A balanced network total does not establish that local forecasts were reliable.
Define exactly what is predicted: weekly packs issued, requests submitted, or units consumed. Those events are related but not interchangeable. A request can remain unfilled, and an issue to a supply room may precede consumption.
Keep the item, facility, forecast creation date, forecast period, quantity, unit, and forecast version. Preserve the version actually available when the business decision was made. Replacing it with an updated estimate after usage is known would make the test misleading.
Use consistent intervals. NIST's introduction to univariate time series describes sequential observations at equal time increments. A weekly supply series should therefore keep its weekly calendar intact; an absent export needs investigation rather than being silently treated as no activity.
The broader predictive supply chain analytics guide covers prediction across operations. Here the specific question is whether supplied healthcare forecasts remain useful once local usage constraints and expiration are visible.
Use the saved forecast export and a finalized usage register for the same periods. Aggregate each to item, facility, and week before matching. Multiple usage transactions can belong to one forecast row; joining raw transactions directly and then summing the repeated forecast would overstate it.
Carry a status for unmatched rows. A new item with actual usage and no saved forecast is a coverage exception. A forecast with no actual row may reflect a missing file, a discontinued item, or no usage; identify which before calculating an error.
For the current inventory and supplier review, use healthcare supply chain analytics. Forecast validation builds on those reconciled operational records; it should not redefine inventory movements to make predictions look better.
A fictional US operator forecasts one supply pack for East and West facilities. The actuals are finalized weekly usage in packs. These are illustrative business records, with no patient information.
Signed error = Forecast − Actual usage
Weighted absolute percentage error, or WAPE = Sum of absolute errors / Sum of actual usage
Positive signed error means an overestimate under this convention. State that convention because some teams use the opposite sign.
| Facility | Week ending | Forecast packs | Actual packs | Signed error | Absolute error |
|---|---|---|---|---|---|
| East | October 9, 2026 | 1,000 | 900 | 100 | 100 |
| East | October 16, 2026 | 1,000 | 1,100 | -100 | 100 |
| West | October 9, 2026 | 600 | 500 | 100 | 100 |
| West | October 16, 2026 | 600 | 700 | -100 | 100 |
| Total | 3,200 | 3,200 | 0 | 400 |
The summed signed error is 100 − 100 + 100 − 100 = 0 packs. Aggregate bias, defined here as summed signed error divided by actual usage, is 0 / 3,200 = 0.0%. Nevertheless, absolute error totals 400 packs, so WAPE is 400 / 3,200 = 12.5%.
Reconcile the forecast total to the saved forecast file and the actual total to the usage register. Then confirm that absolute errors were calculated before aggregation. Taking the absolute value of the total signed error would incorrectly return zero.
To compare another saved forecasting approach on these same rows, suppose its forecasts were 950, 1,050, 550, and 650 packs, respectively. Its absolute errors are 50, 50, 50, and 50, totaling 200. WAPE becomes 200 / 3,200 = 6.25%, displayed as 6.3% when rounded to one decimal place.
That comparison favors the second approach on this small test. It is not enough evidence to approve it generally: repeat the comparison across completed periods, items, and facilities before changing the review process.
Forecast error measures prediction performance. Expiration exposure asks a different question: how much existing stock may remain unused before its expiration? Keep the two measures alongside each other.
At a separate October 16 snapshot, the operator reviews distinct item and facility combinations below. Each row represents the earliest-expiring lot for that combination. Expected demand before its expiration is compared with the lot quantity, with no competing lots or incoming receipts assumed.
| Facility and item | Lot packs | Expected demand before expiration (packs) | Exposed packs | Cost per pack (USD) | Exposure (USD) |
|---|---|---|---|---|---|
| East, pack A | 180 | 120 | 60 | 5 | 300 |
| East, pack B | 120 | 100 | 20 | 5 | 100 |
| West, pack A | 140 | 80 | 60 | 5 | 300 |
| West, pack B | 100 | 120 | 0 | 5 | 0 |
| Total | 540 | 140 | 700 |
Exposed packs are the greater of zero or lot packs minus expected demand. West pack B cannot have negative exposure: MAX(0, 100 − 120) = 0. Total exposure is (60 + 20 + 60) × 5 = $700. This is a planning estimate, with the demand assumption and inventory status held constant.
The table does not establish that a loss will occur. Facility operations must confirm restrictions, actual movements, and feasible actions; finance determines how recorded losses are recognized. Predictions do not authorize clinical substitutions or purchasing changes.
If supplies were unavailable, recorded usage may understate what the facility needed. Put unfilled requests and unavailable periods beside the error review. Otherwise, a forecast that matches constrained usage can receive an undeserved favorable assessment.
External context also has limits. The FDA medical device shortage list covers device categories FDA has determined to be in shortage. It cannot replace the local request, stock, and receipt records needed to explain an individual facility's usage.
Procurement owns supplier discussions, facility supply operations owns local exceptions, and the planning owner maintains forecast versions and evaluation periods. Review broader continuity issues through supply chain risk analytics.
Bring a sample of your own saved forecast, finalized usage, and lot inventory files to Covirage, which maps the columns and uses its tools to compute and check the comparisons. The external AI model explains the results with every figure traced to its rows. Create a free Covirage account, which takes about a minute, then set up your business and bring your export.
See forecast analysis for the supporting approach to reviewing predictions against business outcomes.
Compare the forecast saved before the period with actual usage for the same item, facility, and interval. Calculate signed error to identify persistent overestimation or underestimation, and absolute error to measure the size of misses. Reconcile the totals before comparing forecasting approaches.
Yes. Overestimates at one facility or in one week can offset underestimates elsewhere. Total forecast and total usage may match even when local errors are substantial. Review absolute errors and signed errors at the level where staff make replenishment decisions.
No. Exposure estimates stock that may remain unused before its expiration, given stated usage assumptions and restrictions. Actual losses depend on later movements and authorized actions. Keep the forecast exposure separate from recorded expired inventory and finance's booked adjustments.