Definition
Percentage of AI-predicted discharges that actually occur within the predicted time window, measuring the reliability of discharge predictions used to trigger bed assignments and patient flow decisions.
The Measurement Gap
The Decision That Legacy Metrics Cannot Support
When the model says Patient A will discharge by 2 PM, is that reliable enough to commit the bed to Patient B who is boarding in the ED?
What Goes Unmeasured Today
Predictive patient flow tools promise to solve one of healthcare's most persistent operational problems: the discharge timing uncertainty that creates cascading delays from admission through transfer through ED boarding. These tools predict which patients will discharge within a given window, enabling bed management teams to pre-assign beds and initiate admission processes before the discharge physically occurs. The problem is that most organizations track prediction accuracy in aggregate (the model was 78% accurate this month) without measuring the operational consequence of the predictions that failed.
A failed discharge prediction (the model said discharge by 2 PM, but the patient was still in the bed at 6 PM) triggers a cascade of wasted effort: the bed was pre-assigned to an ED boarding patient, the admission paperwork was started, transport was arranged, and then everything had to be unwound. Each phantom discharge consumes bed management time, disrupts ED flow expectations, and erodes trust in the prediction system. Legacy flow metrics (ED boarding time, bed turnaround time, length of stay) capture the downstream symptoms but cannot attribute them to prediction failures.
The measurement gap is particularly acute because discharge timing depends on factors the model may not fully capture: social work disposition, insurance authorization, family transport, medication reconciliation delays. A model that accurately predicts medical readiness for discharge but fails to account for disposition barriers will show good clinical accuracy and poor operational reliability.
The Invisible Failure Mode
Phantom discharges (predictions that trigger flow actions that never materialize) are individually minor but cumulatively corrosive. Each one teaches the bed management team that the predictions are unreliable, driving them back to phone-call-based confirmation with charge nurses. The prediction system becomes a dashboard no one trusts, consuming IT investment without delivering flow improvement.
Formula
Primary Formula
DPRI = Discharged_within_window / Total_discharge_predictions
Companion Metric: False Commitment Rate:
FCR_bed = Bed_assignments_made_on_prediction × (1 − DPRI)
Variable Definitions
| Variable | Definition | Unit |
|---|---|---|
| Discharged_within_window | Patients predicted to discharge within window W who actually discharged within W | Count |
| Total_discharge_predictions | All patients the model flagged as likely-to-discharge within W | Count |
| W | Prediction window (e.g., within 4 hours, by end of shift, today) | Hours |
| FCR_bed | Expected number of flow actions taken on predictions that will not materialize | Count |
| Bed_assignments_made_on_prediction | Flow actions (bed pre-assignments, transport orders) triggered by discharge predictions | Count |
Data Requirements
| Data Element | Source System | Classification |
|---|---|---|
| Discharge predictions (patient, predicted time, confidence) | Predictive flow platform / EHR ML model | NATIVE |
| Actual discharge timestamps | EHR ADT feed | NATIVE |
| Bed assignment decisions triggered by predictions | Bed management / patient flow system | NATIVE (if logged) / NOT CAPTURED (if verbal) |
| Discharge delay reasons | EHR discharge disposition, care coordination notes | NATIVE |
Edge Cases and Exclusions
Medical hold discharges (patients medically ready but delayed by disposition (SNF placement, insurance authorization, family transport)) should be tracked separately. The prediction was clinically accurate but operationally wrong. If the model provides probability scores, track DPRI at each probability band. Minimum sample: 100 predicted discharges per unit type per month.
Proposed formula, pending validation pilots.