Hybrid nonparametric neural forecasts ADC dispensing, overrides, errors

• The NPAR-ANN hybrid produced the lowest forecasting errors across ADC metrics: RMSE 71.50 items issued, 15.43 overrides, and 20.92 integration errors.

• Accurate short-term forecasts enable pharmacy leaders to prioritize staffing, target override-reduction training, schedule ADC–HIS maintenance, and strengthen patient-safety oversight at the point of care.

• Automated dispensing cabinets (ADCs) are widely deployed for medication dispensing, but key indicators—items dispensed, overrides (nurse access without prior pharmacist verification), and ADC–HIS integration errors (synchronization failures)—are usually monitored reactively rather than forecasted, leaving safety and continuity vulnerable to abrupt changes.

• Reactive monitoring forces pharmacy teams into time-consuming audits and troubleshooting, strains limited informatics and clinical resources, and can allow transient spikes in overrides or synchronization failures to go undetected during high-demand periods.

• Accurate short-term forecasts would allow clinical decision-support and operations teams to preemptively target interventions and allocate constrained pharmacy and IT resources, improving medication safety and operational resilience.

• Used retrospective monthly ADC logs from the medical intensive care unit (MICU) at Almoosa Specialist Hospital (January 2023–December 2024

• 24 months) for items issued, overrides, and ADC–HIS integration errors.

• Built and tuned a suite of forecasting approaches in R/RStudio: conventional linear models (ARIMA, exponential smoothing model [ESM], Theta), nonparametric LOESS/NPAR, artificial neural networks (ANN), and hybrid pipelines (ARIMA-ANN, ESM-ANN, NPAR-ANN).

• Compared models using RMSE, MAE, MAPE and RMSLE with 95% bootstrap confidence intervals, rolling-origin time-series cross-validation, and Diebold–Mariano tests for pairwise accuracy differences.

• Conducted a single-center, retrospective time-series study (24 monthly points)

• data were checked for stationarity and seasonality, models were pragmatically tuned and validated, and the work proceeded under institutional ethics approval.

• The NPAR-ANN hybrid produced the lowest forecasting errors with RMSEs of 71.50 items (95% confidence interval [CI] 60.89–81.34), 15.43 overrides (95% CI 12.81–18.56), and 20.92 integration errors (95% CI 17.48–24.67)

• corresponding MAEs were 53.78, 11.90, and 15.46.

• Cross-validation confirmed robustness: NPAR-ANN cross-validation RMSEs were 59.30 (items), 12.08 (overrides

• MAE 8.58, MAPE 8.05%), and 15.23 (integration errors

• MAE 10.27), outperforming ARIMA, ESM, ANN and other hybrid pipelines.

• Observed peaks aligned with model fit (items max 10,686 in Oct 2023

• overrides 194 in Jan 2024

• integration errors 182 in Mar 2024). Two-year forecasts predict November–February item surges and later decline, with overrides and integration errors stabilizing toward end-2026—findings that support planning for verification staffing, targeted training, and scheduled integration fixes.

• Performance gains derived from NPAR smoothing of trend components combined with ANN modeling of nonlinear residuals.

• Integrate short-term NPAR-ANN forecasts into ADC dashboards or business-intelligence feeds to generate daily/weekly workqueues for verification teams and pre-schedule pharmacist coverage and ADC–HIS maintenance during predicted high-volume periods.

• Translate override and integration-error forecasts into targeted interventions: trigger unit-specific override-reduction training, temporarily redeploy verification staff or pharmacists, and prioritize troubleshooting tickets for integration failures instead of relying on retrospective audits.

• Operationalize governance around forecasts: set alert thresholds, require human-in-the-loop rules (for example, pharmacist sign-off for schedule or verification changes), monitor model performance with rolling cross-validation and bootstrap CIs, and allocate informatics support for periodic retraining.

• Treat forecasts as operational situational awareness—they flag upcoming high-risk periods so teams can move people, training, and system fixes in advance while preserving pharmacists as the final clinical arbiters.

Strengths:

• Robust comparative evaluation using multiple accuracy metrics, 95% bootstrap confidence intervals, rolling-origin cross-validation, and Diebold–Mariano tests.

• NPAR-ANN architecture explicitly decomposed smoothed trends (NPAR) and nonlinear residuals (ANN), improving short-term stability and mitigating overfit in a small-sample context.

Limitations:

• Single-center MICU monthly series limits external generalizability across other hospitals, units, ADC configurations, and medication mixes.

• Twenty-four months of monthly aggregation constrains ANN training and temporal resolution

• operational validation and periodic recalibration are required before broader deployment.

The NPAR-ANN hybrid delivers accurate short-term ADC forecasts and is suitable for pilot dashboard deployment to guide staffing, targeted training, and ADC–HIS integration fixes, provided local validation and routine retraining are implemented.